JOURNAL OF THE

AMERICAN MATHEMATICAL SOCIETY

Volume 14, Number 4, Pages 843939
S 0894-0347(01)00370-8
Article electronically published on May 15, 2001
ON THE MODULARITY OF ELLIPTIC CURVES OVER Q:
WILD 3-ADIC EXERCISES
CHRISTOPHE BREUIL, BRIAN CONRAD, FRED DIAMOND, AND RICHARD TAYLOR
Introduction
In this paper, building on work of Wiles [Wi] and of Taylor and Wiles [TW], we
will prove the following two theorems (see 2.2).
Theorem A. If E
/Q
is an elliptic curve, then E is modular.
Theorem B. If : Gal(Q/Q) GL
2
(F
5
) is an irreducible continuous represen-
tation with cyclotomic determinant, then is modular.
We will rst remind the reader of the content of these results and then briey
outline the method of proof.
If N is a positive integer, then we let
1
(N) denote the subgroup of SL
2
(Z)
consisting of matrices that modulo N are of the form
_
1
0 1
_
.
The quotient of the upper half plane by
1
(N), acting by fractional linear transfor-
mations, is the complex manifold associated to an ane algebraic curve Y
1
(N)
/C
.
This curve has a natural model Y
1
(N)
/Q
, which for N > 3 is a ne moduli scheme
for elliptic curves with a point of exact order N. We will let X
1
(N) denote the
smooth projective curve which contains Y
1
(N) as a dense Zariski open subset.
Recall that a cusp form of weight k 1 and level N 1 is a holomorphic
function f on the upper half complex plane H such that
for all matrices
_
a b
c d
_

1
(N)
and all z H, we have f((az +b)/(cz +d)) = (cz +d)
k
f(z);
and [f(z)[
2
(Imz)
k
is bounded on H.
Received by the editors February 28, 2000 and, in revised form, January 1, 2001.
2000 Mathematics Subject Classication. Primary 11G05; Secondary 11F80.
Key words and phrases. Elliptic curve, Galois representation, modularity.
The rst author was supported by the CNRS. The second author was partially supported by
a grant from the NSF. The third author was partially supported by a grant from the NSF and an
AMS Centennial Fellowship, and was working at Rutgers University during much of the research.
The fourth author was partially supported by a grant from the NSF and by the Miller Institute
for Basic Science.
c 2001 American Mathematical Society
843
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
844 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
The space S
k
(N) of cusp forms of weight k and level N is a nite-dimensional
complex vector space. If f S
k
(N), then it has an expansion
f(z) =

n=1
c
n
(f)e
2inz
and we dene the L-series of f to be
L(f, s) =

n=1
c
n
(f)/n
s
.
For each prime p,[ N there is a linear operator T
p
on S
k
(N) dened by
(f[T
p
)(z) = p
1
p1

i=0
f((z +i)/p) +p
k1
(cpz +d)
k
f((apz +b)/(cpz +d))
for any
_
a b
c d
_
SL
2
(Z)
with c 0 mod N and d p mod N. The operators T
p
for p,[ N can be simulta-
neously diagonalised on the space S
k
(N) and a simultaneous eigenvector is called
an eigenform. If f is an eigenform, then the corresponding eigenvalues, a
p
(f), are
algebraic integers and we have c
p
(f) = a
p
(f)c
1
(f).
Let be a place of the algebraic closure of Qin Cabove a rational prime and let
Q

denote the algebraic closure of Q

thought of as a Q algebra via . If f S

k
(N)
is an eigenform, then there is a unique continuous irreducible representation

f,
: Gal(Q/Q) GL
2
(Q

)
such that for any prime p,[ Nl,
f,
is unramied at p and tr
f,
(Frob
p
) = a
p
(f).
The existence of
f,
is due to Shimura if k = 2 [Sh2], to Deligne if k > 2 [De]
and to Deligne and Serre if k = 1 [DS]. Its irreducibility is due to Ribet if k > 1
[Ri] and to Deligne and Serre if k = 1 [DS]. Moreover is odd in the sense that
det of complex conjugation is 1. Also,
f,
is potentially semi-stable at in the
sense of Fontaine. We can choose a conjugate of
f,
which is valued in GL
2
(O
Q

),
and reducing modulo the maximal ideal and semi-simplifying yields a continuous
representation

f,
: Gal(Q/Q) GL
2
(F

),
which, up to isomorphism, does not depend on the choice of conjugate of
f,
.
Now suppose that : G
Q
GL
2
(Q

) is a continuous representation which

is unramied outside nitely many primes and for which the restriction of to
a decomposition group at is potentially semi-stable in the sense of Fontaine. To
[
Gal(Q

/Q

)
we can associate both a pair of Hodge-Tate numbers and a Weil-Deligne
representation of the Weil group of Q

. We dene the conductor N() of to be

the product over p ,= of the conductor of [
Gal(Q
p
/Qp)
and of the conductor of
the Weil-Deligne representation associated to [
Gal(Q

/Q

)
. We dene the weight
k() of to be 1 plus the absolute dierence of the two Hodge-Tate numbers of
[
Gal(Q

/Q

)
. It is known by work of Carayol and others that the following two
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 845
conditions are equivalent:

f,
for some eigenform f and some place [;

f,
for some eigenform f of level N() and weight k() and some place
[.
When these equivalent conditions are met we call modular. It is conjectured
by Fontaine and Mazur that if : G
Q
GL
2
(Q

) is a continuous irreducible
representation which satises
is unramied outside nitely many primes,
[
Gal(Q

/Q

)
is potentially semi-stable with its smaller Hodge-Tate number
0,
and, in the case where both Hodge-Tate numbers are zero, is odd,
then is modular [FM].
Next consider a continuous irreducible representation : Gal(Q/Q) GL
2
(F

).
Serre [Se2] denes the conductor N() and weight k() of . We call modular
if
f,
for some eigenform f and some place [. We call strongly modular
if moreover we may take f to have weight k() and level N(). It is known from
work of Mazur, Ribet, Carayol, Gross, Coleman, Voloch and others that for 3,
is strongly modular if and only if it is modular (see [Di1]). Serre has conjectured
that all odd, irreducible are strongly modular [Se2].
Now consider an elliptic curve E
/Q
. Let
E,
(resp.
E,
) denote the represen-
tation of Gal(Q/Q) on the -adic Tate module (resp. the -torsion) of E(Q). Let
N(E) denote the conductor of E. It is known that the following conditions are
equivalent:
(1) The L-function L(E, s) of E equals the L-function L(f, s) for some eigenform
f.
(2) The L-function L(E, s) of E equals the L-function L(f, s) for some eigenform
f of weight 2 and level N(E).
(3) For some prime , the representation
E,
is modular.
(4) For all primes , the representation
E,
is modular.
(5) There is a non-constant holomorphic map X
1
(N)(C) E(C) for some
positive integer N.
(6) There is a non-constant morphism X
1
(N(E)) E which is dened over Q.
The implications (2) (1), (4) (3) and (6) (5) are tautological. The impli-
cation (1) (4) follows from the characterisation of L(E, s) in terms of
E,
. The
implication (3) (2) follows from a theorem of Carayol [Ca1]. The implication (2)
(6) follows from a construction of Shimura [Sh2] and a theorem of Faltings [Fa].
The implication (5) (3) seems to have been rst noticed by Mazur [Maz]. When
these equivalent conditions are satised we call E modular.
It has become a standard conjecture that all elliptic curves over Q are modular,
although at the time this conjecture was rst suggested the equivalence of the
conditions above may not have been clear. Taniyama made a suggestion along the
lines (1) as one of a series of problems collected at the Tokyo-Nikko conference in
September 1955. However his formulation did not make clear whether f should
be a modular form or some more general automorphic form. He also suggested
that constructions as in (5) and (6) might help attack this problem at least for
some elliptic curves. In private conversations with a number of mathematicians
(including Weil) in the early 1960s, Shimura suggested that assertions along the
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
846 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
lines of (5) and (6) might be true (see [Sh3] and the commentary on [1967a] in
[We2]). The rst time such a suggestion appears in print is Weils comment in
[We1] that assertions along the lines of (5) and (6) follow from the main result of
that paper, a construction of Shimura and from certain reasonable suppositions
and natural assumptions. That assertion (1) is true for CM elliptic curves follows
at once from work of Hecke and Deuring. Shimura [Sh1] went on to check assertion
(5) for these curves.
Our approach to Theorem A is an extension of the methods of Wiles [Wi] and
of Taylor and Wiles [TW]. We divide the proof into three cases.
(1) If
E,5
[
Gal(Q/Q(

5))
is irreducible, we show that
E,5
is modular.
(2) If
E,5
[
Gal(Q/Q(

5))
is reducible, but
E,3
[
Gal(Q/Q(

3))
is absolutely irre-
ducible, we show that
E,3
is modular.
(3) If
E,5
[
Gal(Q/Q(

5))
is reducible and
E,3
[
Gal(Q/Q(

3))
is absolutely re-
ducible, then we show that E is isogenous to an elliptic curve with j-invariant
0, (11/2)
3
, or 5(29)
3
/2
5
and so (from tables of modular elliptic curves of
low conductor) is modular.
In each of cases (1) and (2) there are two steps. First we prove that
E,
is modular
and then that
E,
is modular. In case (1) this rst step is our Theorem B and in
case (2) it is a celebrated theorem of Langlands and Tunnell [L], [T]. In fact, in
both cases E obtains semi-stable reduction over a tame extension of Q

and the
deduction of the modularity of
E,
from that of
E,
was carried out in [CDT] by
an extension of the methods of [Wi] and [TW]. In the third case we have to analyse
the rational points on some modular curves of small level. This we did, with Elkies
help, in [CDT].
It thus only remained to prove Theorem B. Let be as in that theorem. Twisting
by a quadratic character, we may assume that [
Gal(Q
3
/Q3)
falls into one of the
following cases (see 2.2):
(1) is unramied at 3.
(2) (I
3
) has order 5.
(3) (I
3
) has order 4.
(4) (I
3
) has order 12 and [
Gal(Q
3
/Q3)
has conductor 27.
(5) (I
3
) has order 3.
(6) [
Gal(Q
3
/Q3)
is induced from a character : Gal(Q
3
/Q
3
(

3)) F

25
such
that (1) = 1 and
(

3) = (1 + 3

3) (1 3

3),
where we use the Artin map (normalised to take uniformisers to arithmetic
Frobenius) to identify with a character of Q
3
(

3)

.
We will refer to these as the f = 1, 3, 9, 27, 81 and 243 cases respectively.
Using the technique of Minkowski and Klein (i.e. the observation that the moduli
space of elliptic curves with full level 5 structure has genus 0; see for example [Kl]),
Hilbert irreducibility and some local computations of Manoharmayum [Man], we
nd an elliptic curve E
/Q
with the following properties (see 2.2):

E,5
,

E,3
is surjective onto GL
2
(F
3
),
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 847
and
(1) in the f = 1 case, either
E,3
[
I3
F
9

2

3
2
or

E,3
[
I3

_

0 1
_
and is peu ramie;
(2) in the f = 3 case,

E,3
[
I3

_

0 1
_
;
(3) in the f = 9 case,
E,3
[
I3
F
9

2

3
2
;
(4) in the f = 27 case,

E,3
[
I3

_

0 1
_
and is tr`es ramie;
(5) in the f = 81 case,

E,3
[
I3

_
1
0
_
and is tr`es ramie;
(6) in the f = 243 case,

E,3
[
Gal(Q
3
/Q3)

_

0 1
_
is non-split over Q
ker
3
and is tr`es ramie.
(We are using the terms tr`es ramie and peu ramie in the sense of Serre [Se2].
We are also letting denote the mod3 cyclotomic character and
2
the second
fundamental character I
3
F

9
, i.e.

2
() (
8

3)/
8

3 mod
8

3.
We will often use the equality =
1
without further remark.) We emphasise
that for a general elliptic curve over Q with
E,5

= , the representation
E,3
does
not have the above form, rather we are placing a signicant restriction on E.
In each case our strategy is to prove that
E,3
is modular and so deduce that

E,5
is modular. Again we use the Langlands-Tunnell theorem to see that
E,3
is modular and then an analogue of the arguments of [Wi] and [TW] to conclude
that
E,3
is modular. This was carried out in [Di2] in the cases f = 1 and f = 3,
and in [CDT] in the case f = 9. (In these cases the particular form of
E,3
[
I3
is not
important.) This leaves the cases f = 27, 81 and 243, which are complicated by
the fact that E now only obtains good reduction over a wild extension of Q
3
. In
these cases our argument relies essentially on the particular form we have obtained
for
E,3
[
Gal(Q
3
/Q3)
(depending on
E,5
[
I3
). We do not believe that our methods
for deducing the modularity of
E,3
from that of
E,3
would work without this key
simplication. It seems to be a piece of undeserved good fortune that for each
possibility for [
I3
we can nd a choice for
E,3
[
Gal(Q
3
/Q3)
for which our methods
work.
Following Wiles, to deduce the modularity of
E,3
from that of
E,3
, we consider
certain universal deformations of
E,3
and identify them with certain modular de-
formations which we realise over certain Hecke algebras. The key problem is to
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
848 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
nd the right local condition to impose on these deformations at the prime 3. As
in [CDT] we require that the deformations lie in the closure of the characteristic
zero points which are potentially Barsotti-Tate (i.e. come from a 3-divisible group
over the ring of integers of a nite extension of Q
3
) and for which the associated
representation of the Weil group (see for example Appendix B of [CDT]) is of some
specied form. That one can nd suitable conditions on the representation of the
Weil group at 3 for the arguments of [TW] to work seems to be a rare phenomenon
in the wild case. It is here we make essential use of the fact that we need only treat
our specic pairs (
E,5
,
E,3
).
Our arguments follow closely the arguments of [CDT]. There are two main new
features. Firstly, in the f = 243 case, we are forced to specify the restriction of our
representation of the Weil group completely, rather than simply its restriction to
the inertia group as we have done in the past. Secondly, in the key computation
of the local deformation rings, we now make use of a new description (due to
Breuil) of nite at group schemes over the ring of integers of any p-adic eld in
terms of certain (semi-)linear algebra data (see [Br2] and the summary [Br1]). This
description enables us to make these computations. As the persistent reader will
soon discover they are lengthy and delicate, particularly in the case f = 243. It
seems miraculous to us that these long computations with nite at group schemes
in 7, 8 and 9 give answers completely in accord with predictions made from much
shorter computations with the local Langlands correspondence and the modular
representation theory of GL
2
(Q
3
) in 3. We see no direct connection, but cannot
help thinking that some such connection should exist.
Notation. In this paper denotes a rational prime. In 1.1, 4.1, 4.2 and 4.3 it
is arbitrary. In the rest of 1 and in 5 we only assume it is odd. In the rest of the
paper we only consider = 3.
If F is a eld we let F denote a separable closure, F
ab
the maximal subextension
of F which is abelian over F and G
F
the Galois group Gal(F/F). If F
0
is a p-adic
eld (i.e. a nite extension of Q
p
) and F

/F
0
a (possibly innite) Galois extension,
then we let I
F

/F0
denote the inertia subgroup of Gal(F

/F
0
). We also let I
F0
denote
I
F0/F0
, Frob
F0
G
F0
/I
F0
denote the arithmetic Frobenius element and W
F0
denote
the Weil group of F
0
, i.e. the dense subgroup of G
F0
consisting of elements which
map to an integer power of Frob
F0
. We will normalise the Artin map of local class
eld theory so that uniformisers and arithmetic Frobenius elements correspond.
(We apologise for this convention, which now seems to us a bad choice. However
we feel it is important to stay consistent with [CDT].) We let O
F0
denote the ring
of integers of F
0
,
F0
the maximal ideal of O
F0
and k
F0
the residue eld O
F0
/
F0
.
We write simply G
p
for G
Qp
, I
p
for I
Qp
and Frob
p
for Frob
Qp
. We also let Q
p
n
denote the unique unramied degree n extension of Q
p
in Q
p
. If k is any perfect
eld of characteristic p we also use Frob
p
to denote the p
th
-power automorphism of
k and its canonical lift to the Witt vectors W(k).
We write for the -adic cyclotomic character and sometimes for the reduction
of modulo . We write
2
for the second fundamental character I

2
, i.e.

2
() (
1/(
2
1)
)/
1/(
2
1)
mod
1/(
2
1)
.
We also use and
2
to denote the Teichmuller lifts of and
2
.
We let 1 denote the trivial character of a group. We will denote by V

the dual
of a vector space V .
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 849
If g : A B is a homomorphism of rings and if X
/ Spec A
is an A-scheme, then
we sometimes write
g
X for the pullback of X by Spec g. We adopt this notation
so that
g
(
h
X) =
gh
X. Similarly if : X Y is a morphism of schemes over A we
will sometimes write
g
for the pullback of by Spec g.
By nite at group scheme we always mean commutative nite at group scheme.
If F
0
is a eld of characteristic 0 with xed algebraic closure F
0
we use without
comment the canonical identication of nite at F
0
-group schemes with nite
discrete Gal(F
0
/F
0
)-modules, and we will say that such objects correspond. If R is
a Dedekind domain with eld of fractions F of characteristic 0, then by a model of
a nite at F-group scheme G we mean a nite locally free R-group scheme G and
an isomorphism i : G

G F

. As in Proposition 2.2.2 of [Ra] the isomorphism

classes of models for G form a lattice ((G, i) (G

, i

) if there exists a map of nite

at group schemes G G

compatible with i and i

) and we can talk about the inf

and sup of two such models. If R is also local we call the model (G, i) local-local if
its special bre is local-local. When the ring R is understood we sometimes simply
refer to (G, i), or even just G, as an integral model of G.
We use Serres terminology peu ramie and tr`es ramie; see [Se2].
Contents
Introduction 843
Notation 848
1. Types 850
1.1. Types of local deformations 850
1.2. Types for admissible representations 852
1.3. Reduction of types for admissible representations 854
1.4. The main theorems 857
2. Examples and applications 860
2.1. Examples 860
2.2. Applications 862
2.3. An extension of a result of Manoharmayum 865
3. Admittance 866
3.1. The case of
1
866
3.2. The case of
1
866
3.3. The case of
3
867
3.4. The case of

i
867
4. New deformation problems 867
4.1. Some generalities on group schemes 868
4.2. Filtrations 870
4.3. Generalities on deformation theory 874
4.4. Reduction steps for Theorem 2.1.2 875
4.5. Reduction steps for Theorem 2.1.4 876
4.6. Reduction steps for Theorem 2.1.6 877
4.7. Some Galois cohomology 880
5. Breuil modules 883
5.1. Basic properties of Breuil modules 883
5.2. Examples 886
5.3. Relationship to syntomic sheaves 888
5.4. Base change 890
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
850 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
5.5. Reformulation 894
5.6. Descent data 896
5.7. More examples 898
6. Some local elds 899
6.1. The case of F

1
899
6.2. The case of F

1
900
6.3. The case of F

3
900
6.4. The case of F

3
901
6.5. The case of F

i
902
7. Proof of Theorem 4.4.1 904
7.1. Rank one calculations 904
7.2. Rank two calculations 904
7.3. Rank three calculations 908
7.4. Conclusion of the proof of Theorem 4.4.1 909
8. Proof of Theorem 4.5.1 911
8.1. Rank one calculations 911
8.2. Models for 912
8.3. Completion of the proof of Theorem 4.5.1 917
9. Proof of Theorems 4.6.1, 4.6.2 and 4.6.3 919
9.1. Rank one calculations 919
9.2. Models for 920
9.3. Further rank two calculations 927
9.4. Completion of the proof of Theorem 4.6.1 931
9.5. Completion of the proof of Theorem 4.6.3 932
9.6. Completion of the proof of Theorem 4.6.2 935
10. Corrigenda for [CDT] 937
Acknowledgements 938
References 938
1. Types
1.1. Types of local deformations. By an -type we mean an equivalence class
of two-dimensional representations
: I

GL(D)
over Q

which have open kernel and which can be extended to a representation

of W
Q

. By an extended -type we shall simply mean an equivalence class of two-

dimensional representations

: W
Q

GL(D

)
over Q

with open kernel.

Suppose that is an -type and that K is a nite extension of Q

in Q

. Recall
from [CDT] that a continuous representation of G

on a two-dimensional K-vector
space M is said to be of type if
(1) is Barsotti-Tate over F for any nite extension F of Q

such that [
IF
is
trivial;
(2) the restriction of WD() to I

is in ;
(3) the character
1
det has nite order prime to .
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 851
(For the denition of Barsotti-Tate and of the representation WD() associated
to a potentially Barsotti-Tate representation, see 1.1 and Appendix B of [CDT].)
Similarly if

is an extended -type, then we say that is of extended type

if
(1) is Barsotti-Tate over F for any nite extension F of Q

such that

[
IF
is
trivial;
(2) WD() is equivalent to

;
(3) the character
1
det has nite order prime to .
Note that no representation can have extended type

unless det

is of the form

2
where
1
has nite order prime to and where
2
is unramied and takes an
arithmetic Frobenius element to ; see Appendix B of [CDT]. (Using Theorem 1.4
of [Br2], one can show that for odd one obtains equivalent denitions of type
and extended type

if one weakens the rst assumption to simply require that

is potentially Barsotti-Tate.)
Now x a nite extension K of Q

in Q

over which (resp.

) is rational. Let
O denote the integers of K and let k denote the residue eld of O. Let
: G

GL(V )
be a continuous representation of G

on a two-dimensional k-vector space V and

suppose that End
k[G

]
V = k. One then has a universal deformation ring R
V,O
for
(see, for instance, Appendix A of [CDT]).
We say that a prime ideal p of R
V,O
is of type (resp. of extended type

)
if there exist a nite extension K

of K in Q

and an O-algebra homomorphism

R
V,O
K

with kernel p such that the pushforward of the universal deformation

of over R
V,O
to K

is of type (resp. of extended type

).
Let be an -type and

an irreducible extended -type. If there do not exist

any prime ideals p of type (resp. of extended type

), we dene R
D
V,O
= 0 (resp.
R
D

V,O
= 0). Otherwise, dene R
D
V,O
(resp. R
D

V,O
) to be the quotient of R
V,O
by
the intersection of all p of type (resp. of extended type

). We will sometimes
write R

V,O
(resp. R

V,O
) for R
D
V,O
(resp. R
D

V,O
). We say that a deformation of is
weakly of type (resp. weakly of extended type

) if the associated local O-algebra

map R
V,O
R factors through the quotient R
D
V,O
(resp. R
D

V,O
). We say that
(resp.

) is weakly acceptable for if either R

D
V,O
= 0 (resp. R
D

V,O
= 0) or there
is a surjective local O-algebra map O[[X]] R
D
V,O
(resp. O[[X]] R
D

V,O
). We say
that (resp.

) is acceptable for if R
D
V,O
,= 0 (resp. R
D

V,O
,= 0) and if there is a
surjective local O-algebra map O[[X]] R
D
V,O
(resp. O[[X]] R
D

V,O
).
If K

is a nite extension of K in Q

with valuation ring O

and residue eld

k

, then O

O
R
D
V,O
(resp. O

O
R
D

V,O
) is naturally isomorphic to R
D
V
k
k

,O
(resp.
R
D

V
k
k

,O
). Thus (weak) acceptability depends only on (resp.

) and , and not

on the choice of K. Moreover (resp.

) is acceptable for if and only if (resp.

) is acceptable for
k
k

.
Although it is of no importance for the sequel, we make the following conjecture,
part of which we already conjectured as Conjecture 1.2.1 of [CDT].
Conjecture 1.1.1. Suppose that is an -type and

an absolutely irreducible
extended -type. A deformation : G

GL(M) of to the ring of integers O

of
a nite extension K

/K in Q

is weakly of type (resp. weakly of extended -type

) if and only if M is of type (resp. of extended type

).
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
852 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
If is a tamely ramied -type, then we expect that it is frequently the case that
is acceptable for residual representations , as in Conjectures 1.2.2 and 1.2.3 of
[CDT]. On the other hand if (resp.

) is a wildly ramied -type (resp. wildly

ramied extended -type), then we expect that it is rather rare that (resp.

) is
acceptable for a residual representation . In this paper we will be concerned with
a few wild cases for the prime = 3 which do turn out to be acceptable.
1.2. Types for admissible representations. From now on we assume that
is odd. If F is a nite extension of Q

we will identify F

with W
ab
F
via the
Artin map. Let U
0
(
r
) denote the subgroup of GL
2
(Z

) consisting of elements with

upper triangular mod
r
reduction. Also let

U
0
() denote the normaliser of U
0
()
in GL
2
(Q

). Thus

U
0
() is generated by U
0
() and by
w

=
_
0 1
0
_
. (1.2.1)
If is an -type, set U

= GL
2
(Z

) if is reducible and U

= U
0
() if is
irreducible. If

is an extended -type with

[
I

irreducible, set U

=

U
0
(). In
this subsection we will associate to an -type an irreducible representation

of
U

over Q

with open kernel, and to an extended -type

with

[
I

irreducible an
irreducible representation

of U

over Q

with open kernel. We need to consider

several cases, which we treat one at a time.
First suppose that =
1
[
I

2
[
I

where each
i
is a character of W
Q

. Let a
denote the conductor of
1
/
2
. If a = 0, then set

= St (
1
det) = St (
2
det),
where St denotes the Steinberg representation of PGL
2
(F

). Now suppose that

a > 0. Let

denote the induction from U

0
(
a
) to GL
2
(Z

) of the character of
U
0
(
a
) which sends
_

a

_
(
1
/
2
)()
2
(
a
).
This is irreducible and does not depend on the ordering of
1
and
2
.
For the next case, let F denote the unramied quadratic extension of Q

and s
the non-trivial automorphism of F over Q

. Suppose that is the restriction to

I

of the induction from W

F
to W
Q

of a character of W
F
with ,=
s
. Let a
denote the conductor of /
s
, so that a > 0. Choose a character

of W
Q

such
that

[
1
WF
has conductor a. If a = 1 we set

= (

[
1
WF
) (

det),
where () is the irreducible representation of GL
2
(F

) dened on page 532 of

[CDT].
To dene

for a > 1 we will identify GL

2
(Z

) with the automorphisms of the Z

-
module O
F
. If a is even, then we let

denote the induction from O

F
(1+
a/2
O
F
s)
to GL
2
(Z

) of the character of O

F
(1 +
a/2
O
F
s), where, for O

F
and
(1 +
a/2
O
F
s),
() = (

[
1
WF
)()

(det ).
If a > 1 is odd, then we let

denote the induction from O

F
(1 +
(a1)/2
O
F
s)
to GL
2
(Z

) of , where is the -dimensional irreducible representation of

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 853
O

F
(1 +
(a1)/2
O
F
s) such that [
O

F
(1+
(a+1)/2
OF s)
is the direct sum of the char-
acters
(

[
1
WF

)()

(det )
for O

F
and (1 +
(a+1)/2
O
F
s), where

runs over the non-trivial

characters of O

F
/Z

(1 +O
F
).
Now suppose

is an extended type such that

[
I

is irreducible. There is a
ramied quadratic extension F/Q

and a character of W
F
such that the induction
from W
F
to W
Q

of is

(see 2.6 of [G]). Let s denote the non-trivial eld

automorphism of F over Q

and also let

F
denote the maximal ideal of the ring
of integers O
F
of F. Let a denote the conductor of /
s
, so a is even and a 2.
We may choose a character

of W
Q

such that

[
1
WF
has conductor a. We will
identify GL
2
(Q

) with the automorphisms of the Q

vector space F. We will also

identify U
0
() with the stabiliser of the pair of lattices
1
F
O
F
. We dene

to
be the induction from F

(1 +
a/2
F
s) to

U
0
() of the character of F

(1 +
a/2
F
s),
where
() = (

[
1
WF

)()

(det ),
with F

and (1 +
a/2
F
s), where

is a character of F

/(O

F
)
2
dened as
follows. Let be a character of Q

with kernel Z

. Choose F

such that for

x
a1
F
we have
(

[
1
WF
)(1 +x) = (tr
F/Q

(x)).
We impose the following conditions which determine

is a character of F

/(O

F
)
2
;

[
O

F
is non-trivial;
and

((N
F/Q

)
a/2
) =

xZ/Z
(x
2
/N
F/Q

),
where is a uniformiser in O
F
.
Finally if is an irreducible -type, choose an extended -type

which restricts
to on I

and set

[
U0()
.
We remark that these denitions are independent of any choices (see [G]).
Recall that by the local Langlands conjecture we can associate to an irreducible
admissible representation of GL
2
(Q

) a two-dimensional representation WD()

of W
Q

. (See 4.1 of [CDT] for the normalisation we use.)

Lemma 1.2.1. Suppose that is an -type and that

is an extended -type with

[
I

irreducible. Suppose also that is an innite-dimensional irreducible admis-

sible representation of GL
2
(Q

) over Q

. Then:
(1)

and

are irreducible.
(2) If WD()[
I

(resp. WD()

), then
Hom
U
(

, )

= Q

(resp.
Hom
U

, )

= Q

).
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
854 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
(3) If WD()[
I

= (resp. WD() ,

), then
Hom
U
(

, ) = (0)
(resp.
Hom
U

, ) = (0)).
Proof. The case that extends to a reducible representation of W
Q

follows from
the standard theory of principal series representations for GL
2
(Q

). The case that

is reducible but does not extend to a reducible representation of W
Q

follows from
Theorem 3.7 of [G]. The case of

follows from Theorem 4.6 of [G].

Thus, suppose that is an irreducible -type and that

is an extension of to
an extended -type. If denotes the unramied quadratic character of W
Q

, then

and so we deduce that

( det).
Thus

[
Q

U0()
is irreducible. It follows that

is irreducible. The second and

third part of the lemma for follow similarly.
1.3. Reduction of types for admissible representations. We begin by re-
viewing some irreducible representations of GL
2
(Z

), U
0
() and

U
0
(). Let
1,0
denote the standard representation of GL
2
(F

) over F

. If n = 0, 1, ..., 1 and
if m Z/( 1)Z, then we let
n,m
= Symm
n
(
1,0
) det
m
. We may think of

n,m
as a continuous representation of GL
2
(Z

) over F

. These representations
are irreducible, mutually non-isomorphic and exhaust the irreducible continuous
representations of GL
2
(Z

) over F

.
If m
1
, m
2
Z/( 1)Z we let

m1,m2
denote the character of U
0
() over F

determined by
_
a b
c d
_
a
m1
d
m2
.
These representations are irreducible, mutually non-isomorphic, and exhaust the
irreducible continuous representations of U
0
() over F

.
If m
1
, m
2
Z/( 1)Z, a F

and m
1
,= m
2
, then we let

{m1,m2},a
denote
the representation of

U
0
() over F

obtained by inducing the character of Q

U
0
()
which restricts to

m1,m2
on U
0
() and which sends to a. If m Z/( 1)Z and
a F

, then we let

{m},a
denote the character of

U
0
() over F

which restricts to

m,m
on U
0
() and which sends w

to a. These representations are irreducible, mu-

tually non-isomorphic and exhaust the irreducible, nite-dimensional, continuous
representations of

U
0
() over F

.
We will say that a reducible -type (resp. irreducible -type, resp. extended
-type with irreducible restriction to I

) admits an irreducible representation of

GL
2
(Z

) (resp. U
0
(), resp.

U
0
()) over F

, if

(resp.

, resp.

) contains an
invariant O
Q

-lattice and if is a Jordan-H older constituent of F

. We will
say that (resp. , resp.

) simply admits if is a Jordan-H older constituent

of F

of multiplicity one.
For each of the F

-representations of GL
2
(Z

), U
0
() and

U
0
() just dened,
we wish to dene notions of admittance and simple admittance with respect
to a continuous representation : G

GL
2
(F

). Let be a xed continuous

representation G

GL
2
(F

).
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 855
The representation
n,m
admits if either
[
I

_

1nm(+1)
2
0
0
nm(+1)
2
_
or
[
I

_

1m

0
nm
_
,
which in addition we require to be peu-ramie in the case n = 0. (Note that

n,0
admits if and only if the Serre weight (see [Se2]) of

is n + 2.)
The representation
n,m
simply admits if
n,m
admits .
The representation

m1,m2
admits if either
[
I

_

1mimj
2
0
0
mimj
2
_
,
where m
i
, m
j
= m
1
, m
2
and m
i
m
j
, or
[
I

_

1m1

0
m2
_
,
or
[
I

_

1m2

0
m1
_
.
(Note that

m1,m2
admits if and only if some irreducible constituent of
Ind
GL2(Z

)
U0()

m1,m2
admits .)
The representation

m1,m2
with m
1
,= m
2
simply admits if either
[
I

_

1m1

0
m2
_
or
[
I

_

1m2

0
m1
_
.
The representation

m,m
simply admits if
[
I

_

1m

0
m
_
is tr`es ramie.
The representation

{m1,m2},a
with m
1
,= m
2
admits if either

m1,m2
or

m2,m1
admits and if (
1
det )[
WQ

equals the central character of

{m1,m2},a
. (Note that in this case

{m1,m2},a
[
U0()
=

m1,m2

m2,m1
.)
The representation

{m1,m2},a
with m
1
,=m
2
simply admits if (
1
det )[
WQ

equals the central character of

{m1,m2},a
and either
[
I

_

1m1

0
m2
_
or
[
I

_

1m2

0
m1
_
.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
856 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
The representation

{m},a
admits if

m,m
admits ,
(
1
det )[
WQ

equals the central character of

{m},a
,
and, if
[
I

_

1m

0
m
_
is tr`es ramie, then

_

0
m

_
,
where is unramied and sends Frobenius to a.
(Note that

{m},a
[
U0()
=

m,m
.)
The representation

{m},a
simply admits if

{m},a
admits .
We remark that the denition of admits the Cartier dual of might look more
natural to the reader. We are forced to adopt this version of the denition by some
unfortunate choices of normalisations in [CDT].
We say that a reducible -type (resp. irreducible -type , resp. extended -
type

with

[
I

irreducible) admits a continuous representation : G

GL
2
(F

)
if (resp. , resp.

) admits an irreducible representation of GL

2
(Z

) (resp. U
0
(),
resp.

U
0
()) over F

which in turn admits . We say that (resp. , resp.

)
simply admits if
(resp. , resp.

) admits a unique irreducible representation of GL

2
(Z

)
(resp. U
0
(), resp.

U
0
()) over F

which admits ,
(resp. , resp.

) simply admits ,
and simply admits .
Note that the concept of simply admits is strictly stronger than the concept
admits.
The starting point for this work was the following conjecture, of which a few
examples will be veried in 2.1.
Conjecture 1.3.1. Let k be a nite subeld of F

, : G

GL
2
(k) a continuous
representation, an -type and

an extended -type with irreducible restriction

to I

. Suppose that det and det

are tamely ramied, that the centraliser of the

image of is k and that the image of is not contained in the centre of GL
2
(Q

).
(1) (resp.

) admits if and only if R

D
V,O
,= (0) (resp. R
D

V,O
,= (0)), i.e.
if and only if there is a nite extension K

of Q

in Q

and a continuous
representation : G

GL
2
(O
K
) which reduces to and has type (resp.
has extended type

).
(2) (resp.

) simply admits if and only if (resp.

) is acceptable for .
We remark that to check if or

simply admits is a relatively straightforward

computation. On the other hand to show that or

is acceptable for is at present

a non-trivial undertaking. (The reader who doubts us might like to compare 3 with
4, 5, 6, 7, 8 and 9. All the latter sections are devoted to verifying some very
special cases of this conjecture.)
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 857
1.4. The main theorems. With these denitions, we can state our two main
theorems. The proofs very closely parallel the proof of Theorem 7.1.1 of [CDT].
Theorem 1.4.1. Let be an odd prime, K a nite extension of Q

in Q

and k
the residue eld of K. Let
: G
Q
GL
2
(K)
be an odd continuous representation ramied at only nitely many primes. Assume
that its reduction
: G
Q
GL
2
(k)
is absolutely irreducible after restriction to Q(
_
(1)
(1)/2
) and is modular. Fur-
ther, suppose that
[
G

has centraliser k,
[
G

is potentially Barsotti-Tate with -type ,

admits ,
and is weakly acceptable for .
Then is modular.
Proof. Note that the existence of shows that is acceptable for . Now the proof
is verbatim the proof of Theorem 7.1.1 of [CDT] (see 1.3, 1.4, 3, 4, 5 and 6
of that paper, and the corrigendum at the end of this paper), with the following
exceptions.
On page 539 one should take U
S,
= U

, V
S,
= ker

and
S,l
=

.
In the proof of Lemma 5.1.1 one must use Lemma 1.2.1 of this paper, in
addition to the results recalled in 4 of [CDT].
On page 546 replace Setting S = T() r ... to the end of the rst
paragraph by the following. (Again the key component of this argument is
very similar to the main idea of [Kh].)
Set S = T() r; U

S
=

p
U

S,p
where U

S,p
= U
1
(p
cp
) if p T() and
U

S,p
= U
S,p
otherwise; V

S
=

p
V

S,p
where V

S,p
= U
1
(p
cp
) if p T()
and V

S,p
= V
S,p
otherwise; and L

S
= Hom
O[U

S
/V

S
]
(M

, H
1
(X
V

S
, O))[I

S
].
Then = SL
2
(Z) (U

S
GL
2
(Z

)) satises the hypotheses of Theorem 6.1.1.

Furthermore
H
1
(Y
U

S
GL2(Z

)
, F
M
)

= H
1
(, L
n
k)
as a

T

S
-module, where M is the module for U
S,
= GL
2
(Z

) dened by
the action of GL
2
(F

) on L
n
k. Therefore m
S
is in the support of
H
1
(Y
U

S
GL2(Z

)
, F
M
).
We now drop the special assumption on [
I

made in the last paragraph.

Twisting we see that if is an irreducible representation of GL
2
(Z

) over F

admitting [
G

, then
H
1
(Y
U

S
GL2(Z

)
, F

)
mS
,= (0).
Moreover if is irreducible and if

is an irreducible representation of U
0
()
over F

which admits [
G

, then we see using the denition of admits and

Lemmas 3.1.1 and 6.1.2 of [CDT] that
H
1
(Y
U

S
, F

)
mS

= H
1
(Y
U

S
GL2(Z

)
, F
Ind
GL
2
(Z

)
U
0
()

)
mS
,= (0).
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
858 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
It follows from the denition of admits and Lemma 6.1.2 of [CDT] that m
S
is in the support of H
1
(Y
U

S
, F
Hom
O
(M

,O)
), so L

S
is non-zero. Using the fact
that Lemma 5.1.1 holds with U

S
replacing U
S
and

replacing
S
and the
discussion on page 541 we conclude that N
S
is non-empty.
Theorem 1.4.2. Let be an odd prime, K a nite extension of Q

in Q

and k
the residue eld of K. Let
: G
Q
GL
2
(K)
be an odd continuous representation ramied at only nitely many primes. Assume
that its reduction
: G
Q
GL
2
(k)
is absolutely irreducible after restriction to Q(
_
(1)
(1)/2
) and is modular. Fur-
ther, suppose that
[
G

has centraliser k,
[
G

is potentially Barsotti-Tate with extended -type

admits ,
and

is weakly acceptable for .

Then is modular.
Proof. The existence of shows that

is in fact acceptable for . Again the proof

now follows very closely that of Theorem 7.1.1 of [CDT]. In this case we have to
make the following changes. All references are to [CDT] unless otherwise indicated.
On page 539 one should take U
S,
= U
0
(), V
S,
= ker

[
U0()
and
S,
=

[
U0()
. One should also dene

U
S
to be the group generated by U
S
and
w

GL
2
(Q

) and
S
to be the extension of
S
to

U
S
which restricts to

on

U
0
().
In the statement of Lemma 5.1.1 one should replace Hom
US
(
S
,

) by
Hom

US
(
S
,

).
In the proof of Lemma 5.1.1 one must use Lemma 1.2.1 above in addition to
the results recalled in 4 of [CDT].
Because

is acceptable for , we know that det of a Frobenius lift is for

some root of unity . Thus,

(
s
) = 1 for some s > 0. Hence,
S
factors
through the nite group

G
S
=

U
S
/V
S

sZ
, where GL
2
(Q

).
In 5.3 choose M

so that it is invariant for the action of

U
0
()/V
S,

sZ
. Also,
in the denition of L
S
replace G
S
by

G
S
.
In the proof of Lemma 5.3.1 replace U
S
by

U
S
and
S
by
S
.
Note that w

acts naturally on Y
S
and F
S
. In Lemma 6.1.3 we should
replace the group H
1
c
(Y
S
, F
S
) by H
1
c
(Y
S
, F
S
)
w

=1
and the group H
1
(Y
S
, F
S
)
by H
1
(Y
S
, F
S
)
w

=1
.
Replace 6.2 with the proof of the required extension of Proposition 5.4.1
given below.
On page 547 the isomorphism
H
1
c
(Y
S
, F
S
) Hom
O
(H
1
(Y
S
, F
S
), O)
on line 6 is

T

S
[w
l
]-linear. In the next line one should not only localise at m
but restrict to the kernel of w

1. Because w
2

= 1 on H
1
(Y
S
, F
S
)
m
we see
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 859
that the natural map
H
1
(Y
S
, F
S
)
w

=1
m
H
1
(Y
S
, F
S
)
m
/(w

1)
is an isomorphism, and so the map
L
S
Hom
O
(L
S
, O)
is also an isomorphism.
On page 547 the groups H
1
(Y
S
, F
S
)
m
(p)
S
and H
1
(Y
S
, F
S
)
m
(p)
S
should be re-
placed by their maximal subgroups on which w

= 1.
On page 549 one should also dene

V
0
(resp.

V
1
) to be the group generated
by V
0
(resp. V
1
) and w

GL
2
(Q

). Similarly dene to be

2
r
.
In Lemma 6.4.1 replace V
0
by

V
0
, V
1
by

V
1
and by . In the proof of Lemma
6.4.1 also replace U
{r,r

}
(resp. U
S{r,r

}
) by

U
{r,r

}
(resp.

U
S{r,r

}
) and

{r,r

}
(resp.
S{r,r

}
) by
{r,r

}
(resp.
S{r,r

}
).
On line 20 of page 550 M should be chosen as a model of . This is possible
because ker has nite index in

V
0
, because in turn

(
s
) = 1 for some
s > 0. One should also set L
i
= H
1
(Y
Vi
, F
M
)
w

=1
m
. On line 25, we must
replace V
i
by

V
i
.
In the proof of Lemma 6.4.2, one must replace V
1
by

V
1
and by .
In line 2 of the proof of Lemma 6.4.3, to see that L
1
is a direct summand of
H
1
(Y
V1
, F
M
) as an O[
S
]-module, one needs to note that H
1
(Y
V1
, F
M
)
w

=1
m
is a direct summand of H
1
(Y
V1
, F
M
)
m
, because w
2

= 1 on H
1
(Y
V1
, F
M
)
m
.
On line 12 of page 551 replace R
,D
V,O
by R
,D

V,O
.
Proof of extension of Proposition 5.4.1 of [CDT]. Let =

pT()
M
p
.
First suppose that

admits

{m1,m2},a
with m
1
,= m
2
and that

{m1,m2},a
admits . As in the proof of Theorem 1.4.1 (especially 6.2 of [CDT] as modied
above), we have
H
1
(Y
{r}
, F

F
(

m
1
,m
2
)
)
m

{r}
,= (0).
On the other hand
H
1
(Y
{r}
, F

F
(

m
1
,m
2
)
)
m

{r}

= H
1
(Y
{r}
, F

F
(

m
1
,m
2
)
)
w
2

=a
m

{r}

= H
1
(Y
{r}
, F

F
(

{m
1
,m
2
},a
)
)
w

=1
m

{r}
.
Thus, using the denition of admits and Lemma 6.1.2 of [CDT], we see that
H
1
(Y
{r}
, F
{r}
)
w

=1
m

{r}
,= (0),
so N

= N
{r}
,= .
Next suppose that

admits

{m},a
which in turn admits . Assume that [
G

is irreducible or peu ramie. By twisting we may reduce to the case m = 0. As in

the proof of Theorem 1.4.1 (especially 6.2 of [CDT] as modied above), we have
H
1
(Y
U
{r}
GL2(Z

)
, F

)
m

{r}
,= (0).
Thus
H
1
(Y
U
{r}
GL2(Z

)
, F

)
w
2

= a
2
m

{r}
,= (0),
where a is the Teichm uller lift of a. Using the embedding
a +w

: H
1
(Y
U
{r}
GL2(Z

)
, F

) Q

H
1
(Y
{r}
, F

) Q

,
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
860 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
we deduce that
H
1
(Y
{r}
, F

)
w

= a
m

{r}
,= (0),
and so
H
1
(Y
{r}
, F

F
(

{0},a
)
)
w

=1
m

{r}
,= (0).
Thus, using the denition of admits and Lemma 6.1.2 of [CDT], we see that
H
1
(Y
{r}
, F
{r}
)
w

=1
m

{r}
,= (0),
and so N

= N
{r}
,= .
Finally suppose that

admits
{m},a
which in turn admits , and that [
G

is
reducible and tr`es ramie. By twisting we may reduce to the case m = 0. Note
that
I

(Frob

) = a. As in the proof of Theorem 1.4.1 (especially 6.2 of [CDT]

as modied above), we have
H
1
(Y
{r}
, F

)
m

{r}
,= (0).
Suppose that is a cuspidal automorphic representation which contributes to
H
1
(Y
{r}
, F

)
m

{r}
, so is a cuspidal automorphic representation of GL
2
(A) such
that

is the lowest discrete series with trivial innitesimal character,

is a lift
of of type (r, 1), and hence of type (, 1), and dim
U0()

= 1. As
1
det

has
order prime to , we see that w
2

acts on
U0()

by the Teichm uller lift of a

2
. As

has a U
0
()-xed vector but no GL
2
(Z

)-xed vector, we see that 1 + U

w
1

= 0
on
U0()

. On the other hand, the eigenvalue of U

on
U0()

reduces to a. Thus,
w

acts on
U0()

by the Teichm uller lift of a, so w

acts on H
1
(Y
{r}
, F

)
m

{r}
by
the Teichm uller lift of a. We deduce that
H
1
(Y
{r}
, F

F
(

0,0
)
)
w

=a
m

{r}
,= (0).
Using the denition of admits and Lemma 6.1.2 of [CDT], we see that
H
1
(Y
{r}
, F
{r}
)
w

=1
m

{r}
,= (0),
so N

= N
{r}
,= .
2. Examples and applications
2.1. Examples. Now we will specialise to the case = 3. Fix an element
GL
2
(Z
3
) with
3
= 1 but ,= 1. The following denitions, which concern isomor-
phism classes of 2-dimensional representations into GL
2
(Q
3
), do not depend on this
choice. We will consider the following -types. (These are in fact, up to twist, a
complete list of the wildly ramied types which can arise from elliptic curves over
Q
3
, or, in the case of conductor 243, the extended types. We will not need this
fact. Rather the justication for studying these particular types can be found in
2.2. More detailed information about the xed elds of these types can be found
in 6.)

1
corresponds to the order 3 homomorphism
Z

3
Z
3
[]

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

THE MODULARITY OF ELLIPTIC CURVES OVER Q 861
determined by
1 1
4 .

1
corresponds to the order 3 homomorphism
Z
3
[

1]

Z
3
[]

determined by
4

1 1
4 1
1 + 3

1 .

3
is the unique 3-type such that
3
[
I
Q
3
(

3)
corresponds to the order 6 ho-
momorphism
Z
3
[

3]

Z
3
[]

determined by
1 1
4 1
1 +

3 .

3
is the unique 3-type such that
3
[
I
Q
3
(

3)
corresponds to the order 6
homomorphism
Z
3
[

3]

Z
3
[]

determined by
1 1
4 1
1 + 3

3 1
1 +

3 .
For i Z/3Z, we will also consider the unique extended 3-types

i
whose restrictions
to G
Q3(

3)
correspond to the homomorphisms
Q
3
(

3)

Q
3
()

determined by

3
1
1 1
4 1
1 + 3

3
1 +

3
i
.
(2.1.1)
Subsequent sections of this paper will be devoted to checking the following special
cases of Conjecture 1.3.1.
Lemma 2.1.1. Suppose that : G
3
GL
2
(F
3
) and
[
I3

_
1
0
_
is tr`es ramie. Both
1
and
1
simply admit .
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
862 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
Theorem 2.1.2. Suppose that : G
3
GL
2
(F
3
) and
[
I3

_
1
0
_
is tr`es ramie. Both
1
and
1
are weakly acceptable for .
Lemma 2.1.3. Suppose that : G
3
GL
2
(F
3
) and
[
I3

_

0 1
_
is tr`es ramie. Both
3
and
3
simply admit .
Theorem 2.1.4. Suppose that : G
3
GL
2
(F
3
) and
[
I3

_

0 1
_
is tr`es ramie. Both
3
and
3
are weakly acceptable for .
Lemma 2.1.5. Let i Z/3Z. Suppose that : G
3
GL
2
(F
3
) and

_

0 1
_
is tr`es ramie. The extended 3-type

i
simply admits .
Theorem 2.1.6. Let i Z/3Z. Suppose that : G
3
GL
2
(F
3
) and

_

0 1
_
is tr`es ramie. Then

i
is weakly acceptable for .
We remark that in Theorems 2.1.2, 2.1.4 and 2.1.6 we could replace weakly
acceptable by acceptable. This can be shown by using elliptic curves to construct
explicit liftings of the desired type. For Theorems 2.1.2 and 2.1.4 the results of [Man]
suce for this. For Theorem 2.1.6 a slightly more rened analysis along the lines
of 2.3 is required.
We also remark that it was Lemmas 2.1.1, 2.1.3, 2.1.5 and Conjecture 1.3.1 which
originally suggested to us that we try to prove Theorems 2.1.2, 2.1.4 and 2.1.6.
2.2. Applications. Conditional on the results stated in 2.1, which we will prove
below, we prove the following results.
Theorem 2.2.1. Any continuous absolutely irreducible representation : G
Q

GL
2
(F
5
) with cyclotomic determinant is modular.
Proof. Choose an element GL
2
(F
5
) with
3
= 1 but ,= 1. (The following
classication will be independent of the choice of .) Then up to equivalence and
twisting by a quadratic character, one of the following possibilities can be attained.
(1) is tamely ramied at 3.
(2) [
G3
is given by the character
Q

3
F
5
()

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

THE MODULARITY OF ELLIPTIC CURVES OVER Q 863
determined by
3
i
(
1
)
1 1
4 ,
where i Z/3Z.
(3) [
G
Q
3
(

1)
is given by the character
Q
3
(

1)

F
5
()

determined by
3 2
4

1 1
4 1
1 + 3

1 .
(4) [
G
Q
3
(

3)
is given by the character
Q
3
(

3)

F
5
()

determined by

3
1
1 1
4 1
1 +

3 .
(5) [
G
Q
3
(

3)
is given by the character
Q
3
(

3)

F
5
()

determined by

3
1
1 1
4 1
1 + 3

3 1
1 +

3 .
(6) [
G
Q
3
(

3)
is given by the character
Q
3
(

3)

F
5
()

determined by

3
1
1 1
4 1
1 + 3

3
1 +

3
i
,
where i Z/3Z.
To see that one of these cases can be attained, use the following facts, all of which
are easy to verify.
A subgroup of GL
2
(F
5
) with a non-trivial normal subgroup of 3-power order
is, up to conjugation, contained in the normaliser of F
5
()

.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
864 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
The intersection of SL
2
(F
5
) with the normaliser of in GL
2
(F
5
) is generated
by and an element such that
2
= 1 and
1
=
1
.
If F
5
()

, det = 3, and
1
= , then = (
1
).
In each case, we may choose an elliptic curve E
1/Q3
such that the representation

E1,5
of G
3
on E
1
[5](Q
3
) is isomorphic to [
G3
and such that the representation

E1,3
of G
3
on E
1
[3](Q
3
) has the following form (where we use the same numbering
as above).
(1) We place no restriction on
E1,3
.
(2) The restriction of
E1,3
to I
3
has the form
_
1
0
_
and is tr`es ramie. (Use Theorem 5.3.2 of [Man].)
(3) The restriction of
E1,3
to I
3
has the form
_
1
0
_
and is tr`es ramie. (Use Theorem 5.3.2 of [Man].)
(4) The restriction of
E1,3
to I
3
has the form
_

0 1
_
and is tr`es ramie. (Use 5.4 of [Man].)
(5) The restriction of
E1,3
to I
3
has the form
_

0 1
_
and is tr`es ramie. (Use 5.4 of [Man].)
(6)
E1,3
has the form
_

0 1
_
,
is tr`es ramie and remains indecomposable when restricted to the splitting
eld of . (Use Corollary 2.3.2 below.)
In each case choose such an E
1
and x an isomorphism : F
2
5

E
1
[5](Q
3
), such
that the Weil pairing on E
1
[5] corresponds to the standard alternating pairing on
F
2
5
, following the conventions in 1 of [SBT]. The pair (E
1
, ) denes a Q
3
-rational
point on the smooth curve denoted X

in [SBT]. We can nd a 3-adic open set

U X

(Q
3
) containing (E
1
, ) such that if (E
2
, ) denes a point in U, then
E
2
[3]

= E
1
[3] as F
3
[G
3
]-modules.
Using Ekedahls version of the Hilbert Irreducibility Theorem (see Theorem 1.3
of [E]) and the argument of 1 of [SBT] we may nd an elliptic curve E
/Q
and an
F
5
[G
Q
]-module isomorphism of with E[5](Q) such that (see also 2 of [Man])
under , the standard alternating pairing on F
2
5
and the Weil pairing on
E[5] agree;
the representation
E,3
of G
Q
on E[3](Q) is surjective onto Aut(E[3](Q));
and (E, ) denes a point of U.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 865
Corresponding to the six types of considered above, Proposition B.4.2 of [CDT]
ensures that the representation
E,3
of G
Q
on the 3-adic Tate module of E is
(1) either, up to quadratic twist, ordinary in the sense of Wiles [Wi] or poten-
tially Barsotti-Tate of some tamely ramied type;
(2) potentially Barsotti-Tate of type
1
;
(3) potentially Barsotti-Tate of type
1
;
(4) potentially Barsotti-Tate of type
3
;
(5) potentially Barsotti-Tate of type
3
;
(6) potentially Barsotti-Tate of extended type

i
.
In the rst case, E is modular by Theorem 7.2.1 of [CDT]. In the other cases we will
simply write for the type/extended type. We see that
E,3
(G
3
) has centraliser F
3
and the results of 2.1 show that admits
E,3
and that is weakly acceptable for

E,3
. Moreover
E,3
[
Gal(Q/Q(

3))
is absolutely irreducible and, by the Langlands-
Tunnell theorem (see [Wi]), modular. Thus by Theorems 1.4.1 and 1.4.2 we see
that
E,3
is modular. We deduce that E is modular, so

=
E,5
is modular.
Combining this theorem with Theorem 7.2.4 of [CDT] we immediately obtain
the following corollary.
Theorem 2.2.2. Every elliptic curve dened over the rational numbers is modular.
2.3. An extension of a result of Manoharmayum. The following facts follow
at once from [Man], particularly the classication given just before Theorem 5.4.2 of
that paper. Consider elliptic curves E over Q
3
with minimal Weierstrass equation
Y
2
= X
3
+AX +B, where
A B + 3 0 mod 9,
so
E,3
has the form
_

0 1
_
and is tr`es ramie. This leaves three possibilities
for the equivalence class of
E,3
. Fix in GL
2
(F
5
) with
3
= 1 but ,= 1. The
action of G
Q3(

3)
on E[5](Q
3
) is via a representation of the form

3 (
1
)
1 1
4 1
1 +

3
i
1 + 3

3 ,
for some = 1 and some i Z/3Z. All nine possibilities for the pair (
E,3
, i)
satisfying these conditions can arise for some such choice of A and B.
Lemma 2.3.1. With the above notation and assumptions, we have = 1.
Proof. Let F = Q
3
(

3, , ), where
2
=

3 and

3
+A +B = 9

3.
F is a totally ramied abelian extension of Q
3
(

3) of degree 6, with uniformiser

= /. The change of coordinates Y
15
Y , X
10
X + shows that E
has good reduction over F, and the reduction is isomorphic to
Y
2
= X
3
X 1.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
866 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
The arithmetic Frobenius of W
F
therefore has trace 3 on E[5]. Since
N
F/Q3(

3)
()

3(1 3

3) mod 9

3,
we conclude that
tr (
1
)
1
= 3,
so = 1.
Twisting by quadratic characters we immediately deduce the following corollary.
Corollary 2.3.2. Let
3
: G
3
GL
3
(F
3
) have the form
_

0 1
_
or
_
1
0
_
and be tr`es ramie. Let
5
: G
3
GL
2
(F
5
) have cyclotomic determinant and
restriction to G
Q3(

3)
given by a character
Q
3
(

3)

F
5
()

determined by

3 (
1
)
1 1
4 1
1 +

3
i
1 + 3

3 ,
for some i Z/3Z. There is an elliptic curve E
/Q3
, with E[3](Q
3
)
3
and
E[5](Q
3
)
5
. In particular, the action of I
3
on T
5
E factors through a nite
group and so E has potentially good reduction.
3. Admittance
In this section we will check Lemmas 2.1.1, 2.1.3 and 2.1.5. We freely use the
terminology introduced in 1.2 and 1.3.
3.1. The case of
1
. In this case
1
is the induction from U
0
(9) to GL
2
(Z
3
)
of a character of order 3. Its reduction modulo a prime above 3 has the same
Jordan-H older constituents as the reduction modulo 3 of Ind
GL2(Z3)
U0(9)
1. Using Brauer
characters, we see that the reduction modulo 3 of Ind
U0(3)
U0(9)
1 has Jordan-H older
constituents

0,0
,

0,0
and

1,1
. Thus,
1
admits
0,0
,
2,0
,
0,1
and
2,1
, the latter
two simply. Lemma 2.1.1 follows in this case.
3.2. The case of
1
. Let U denote the subgroup of GL
2
(Z
3
) consisting of ma-
trices
_
a b
c d
_
with a d mod 3 and b +c 0 mod 3, so
1
is the induction from U to GL
2
(Z
3
)
of a character of order 3. Upon reduction modulo a prime above 3 this will have the
same Jordan-Holder constituents as the reduction modulo 3 of Ind
GL2(Z3)
U
1. If
denotes the non-trivial character of F

3
and a character of F

9
of order 4, then this
latter induction splits up as the sum of the representations of GL
2
(Z
3
) GL
2
(F
3
)
denoted 1, sp

and () in 3.1 of [CDT]. By Lemma 3.1.1 of [CDT] we see that

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 867

1
admits
0,0
,
2,1
and
0,1
, the latter two simply. Lemma 2.1.1 follows in this
case.
3.3. The case of
3
. Let U denote the subgroup of GL
2
(Z
3
) consisting of ma-
trices
_
a b
c d
_
with a d mod 3 and c 0 mod 3. Then
3
is the induction from U to U
0
(3) of
a character of order 3. Upon reduction modulo a prime above 3 this will have the
same Jordan-Holder constituents as the reduction modulo 3 of Ind
U0(3)
U
1. Thus,

3
simply admits

0,0
and

1,1
. Lemma 2.1.3 follows.
3.4. The case of

i
. Let be the character of Q
3
(

3)

as in (2.1.1). Let be
a character of Q
3
with kernel Z
3
and which sends 1/3 to . If x (3

3)Z
3
[

3]
we have
(1 +x) = (tr
Q3(

3)/Q3
(x

3/54)).
We deduce that if

is the character used to dene

i
in 1.2, then

3) =
(
1
)
1
.
Let U denote the subgroup of GL
2
(Z
3
) consisting of matrices
_
a b
3c d
_
with a d mod 3 and b + c 0 mod 3. Let

U be the group generated by w
3
(see
(1.2.1)) and U, so

i
is the representation of

U
0
(3) induced from a character of

U
which sends w
3
to 1 and has order 3 when restricted to U. Thus, the Jordan-H older
constituents of the reduction of

i
modulo a prime above 3 are the same as the
Jordan-H older constituents of the reduction modulo 3 of Ind

U0(3)

U
1.
Let V denote the subgroup of GL
2
(Z
3
) consisting of matrices
_
a b
3c d
_
with a d mod 3. Let

V be the group generated by w
3
and V , and let denote
the character of

V /V which sends w
3
to 1. We have
Ind

U
(1) 1 Ind

V
V Q

3
,
where is a non-trivial character of V/U = (V Q

3
)/(UQ

3
). The reduction modulo
a prime above 3 of this (3-dimensional) representation has the same Jordan-H older
constituents as the reduction modulo 3 of 11. Thus,

i
admits

{0},1
,

{1},1
,

{0},1
and

{1},1
, the latter two simply. Lemma 2.1.5 follows.
4. New deformation problems
In this section we begin the proof of Theorems 2.1.2, 2.1.4 and 2.1.6. One could
approach this directly by using the results of [Br2] to attempt to describe R
D
V,O
(resp. R
D

V,O
). At least one of the authors of this paper (Taylor) thinks that such an
approach holds out more promise of attacking the non-acceptable case, and another
author (Breuil) has indeed made several computations along these lines. However
in the present case it seems to be easier to proceed less directly.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
868 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
To this end we will use ad hoc arguments to dene deformation problems, which
will be represented by O-algebras S such that
dim
k
m
S
/(
K
, m
2
S
) 1,
and the map R
V,O
R
D
V,O
(resp. R
D

V,O
) factors through S.
An important advantage of this approach is that to calculate m
S
/(
K
, m
2
S
) one
need only work in the category of nite at group schemes killed by a prime. Breuil
modules (see 5) for nite at group schemes killed by an odd prime are signicantly
simpler than the general case (of prime power torsion). This is particularly true
when we also use descent data. On the other hand, to suitably dene the new
deformation problems is rather delicate. That is what we will do in this section.
4.1. Some generalities on group schemes. In this section, and in 4.2, will
again be an arbitrary rational prime. Moreover R will denote a complete discrete
valuation ring with fraction eld F

of characteristic zero and perfect residue eld k

of characteristic . We will let denote a nite group of continuous automorphisms
of R and we will let F
0
denote the subeld of F

consisting of elements xed by

. Thus F

/F
0
will be nite and Galois with group . In our applications of these
results it suces to consider the case where F
0
is a nite extension of Q
3
(although
we will occasionally pass to the completion of the maximal unramied extension of
F).
Lemma 4.1.1. Let G be a nite at R-group scheme. Scheme theoretic closure
gives a bijection between subgroup schemes of GF

and nite at closed subgroup

schemes of G.
(See for instance 1.1 of [Co].)
Lemma 4.1.2. Let G
1
and G
2
be nite at group schemes over R which have local-
local closed bre. Suppose that G
1
and G
2
are the only nite at R-group schemes
with local-local closed bre which have generic bres G
1
F

and G
2
F

respectively.
Suppose also that we have an exact sequence of nite at R-group schemes
(0) G
1
G G
2
(0).
Then G is the unique nite at R-group scheme with local-local closed bre and with
generic bre G F.
Proof. Let G
+
and G

denote the maximal and minimal local-local models for GF.

The proof that these exist follows the proof of Proposition 2.2.2 of [Ra] and uses
the fact that the Cartier dual of a local-local nite at group scheme is local-local.
We must show that the canonical map G
+
G

is an isomorphism. The scheme-

theoretic closure of G
1
F in G

must be isomorphic to G
1
(by uniqueness), so we
have closed immersions G
1
G

extending G
1
F G

F. Similarly G

/G
1
must be isomorphic to G
2
. This gives a commutative diagram with exact rows:
0 G
1
G
+
G
2
0

0 G
1
G

G
2
0
The vertical maps G
1
G
1
and G
2
G
2
induce isomorphisms on the generic bre
and hence are isomorphisms. This is because some power of them is the identity on
the generic bre and hence is the identity. Working in the abelian category of fppf
abelian sheaves over Spec R, the middle map must also be an isomorphism.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 869
When G has -power order, we will let D(G) denote the classical (contravariant)
Dieudonne module of Gk. It is a W(k)-module equipped with a Frobenius operator
F and a Verschiebung operator V. We have FV = VF = and for all x W(k),
Fx = (Frob

x)F and Vx = (Frob

1

x)V.
If G is a nite at R-group scheme, then by descent data for G over F
0
we mean
a collection [g] of group scheme isomorphisms over R
[g] : G

g
G
for g such that for all g, h we have
[gh] = (
g
[h]) [g].
Note that this is not descent data in the sense of Grothendieck, since R/R

might
be ramied. However, Spec F

/ Spec F
0
is etale, so by etale descent we obtain a
nite at group scheme (G, [g])
F0
over F
0
together with an isomorphism
(G, [g])
F0

F0
F

= G
R
F

compatible with descent data. We also obtain a natural left action of on the
Dieudonne module D(G), semi-linear with respect to the W(k)-module structure
and commuting with F and V. We refer to the pair (G, [g]) as an R-group scheme
with descent data relative to F
0
. One denes morphisms of such objects to be
morphisms of R-group schemes which commute with the descent data. By a closed
nite at subgroup scheme with descent data we mean a closed nite at subgroup
scheme such that the descent data on the ambient scheme takes the subscheme to
itself. Quotients by such subobjects are dened in the obvious way. Thus we obtain
an additive category with a notion of short exact sequence. Suppose that G is a
nite at F
0
-group scheme. By a model with descent data (or simply model ) for G
over R we shall mean a triple (G, [g], i), where (G, [g]) is an R-group scheme
with descent data relative to F
0
and where i : (G, [g])
F0

G. Sometimes we will
suppress i from the notation. It is easy to check that isomorphism classes of models
admitting descent data for G over R form a sublattice of the lattice of models for
G
/F
over R. The following lemma follows without diculty from Lemma 4.1.1.
Lemma 4.1.3. Let F

/F
0
be a nite Galois extension as above, and let (G, [g])
be a nite at R-group scheme with descent data relative to F
0
. Base change from
F
0
to F

, followed by scheme theoretic closure, gives a bijection between subgroup

schemes of (G, [g])
F0
and closed nite at subgroup schemes with descent data in
(G, [g]).
We let FF
F
denote the category of nite at group schemes over R and FD
F

/F0
the category of nite at group schemes over R with descent data over F
0
. Let
W(k)[F, V][] denote the (non-commutative) W(k)-algebra generated by elements
F, V and [g] for g satisfying
[gh] = [g][h] for all g, h ;
[g]F = F[g] and [g]V = V[g] for all g ;
FV = VF = ;
[g]x = (gx)[g] for all x W(k) and g ;
Fx = (Frob

x)F and Vx = (Frob

1

x)V for all x W(k).

If I is a two-sided ideal in W(k)[F, V][], we will let FD
F

/F0,I
denote the full
subcategory of FD
F

/F0
consisting of objects (G, [g]) such that I annihilates D(G).
If (G, [g]) is an object of FD
F

/F0,I
and if (H, [g]) (G, [g]) is a closed nite
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
870 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
at subgroup scheme with descent data, then (H, [g]) and (G, [g])/(H, [g])
are again objects of FD
F

/F0,I
.
Lemma 4.1.4. For I a two-sided ideal of the ring W(k)[F, V][], choose objects
(G
1
, [g]) and (G
2
, [g]) in FD
F

/F0,I
so that (G
1
, [g])
F0

= (G
2
, [g])
F0
. Let G
denote the base change of this F
0
-group scheme to F

, so G has canonical descent

data relative to F

/F
0
. Then the sup and inf of G
1
and G
2
in the lattice of integral
models for G are stable under the descent data on G and with this descent data are
objects of FD
F

/F0,I
.
Proof. By uniqueness of the inf and sup, they are stable under the descent data
on the generic bre. It follows from Raynauds construction of the inf and sup
(Proposition 2.2.2 of [Ra]) in terms of subgroup schemes and quotients of G
1
G
2
that the sup and inf are objects of FD
F

/F0,I
.
Corollary 4.1.5. Let I be a two-sided ideal of the ring W(k)[F, V][]. Let
(0) G
1
G G
2
(0) (4.1.1)
be an exact sequence of nite at group schemes over F
0
. Let (G
1
, [g]) and
(G
2
, [g]) be objects of FD
F

/F0,I
such that (G
1
, [g])
F0

= G
1
and (G
2
, [g])
F0

=
G
2
. Suppose that for all objects (G, [g]) of FD
F

/F0,I
with (G, [g])
F0

= G, the
ltration on (G, [g]) induced by the ltration on G has subobject isomorphic to
(G
1
, [g]) and quotient isomorphic to (G
2
, [g]) (without any assumed compatibil-
ity with (4.1.1)). Then there is at most one model for G in FD
F

/F0,I
.
Proof. By Lemma 4.1.4, it suces to prove that if (G
+
, [g], i
+
) and (G

, [g], i

)
are two such models with a morphism between them, then the morphism between
them must be an isomorphism. In such a case we have a commutative diagram
with exact rows:
0 G
1
G
+
G
2
0

0 G
1
G

G
2
0
The vertical maps G
1
G
1
and G
2
G
2
induce isomorphisms on the generic bre
and hence are isomorphisms. This is because some power of them is the identity on
the generic bre and hence is the identity. Working in the abelian category of fppf
abelian sheaves over Spec R, the middle map must also be an isomorphism.
4.2. Filtrations. We keep the notation and assumptions of the previous section.
Let be a nite non-empty set of objects (G
i
, [g]) of FD
F

/F0,(I,)
. (Note the
in the subscript (I, ), which denotes the two-sided ideal generated by I and .)
Suppose that
Hom((G
i
, [g]), (G
j
, [g])) = Hom((G
i
, [g])
F0
, (G
j
, [g])
F0
)
=
_
0 if i ,= j,
nite eld if i = j
(4.2.1)
(in particular, the objects in are non-zero and pairwise non-isomorphic). By a -
ltration on a nite at F
0
-group scheme G we mean an increasing ltration Fil
j
G
such that for all j the graded piece Fil
j
G/ Fil
j1
G is isomorphic to (G
i(j)
, [g])
F0
for a (unique) (G
i(j)
, [g]) . The following lemma is proved by the standard
Jordan-H older argument.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 871
Lemma 4.2.1. If G is a nite at F
0
-group scheme which admits a -ltration and
if H is a quotient or subobject of G which admits a -ltration, then any -ltration
of H can be extended to a -ltration of G. In addition, all -ltrations of G have
the same length and the same set of successive quotients (with multiplicities).
We say that an object (G, [g]) of FD
F

/F0,I
is weakly ltered by if there is
some increasing ltration Fil
j
(G, [g]) of (G, [g]) by closed subobjects such that
for all j, the graded piece
Fil
j
(G, [g])/ Fil
j1
(G, [g])
is isomorphic to an element of . We say that an object (G, [g]) of FD
F

/F0,I
is
strongly ltered by if (G, [g]) is weakly ltered by and if for every -ltration
of (G, [g])
F0
the corresponding ltration of (G, [g]) satises
Fil
j
(G, [g])/ Fil
j1
(G, [g])
is isomorphic to an element of for all j. The following lemma follows at once
from the denitions and from Lemma 4.2.1.
Lemma 4.2.2. (1) If (G, [g]) and (G

, [g]) are objects of FD

F

/F0,I
which
are weakly ltered by , then (G, [g]) (G

, [g]) is also weakly ltered by

.
(2) Let (G, [g]) and (G

, [g]) be objects of FD
F

/F0,I
with (G

, [g]) a closed
subobject or quotient of (G, [g]). Suppose that (G, [g]) is strongly ltered
by and that (G

, [g])
F0
admits a -ltration. Then (G

, [g]) is strongly
ltered by .
If any object of FD
F

/F0,I
which is weakly ltered by is strongly ltered by
, then we will let FD
F

/F0,I,
denote the full subcategory of FD
F

/F0,I
consisting
of objects which are weakly (and therefore strongly) ltered by .
Lemma 4.2.3. Suppose that any object of FD
F

/F0,I
which is weakly ltered by
is strongly ltered by . Let G be a nite at F
0
-group scheme. If (G
1
, [g]) and
(G
2
, [g]) are two objects of FD
F

/F0,I,
with isomorphisms
i
j
: G

(G
j
, [g])
F0
for j = 1, 2, then there is a unique isomorphism
: (G
1
, [g])

(G
2
, [g])
such that on the generic bre i
2
= i
1
.
Proof. It follows from Raynauds construction of sup and inf that the sup and inf
of ((G
1
, [g]), i
1
) and ((G
2
, [g]), i
2
) are again objects of FD
F

/F0,I,
. Thus we
may suppose that there exists a map : (G
1
, [g]) (G
2
, [g]) such that on the
generic bre i
2
= i
1
. We will argue by induction on the rank of G that is an
isomorphism.
If (G
1
, [g]) is isomorphic to an element of , then the result follows by our
assumption on .
If (G
1
, [g]) is not isomorphic to an element of , then choose an exact sequence
(0) (G
11
, g) (G
1
, g) (G
12
, g) (0),
where (G
11
, g) and (G
12
, g) are weakly ltered by . Let (G
21
, [g]) de-
note the closed subobject of (G
2
, [g]) corresponding to (G
11
, [g])
F0
and dene
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
872 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
(G
22
, [g]) = (G
2
, [g])/(G
21
, [g]). Then we have a commutative diagram with
exact rows
0 G
11
G
1
G
12
0

0 G
21
G
2
G
22
0
compatible with descent data, where the central vertical arrow is and where
by inductive hypothesis the outside vertical arrows are isomorphisms. Working
in the abelian category of fppf abelian sheaves over Spec R, we see that is an
isomorphism.
The following lemma and its corollary give criteria for the equivalence of the
notions of being weakly ltered by and of being strongly ltered by .
Lemma 4.2.4. Fix I and as above. Suppose that for any pair of (possibly equal )
elements (G

, [g]) and (G

, [g]) in , the natural map

Ext
1
FD
F

/F
0
,(I,)
((G

, [g]), (G

, [g])) Ext
1
F

[GF
0
]
((G

, [g])
F0
, (G

, [g])
F0
)
is injective. Then any object (G, [g]) of FD
F

/F0,I
which is weakly ltered by is
also strongly ltered by .
Proof. For brevity, we say weakly/strongly ltered rather than weakly/strongly
ltered by since the data is xed for the entire proof. Also, we omit the
specication of descent data from the notation, but it should not be forgotten.
Suppose G is weakly ltered. In order to prove that G is strongly ltered, we argue
by induction on the length of a -ltration of G
F0
, this length being well-dened
by Lemma 4.2.1. The case of length 1 is clear. Otherwise, by the denition of
being weakly ltered, there is a short exact sequence of nite at R-group schemes
(with descent data relative to F
0
)
(0) G

G G

(0)
with G

and G

weakly ltered (and hence, by inductive hypothesis, strongly

ltered). Let H be any closed subgroup scheme of G (with compatible descent data
relative to F
0
) such that H
F0
G
i0,F0
for some G
i0
and such that (G/H)
F0
admits a -ltration. We need to prove (in the category of nite at group schemes
with descent data relative to F

/F
0
) that
H G
i0
,
and G/H is weakly ltered.
If the composite map
H G G

is zero, then H = G

as closed subgroup schemes of G (with descent data) and

likewise G/H = G

, so we are done. The interesting case is when the composite

map is non-zero. The map G
i0,F0

= H
F0
G

F0
is then non-zero and therefore
must be a closed immersion by the assumption (4.2.1) on and a devissage with
respect to a -ltration of G

F0
. We conclude that the map of generic bre etale
group schemes HF

is a closed immersion.
Taking scheme-theoretic closures, we obtain a closed subgroup scheme H

(with unique compatible descent data over F

0
) tting into a commutative diagram
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 873
of group schemes with descent data
H

H

(0) G

G G

(0)
in which the lower row is short exact, the vertical maps are closed immersions and
the top map H H

induces an isomorphism on generic bres. By Lemma 4.2.1

we may extend H
F0
G

F0
to a -ltration on G

F0
and so, because G

is strongly
ltered by induction, we may extend H

to a -ltration. In particular H

is isomorphic to an object in and G

/H

is strongly ltered.
Pulling back the short exact sequence
0 G

G G

0
by H

, we get a diagram
H

0 G

G
G
H

0
in which the row is a short exact sequence of fppf abelian sheaves and all of the
terms are nite at group schemes (for the middle, this follows from the atness
of G G

). Thus, this bottom row is a short exact sequence of nite at group

schemes (with descent data). As H
F0

F0
, the sequence
0 G

F0
(G
G
H

)
F0
H

F0
0
is split. In particular (G
G
H

)
F0
and hence G
G
H

are killed by . By the

hypothesis of the lemma
0 G

G
G
H

0
is also split, i.e. we have an isomorphism
G
G
H

= G

R
H

such that G

G
G
H

corresponds to injection to the rst factor of G

R
H

and G
G
H

corresponds to projection onto the second factor. By our

hypotheses on we can nd a morphism : H

extending
H

F0

H
F0
G

F0
H

F0
pr
G

F0
.
Then our closed immersion H G

R
H

factors as
H H

1
G

R
H

.
As H G

R
H

is a closed immersion, : H H

must be a closed immersion

and hence an isomorphism. Thus H is isomorphic to an object in .
Now we turn to the proof that G/H is weakly ltered. Since : H H

is an
isomorphism, it is clear that the natural map
H
R
G

G
G
H

is an isomorphism, and hence that

H
R
G

G
is a closed immersion. Thus, the nite at group scheme G/(H G

) makes sense
and the natural map
G/(HG

) G

/H

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

874 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
is an isomorphism (as one sees by using the universal properties of quotients to
construct an inverse map). We therefore arrive at a short exact sequence
0 G

G/H G

/H

0
(compatible with descent data). Since G

/H

is strongly ltered, as we noted

above, and G

, it follows that G/H is weakly ltered.

Corollary 4.2.5. Fix I and as above. Suppose that = (G, [g]) is a single-
ton. Suppose also that we have a short exact sequence
(0) (G
1
, [g]) (G, [g]) (G
2
, [g]) (0)
in FD
F

/F0,I
, where for any i, j (possibly equal )
Hom((G
i
, [g]), (G
j
, [g])) = Hom((G
i
, [g])
F0
, (G
j
, [g])
F0
)
=
_
0 if i ,= j,
nite eld if i = j,
and the natural map
Ext
1
FD
F

/F
0
,(I,)
((G
i
, [g]), (G
j
, [g])) Ext
1
F

[GF
0
]
((G
i
, [g])
F0
, (G
j
, [g])
F0
)
(4.2.2)
is injective. Then any object (H, [g]) of FD
F

/F0,I
which is weakly ltered by
is also strongly ltered by .
Proof. As (H, [g]) is weakly ltered by , it is weakly ltered by (G
1
, [g]),
(G
2
, [g]), and so by Lemma 4.2.4 is strongly ltered by (G
1
, [g]), (G
2
, [g]).
Any -ltration of (H, [g])
F0
extends to a (G
1
, [g]), (G
2
, [g])-ltration of
(H, [g])
F0
, which in turn gives rise to a (G
1
, [g]), (G
2
, [g])-ltration of
(H, [g]). By the injectivity of (4.2.2) we see that this yields a -ltration of
(H, [g]) that induces to our chosen -ltration of (H, [g])
F0
.
4.3. Generalities on deformation theory. Again in this section denotes an
arbitrary rational prime. We let K denote a nite extension of Q

, O the ring of
integers K,
K
the maximal ideal of O and k its residue eld. Note that k has
a dierent meaning from the previous two sections. Let V be a two-dimensional
k-vector space and : G

Aut
k
(V ) a continuous representation. Suppose that
the centraliser of G

in End
k
(V ) is k. Let : G

denote a continuous
character such that ( mod
K
)

= det . Let S() denote the full subcategory of
the category of nite length (discrete) O-modules with a continuous O-linear action
of G

consisting of objects which admit a nite ltration so that each successive

quotient is isomorphic to V . Because End
k[G

]
(V ) = k, it follows from the usual
Jordan-H older argument that S() is an abelian category.
Let S be a full subcategory of S() stable under isomorphisms and which is closed
under nite products, S()-subobjects and S()-quotients, and which contains V .
We will consider the following set-valued functors on the category of complete
noetherian local O-algebras R with nite residue eld k.
D
V,O
(R) is the set of conjugacy classes of continuous representations :
G

GL
2
(R) such that mod m
R
is conjugate to .
D

V,O
(R) is the set of conjugacy classes of continuous representations :
G

GL
2
(R) such that mod m
R
is conjugate to and det = .
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 875
D
S
V,O
(R) is the set of conjugacy classes of continuous representations :
G

GL
2
(R) such that mod m
R
is conjugate to and such that for each
open ideal a R the action makes (R/a)
2
into an object of S.
D
,S
V,O
(R) is the set of conjugacy classes of continuous representations :
G

GL
2
(R) such that mod m
R
is conjugate to , such that det = ,
and such that for each open ideal a R the action makes (R/a)
2
into an
object of S.
Each of these deformation problems is representable by objects which we will denote
R
V,O
, R

V,O
, R
S
V,O
and R
,S
V,O
, respectively.
Recall that the following sets are in natural (k-linear) bijection with each other.
(m
R
V,O
/(
K
, m
2
R
V,O
))

.
The set of deformations of to k[]/(
2
).
Ext
1
k[G

]
(V, V ).
H
1
(G

, ad ).
These bijections give rise to an isomorphism
(m
R

V,O
/(
K
, m
2
R

V,O
))

= H
1
(G

, ad
0
),
as well as bijections between
(m
R
S
V,O
/(
K
, m
2
R
S
V,O
))

,
the set of deformations of to k[]/(
2
) which make (k[]/(
2
))
2
into an
object of S,
Ext
1
k[G

],S
(V, V ), i.e. Ext
1
in the category of discrete k[G

]-modules which
are also objects of S,
the subgroup H
1
S
(G

, ad) H
1
(G

, ad) corresponding to Ext

1
k[G

],S
(V, V ).
We will set H
1
S
(G

, ad
0
) = H
1
S
(G

, ad) H
1
(G

, ad
0
), so that we get an iso-
morphism
(m
R
,S
V,O
/(
K
, m
2
R
,S
V,O
))

= H
1
S
(G

, ad
0
).
4.4. Reduction steps for Theorem 2.1.2. We now begin the proof of Theorem
2.1.2. Making an unramied twist we may suppose that has the form
_
1
0
_
.
We may also suppose that O = Z
3
.
Let F
1
= F

1
denote a totally ramied cubic Galois extension of Q
3
. Let F

1
denote the unique cubic extension of Q
3
(

1) such that F

1
/Q
3
is Galois but not
abelian, and let F
1
denote a cubic subeld of F

1
, so F

1
/F
1
is unramied.
Let S
1
denote the full subcategory S() consisting of Z
3
[G
3
]-modules X for
which there exists a nite at O
F

1
-group scheme (G, [g]) with descent data for
F

1
/Q
3
such that X

= (G, [g])
Q3
(Q
3
) as a Z
3
[G
3
]-module. By Lemma 4.1.3 we
see that S
1
is closed under nite products, subobjects and quotients. Using Tates
theorem on the uniqueness of extensions of 3-divisible groups from F

1
to O
F

1
(Theorem 4 of [T]), we see that the map R
V,Z3
R
1
V,Z3
factors through R
,S1
V,Z3
.
Thus, Theorem 2.1.2 follows from the following result which we will prove in 7.
Theorem 4.4.1. dimH
1
S1
(G
3
, ad
0
) 1.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
876 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
4.5. Reduction steps for Theorem 2.1.4. We now begin the proof of Theorem
2.1.4. Making an unramied twist, we may suppose that has the form
_

0 1
_
.
We may also suppose that O = Z
3
.
Let F

3
denote the degree 12 abelian extension of Q
3
(

3) with norm subgroup

in Q
3
(

3)

topologically generated by 3, 4 and 1 + 3

3. Note that F

3
/Q
3
is Galois. We have an isomorphism
Gal(F

3
/Q
3
(

3))

= C
2
C
2
C
3
.
Let
2
4
I
F

3
/Q3(

3)
be the unique element of order 2. (In later applications
this will be the square of an element of order 4 in Gal(F

3
/Q
3
).) We also let F
3
denote the xed eld of a Frobenius lift of order 2, so F
3
/Q
3
is totally ramied.
We will let I
3
denote the two-sided ideal of W(F
9
)[F, V][Gal(F

3
/Q
3
)] gener-
ated by
F+V,
and [
2
4
] + 1.
Let S
3
denote the full subcategory of S() consisting of objects X for which we
can nd an object (G, [g]) of FD
F

3
/Q3,I3
such that X

= (G, [g])
Q3
(Q
3
) as a
Z
3
[G
3
]-module. By Lemma 4.1.3, we see that S
3
is closed under nite products,
subobjects and quotients.
Now choose a nite extension K/Q
3
and continuous map of rings f : R
V,Z3
Q
3
such that the corresponding representation : G
3
GL
2
(O
K
) is of type
3
. Let
G be the corresponding 3-divisible group over Q
3
. By Tates theorem (Theorem
4 of [T]), the base change of G to F

3
has a unique extension to a 3-divisible
group G over O
F

3
. By the uniqueness of this extension, it is also equipped with
descent data [g] relative to F

3
/Q
3
and with an action of O
K
, compatible with
the canonical structure on the generic bre.
Let
2
Gal(Q
3
(

3)
ab
/Q
3
(

3)) correspond to

3. We will use the no-
tation of Appendix B of [CDT] (in particular WD and D

(G)), except that we will

write F and F

in place of and

. Then
WD()(
2
4
) = 1,
WD()(
2
2
), but not WD()(
2
), is a scalar,
and det WD()(
2
) = 3.
Thus WD()(
2
2
) = 3. Hence on D

(G) Q
3
we have
[
2
4
] = WD()(
2
4
) = 1,
and (F

)
2
= [
2
2
] WD()(
2
2
) = 1/3.
We conclude that on D(G) we have
[
2
4
] = 1,
F
2
= 3,
and so F = V.
In particular I
3
annihilates D(G) and for all m 1 the map (f mod 3
m
) : R
V,Z3

O
K
/(3
m
) factors through R
,S3
V,Z3
. Hence, the map R
V,Z3
R
3
V,Z3
factors through
R
,S3
V,Z3
and Theorem 2.1.4 follows from the following result which we will prove in
8.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 877
Theorem 4.5.1. dimH
1
S3
(G
3
, ad
0
) 1.
4.6. Reduction steps for Theorem 2.1.6. We now begin the proof of Theorem
2.1.6. We may suppose that O = Z
3
.
Let F

i
denote the degree 12 abelian extension of Q
3
(

3) with norms the

subgroup of Q
3
(

3)

topologically generated by 3, 4, 1 + 9

3 and 1 +
(1 3)

3, where is the unique lift of i to Z with 0 < 3. Note that

F

i
/Q
3
is Galois. We identify
Gal(F

i
/Q
3
(

3))

=
2
)
3
)
2
4
),
where
2
corresponds to

3 and has order 2,
3
corresponds to 1 3

3 and
has order 3, and
2
4
corresponds to 1 and has order 2. We also let F
i
denote the
xed eld of 1,
2
, so F
i
/Q
3
is totally ramied.
We will let I
i
denote the two-sided ideal of W(F
9
)[F, V][Gal(F

i
/Q
3
)] generated
by
F+V,
[
2
4
] + 1,
and ([
3
] [
1
3
])[
2
] F.
We remark that the ideal I
i
is unchanged if we change our choice of

3.
In 9 we will prove the following result (and explain the unusual looking nota-
tion).
Theorem 4.6.1. There are objects (G, [g])
(2,6)
, (G, [g])
(6,10)
, (G, [g])
(2,10)
and (G, [g])
(6,6)
in the category FD
F

i
/Q3,Ii
with the following properties.
(1) For (r, s)=(2, 6), (6, 10), (2, 10) and (6, 6) we have

= ((G, [g])
(r,s)
)
Q3
(Q
3
)
as G
3
-modules.
(2) For (r, s) = (2, 6), (6, 10), (2, 10) and (6, 6) there is a short exact sequence
in FD
F

i
/Q3,Ii
,
(0) (G
1
, [g])
(r,s)
(G, [g])
(r,s)
(G
2
, [g])
(r,s)
(0),
such that (G
1
, [g])
(r,s)
and (G
2
, [g])
(r,s)
have order 3 and for all a, b
1, 2 (possibly equal ) the natural map
Ext
1
FD
F

/Q
3
,(I
i
,3)
((G
a
, [g])
(r,s)
, (G
b
, [g]))
(r,s)
Ext
1
F3[GQ
3
]
((G
a
, [g])
(r,s),Q3
, (G
b
, [g])
(r,s),Q3
)
is injective.
(3) If k/F
3
is a nite eld extension and if (G, [g]) is an object of FD
F

i
/Q3,Ii
with an action of k such that (G, [g])
Q3
(Q
3
) is isomorphic to k, then
for some (r, s) = (2, 6), (6, 10), (2, 10) or (6, 6) the object (G, [g]) of
FD
F

i
/Q3,Ii
is weakly ltered by (G
1
, [g])
(r,s)
, (G
2
, [g])
(r,s)
.
(4) For (r, s) = (2, 6), (6, 10) and (2, 10) we have F = 0 on D(G
(r,s)
), while
F ,= 0 on D(G
(6,6)
).
Note that for all a, b (possibly equal), we must have
Hom((G
a
, [g])
(r,s)
, (G
b
, [g])
(r,s)
) = Hom((G
a
, [g])
(r,s),Q3
, (G
b
, [g])
(r,s),Q3
)
=
_
0 if a ,= b,
F
3
if a = b.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
878 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
For (r, s) = (2, 6), (6, 10), (2, 10) and (6, 6), we let S
i,(r,s)
denote the full subcat-
egory of S() consisting of objects X which are isomorphic to (H, [g])
Q3
for some
object (H, [g]) of FD
F

i
/Q3,Ii,{(G,{[g]})
(r,s)
}
. By Lemma 4.2.2, Corollary 4.2.5 and
Theorem 4.6.1 we see that S
i,(r,s)
is closed under nite products, S()-subobjects
and S()-quotients. In 9 we will also prove the following two results.
Theorem 4.6.2. For (r, s) = (2, 6), (6, 10), (2, 10) and (6, 6) we have
dimH
1
S
i,(r,s)
(G
3
, ad
0
) 1.
Theorem 4.6.3. For (r, s) = (2, 6), (6, 10) and (2, 10) and for any N 1 there
exists a continuous representation

N
: G
Q3
GL
2
(F
3
[[T]]/(T
N
))
such that
det
N
= ,
for some object (G
N
, [g]) of FD
F

i
/Q3,(Ii,F),{(G,{[g]})
(r,s)
}
we have

N

= (G
N
, [g])
Q3
(Q
3
)
(where (I
i
, F) denotes the two-sided ideal of W(F
9
)[F, V][Gal(F

i
/Q
3
)] gen-
erated by I
i
and F),
and mod (T
2
) ,

= k[[T]]/(T
2
).
(We are not asserting that
N
and G
N
are independent of the choice of (r, s),
though in fact we believe that
N
is independent of this choice.)
From these results we can easily draw the following consequence.
Corollary 4.6.4. For (r, s) = (2, 6), (6, 10) and (2, 10) we have
R
,S
i,(r,s)
V,Z3

= F
3
[[T]].
Proof. By Theorems 4.6.2 and 4.6.3 we see that R
,S
i,(r,s)
V,Z3
/(3)

= F
3
[[T]] and that if
R is an Artinian quotient of R
,S
i,(r,s)
V,Z3
/(3) corresponding to a (necessarily unique;
see Lemma 4.2.3) object (G, [g]) of FD
F

i
/Q3,(Ii,3),{(G,{[g]})
(r,s)
}
, then F = 0 on
D(G).
Now suppose R is any Artinian quotient of R
,S
i,(r,s)
V,Z3
which corresponds to an
object (G, [g]) of the category FD
F

i
/Q3,Ii,{(G,{[g]})
(r,s)
}
. Let G = (G, [g])
Q3
and
consider the exact sequences
(0) G[3] G 3G (0)
and
(0) 3G G G/3G (0).
By Lemma 4.2.3, we have exact sequences
(G, [g]) (K, [g]) (0)
and
(0) (K, [g]) (G, [g]) (H, [g]) (0)
in FD
F

i
/Q3,Ii,{(G,{[g]})
(r,s)
}
such that the composite
(G, [g]) (K, [g]) (G, [g])
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 879
is multiplication by 3. In particular we have exact sequences
(0) D(K) D(G)
and
(0) D(H) D(G) D(K) (0),
such that the composite
D(G) D(K) D(G)
is multiplication by 3. As F = V = 0 on D(H) we see that F and V factor
through maps D(K) D(G), i.e. we can write F = 3F

and V = 3V

for some
endomorphisms F

and V

of D(G). Thus 3 = 9F

equals zero on D(G)/9D(G)

and so D(K) = 0. We conclude that K = (0), so that 3G = (0) and 3R = (0).
Thus
R
,S
i,(r,s)
V,Z3
= R
,S
i,(r,s)
V,Z3
/(3) = F
3
[[T]].
We now modify the argument in 4.4. Choose a nite extension K/Q
3
and
continuous map of rings f : R
V,Z3
Q
3
such that the corresponding representation
: G
3
GL
2
(O
K
) is of extended type

i
. Let G be the corresponding 3-divisible
group over Q
3
. By Tates theorem (Theorem 4 of [T]) G has a unique extension
to a 3-divisible group G over O
F

i
. By the uniqueness of this extension, G comes
equipped with descent data [g] relative to F

i
/Q
3
and with an action of O
K
,
compatible with the canonical structure on the generic bre.
Let
2
Gal(Q
3
(

3)
ab
/Q
3
(

3)) correspond to

3. We will use the no-
tation of Appendix B of [CDT] (in particular WD and D

(G)), except that we will

write F and F

in place of and

. Then
WD()(
2
4
) = 1,
WD()(
2
2
) = 3,
and WD()(
2
)(WD()(
3
) WD()(
3
)
1
) = 3.
Thus on D

(G) Q
3
we have
[
2
4
] = WD()(
2
4
) = 1,
(F

)
2
= [
2
2
] WD()(
2
2
) = 1/3,
and [
2
]([
3
] [
1
3
]) = 3F

.
We conclude that on D(G) we have
[
2
4
] = 1,
F
2
= 3,
and [
2
]([
1
3
] [
3
]) = 3F
1
.
Hence also
F = V,
and [
2
]([
3
] [
1
3
]) = F.
In particular I
i
annihilates D(G).
Thus (G[
K
], [g]) is an object of FD
F

i
/Q3,Ii
such that (G[
K
], [g])
Q3
corre-
sponds to O
K
/
K
. By Theorem 4.6.1 we see that (G[
K
], [g]) is weakly ltered
by (G
1
, [g])
(r,s)
, (G
2
, [g])
(r,s)
for some (r, s) = (2, 6), (6, 10), (2, 10) or (6, 6).
We will prove (r, s) = (6, 6). By Theorem 4.6.1 and Lemma 4.2.4, (G[
K
], [g])
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
880 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
is strongly ltered by (G
1
, [g])
(r,s)
, (G
2
, [g])
(r,s)
. As (G[
K
], [g])
Q3
is l-
tered by , using Theorem 4.6.1, we see that (G[
K
], [g]) is weakly ltered by
(G, [g])
(r,s)
. For all m 1 we have
(G[
m
K
]/G[
m1
K
], [g])

(G[
K
], [g]),
so for all m 1 the object (G[
m
K
], [g]) is also weakly and hence strongly ltered
by (G, [g])
(r,s)
for the same (r, s). Thus, for all m 1, the map (f mod p
m
) :
R
V,Z3
O
K
/(3
m
) factors through R
,S
i,(r,s)
V,Z3
. By Corollary 4.6.4 we see that
(r, s) = (6, 6), so the map R
V,Z3
R

i
V,Z3
factors through R
,S
i,(6,6)
V,Z3
and Theo-
rem 2.1.6 follows from Theorem 4.6.2.
4.7. Some Galois cohomology. In this section we will begin the proofs of The-
orems 4.4.1, 4.5.1 and 4.6.2. We will let S denote one of the categories S
1
, S
3
or
S
i,(r,s)
. We will let = in the cases S
1
and = 1 otherwise. In all cases

_

0
_
is tr`es ramie.
The maps and induce a commutative diagram with exact rows
and columns:
(0)

Ext
1
F3[G3]
( , )

Ext
1
F3[G3]
(, ) Ext
1
F3[G3]
( , )

(0) Ext
1
F3[G3]
(, ) Ext
1
F3[G3]
(, ) Ext
1
F3[G3]
( , )
We will let
0
denote the composite map
Ext
1
F3[G3]
(, ) Ext
1
F3[G3]
( , ),
and
1
(resp.

) the induced mapping

ker
0
Ext
1
F3[G3]
(, )
(resp.
ker
0
Ext
1
F3[G3]
( , )).
We will also let
1
(resp.

) denote the induced mapping

ker
0
Ext
1
F3[G3]
(, ) Ext
1
F3[I3]
(, )
(resp.
ker
0
Ext
1
F3[G3]
( , ) Ext
1
F3[I3]
( , )).
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 881
If we reinterpret our Ext-groups as cohomology groups and use the isomorphism

, our diagram becomes:

(0)

H
1
(G
3
, F
3
)

H
1
(G
3
, ad ) H
1
(G
3
, )

(0) H
1
(G
3
, F
3
) H
1
(G
3
, ) H
1
(G
3
, )
Fix a basis of F
2
3
so that takes the form
_

0
_
.
Then any extension of by in characteristic 3 may be represented by a matrix
_

0
_
,
where the cocycle
=
_

11

12

21

22
_
Z
1
(G
3
, ad )
represents the class of this extension in Ext
1
F3[G3]
(, )

= H
1
(G
3
, ad ). Moreover

0
([]) = [
21
] H
1
(G
3
, ),
if
21
=0, then
1
([])=[
22
] H
1
(G
3
, F
3
) and

([])=[
11
] H
1
(G
3
, F
3
),
and [] H
1
(G
3
, ad
0
) if and only if 0 = [
11
+
22
] H
1
(G
3
, F
3
).
In particular we have
1
=

on H
1
(G
3
, ad
0
) ker
0
.
We have an exact sequence
(0) ad
0
(0),
where the rst map sends
_
x
y
_

_
y/2 x
0 y/2
_
and the second map sends
_
a b
c a
_
c.
Thus we get an exact sequence
(0) H
1
(G
3
, ) H
1
(G
3
, ad
0
)
0
H
1
(G
3
, )
and so we may identify H
1
(G
3
, ad
0
) ker
0
with H
1
(G
3
, ). We also have an
exact sequence
(0) 1 (0), (4.7.1)
which gives rise to an exact sequence
(0) F
3
H
1
(G
3
, ) H
1
(G
3
, ) H
1
(G
3
, F
3
) H
2
(G
3
, ).
If we identify H
1
(G
3
, ) with H
1
(G
3
, ad
0
) ker
0
, then the latter map
H
1
(G
3
, ) H
1
(G
3
, F
3
) is identied with

=
1
.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
882 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
Lemma 4.7.1. The sequence
(0) F
3
H
1
(G
3
, ) H
1
(G
3
, ) H
1
(I
3
, F
3
)
is exact.
Proof. The key point is that is tr`es ramie (compare with Proposition 6.1 of
[Di1]). It suces to show that the composite
H
1
(G
F3
, F
3
) H
1
(G
3
, F
3
) H
2
(G
3
, )
is injective. Suppose that x H
1
(G
3
, F
3
) maps to zero in H
2
(G
3
, ); then by Tate
duality x is annihilated by the image of the map H
0
(G
3
, F
3
) H
1
(G
3
, ) coming
from the short exact sequence
(0) ( )

1 (0)
Cartier dual to (4.7.1). As ( )

is tr`es ramie we see that the image of

H
0
(G
3
, F
3
) H
1
(G
3
, )

= Q

3
/(Q

3
)
3
is not contained in Z

3
/(Z

3
)
3
. Thus
x Hom(Q

3
/Z

3
, F
3
)

= H
1
(G
F3
, F
3
) H
1
(G
3
, F
3
)

= Hom(Q

3
, F
3
)
must be zero (see Proposition 3 of 1 of Chapter XIV of [Se1]).
Corollary 4.7.2. The maps

1
: H
1
(G
3
, ad
0
) ker
0
H
1
(I
3
, F
3
)
and

: H
1
(G
3
, ad
0
) ker
0
H
1
(I
3
, F
3
)
have the same kernel and this has dimension 1 over F
3
.
Theorems 4.4.1, 4.5.1 and 4.6.2 now follow from the following results, which
we will prove later. One advantage of these new formulations is that, with one
exception, they refer only to Ext
1
S
(, ) and make no mention of the determinant
or ad
0
, concepts which we found tricky to translate into the language of integral
models.
Theorem 4.7.3. (1)
0
: Ext
1
S1
(, ) H
1
(G
3
, ) is the zero map.
(2)

: Ext
1
S1
(, ) H
1
(I
3
, F
3
) is the zero map.
(3)

: H
1
S1
(G
3
, ad
0
) H
1
(I
3
, F
3
) is the zero map.
Theorem 4.7.4. (1)
0
: Ext
1
S3
(, ) H
1
(G
3
, ) is the zero map.
(2)

: Ext
1
S3
(, ) H
1
(I
3
, F
3
) is the zero map.
Theorem 4.7.5. Suppose that i Z/3Z and (r, s) = (2, 6), (6, 10), (2, 10) or
(6, 6).
(1)
0
: Ext
1
S
i,(r,s)
(, ) H
1
(G
3
, ) is the zero map.
(2) Either

: Ext
1
S
i,(r,s)
(, ) H
1
(I
3
, F
3
) or
1
: Ext
1
S
i,(r,s)
(, ) H
1
(I
3
, F
3
)
is the zero map.
The deduction of Theorems 4.4.1, 4.5.1 and 4.6.2 from these results is immediate.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 883
5. Breuil modules
In this section we recall some results from [Br2] (see also the summary [Br1]) and
give some slight extensions of them. Three of the authors apologise to the fourth
for the title of this section, but they nd that the term Breuil module is much
more convenient than ltered
1
-module.
Throughout this section, will be an odd rational prime and R will be a com-
plete discrete valuation ring with fraction eld F

of characteristic zero and perfect

residue eld k of characteristic .
5.1. Basic properties of Breuil modules. We will x a choice of uniformiser
of R and let
E

(u) = u
e
G

(u)
be the Eisenstein polynomial which is the minimal polynomial of over the fraction
eld of W(k), so G

(u) W(k)[u] is a polynomial with unit constant term G

(0)
W(k)

(and degree at most e 1). The

th
power map on k[u]/u
e
is denoted ,
and we dene
c

= (G

(u)) (k[u]/u
e
)

. (5.1.1)
It is very important to keep in mind that these denitions, as well as many of the
denitions below, depend on the choice of the uniformiser .
The category of -torsion Breuil modules (or -torsion Breuil modules over R,
or simply Breuil modules or Breuil modules over R) is dened to be the cate-
gory of triples (M, M
1
,
1
), where
M is a nite free k[u]/u
e
-module,
M
1
is a k[u]/u
e
-submodule of M containing u
e
M,

1
: M
1
M is -semi-linear and has image whose k[u]/u
e
-span is all of
M.
(A morphism (M, M
1
,
1
) (N, N
1
,
1
) is a morphism f : M N of k[u]/u
e
-
modules such that f M
1
N
1
and
1
f = f
1
on M
1
.) We dene the rank of
(M, M
1
,
1
) to be the rank of M over k[u]/u
e
. Breuil modules form an additive
category (not abelian in general) in the obvious manner and this category does not
depend on the choice of . It is denoted
1
mod
R
or
1
mod
F

. The induced
-semi-linear map of k-vector spaces

1
: M
1
/u M
1
M/u M
is bijective (because it is onto and #M
1
/u M
1
= #M
1
[u] #M[u] = #M/u M).
In particular, a map of Breuil modules
(M, M
1
,
1
) (M

, M

1
,

1
)
is an isomorphism if and only if the map MM

on underlying k[u]/u
e
-modules
is an isomorphism.
Lemma 5.1.1. Suppose that
0 M

MM

0
is a complex of Breuil modules. The following are equivalent.
(1) The underlying sequence of k[u]/u
e
-modules is exact.
(2) The underlying sequence of k[u]/u
e
-modules is exact as is the sequence
0 M

1
M
1
M

1
0.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
884 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
(3) The complex of vector spaces
0 M

/u M/u M

/u 0
is exact.
Proof. The second statement clearly implies the rst. The rst implies the third as
Breuil modules are free over k[u]/u
e
. It remains to show that the third condition
implies the second. Using Nakayamas lemma and the freeness of Breuil modules
we see that
0 M

MM

0
is an exact sequence of k[u]/u
e
-modules. Using the bijectivity of
1
, we see that
the natural map
f
1
: M
1
M

1
is surjective modulo u and therefore is surjective. It remains to check that the
inclusion of k[u]/u
e
-modules M

1
ker(f
1
) is an equality. Since f
1
is compatible
with f : MM

via the inclusions M

1
M, M

1
M

and also via the maps

1
and

1
, it is obvious that ker(f
1
) ker(f) = M

and that
1
(ker(f
1
)) M

. Since
ker(f
1
) contains M

1
, which in turn contains u
e
M

, we see that (M

, ker(f
1
),
1
) is
a Breuil module! Then (M

, M

1
,

1
) (M

, ker(f
1
),
1
) dened via the identity
map on M

is a map of Breuil modules which is an isomorphism on underlying

k[u]/u
e
-modules, so it must be an isomorphism of Breuil modules. This forces
ker(f
1
) = M

1
.
When the equivalent conditions of this lemma are met we call the sequence of
Breuil modules
0 M

MM

0
exact.
For any Breuil module (M, M
1
,
1
), we dene the Frobenius endomorphism :
MM by
(m) =
1
c

1
(u
e
m), (5.1.2)
where c

is dened as in (5.1.1). Note that this depends on our choice of uniformiser.

We let N : W(k)[[u]] W(k)[[u]] denote the unique continuous W(k)-linear
derivation satisfying Nu = u, i.e. N = u
d
du
. This operator extends to any Breuil
module. More precisely, we have the following lemma.
Lemma 5.1.2. Let M be an object of
1
mod
R
. There is a unique additive opera-
tor N : MM(the monodromy operator) satisfying the following three conditions:
(1) N(sx) = N(s)x +sN(x), s k[u]/u
e
, x M,
(2) N
1
= N on M
1
,
(3) N(M) uM.
Moreover, any morphism of Breuil modules MM

automatically commutes with

N.
Proof. Lets start with unicity. Recall we have an isomorphism k[u]/u
e

k[u

]/u
e

1
(M
1
)

M([Br2], 2.1.2.1). Suppose there are two operators N and N

satisfying
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 885
(1), (2) and (3) above, so = N N

is k[u]/u
e
-linear and satises
1
=
and (M) uM. Thus,

1
(M
1
) = (M
1
) (uM) u

M,
so (M) = (k[u]/u
e

k[u

]/u
e
1
(M
1
)) u

M. Iterating
1
(M
1
) (M)
u

2
M so (M) u

2
M, and so on. As u
e
= 0, we get = 0. For the existence,
let N
0
= N 1 on
k[u]/u
e

k[u

]/u
e
1
(M
1
) M,
and note N
0
satises N
0
(sx) = N(s)x + sN
0
(x). Call a derivation of M any
additive operator satisfying this relation and dene successive derivations of M by
the formula
N
j+1
(s
1
(x)) = N(s)
1
(x) +s(N
j
(x)),
for j 0. Note that N
j+1
is well dened by the following observations.
N(u

s) = u

N(s) and N
j
(ux) = ux+uN
j
(x) imply that N
j+1
(u

s
1
(x)) =
N
j+1
(s
1
(ux)).
If
1
(x) = 0, then x u
e
M (see (1) of Lemma 2.1.2.1 of [Br1]) and so
N
j
(x) u
e
M and (N
j
(x)) = 0.
As N
0
(M) uM, we have (N
j+1
N
j
)(M) u

j+1
M, so N
j
= N
j+1
for j 0.
This N
j
satises (1), (2) and (3).
The reason for introducing Breuil modules (and putting the factor c
1

in the
denition of ) is the following theorem.
Theorem 5.1.3. (1) Given the choice of uniformiser for R there is a con-
travariant functor M

from nite at R-group schemes which are killed by

to
1
mod
R
and a quasi-inverse functor G

.
(2) If G is a nite at R-group scheme killed by , then G has rank
r
if and only
if M

(G) has rank r.

(3) If G is a nite at R-group scheme killed by , then there is a canonical
k-linear isomorphism
D(G)
k,Frob

k

= M

(G)/uM

(G).
Under this identication, F Frob

corresponds to and V Frob

1

cor-
responds to the composite
V
M
: M/u M

1
1
M
1
/u M
1
M/u M.
(4) If
0 G

G G

0
is a diagram of nite at group schemes over R which are killed by and if
0 M

(G

) M

(G) M

(G

) 0
is the corresponding diagram of Breuil modules, then the diagram of nite
at group schemes is a short exact sequence if and only if the diagram of
Breuil modules is a short exact sequence.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
886 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
Proof. See 2.1.1, Proposition 2.1.2.2, Theorem 3.3.7, Theorem 4.2.1.6 and the
proof of Theorem 3.3.5 of [Br2]. In 3.3.5 of [Br2] it is shown that M

(G)/u M

(G)
can be k-linearly identied with the crystalline Dieudonne module of Gk. In 4.2.14
of [BBM] the crystalline Dieudonne module of Gk is identied with D(G)
k,Frob

k.
The equivalence of the two notions of exactness can be deduced from the compat-
ibility of M

with Dieudonne theory, from Lemma 5.1.1, and from the fact that a
complex of nite at group schemes over R is exact if and only if its special bre
is exact (see for example Proposition 1.1 of [deJ]).
5.2. Examples. For 0 r e an integer and for a k

, dene a Breuil module

M(r, a) by
M(r, a) = (k[u]/u
e
)e,
M(r, a)
1
= (k[u]/u
e
)u
r
e,

1
(u
r
e) = ae.
It is easy to check that
1
is well dened (and uniquely determined by the given
conditions). We will refer to e as the standard generator of M(r, a) and write G(r, a)
for G

(M(r, a)). The following lemma is easy to check.

Lemma 5.2.1. (1) Any Breuil module of rank 1 over k[u]/u
e
is isomorphic to
some M(r, a).
(2) There is a non-zero morphism M(r, a) M(r

, a

) if and only if r

r,
r

r mod 1 and a/a

(k

)
1
. All such morphisms are then of the
form e bu
(r

r)/(1)
e

, where b
1
= a/a

.
(3) The modules M(r, a) and M(r

, a

) are isomorphic if and only if r = r

and
a/a

(k

)
1
, or equivalently if and only if there are non-zero morphisms
M(r, a) M(r

, a

) and M(r

, a

) M(r, a).
(4) If we order the M(r, a) by setting M(r, a) M(r

, a

) if there is a non-
zero morphism M(r

, a

) M(r, a), then the set of isomorphism classes of

M

(G)s as G runs over models of a xed nite at F

-group scheme G of
order is well ordered.
(5) On M(r, a) we have Ne = 0, so N
1
= 0.
(6) G(r, a) is etale (resp. multiplicative) if and only if r = e (resp. r = 0).
(7) G(0, 1)

and G(e, G

(0))

= Z/Z.
(8) The Cartier dual of G(r, G

(0)) is G(e r, 1).

Proof. The rst three parts are easy computations. For the fourth part note that
two nite at group schemes G and G

of order over R have isomorphic generic

bres if and only if there is a non-zero morphismG G

or G

G. The fth part is

another easy computation and the sixth part follows on computing the Dieudonne
module using Theorem 5.1.3.
By 3.1.2 of [Br2] we see that the ane R-algebra of the group scheme attached
to M(r, a) is
R[X]/(X

+

er
a
G

()
X),
where a is a lift of a to W(k). This has constant generic bre if and only if

er
a/G

() F

is an (1)
th
power. This occurs if and only if r e mod 1
and a/G

(0) k is an ( 1)
th
power. Thus M(e, G

(0)) corresponds to the

etale group scheme Z/Z over R.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 887
Next, we show that the group scheme G corresponding to the Breuil module
M(0, 1) is isomorphic to

. By using the relation between Breuil modules and

Dieudonne modules (see Theorem 5.1.3) we see that the Dieudonne module of the
closed bre of G is isomorphic to the Dieudonne module of the closed bre of

.
This forces G

, since we may consider Cartier duals and observe that a nite

at R-group scheme G is etale if and only if its special bre is etale, and then
18.5.15 of book IV
4
of [EGA] may be used.
This establishes the seventh part. The nal part follows from parts four and
seven.
Now suppose that 0 r, s e are integers and choose a, b k

and f
u
max(0,r+se)
k[u]/u
e
. We can dene an extension class
(0) M(s, b) M(s, b; r, a; f) M(r, a) (0)
in
1
mod
R
by
M(s, b; r, a; f) = (k[u]/u
e
)e (k[u]/u
e
)e

,
M(s, b; r, a; f)
1
= u
s
e, u
r
e

+fe),

1
(u
s
e) = be and
1
(u
r
e

+fe) = ae

,
the standard generator of M(s, b) maps to e,
e maps to 0 and e

maps to the standard generator in M(r, a).

The following lemma is also easy to check.
Lemma 5.2.2. (1) Any extension of M(r, a) by M(s, b) in
1
mod
R
is isomor-
phic to M(s, b; r, a; f) for some f u
max(0,r+se)
k[u]/u
e
.
(2) Two such extensions M(s, b; r, a; f) and M(s, b; r, a; f

) are isomorphic as
extension classes if and only if
f

f = u
s
h (b/a)u
r
h

for some h k[u]/u

e
, in which case one such isomorphism xes e and sends
e

to e

(b/a)h

e.
We remark that f u
max(0,r+se)
k[u]/u
e
is required so that M(s, b; r, a; f)
1

u
e
M(s, b; r, a; f). We will write G(s, b; r, a; f) for G

(M(s, b; r, a; f)).
We will also need some slight extensions of these results to allow for coecients.
To this end let k

/F

be a nite extension linearly disjoint from k and write k

k
for the eld k

k. For 0 r e an integer and for a (k

k)

, dene a Breuil
module, M(k

; r, a), with an action of k

by
M(k

; r, a) = ((k

k)[u]/u
e
)e,
M(k

; r, a)
1
= ((k

k)[u]/u
e
)u
r
e,

1
(u
r
e) = ae.
We will let denote the automorphism of k

k[u], which is the identity on k

and
which raises elements of k[u] to the
th
power. The following lemma is easy to
check.
Lemma 5.2.3. (1) Any Breuil module with an action of k

which is free of rank

[k

: k] over k[u]/u
e
is isomorphic to some M(k

; r, a).
(2) There is a non-zero morphism M(k

; r, a) M(k

; r

, a

) if and only if r

r,
r

r mod 1 and a/a

(b)/b for some b (k

k)

. All such morphisms

are then of the form e b

u
(r

r)/(1)
e

, where b (k

k)

and (b

)/b

=
a/a

.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
888 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
(3) The modules M(k

; r, a) and M(k

; r

, a

) are isomorphic if and only if r = r

and a/a

((k

k)

)
1
.
(4) On M(k

; r, a) we have Ne = 0 and so N
1
= 0.
(5) G

(M(k

; r, a)) is etale (resp. multiplicative) if and only if r = e (resp.

r = 0).
Now choose 0 r, s e integers, a, b (k

k)

and f u
max(0,r+se)
(k

k)[u]/u
e
.
We dene an extension class
(0) M(k

; s, b) M(k

; s, b; r, a; f) M(k

; r, a) (0)
in
1
mod
R
with an action of k

by
M(k

; s, b; r, a; f) = ((k

k)[u]/u
e
)e ((k

k)[u]/u
e
)e

,
M(k

; s, b; r, a; f)
1
= u
s
e, u
r
e

+fe),

1
(u
s
e) = be and
1
(u
r
e

+fe) = ae

,
the standard generator of M(k

; s, b) maps to e,
e maps to 0 and e

to the standard generators in M(k

; r, a).
Then the following lemma is easy to check.
Lemma 5.2.4. (1) Any extension of M(k

; r, a) by M(k

; s, b) in
1
mod
R
with
a compatible action of k

is isomorphic to M(k

; s, b; r, a; f) for some f
u
max(0,r+se)
(k

k)[u]/u
e
.
(2) Two such extensions M(k

; s, b; r, a; f) and M(k

; s, b; r, a; f

) are isomorphic
(as extensions) if and only if
f

f = u
s
h (b/a)u
r
(h)
for some h (k

k)[u]/u
e
, in which case one such isomorphism xes e and
sends e

to e

(b/a)(h)e.
We will write G(k

; r, a; s, b; f) and G(k

; r, a) for G

(M(k

; r, a; s, b; f)) and
G

(M(k

; r, a)) respectively.
5.3. Relationship to syntomic sheaves. Let us rst recall some of the notations
of [Br1] and [Br2]. Let Spf(R)
syn
be the small -adic formal syntomic site over R,
S the -adic completion of W(k)[u,
u
ie
i!
]
iN
, S
n
= S/
n
S, E
n
= Spec(S
n
) and for
any X Spf(R)
syn
:
O
cris
n,
(X) = H
0
((X
n
/E
n
)
cris
, O
Xn/En
),
where X
n
= X
R
R/
n
is viewed over E
n
via the thickening (Spec(R/
n
)
E
n
, u ). It turns out O
cris
n,
is the sheaf of S
n
-modules on Spf(R)
syn
associated
to the presheaf (cf. the proof of Lemma 2.3.2 in [Br2]):
X
_
W
n
(k)[u]

n
,Wn(k)
W
n
((X
1
, O
X1
))
_
DP
=
_
(W
n
(k)[u]/u
e
n
)

n
,Wn(k)
W
n
((X
1
, O
X1
))
_
DP
.
(5.3.1)
Here, the subscript
n
means we twist by the n
th
power of the Frobenius when
sending W
n
(k) to W
n
(k)[u] and the exponent DP means we take the divided
power envelope with respect to the kernel of the canonical map:
W
n
(k)[u]

n
,Wn(k)
W
n
((X
1
, O
X1
)) (X
n
, O
Xn
)
s(u) (w
0
, ..., w
n1
) s()( w

n
0
+ w

n1
1
+... +
n1
w

n1
),
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 889
where w
i
is a local lifting of w
i
, these divided powers being required to be compatible
with the usual divided powers
i
(x) =

i
i!
x
i
(i.e. we take the divided power envelope
relative to the usual divided power structure on the maximal ideal of W
n
(k)). Note
that the latter map induces a canonical surjection of sheaves of S
n
-modules on
Spf(R)
syn
:
O
cris
n,
O
n
,
where O
n
(X) = (X
n
, O
Xn
). We denote by J
cris
n,
the kernel of this surjection. For
any n, let : S
n
S
n
be the unique lifting of Frobenius such that (u) = u

and (
u
ie
i!
) =
u
ie
i!
. The sheaf O
cris
n,
is equipped with the crystalline Frobenius ,
which is also induced by the map s(u) (w
0
, ..., w
n1
) (s(u)) (w

0
, ..., w

n1
)
on the above presheaf (5.3.1). (Here on W
n
(k)[u] is Frobenius on W
n
(k) and
takes u to u

.) Since divides (x) x

, we get (J
cris
n,
) O
cris
n,
for all n, so we
can dene an S
1
-linear
1
=

[
J
cris
n,
by the usual atness trick (see 2.3 of [Br2]).
Let N : S
n
S
n
be the unique W
n
(k)-linear derivation such that N(u) = u and
N(
i
(u
e
)) = eu
e

i1
(u
e
) = ie
i
(u
e
). Finally dene:
N : O
cris
n,
O
cris
n,
to be the unique W
n
(k)-linear morphism of sheaves which on the presheaf (5.3.1) is
given by N(
i
(

s w)) = (

N(s) w)
i1
(

s w). Note that N = N,

so N
1
= N on J
cris
n,
.
Let G be a nite at group scheme over R, which is killed by . Viewing G as a
formal scheme over R, it is an object in Spf(R)
syn
. Viewing it as a sheaf of groups
on Spf(R)
syn
, its associated Breuil module is dened as:
(1) M

(G) = Hom
sheaves of groups
(G, O
cris
1,
)
S1
k[u]/u
e
,
(2) M

(G)
1
= image of Hom
sheaves of groups
(G, J
cris
1,
)
S1
k[u]/u
e
in M

(G),
(3)
1
is induced by
1
,
where the S
1
-module structures are induced by the compatible S
1
actions on O
cris
1,
and J
cris
1,
(see 3.2 and 2.1.2.2 of [Br2]). Here S
1
k[u]/u
e
is the surjection that
sends u to u,
i
(u
e
) to
i
(u
e
) for i < l and
i
(u
e
) to 0 for i l.
We record for future reference the following straightforward observation.
Lemma 5.3.1. If we denote by (resp. pr
i
, i 1, 2) the coproduct (resp. the
two projections)
G
Spec(R)
G G,
then for any sheaf of commutative groups F on Spf(R)
syn
we have:
Hom
sheaves of groups
(G, F) = x F(G) [ (

pr

1
pr

2
)(x) = 0.
The operator N on O
cris
1,
induces an operator N on Hom
sheaves of groups
(G, O
cris
1,
),
hence on M

(G).
Lemma 5.3.2. The above operator N on M

(G) coincides with the operator N

dened in Lemma 5.1.2.
Proof. By unicity in Lemma 5.1.2, we only have to prove that N satises N(M

(G))
uM

(G), since the other conditions are automatically satised. Its enough to
prove that N(
1
(x)) = (N)(x) uM

(G) for any x M

(G)
1
. But u
e
N =
0 on O
cris
1,
because it is so on (k[u](X
1
, O
X1
))
DP
. Thus one also has u
e
N =
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
890 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
0 on Hom
groups
(G, O
cris
1,
), hence on M

(G). This implies N(M

(G)) u

(G)
uM

(G) since M

(G) is free over k[u]/u

e
.
5.4. Base change. In this section we will examine the relationship of the functor
M

with two instances of base change. First we consider unramied base change.
Let k

be a perfect eld of characteristic which is an extension of k and R

=
R
W(k)
W(k

). Choose

= 1 as uniformiser in R

. If X Spf(R)
syn
, let
X

= Spf(R

)
Spf(R)
X and dene:
O
cris

n,
(X) = O
cris
n,
(X

) and J
cris

n,
(X) = J
cris
n,
(X

).
As in the proof of 2.3.2 of [Br2], we have that O
cris

n,
is the sheaf on Spf(R)
syn
associated to the presheaf:
X
_
W
n
(k

)[u]

n
,Wn(k

)
W
n
((X

1
, O
X

1
))
_
DP
=
_
W
n
(k

)[u]

n
,Wn(k

)
W
n
(k

k
(X
1
, O
X1
))
_
DP
.
Dene S

n
as S
n
but with k

instead of k. There is a canonical isomorphism of

sheaves:
O
cris
n,

Sn
S

n
= O
cris
n,

Wn(k)
W
n
(k

)

O
cris

n,
coming from the obvious isomorphism:
(W
n
(k

)[u]/u
e
n
)

n
,Wn(k)
W
n
((X
1
, O
X1
))

(W
n
(k

)[u]/u
e
n
)

n
,Wn(k

)
W
n
(k

k
(X
1
, O
X1
))
and one easily sees it induces an isomorphism J
cris
n,

Wn(k)
W
n
(k

)

J
cris

n,
. More-
over, we have the following obvious lemma.
Lemma 5.4.1. The diagram of sheaves on Spf(R)
syn
:
J
cris
n,

Wn(k)
W
n
(k

)

J
cris

n,
1 1
O
cris
n,

Wn(k)
W
n
(k

)

O
cris

n,
is commutative.
Using the identication from 5.3, Lemma 5.3.1 and Lemma 5.4.1 (for n = 1),
together with obvious functorialities, we obtain after tensoring by k[u]/u
e
the
following corollary.
Corollary 5.4.2. Let G be a nite at group scheme over R, which is killed by .
Let k

/k be an extension of elds with k

perfect and let

= 1, a uniformiser
for R

= R
W(k)
W(k

). Then there is a canonical isomorphism in the category

1
mod
R
_
M

(G)
k
k

, M

(G)
1

k
k

,
1

_
M

(G

), M

(G

)
1
,
1
_
compatible with composites of such residue eld extensions.
We will now turn to the case of base change by a continuous automorphism
g : R

R. For any s =

w
i
u
i
W(k)[[u]], let
(g)
s =

g(w
i
)u
i
and
()
s =

(w
i
)u
i
, where g and act on W(k) through their action on k. Choose H
g
(u)
W(k)[[u]] such that g() = H
g
(). Notice that H
g
(u) W(k)[[u]]

. Dene
g : W(k)[[u]]

W(k)[[u]] by g(

w
i
u
i
) =

g(w
i
)u
i
H
g
(u)
i
.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 891
Lemma 5.4.3. There is a unique element
g
t(u) W(k)[[u]] such that, if
g
is
dened by
g
(

w
i
u
i
) =

(w
i
)(u

(1 +
g
t(u)))
i
, one has g
g
= g.
Proof. One has to solve in W(k)[[u]]:
1 +
(g)
g
t(uH
g
(u)) =
_
H
g
(u)
1
_

()
H
g
(u

)
(where the two sides clearly belong to 1 + W(k)[[u]]). As H
g
(u) W(k)[[u]]

,
there is a unique K
g
uW(k)[[u]]

such that K
g
(u)H
g
(K
g
(u)) = u, so we have
1 +
(g)
g
t(u) =
_
H
g
(K
g
(u))
1
_

()
H
g
(K
g
(u)

).
For any object M of
1
mod
R
, dene
g

1
: M
1
M by the following formula:
g

1
(x) =
1
(x) +
g
t(u)N(
1
(x)) (5.4.1)
where N is as in Lemma 5.1.2.
For any X Spf(R)
syn
, let
g
X = Spf(R)
g

,Spf(R)
X and dene:
O
cris,(g)
n,
(X) = O
cris
n,
(
g
X) and J
cris,(g)
n,
(X) = J
cris
n,
(
g
X).
Then O
cris,(g)
n,
is the sheaf on Spf(R)
syn
associated to the presheaf:
X
_
W
n
(k)[u]

n
,Wn(k)
W
n
((
g
X
1
, Og
X
1
))
_
DP
=
_
W
n
(k)[u]

n
,Wn(k)
W
n
(R
g,R
(X
1
, O
X1
))
_
DP
.
Let g : S
n
S
n
be the unique ring isomorphism such that
g
_
w
i
u
ei+j
i!
_
= g(w
i
)
u
ei+j
i!
H
g
(u)
ei+j
for 0 j < e, i 0. There is a canonical isomorphism of sheaves:
O
cris
n,

Sn, g
S
n

O
cris,(g)
n,
coming from the obvious g-semi-linear isomorphism:
(W
n
(k)[u]/u
e
n
)

n
,Wn(k)
W
n
((X
1
, O
X1
))

(W
n
(k)[u]/u
e
n
)

n
,Wn(k)
W
n
(R
g,R
(X
1
, O
X1
))
s (w
0
, ..., w
n1
) g(s) (1 w
0
, ..., 1 w
n1
)
and one easily sees it induces an isomorphism J
cris
n,

Sn, g
S
n

J
cris,(g)
n,
.
Dene
g
: S
n
S
n
as in Lemma 5.4.3 and dene:
g
: O
cris
n,
O
cris
n,
to be the unique morphism of sheaves which is induced by
g
(
i
(

s w)) =

i
(

g
(s) (w)) on the presheaf (5.3.1) (see 5.3 and note that this is well
dened). Since
g
(J
cris
n,
) O
cris
n,
, we can dene
g

1
=
g

[
J
cris
n,
.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
892 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
Lemma 5.4.4. The diagram of sheaves on Spf(R)
syn
:
J
cris
n,

Sn, g
S
n

J
cris,(g)
n,
g
1
1
O
cris
n,

Sn, g
S
n

O
cris,(g)
n,
is commutative. Moreover we have on J
cris
n,
:
g

1
=

i=0
_
log(1 +
g
t(u))

_
i
N
i
i!

1
,
where N is dened as in 5.3.
Proof. By working modulo
n+1
, i.e. with J
cris
n+1,
and
g
, and looking on the above
presheaves, the proof is completely straightforward.
Let G be a nite at group scheme over R which is killed by . Note that
thanks to Lemma 5.3.2 and the formula for
g

1
in Lemma 5.4.4, the operator
M

(G)
1
M

(G) induced by the map

g

1
: J
cris
n,
O
cris
n,
is precisely the operator
denoted
g

1
earlier in this section (see (5.4.1)). Using this, together with Lemma
5.3.1, Lemma 5.4.4 (for n = 1) and obvious functorialities, we obtain, after tensoring
by k[u]/u
e
, the following corollary.
Corollary 5.4.5. Let g : R R be a continuous automorphism.
(1) Let G be a nite at group scheme over R, which is killed by . Then there
is a canonical isomorphism in the category
1
mod
R
:
_
M

(G)
k[u]/u
e
, g
k[u]/u
e
, M

(G)
1

k[u]/u
e
, g
k[u]/u
e
,
g

_
M

(
g
G), M

(
g
G)
1
,
1
_
.
(2) If f : G G

is a morphism of nite at R-group schemes killed by

and M

(f) is the corresponding morphism in

1
mod
R
, then M

(f) also
commutes with the
g

1
and there is a commutative diagram in
1
mod
R
:
M

(G

)
k[u]/u
e
, g
(k[u]/u
e
)
M(f)1
M

(G)
k[u]/u
e
, g
(k[u]/u
e
)

M

(
g
G

)
M(
g
f)
M

(
g
G)
(3) If g
1
, g
2
are two continuous automorphisms of R and if we choose the unique
H
g2g1
W(k)[[u]] such that g
2
g
1
= g
2
g
1
on W(k)[[u]], then on
(M

(G)
k[u]/u
e
, g1
k[u]/u
e
)
k[u]/u
e
, g2
k[u]/u
e
M

(G)
k[u]/u
e
, g2g1
k[u]/u
e
,
one has
g2
(
g1

1
) =
g2g1

1
.
Corollary 5.4.6. Let G be a nite at group scheme over R, which is killed by .
To give a morphism of schemes [g] : G G such that the diagram of schemes
G
[g]
G

Spec(R)
Spec(g)
Spec(R)
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 893
is commutative and the induced morphism G Spec(R)
g,Spec(R)
G is a morphism
of group schemes over R, is equivalent to giving an additive map g : M

(G)
M

(G) such that both of the following hold:

(1) For all s k[u]/u
e
and x M

(G), g(sx) = g(s) g(x).

(2) g(M

(G)
1
) M

(G)
1
and
1
g = g
1
+ g(
g
t(u)) g N
1
with
g
t as in
Lemma 5.4.3 and N as in Lemma 5.1.2.
Proof. Note that the last condition is equivalent to
1
g = g
g

1
. The rst two
conditions are equivalent to giving a morphism g : M

(
g
G) M

(G) in
1
mod
R
,
which is equivalent to the last two by Corollary 5.4.5.
Finally we make some computations that concern the dependence of the above
compatibilities on the choice of H
g
(u). Let f(u) be an element of (k[u]/u
e
)
1
=
u
e
(k[u]/u
e
) and dene, for any Min
1
mod
R
, the additive map 1
f
:
1
(M
1
) M
via
1
f
= 1 +
_
1

i=1
(1)
i1
i
f(u)
i
_
N,
where N is as in Lemma 5.1.2. Using k[u]/u
e

k[u

]/u
e
1
(M
1
) M, we extend
1
f
to all of M by the formula:
1
f
(u
i
x) = u
i
(1 +f(u))
i
1
f
(x)
for x
1
(M
1
). If x M
1
, one checks that:
1
f
(
1
(u
i
x)) = u
i
1
f
(
1
(x)) = 1
f
(u
i

1
(x))
so 1
f
is well dened. Moreover, it is clear that 1
f
(M
1
) M
1
. Let
1
f
: O
cris
1,

O
cris
1,
be the unique isomorphism of sheaves coming from the semi-linear isomorphism of
presheaves:
(k[u]/u
e
)
,k
(X
1
, O
X1
)

(k[u]/u
e
)
,k
((X
1
, O
X1
))
s(u) (w
0
, ..., w
n1
) s(u(1 +f(u))) (w
0
, ..., w
n1
)
(see 5.3.1).
Let G be a nite at group scheme over R killed by and recall that
M

(G) = Hom
sheaves of groups
(G, O
cris
1,
) k[u]/u
e
.
Lemma 5.4.7. The operator 1
f
on M

(G) is induced by the operator 1

f
on O
cris
1,
.
Proof. One can check that the operator 1
f
on O
cris
1,
satises 1
f

1
=
1
+
log(1 + f)N
1
, where N is dened as in 5.3 and log(1 + f) is the usual ex-
pansion of log in S
1
, which makes sense because of the assumption that u
e
[f and
because of the divided powers
i
(u
e
) =
u
ei
i!
. After tensoring with k[u]/u
e
, we get
1
f
= 1 + (

1
i=1
(1)
i1
i
f(u)
i
)N on
1
(M

(G)
1
) which clearly implies the two 1
f
s
are the same.
Let g = 1 and choose H
g
(u) = 1 + f(u) for some f E

(u)W(k)[[u]] (see the

start of 5.1 for the denition of E

(u)). Recall from Corollary 5.4.5 that we have

a canonical isomorphism M

(G)
k[u]/u
e
, g
(k[u]/u
e
)

M

(
g
G).
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
894 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
Lemma 5.4.8. The map 1
f
is the composite M

(G)

M

(
g
G)

M

(G), where
the rst map is the one in Corollary 5.4.5 and the second comes from the obvious
isomorphism G

g
G. In other words, once H
g
(u) = 1 + f(u) has been chosen,
1
f
: M

(G) M

(G) is the map corresponding to the identity 1

G
: G G under
the equivalence of Corollary 5.4.6.
The proof is straightforward by looking at the usual presheaves and using Lemma
5.4.7. We remark that 1
f
is not necessarily the identity even though 1
G
is. However,
with f = 0, 1
f
is the identity.
5.5. Reformulation. In this section, we will reformulate Corollary 5.4.6.
Lemma 5.5.1. There is a unique element t
g
(u) W(k)[[u]] such that if
g
is
dened by
g
(

w
i
u
i
) =

(w
i
)(u

(1 +t
g
(u)))
i
, one has g =
g
g.
Proof. One has to solve in W(k)[[u]]:
u

H
g
(u)

= u

(1 +t
g
(u))
()
H
g
(u

(1 +t
g
(u))).
As H
g
(u) W(k)[[u]]

, there is a unique L
g
uW(k)[[u]]

such that L
g
(uH
g
(u)) =
u. Applying L
g
to u = K
g
(u)H
g
(K
g
(u)) (cf. the proof of Lemma 5.4.3), we get
L
g
(u) = K
g
(u). We must solve:
1 +t
g
(u) =
()
K
g
(u

H
g
(u)

)
u

.
Lemma 5.5.2. There is a unique
g
(u) 1 + uW(k)[[u]] such that if N
g
=
g
N,
then N
g
g = g N. Similarly, there is a unique
g
(u) 1 + uW(k)[[u]] such
that if
g
N =
g
N, then g
g
N = N g. Moreover, N
g

g
=
g
N
g
and
g
N
g
=
g

g
N.
Proof. Since N is a derivation, so is N for any W(k)[[u]]. One has to solve

g
(u)N(uH
g
(u)) = uH
g
(u) and
(g)
g
(uH
g
(u)) = 1 +
N(Hg(u))
Hg(u)
, which amounts to:

g
(u) =
_
1 +
N(H
g
(u))
H
g
(u)
_
1
,
(g)
g
(u) = 1 +
N(H
g
)(K
g
(u))
H
g
(K
g
(u))
,
where K
g
is as in the proof of Lemma 5.4.3. The commutation relations with the
Frobenius follow from N = N,
g
g = g , N
g
g = g N, g
g
= g,
g
g
N = N g and the fact g is bijective on W(k)[[u]].
We also denote by
g
N =
g
N and N
g
=
g
N the corresponding derivations on
k[u]/u
e
. For any object M of
1
mod
R
, dene
1,g
: M
1
M by the formula

1,g
(x) =
1
(x) +t
g
(u)N(
1
(x)),
where N is as in Lemma 5.1.2, and we recall that we dened
g

1
in (5.4.1). One
checks that
1,g
(u
e
) =
g

1
(u
e
) =
1
(u
e
) = c

(see (5.1.1)). Note that we also have

1,g
g = g
g

1
, N
g
g = g N, g
g

1
=
1
g, g
g
N = N g in k[u]/u
e
.
Lemma 5.5.3. Let M be an object of
1
mod
R
; then there is a unique operator
N
g
: MM satisfying the following three conditions:
(1) N
g
(sx) = N
g
(s)x +sN
g
(x), s k[u]/u
e
, x M,
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 895
(2) N
g

1,g
(x) =
g
N
g
(x), x M
1
, where
g
(y) =
1
c

1,g
(u
e
y) if y M,
(3) N
g
(M) uM.
The same statement holds for
g
N,
g
and
g

1
.
Proof. The proof is the same as for Lemma 5.1.2, using the fact we still have
isomorphisms
k[u]/u
e

k[u

]/u
e
1,g
(M
1
)

M
(resp. with
g

1
replacing
1,g
).
Lemma 5.5.4. For M an object of
1
mod
R
, N
g
=
g
N and
g
N =
g
N, where
N
g
,
g
N are as in Lemma 5.5.3,
g
,
g
as in Lemma 5.5.2 and N as in Lemma
5.1.2.
Proof. By unicity of N
g
, one has to check
g
N satises the three conditions of
Lemma 5.5.3. The rst and last are obvious. Note that N
1
(u
e
x) = N(u
e
x) = 0
so
1,g
(u
e
x) =
1
(u
e
x), which implies =
g
on M(
g
is as in Lemma 5.5.3). One
computes:
(
g
(u)N)
1,g
=
g
(u)(1 +N(t
g
(u))) N,

g
(
g
(u)N) =
()

g
(u

) N.
But the equality N
g

g
(u) =
g
N
g
(u) in W(k)[[u]] (from Lemma 5.5.2) yields

g
(u)(1 +N(t
g
(u)))
()

g
(u

) W(k)[[u]].
We thus get (
g
N)
1,g
=
g
(
g
N), hence condition (2). For
g
N, the proof is
completely similar.
Lemma 5.5.5. Let M be an object of
1
mod
R
and g : MM an additive map
such that for all s k[u]/u
e
and x M, g(sx) = g(s) g(x) and g(M
1
) M
1
.
If g
1
=
1,g
g, then g N = N
g
g. Similarly, if
1
g = g
g

1
, then
N g = g
g
N.
Proof. We prove the rst case, the other one being the same. As in the proof of
Lemma 5.1.2, we dene N
g,0
, N
g,1
,..., with N
g
= N
g,i
for i large enough, using
k[u]/u
e

k[u

]/u
e
1,g
(M
1
)

M. It is enough to show g N
i
= N
g,i
g for all i.
Suppose g N
i1
= N
g,i1
g and let s k[u]/u
e
and x M
1
. Then
N
g,i
g(s
1
(x)) = N
g,i
( g(s)
1,g
( g(x)))
= N
g
( g(s))
1,g
( g(x)) +g(s)
1,g
N
g,i1
( g(x))
= g(N(s)
1
(x)) + g(s)
1,g
g(N
i1
(x))
= g(N(s)
1
(x)) + g(s
1
(N
i1
(x)))
= gN
i
(s
1
(x)),
so g N
i
= N
g,i
g by linearity. One easily checks by a similar computation that
N
g,0
g = g N
0
, hence the result follows by induction.
Lemma 5.5.6. Let M be an object of
1
mod
R
and g : MM an additive map
such that for all s k[u]/u
e
and x M, g(sx) = g(s) g(x) and g(M
1
) M
1
. Then
the following two conditions are equivalent:
(1)
1
g = g
1
+ g(
g
t(u)) g N
1
,
(2) and g
1
=
1
g +t
g
(u)N
1
g.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
896 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
Proof. One has to show
1
g = g
g

1
is equivalent to g
1
=
1,g
g. We prove
(1) (2), the other case being the same. On M, we have g = g, because
=
g
=
g
, as in the proof of Lemma 5.5.4. By Lemmas 5.5.4 and 5.5.5, we have
g N = g(
g

1
)N g. Thus we get from (1), using N
1
= N,
g
1
=
1
g g(
g
t(u)) g(
g
(u)
1
)

N
1
g.
Playing the same game over W
2
(k)[[u]] with the relation g = g +g(
g
t(u)) g
N , which is easily checked to hold in W
2
(k)[[u]], we again end up with g =
g g(
g
t(u)) g(
g
(u)
1
)

N g in W
2
(k)[[u]]. But we also have in W
2
(k)[[u]]
the equality:
g = g +t
g
(u)N g.
Thus g(
g
t(u)) g(
g
(u)
1
)

= t
g
(u) in k[u]/u
e
, so relation (2) holds.
We can now derive the variant of Corollary 5.4.6 which we will use.
Corollary 5.5.7. Let G be a nite at R-group scheme killed by . Let g : R

R
be a continuous automorphism, choose H
g
(u) W(k)[[u]] such that g() = H
g
()
and dene g : k[u]/u
e
k[u]/u
e
by g(

w
i
u
i
) =

g(w
i
)u
i
H
g
(u)
i
. To give a
morphism of schemes [g] : G G such that the diagram of schemes
G
[g]
G

Spec(R)
Spec(g)
Spec(R)
is commutative and the induced morphism G Spec(R)
g,Spec(R)
G is an morphism
of group schemes over R, is equivalent to giving an additive map g : M

(G)
M

(G) such that both of the following hold:

(1) For all s k[u]/u
e
and x M

(G), g(sx) = g(s) g(x).

(2) g(M

(G)
1
) M

(G)
1
and g
1
= (1+t
g
(u)N)
1
g, with t
g
as in Lemma
5.5.1 and N as in Lemma 5.1.2.
Moreover, [g] is an isomorphism if and only if g is. Assume these are isomorphisms.
Choose H
g
1 such that

g
1
(u) = g
1
(u) on W(k)[[u]], i.e. H
g
1 (u) = g
1
(u)/u.
Then the map

g
1
that corresponds to [g]
1
is equal to g
1
. Also, if g
1
, g
2
are two
automorphisms of R and if we choose H
g1
, H
g2
as above, then [g
1
] [g
2
] corresponds
to g
2
g
1
provided we choose H
g2g1
such that g
2
( g
1
(u)) = uH
g2g1
(u).
Proof. The equivalence is clear thanks to Corollary 5.4.6 and Lemma 5.5.6. The
fact that [g
1
] [g
2
] corresponds to g
2
g
1
is automatic using Corollary 5.4.5 and
the functor G M

(G). Applying this to g

1
= g and g
2
= g
1
, we see that
1
G
= [g] [g]
1
corresponds to

g
1
g. But by Lemma 5.4.8, 1
G
corresponds to 1
f
with f dened by (

g
1
g)(u) = u(1 +f) in W(k)[[u]]. We see that f = 0 and that
1
f
is the identity on M

(G). Thus

g
1
= g
1
on M

(G).
5.6. Descent data. Assume now that R is endowed with a continuous left faithful
action of a nite group . Then becomes the Galois group of the fraction eld
F

of R over some subeld. For each g , choose H

g
(u) W(k)[u] so that
g() = H
g
(), with the one condition that H
1
(u) = 1. Recall from Lemma 5.5.1
that this uniquely determines elements t
g
(u) W(k)[[u]] such that
u

H
g
(u)

= u

(1 +t
g
(u))
()
H
g
(u

(1 +t
g
(u))).
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 897
Moreover, for any pair g
1
, g
2
, there is obviously a unique f
g1,g2
(u)
E

(u)W(k)[[u]] such that

g
1
g
2
(u) = g
1
g
2
(u(1 +f
g1,g2
(u))).
If M is an object of
1
mod
R
, then we will denote by 1
g1,g2
the unique k-linear
map MM such that for x M
1
we have
1
g1,g2
(
1
(x)) =
_
1 + (

1
i=1
(1)
i1
i
f
g1,g2
(u)
i
)N
_
(
1
(x)), where N is as in
Lemma 5.1.2,
and 1
g1,g2
(u
i

1
(x)) = u
i
(1 +f
g1,g2
(u))
i
1
g1,g2
(
1
(x)).
(See 5.4 where we denoted 1
g1,g2
by 1
fg
1
,g
2
.)
Suppose that G is a nite at R-group scheme. Recall that by descent data on
G for we mean isomorphisms of nite at group schemes
[g] : G

g
G
for g , such that
[gh] = (
g
[h]) [g]
for all g, h . Equivalently we may think of [g] as a map of schemes G G over
g

: Spec R Spec R which induces an isomorphism of group schemes G

g
G. In
this picture the compatibility condition simply becomes
[gh] = [h][g].
Theorem 5.6.1. Suppose that G is a nite at R-group scheme killed by . Fix
H
g
(u) as above for all g .
(1) To give descent data on G relative to is equivalent to giving additive bijec-
tions g : M

(G) M

(G) for all g so that g takes M

(G)
1
into M

(G)
1
and:
g(wu
i
m) = g(w)(uH
g
(u))
i
g(m) for m M

(G), w k,
g
1
= (1 +t
g
(u)N)
1
g on M

(G)
1
,

= 1 and g
1
g
2
= g
1
g
2
1
g1,g2
.
(2) The above equivalence is functorial in G and is compatible with classical
Dieudonne theory in the following sense: if the action g
g
on M

(G)
corresponds to descent data [g] on G, then the g-semi-linear map D([g])
induced on the contravariant Dieudonne module D(G) and the g-semi-linear
map g mod u induced on M

(G)/uM

(G) are compatible via the isomorphism

of Theorem 5.1.3.
Proof. Part (1) is a consequence of Corollary 5.5.7, Lemma 5.4.8 and the choice
H
1
= 1. The functoriality in (2) follows from Corollary 5.4.5, and the last statement
there comes from
g
(G k)

=
g
Gk, the functoriality of the isomorphismin Theorem
5.1.3 and the reduction modulo u of Corollary 5.4.5.
Suppose that R

. Then we may take H

g
(u) = 1 for all g . With this
choice we see that t
g1
= 0, f
g1,g2
= 0 and 1
g1,g2
= 1 for all g
1
, g
2
. In this
case to give bijections g : M

(G) M

(G) as in the lemma is equivalent to giving

an R-semi-linear -action on M

(G) which commutes with u and

1
and preserves
M

(G)
1
. Thus (M

(G)

, M

(G)

1
,
1
) is a Breuil module over R

from which we can

recover M

(G) by tensoring with W(k) over W(k

). In other words, etale descent

for group schemes translates in the obvious manner for Breuil modules if we choose
to be -invariant.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
898 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
To build an action of on G using Theorem 5.6.1, the conditions g
1
g
2
= g
1
g
2

1
g1,g2
are not very convenient to check in practice since there are too many of them.
It is useful to have the following variant. Choose d Z
>0
and a group surjection
:
d
, where
d
is the free group on d generators
1
, ...,
d
. The group
d
still acts on R (via its quotient ) and for each i 1, ..., d, choose elements
H
i
(u) W(k)[[u]] such that H
i
() =
i
(). This determines isomorphisms
i
on W(k)[[u]] and k[u]/u
e
and, by composition, isomorphisms for all
d
.
Note that if ker(), then H

(u) = u(1 + f

(u)) for some f

(u)W(k)[[u]].
For such , denote by 1

the unique k-vector space endomorphism of any object

M of
1
mod
R
such that for x M
1
we have
1

= 1 +
_

1
i=1
(1)
i1
i
f

(u)
i
_
N on the image of
1
,
and 1

(u
i

1
(x)) = u
i
(1 +f

(u))
i
1

(
1
(x)),
where N is as in Lemma 5.1.2. (See 5.4, where we denoted 1

by 1
f
.) Let R be a
subset of ker() such that ker() is the smallest normal subgroup of
d
containing
R.
Corollary 5.6.2. With the above notation, to give descent data on G for is
equivalent to giving additive bijections
j
: M

(G) M

(G) for j 1, ..., d so

that
j
takes M

(G)
1
into M

(G)
1
and:

j
(wu
i
m) =
j
(w)(uH
j
(u))
i

j
(m) for m M

(G), w k,

j

1
= (1 +t
j
(u)N)
1

j
on M

(G)
1
,
if =
n1
i1

nm
im
R, where i
j
1, . . . , d, n
j
Z, and i
j
,= i
j+1
for
1 j < m, and if we dene =
n1
i1

nm
im
, then = 1

.
Proof. Straightforward from Corollary 5.5.7 and Lemma 5.4.8.
We dene a category
1
DD
F

/(F

of Breuil modules with descent data for in

the obvious way. This category is additive but not necessarily abelian. We call a
complex in
1
DD
F

/(F

exact if the underlying complex in

1
mod
R
is exact. In
the natural way, we extend M

to a functor from FD
F

/(F

)
to
1
DD
F

/(F

.
5.7. More examples. In this section we will determine the possible descent data
on a rank one Breuil module. Let be as in 5.6.
Lemma 5.7.1. Suppose that G is a nite at R-group scheme of order and that
its generic bre admits descent data over (F

. Then there is unique descent data

on G over (F

extending any choice of descent data on G F

over (F

. If
M

(G)

= M(r, a) and if satises ()/ 1 mod (), then

(e) = H

(u)
r/(1)
e,
where H

(u)
r/(1)
denotes the unique (1)
th
root of H

(u)
r
in k[u]/u
e
with
constant term 1.
We remark that since Aut(M(r, a)) = (Z/Z)

by consideration of the geometric

generic bre, the choice of isomorphism M

(G)

= M(r, a) does not matter.

Proof. We rst claim two such nite at group schemes G and G

have isomorphic
generic bres if and only if there is a non-zero morphism G G

or G

G. By
Lemma 5.2.1 we see that if G is a nite at F

-group scheme, then the lattice of

models for G over R is well ordered. Suppose all the integral models are G
1
< <
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 899
G
n
. For , any isomorphism [] : G

G must then induce isomorphisms
[] : G
i

G
i
for all i = 1, ..., n. The rst part of the lemma follows.
Let M = M(r, a), so M is a free k[u]/u
e
-module of rank 1 with the usual basis
element e. The submodule M
1
is spanned by u
r
e and
1
(u
r
e) = ae. From Theorem
5.2.1, we have N
1
= 0, which implies that

1
=
1
.
For
1
, H

(0) 1 mod . Clearly

: cu
i
e cu
i
H

(u)
i
(e)
is a bijection if and only if (e) =

e for some unit

(k[u]/u
e
)

. Evaluating

1
=

1
on the element u
r
e M
1
, we get

= H

(u)
r

in k[u]/u
e
. Thus,

H
r/(1)

for some unit

.
Since Breuil module descent data always induces a k-linear action of the inertia
group on the k-vector space M/u M and in this case dim
k
M/u M= 1, the action
of the element of -power order on M/u M must be trivial. Thus

= 1.
6. Some local fields
In order to apply the methods of 5, we need some more explicit information
about the elds F

introduced in 4. In this section we will collect this essentially

elementary information. In each case we will give an explicit description of the
Galois group Gal(F

/Q
3
). This is needed to carry out the delicate Breuil module
calculations in subsequent sections. We will also specify a uniformiser of F

and
partially calculate the following polynomials and power series (depending on our
choice of ).
G(u) W(k
F
)[u], a polynomial of degree at most e(F

/Q
3
) 1 such that
has minimal polynomial u
e(F

/Q3)
3G(u) over Q
3
.
c

G(u)
3
mod (3, u
3e(F

/Q3)
).
For Gal(F

/Q
3
), the unique polynomial H

(u) W(k
F
)[u] of degree
at most e(F

/Q
3
) 1 such that ()/ = H

().
In some cases power series t

and f
,
as in 5.6.
6.1. The case of F

1
. Recall that
1
corresponds to the order 3 homomorphism
Z

3
GL
2
(Q
3
)
is determined by
1 1
4 ,
where det = 1 and
3
= 1 but ,= 1. Recall also that F

1
= F
1
is any totally
ramied cubic Galois extension of Q
3
. We may take F

1
= F
1
= Q
3
[], where is
a root of X
3
3X
2
+ 3. One may check that the other roots of X
3
3X
2
+ 3 are

2
2 and 3 +
2
, so Gal(F

1
/Q
3
) is generated by one element
3
, which sends
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
900 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
to
2
2 and satises
3
3
= 1. Also, is a uniformiser for F

1
, so
G(u) = u
2
1,
c

1 u
6
mod (3, u
9
),
H
3
(u) = u 2.
6.2. The case of F

1
. Recall that
1
corresponds to the order 3 homomorphism
Z
3
[

1]

GL
2
(Q
3
)
determined by
4

1 1
4 1
1 + 3

1 ,
where det = 1 and
3
= 1 but ,= 1. Recall also that F

1
/Q
3
(

1) is the unique
cubic extension such that F

1
/Q
3
is Galois but not abelian and that F
1
is any
cubic subeld. We may take F
1
= Q
3
() and F

1
= F
1
(

1), where is a root

of X
3
3X
2
+6. The other roots of X
3
3X
2
+6 are (

1
2
+3(1

1))/2
and (

1
2
+ 3(1 +

1))/2. Thus, Gal(F

1
/Q
3
) is generated by two
elements
2
and
3
dened by

2
() = ,

2
(

1) =

1,

3
() = (

1
2
+ 3(1

1))/2,
and
3
(

1) =

1.
We have
2
2
=
3
3
= 1 and
2

3
=
2
3

2
, and is a uniformiser for F

1
. Thus
G(u) = u
2
2,
c

1 u
6
mod (3, u
9
),
H
2
(u) = 1,
H
3
(u) = ((

1 1)u
2
+ (3

1)u 2)/4.
6.3. The case of F

3
. Recall that
3
is the unique 3-type such that
3
[
I
Q
3
(

3)
corresponds to the order 6 homomorphism
Z
3
[

3]

GL
2
(Q
3
)
determined by
1 1
4 1
1 +

3 ,
where det = 1 and
3
= 1 but ,= 1. Recall also that F

3
is the degree 12 abelian
extension of Q
3
(

3) with norm subgroup in Q

3
(

3)

topologically generated by
3, 4 and 1 + 3

3. We also let
2
4
denote the unique element of I
F

3
/Q3(

3)
of order
3 and we let F
3
denote the xed eld of some Frobenius lift of order 2.
We claim that F

3
= Q
3
(

3)(

1, , ), where is a root of X
3
3X + 3
and a root of X
2

3. To verify this, set F

= Q
3
(

3)(

1, , ). We
must check that F

/Q
3
(

3) is abelian and that N

F

/Q3(

3)
(F

contains 3, 4,
and 1 + 3

3. To see that F

/Q
3
(

3) is abelian, note that if is one root of

X
3
3X + 3, then the other roots are (2

3
2
(3

3 +

5) 4

3)/2

5
and (2

3
2
(3

3 +

5) + 4

3)/2

5 (where for deniteness we choose

5 1 + 3Z
3
). Note that N
F

/Q3(

3)
(/) = 3 and N
F

/Q3(

3)
(1 +) = 5
4
.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 901
Note that Gal(F

3
/Q
3
(

3)) is generated by three commuting elements

2
,
2
4
and
3
of respective orders 2, 2 and 3. They may be dened by

2

1 =

1,
2
= and
2
= ;

2
4

1 =

1,
2
4
= and
2
4
= ;

3

1 =

1,
3
= and
3
= (2

3
2
(3

3+

5)+4

3)/2

5.
Choose an element I
F

3
/Q3
I
F

3
/Q3(

3)
. Then
2

2
4
,
3
). As
3

1
=
2
3
we may alter our choice of so that
2

2
4
). As

3 =

3 we see that
=

1, so
2
=
2
4
. We will rename as
4
and suppose it chosen so that

4
=

1. Thus, Gal(F

3
/Q
3
) is generated by elements
2
,
3
and
4
satisfying

2
2
=
3
3
=
4
4
= 1,

2

3
=
3

2
,

4

2
=
2

1
4
,
and
4

3
=
2
3

4
.
The element
2
4
is the unique element of I
F

3
/Q3(

3)
of order 2 and hence coincides
with our previous denition. The element
2
is a Frobenius lift of order 2 and so we
may take F
3
to be its xed eld, i.e. F
3
= Q
3
(), where = / is a uniformiser
for F

3
. (We are not asserting that
2
equals the element denoted
2
in 4.) One
can check that

3
()/ 1 +
2
mod
4
.
Note also that
2
,
4
) projects isomorphically to the quotient of Gal(F

3
/Q
3
) by
the wild inertia subgroup.
We conclude
G(0) = 1,
c

1 mod (3, u),

H
2
(u) = 1,
H
4
(u) =

1,
H
3
(u) 1 +u
2
mod (3, u
4
),
t
g
= f
g,g
= 0 for g, g

2
,
4
).
6.4. The case of F

3
. Recall that
3
is the unique 3-type such that
3
[
I
Q
3
(

3)
corresponds to the order 6 homomorphism
Z
3
[

3]

GL
2
(Q
3
)
determined by
1 1
4 1
1 + 3

3 1
1 +

3 ,
where det = 1 and
3
= 1 but ,= 1. Recall also that F

3
is the degree 12 abelian
extension of Q
3
(

3) with norm subgroup in Q

3
(

3)

topologically generated
by 3, 4 and 1 +3

3. We also let
2
4
denote the unique element of I
F

3
/Q3(

3)
of order 3 and we let F
3
denote the xed eld of some Frobenius lift of order 2.
We claim that F

3
= Q
3
(

3)(

1, , ) where is a root of X
3
4 and
a root of X
2
+

3. To verify this, set F

= Q
3
(

3)(

1, , ). Then
F

/Q
3
(

3) is abelian and so we must check that N

F

/Q3(

3)
(F

contains 3,
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
902 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
4 and 1+3

3. But note that we have the identities N

F

/Q3(

3)
((1)/) = 3,
N
F

/Q3(

3)
() = 4
4
and N
F

/Q3(

3)
(1 ) = (1 +

3)
6
.
Note that Gal(F

3
/Q
3
(

3)) is generated by three commuting elements

2
,
2
4
and
3
of respective orders 2, 2 and 3. They may be dened by

2

1 =

1,
2
= and
2
= ;

2
4

1 =

1,
2
4
= and
2
4
= ;

3

1 =

1,
3
= and
3
= (1

3)/2.
Choose an element I
F

3
/Q3
I
F

3
/Q3(

3)
, so
2

2
4
,
3
). As
3

1
=
2
3
,
we may alter our choice of so that
2

2
4
). As

3 =

3 we see that
=

1, so
2
=
2
4
. We will rename as
4
and suppose it chosen so that

4
=

1. Thus, Gal(F

3
/Q
3
) is generated by elements
2
,
3
and
4
satisfying

2
2
=
3
3
=
4
4
= 1,

2

3
=
3

2
,

4

2
=
2

1
4
,
and
4

3
=
2
3

4
.
The element
2
4
is the unique element of I
F

3
/Q3(

3)
of order 2 and hence coincides
with our previous denition. The element
2
is a Frobenius lift of order 2 and so
we may take F
3
to be its xed eld, i.e. F
3
= Q
3
(), where = / is a
uniformiser for F

3
. (We are not asserting that
2
equals the element denoted
2
in 4.) One can check that

3
()/ 1 +
2
mod
4
.
Note also that
2
,
4
) lifts tame inertia.
We conclude
G(0) = 1,
c

1 mod (3, u),

H
2
(u) = 1,
H
4
(u) =

1,
H
3
(u) 1 +u
2
mod (3, u
4
),
t
g
= f
g,g
= 0 for g, g

2
,
4
).
6.5. The case of F

i
. Here i Z/3Z and we will let denote the unique lifting of i
to Z with 0 < 3. Recall that

i
is the unique extended 3-type whose restrictions
to G
Q3(

3)
correspond to the homomorphism
Q
3
(

3)

GL
2
(Q
3
)
determined by

3
1
1 1
4 1
1 + 3

3
1 +

3
i
,
where det = 1 and
3
= 1 but ,= 1. Recall also that F

i
is the degree 12
abelian extension of Q
3
(

3) with norms the subgroup of Q

3
(

3)

topologically
generated by 3, 4, 1 + 9

3 and 1 + (1 3)

3. We let
2
,
3
and
2
4
denote
the elements of Gal(F

i
/Q
3
) which correspond respectively to

3, 1 3

3 and
1.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 903
We claim that F

i
= Q
3
(

3)(

1, , ), where is a root of X
3
3(1 + 3)
and a root of X
2
+

3. To verify this, set F

= Q
3
(

3)(

1, , ), so
F

/Q
3
(

3) is abelian and we must check that N

F

/Q3(

3)
(F

contains 3,
4, 1+9

3, and 1+(13)

3. But note that N

F

/Q3(

3)
(/) = 3(1+3)
4
,
N
F

/Q3(

3)
(1 +) = (4 + 9)
4
and
N
F

/Q3(

3)
((

3 )/) = (1 +

3 + 3)/(1 + 3)
4
1 + (1 3)

3 mod 9.
Note that
2
4
is an element of I
F

i
/Q3(

3)
of order 2,
2
,=
2
4
but also has order
2, and
3
is an element of I
F

i
/Q3(

3)
of order 3. Thus,

2
4

1 =

1,
2
4
= and
2
4
= ;

2

1 =

1 and
2
= ;

3

1 =

1 and
3
= .
Moreover

3 is a norm fromQ
3
(

3)(, ), because / has norm

3(1+3)
2
,
so

2
() = .
The determination of
3
() is more delicate. Let be a root of X
3

(1 + 3

3), so = 1 +

3, where is a root of Y
3

3Y
2
Y + 1. Thus
Q
3
(

3)()/Q
3
(

3) is unramied and
Frob
3
()/ (1 +

3
3
)/(1 +

3) (1 +

3)/2 mod 3.
The norms from Q
3
(

3)()

to Q
3
(

3)

are generated by Z
3
[

3]

and
3

3. The norms from Q

3
(

3)()

to Q
3
(

3)

are generated by 1+9Z

3
[

3],
1 +(1 3)

3, 4, 1 and

3. The norms from Q
3
(

3)(, )

to Q
3
(

3)

are generated by 1 + 9Z
3
[

3], 1 + (1 3)

3, 4, 1 and 3

3. Thus
(
3
, Frob
3
) Gal(Q
3
(

3)()/Q
3
(

3)) Gal(Q
3
(

3)()/Q
3
(

3))

= Gal(Q
3
(

3)(, )/Q
3
(

3))
corresponds to

3(1 3

3) Q
3
(

3)

. As has norm to Q
3
(

3) the
product of (

3(1 3

3))
2
and (1 +3)(1 +3

3)/(1 3

3)
2
, we conclude
that (
3
, Frob
3
) xes . Thus
3
()/ = / Frob
3
() = (1

3)/2. In other
words

3
() = (1

3)/2.
Choose an element I
F

i
/Q3
I
F

i
/Q3(

3)
. Then
2

2
4
,
3
). As
3

1
=

2
3
we may alter our choice of so that
2

2
4
). As

3 =

3 we see
that =

1 and so
2
=
2
4
. We will rename as
4
and suppose it chosen
so that
4
=

1. Thus, Gal(F

i
/Q
3
) is generated by elements
2
,
3
and
4
satisfying

2
2
=
3
3
=
4
4
= 1,

2

3
=
3

2
,

4

2
=
2

1
4
,
and
4

3
=
2
3

4
.
The element
2
is a Frobenius lift and it has xed eld F
i
= Q
3
(), where = /
is a uniformiser for F

i
. One can check that

1
3
()/ = (1 (1 + 3)
2

6
)/2.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
904 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
We conclude
G(u) = (1 + 3)
4
,
c

1 mod (3, u
36
),
H
2
(u) = 1,
H
4
(u) =

1,
H

1
3
(u) 1 u
6
mod 3,
t

1
3
(u) 1 u
6
mod (3, u
12
),
t
g
= f
g,g
= 0 for g, g

2
,
4
),
f

1
3
,
1
3
(u), f

1
3
,
1
3
(u) 0 mod (3, u
12
).
7. Proof of Theorem 4.4.1
In this section we will keep the notation of 4.4 and either 6.1 or 6.2 (depending
on whether we are working with S
1
or S
1
). We will set = 1 in the case of S
1
.
We will write F for F
1
and F

for F

1
. If G (resp. M) is a nite at O
F
-group
scheme (resp. Breuil module over O
F
) we will write G

(resp. M

) for the unramied

base change to O
F
.
7.1. Rank one calculations. We recall from Lemma 5.2.1 that the only O
F
-
models for (Z/3Z)
/F
are G(3, )

= (Z/3Z)
/OF
and G(1, ), and the only O
F
-models
for (
3
)
/F
are G(0, 1)

= (
3
)
/OF
and G(2, 1). In each case, by Lemma 5.7.1, the base
change to O
F
admits unique descent data over Q
3
compatible with the canonical
descent data on the generic bre of Z/3Z (resp.
3
) over Q
3
. We will refer to this
descent data as the standard descent data on these nite at group schemes.
7.2. Rank two calculations.
Lemma 7.2.1. The group of extensions of M(2, 1) by M(1, ) over O
F
is param-
etrised by c F
3
. The Breuil module M(1, ; 2, 1; c) corresponding to c is free of
rank two over F
3
[u]/u
9
with a basis e
1
, e

such that
M
1
= ue
1
, u
2
e

+ce
1
),

1
(ue
1
) = e
1
,
1
(u
2
e

+ce
1
) = e

,
N(e
1
) = 0, N(e

) = cu
6
e
1
.
The standard descent data on M(2, 1)

and M(1, )

extends uniquely to descent

data on M(1, ; 2, 1; c)

. The corresponding representations G

3
GL
2
(F
3
) are of
the form
_

0 1
_
and are peu ramie. Any such peu-ramie extension arises for a suitable choice of
c.
Proof. The classication of extensions of Breuil modules follows from Lemma 5.2.2.
Next, we compute N on M= M(1, ; 2, 1; c). (We will not in fact need the result of
this computation of N, but the calculation is given here as a representative sample
of calculations needed later in more complicated settings.) By the last part of
Lemma 5.2.1, N(e
1
) = 0 and N(e

) = ge
1
for some g F
3
[u]/u
e
divisible by u.
In F
3
[u]/u
9
we compute
c

= (G

(u)) = (u
2
)
3
= u
6
+,
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 905
so
u
6
c

= u
6
.
Using the dening properties of N, we compute in F
3
[u]/u
9
N(e

) = N
1
(u
2
e

+ce
1
)
= N(u
2
e

+ ce
1
)
= (u
2
e

+u
2
N(e

))
=

1
c

(u
5
e

+u
5
N(e

))
=

1
c

(u
3
(u
2
e

+ce
1
) +cu
3
e
1
+u
5
N(e

))
=

1
c

(cu
3
e
1
)
since u
5
N(e

) u
6
M = u
3
u
3
M u
3
M
1
and the Frobenius-semi-linear
1
must
kill u
3
M
1
. Thus,
N(e

) =

1
c

(cu
2
ue
1
) =
cu
6
c

1
(ue
1
) =
cu
6
c

e
1
= cu
6
e
1
.
To see existence and uniqueness of the descent data on M(1, ; 2, 1; c)

compatible
with the standard descent data on M(1, )

and M(2, 1)

we will work on the side of

nite at group schemes. Because G(1, ; 2, 1; c)

is the unique extension of G(1, )

by G(2, 1)

with generic bre G(1, ; 2, 1; c)

(by Lemma 4.1.2), uniqueness

reduces to the corresponding questions on the generic bre, which follows from the
injectivity of
H
1
(G
3
, ) H
1
(G
F
, ).
For existence it suces to exhibit a continuous representation G
3
GL
2
(F
3
) of
the form
_

0 1
_
which is peu ramie but not split, with restriction to G
F
corresponding to a local-
local nite at O
F
-group scheme G. By Theorem 5.3.2 of [Man] we can nd an
elliptic curve E
/Q3
such that E[3] furnishes the desired example. This also proves
the nal two assertions of the lemma.
Lemma 7.2.2. Suppose that

F
1
is a totally ramied abelian cubic extension of Q
3
and suppose that G is a local-local nite at O

F1
-group scheme killed by 3 such that
G

F
1
is an extension of Z/3Z by
3
. Then G
O

F
1

F
1

= G
Q3

F
1
for some nite
at Q
3
-group scheme G.
Proof. As in the proof of the last lemma we see that M

(G)

= M(1, 1; 2, 1; c) for
some c F
3
. As the only action of Gal(

F
1
/Q
3
) on a one-dimensional F
3
-vector
space is trivial, we see that each such c gives a class in H
1
(G

F1
, ) which is invariant
by Gal(

F
1
/Q
3
). But
H
1
(G
3
, )

H
1
(G

F1
, )
Gal(

F1/Q3)
,
and so the lemma follows.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
906 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
Lemma 7.2.3. The group of extensions of M(1, ) by M(2, 1) over O
F
is iso-
morphic to the group of linear polynomials c + c

u in F
3
[u]. The Breuil module
M(2, 1; 1, ; c +c

u) corresponding to c +c

u is free of rank two over F

3
[u]/u
9
with
a basis e

, e
1
such that
M(2, 1; 1, ; c +c

u)
1
= u
2
e

, ue
1
+ (c +c

u)e

),

1
(u
2
e

) = e

,
1
(ue
1
+ (c +c

u)e

) = e
1
.
Each M(2, 1; 1, ; c +c

u)

admits unique descent data compatible with the standard

descent data on M(1, )

and M(2, 1)

. As c, c

vary over F
3
the corresponding de-
scent to Q
3
of the generic bre of G

(M(2, 1; 1, ; c+c

u)

) runs over all 9 extensions

of
3
by Z/3Z. The corresponding representation of G
3
is peu ramie if and only
if c = 0.
Proof. The classication of extensions of Breuil modules follows from Lemma 5.2.2.
The uniqueness of the descent data on M(2, 1; 1, ; c + c

u)

follows from Lemma

4.1.2 and the injectivity of H
1
(G
3
,
1
) H
1
(G
F
,
1
) as in the proof of Lemma
7.2.1. Note that Frobenius vanishes on the Dieudonne module of G(2, 1; 1, ; c+c

u)
if and only if c = 0. Thus the lemma will follow if for each 3-torsion extension G of

3
by Z/3Z over Q
3
which is tr`es ramie, we can nd a nite at O
F
-group scheme
G such that
the generic bre of G is isomorphic to GF,
the closed bre of G is local-local,
and Frobenius is not identically zero on D(G).
The splitting eld of G contains a cube root of 3v for some v 1 mod 3, where
the three choices of v mod 9 correspond to the three choices of tr`es ramie . The
calculations in 5.3 of [Man] give explicit additive reduction elliptic curves E and
E

over Q
3
with E[3] E

[3] G, where E acquires good supersingular reduction

over the non-Galois cubic ramied extension
Q
3
[X]/(X
3
3X + 2b),
with b
2
= 1 + 3v, and E

acquires good supersingular reduction over the abelian

cubic ramied extension of Q
3
with norm group generated by 3v mod (Q

3
)
3
. The
appropriate G are provided by the 3-torsion on the Neron models of E or E

over
O
F
.
Corollary 7.2.4. Suppose that G is a nite at O
F
-group scheme and that [g]
is descent data on G

= GO
F
such that (G

, [g])
Q3
(Q
3
) corresponds to . Then
M

(G)

= M(2, 1; 1, ; c +c

u)
for some c, c

F
3
with c ,= 0.
Proof. From the connected-etale exact sequence and its dual we see that G F
3
must be local-local. The corollary now follows from Lemma 7.2.3 and the discussion
of 7.1.
Lemma 7.2.5. The group of extensions of M(1, ) by M(1, ) over O
F
is iso-
morphic to the group of linear polynomials b + b

u in F
3
[u]. The Breuil module
M(1, ; 1, ; b +b

u) corresponding to b +b

u is free of rank two over F

3
[u]/u
9
with
a basis e, e

such that
M(1, ; 1, ; b +b

u)
1
= ue, ue

+ (b +b

u)e),

1
(ue) = e,
1
(ue

+ (b +b

u)e) = e

.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 907
This extension splits over an unramied extension if and only if b = 0. If F

/Q
3
is
non-abelian, then any descent data on M(1, 1; 1, 1; b +b

u)

compatible with the

standard descent data on M(1, 1)

satises

2
e = e,
2
e

= e

1
3
(e) = H

1
3
(u)
3
e,

1
3
(e

) = H

1
3
(u)
3
e

+h

1
3
(u)e,
where
h

1
3
(0) = bH

1
3
(0).
Proof. The classication of extensions of Breuil modules follows from Lemma 5.2.2.
The computation of which of these split over an unramied extension follows from
Lemma 5.2.2 and Corollary 5.4.2.
Now suppose that F

/Q
3
is non-abelian. By Lemma 5.7.1, the only issue is to
compute h
3
(0). Since H
3
(0) 1 mod 3, by evaluating the congruence

3

1

3
mod u M(1, 1; 1, 1; b +b

u)

on ue

+ (b + b

u)e and comparing constant terms of the coecients of e on both

sides we get
h
3
(0) = h
3
(0)
3
+b
_
1 H
3
(u)
u
_
3
[
u=0
= h
3
(0)
3
bH

3
(0)
3
= h
3
(0)
3
+bH

3
(0)
in F
9
, where we have used the equality H

3
(0)
2
1 mod 3 (see 6.2).
In other words h
3
(0) is a root of T
3
T +bH

3
(0) = 0. Since H

3
(0)
2
= 1, we
must have h
3
(0) = bH

3
(0)+a for some a F
3
. Since
2
(H
3
(u)) = H

1
3
(u) and

2
(h
3
(u)) = h

1
3
(u) are forced by the identity
2
() = , we see that h

1
3
(0) =
bH

1
3
(0) +a for the same a F
3
. The identity

3

1
3

1
mod u M(1, 1; 1, 1; b +b

u)

then implies h
3
(0) +h

1
3
(0) = 0, so a = 0.
Lemma 7.2.6. The group of extensions of M(2, 1) by M(2, 1) over O
F
is isomor-
phic to the group of quadratic polynomials vanishing at 0, (b +b

u)u, in F
3
[u]. The
Breuil module M(2, 1; 2, 1; (b +b

u)u) corresponding to (b +b

u)u is free of rank two

over F
3
[u]/u
9
with a basis e, e

such that
M(2, 1; 2, 1; (b +b

u)u)
1
= u
2
e, u
2
e

+ (b +b

u)ue),

1
(u
2
e) = e,
1
(u
2
e

+ (b +b

u)ue) = e

.
This extension splits over an unramied extension if and only if b = 0. If F

/Q
3
is
non-abelian, then any descent data on M(2, 1; 2, 1; (b + b

u)u)

compatible with the

standard descent data on M(2, 1)

satises

3
(e) = H
3
(u)
6
e,
3
(e

) = H
3
(u)
6
e

+h
3
(u)e,
where
h
3
(0) = bH

3
(0).
The sign in h
3
(0) = bH

3
(0) will be very important in 7.4. The proof of this
lemma is essentially the same as that of Lemma 7.2.5, but we repeat it anyway.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
908 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
Proof. The classication of extensions of Breuil modules follows from Lemma 5.2.2.
The computation of which of these split over an unramied extension follows from
Lemma 5.2.2 and Corollary 5.4.2.
Now suppose that F

/Q
3
is non-abelian. By Lemma 5.7.1, the only issue is to
compute h
3
(0). Since H
3
(0) 1 mod 3, by evaluating the congruence

3

1

3
mod u M(2, 1; 2, 1; (b + b

u)u)

on ue

+ (b + b

u)e and comparing constant terms of the coecients of e on both

sides we get
h
3
(0) = h
3
(0)
3
+b
_
1 H
3
(u)
u
_
3
[
u=0
= h
3
(0)
3
bH

3
(0)
3
= h
3
(0)
3
+bH

3
(0)
in F
9
, where we have used the equality H

3
(0)
2
= 1 (see 6.2).
In other words h

(0) is a root of T
3
T +bH

3
(0) = 0. Since H

3
(0)
2
= 1, we
must have h
3
(0) = bH

3
(0)+a for some a F
3
. Since
2
(H
3
(u)) = H

1
3
(u) and

2
(h
3
(u)) = h

1
3
(u) are forced by the identity
2
() = , we see that h

1
3
(0) =
bH

1
3
(0) +a for the same a F
3
. The identity

3

1
3

1
mod u M(2, 1; 2, 1; b +b

u)

then implies h
3
(0) +h

1
3
(0) = 0, so a = 0.
7.3. Rank three calculations.
Lemma 7.3.1. Suppose that G is a nite at group scheme over O
F
which is killed
by 3. Suppose that there is a ltration by closed nite at subgroup schemes G
1

G
2
G such that G
1

= G(1, ), G
2
/G
1

= G(2, 1) and G/G
2

= G(1, ). Suppose nally
that G
2

OF
F

descends to Q
3
in such a way that it is a tr`es ramie extension of

3
by Z/3Z. Then
G/G
1

= G(2, 1) G(1, )
compatibly with the extension class structure.
Proof. Let M= M

(G) and N = M

(G/G
1
). Using Lemmas 7.2.1 and 7.2.3 we see
that we can write
M= (F
3
[u]/u
9
)e
1
(F
3
[u]/u
9
)e

(F
3
[u]/u
9
)e

1
,
M
1
= ue
1
, u
2
e

+be
1
, ue

1
+ (c +c

u)e

+fe
1
)
for b, c, c

F
3
with c ,= 0 and with f F
3
[u]/u
9
. It suces to show b = 0. Since
we must have u
3
M M
1
, we see that
(c +c

u)(u
2
e

+be
1
) u
2
(ue

1
+ (c +c

u)e

+fe
1
) +u
3
e

1
= (bc +bc

u u
2
f)e
1
M
1
.
The Breuil module N is spanned as a F
3
[u]/u
9
-module by e
1
and e

, so by Lemma
7.2.1 u must divide bc +bc

u u
2
f. As c ,= 0 we must have b = 0, as desired.
Combining this with Lemma 7.2.1 and the injectivity of H
1
(G
3
, ) H
1
(G
F
, )
we get the following corollary, which is also the rst part of Theorem 4.7.3.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 909
Corollary 7.3.2. The natural map

0
: Ext
1
S1
(, ) H
1
(G
3
, )
is zero.
7.4. Conclusion of the proof of Theorem 4.4.1. Consider rst the case of F
1
.
We still have to explain why

: H
1
S1
(G
3
, ad
0
) H
1
(I
3
, F
3
)
is zero. Suppose x H
1
S1
(G
3
, ad
0
) does not map to zero in H
1
(I
3
, F
3
).
By our hypothesis on x we may choose a totally ramied abelian cubic exten-
sion

F
1
/Q
3
such that x restricts to zero under the natural map H
1
(G
3
, ad
0
)
H
1
(G

F1
, F
3
). Then the image of x under the natural map H
1
(G
3
, ad
0
)
H
1
(G

F1
, ) is the image of some x H
1
(G

F1
, ) under the natural map
H
1
(G

F1
, ) H
1
(G

F1
, ). The element x parametrises a nite at

F
1
-group
scheme H which is an extension of
3
by Z/3Z and which is a subquotient of the
restriction to G

F1
of the extension of by itself parametrised by x. It follows that
H has a nite at model H
/O

F
1
(see Lemma 4.1.1) and the special bre of H must
be local-local (if x = 0, then the extension of by itself parametrised by x splits
over

F
1
and this is clear, while if x ,= 0 we would otherwise get a contradiction
from the connected-etale sequence). By Lemma 7.2.2, we may therefore lift x to
H
1
(G
3
, ). Using the commutative diagram
H
1
(G
3
, ) H
1
(G
3
, )
res res
H
1
(G

F1
, ) H
1
(G

F1
, )
and noting that the right-hand vertical map is injective we conclude that
x H
1
S1
(G
3
, ad
0
) H
1
(G
3
, )
is in the image of H
1
(G
3
, ) H
1
(G
3
, ), a contradiction with the hypothesis
that even the image of x in H
1
(I
3
, F
3
) is non-zero.
Now consider the case F

= F

1
which is non-abelian over Q
3
. We must show
that

: Ext
1
S1
(, ) H
1
(I
3
, F
3
)
is zero.
An element x Ext
1
S1
(, ) gives rise to a nite at O
F1
-group scheme G killed
by 3 and descent data [g] for F

1
/Q
3
on G

= G
O
F

1
F

1
, such that (G

, [g])
Q3
corresponds to the extension of by itself classied by x. Let N denote the Breuil
module for G and let N

= N F
9
. According to Lemmas 7.2.1, 7.2.3, 7.2.5, 7.2.6
and 7.3.1 we see that we can write
N = (F
3
[u]/u
9
)e

(F
3
[u]/u
9
)e
1
(F
3
[u]/u
9
)e

(F
3
[u]/u
9
)e

1
with
(7.4.1) N
1
= u
2
e

, ue
1
+ (c +c

u)e

, u
2
e

+ (au +a

u
2
)e

,
ue

1
+ (c +c

u)e

+ (b +b

u)e
1
+he

),
where h F
3
[u]/u
9
is some polynomial and where a, a

, b, b

, c, c

F
3
with c ,= 0
(as is tr`es ramie). By Lemma 7.2.6 what we must show is that a = 0.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
910 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
Note that H

3
(0) ,= 0 in F
9
by 6.2. By Lemmas 5.7.1 and 7.2.1, the action
3
is determined by

3
(e

) = H
3
(u)
6
e

,
3
(e
1
) = H
3
(u)
3
e
1
+g
3
(u)e

,

3
(e

) = H
3
(u)
6
e

+h
3,
(u)e

,

3
(e

) = H
3
(u)
3
e

1
+g
3
(u)e

+h
3,1
e
1
+G
3
(u)e

,
where g
3
(u), G
3
(u) F
9
[u]/u
9
and h
3,
and h
3,1
are as in Lemmas 7.2.6 and
7.2.5 respectively.
Due to the requirement
3
(N

1
) N

1
, we must have

3
(ue

1
+ (c +c

u)e

+ (b +b

u)e
1
+h(u)e

) N

1
,
and this element is obviously equal to
(uH
3
)(H
3
3
e

1
+g
3
e

+h
3,1
e
1
+G
3
e

) + (c +c

uH
3
)(H
6
3
e

+h
3,
e

)
+(b +b

uH
3
)(H
3
3
e
1
+g
3
e

) +h(uH
3
)H
6
3
e

.
We now try to express this as a linear combination of the generators of N

1
listed
in (7.4.1), while working modulo u
3
N

, u
2
e

) N

1
. Using that H
3
(0) = 1 in F
9
and h(uH
3
) h(u) mod u
2
, we arrive at the expression
H
3
(ue

1
+ (c +c

u)e

+ (b +b

u)e
1
+he

)
+
_
c((1 H
3
)/u) +g
3
u
_
(u
2
e

+ (au +a

u
2
)e

)
+
_
H
3
h
3,1
+b
_
1 H
3
u
__
(ue
1
+ (c +c

u)e

) +F
3
(u)e

,
where
F
3
(u) = uH
3
G
3
+ (c +c

uH
3
)h
3,
+ (b +b

uH
3
)g
3
+h(u)(1 H
3
)
(a +a

u)(c(1 H
3
)/u +g
3
) (c +c

u)(H
3
h
3,1
+b((1 H
3
)/u))
in F
9
[u]/u
9
. In particular, c(1H
3
(u))/u+g
3
0 mod u and F
3
(u) 0 mod u
2
.
The condition c((1 H
3
)/u) +g
3
0 mod u can be reformulated as
g
3
(0) = cH

3
(0).
Since F
3
(u) 0 mod u
2
, we have to have F
3
(0) = 0. But a direct calculation
using g
3
(0) = cH

3
(0) and the denition of F
3
gives
F
3
(0) = 0 +ch
3,
(0) +bg
3
(0) + 0 0 c(h
3,1
(0) bH

3
(0))
= c(h
3,
(0) h
3,1
(0) bH

3
(0)),
so the non-vanishing of c forces
h
3,
(0) h
3,1
(0) = bH

3
(0).
Lemmas 7.2.6 and 7.2.5 give us the values
h
3,
(0) = aH

3
(0), h
3,1
(0) = bH

3
(0).
Thus (a+b)H

3
(0) = bH

3
(0), and so a = 0. This completes the proof of Theorem
4.7.3 and hence of Theorem 4.4.1.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 911
8. Proof of Theorem 4.5.1
In this section we will keep the notation of 4.5 and either 6.3 or 6.4 (depending
on whether we are working with S
3
or S
3
). We will set = 1 in the case of S
3
(so that c

mod (3, u)). Note the signs. We will write F for F

3
, F

for F

3
and I for I
3
. If G (resp. M) is a nite at O
F
-group scheme (resp. Breuil module
over O
F
) we will write G

(resp. M

) for the base change to O

F
.
8.1. Rank one calculations. We remark that with our choice of polynomials
H
g
(u) in 6.3 and 6.4, any object M in
1
DD
F

/Q3
has an action of
2
,
4
) (via

2
and
4
, the action of
2
being Frob
3
-semi-linear). Also, since
3
and
2
commute,
H
2
= 1 and H

1
3
(u) Z
3
[u], we see that
2
must commute with

1
3
by Corollary
5.6.2.
We recall from Lemma 5.2.1 that the only models for (Z/3Z)
/F
over O
F
are
G(r, ) for r = 0, 2, 4, 6, 8, 10, 12 with G(12, )

= (Z/3Z)
/OF
, and the only models
for (
3
)
/F
over O
F
are G(r, 1) for r = 0, 2, 4, 6, 8, 10, 12 with G(0, 1)

= (
3
)
/OF
. In
each case, the base change to O
F
admits unique descent data over Q
3
such that
descent of the generic bre to Q
3
is Z/3Z (resp.
3
). (See Lemma 5.7.1.) We
will write G

r,1
(resp. G

r,
) for the corresponding pair (G(r, ) O
F
, [g]) (resp.
(G(r, 1) O
F
, [g])). We will also let M

r,1
(resp. M

r,
) denote the corresponding
object of
1
DD
F

/Q3
. In particular, for = 1 or , the underlying F
9
[u]/u
36
-
module has the form (F
9
[u]/u
36
)e

with e

the standard generator, though we

write e rather than e

if is understood.
We have the following useful lemma.
Lemma 8.1.1. Let 0 r e = 12 be an even integer. The descent data on M

r,1
is determined by

2
(e) = e,
4
(e) = (

1)
r/2
e,

1
3
(e) = H

1
3
(u)
3r/2
e,
and the descent data on M

r,
is determined by

2
(e) = e,
4
(e) = (

1)
r/2
e,

1
3
(e) = H

1
3
(u)
3r/2
e.
In particular,
2
4
= 1 on D(G

r,1
) if and only if
2
4
= 1 on D(G

r,
) if and only
if r = 2, 6 or 10.
Proof. Certainly
2
(e) = e. We have already seen in Lemma 5.7.1 that descent
data must exist in each case, so our task is to compute the unique units
4
,
3
1
(F
9
[u]/u
36
)

so that

4
(e) =
4
e,

3
1
(e) =
3
1e
corresponds to generic bre descent data for the mod 3 cyclotomic or trivial char-
acter on G
3
. The case of
3
1
follows from Lemma 5.7.1.
From the condition

4

1
(u
r
e) =

1

4
(u
r
e)
we get
2
4
(u) = (

1)
r
, so

4
(u) = (

1)
r/2
.
The non-zero morphisms M
r,1
M
12,1
are given by e u
3(12r)/2
e and the
non-zero morphisms M
0,
M
r,
are given by e u
3r/2
e. Thus, it suces to
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
912 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
check that
4
e = e on M

12,1
and
4
e = e on M

0,
. In both cases we have shown
that
4
e = e and so we only need to check that
4
= 1 on D(G

12,1
) and D(G

0,
).
That is, we have to show that the O
F
-group scheme maps Z/3Z
4
(Z/3Z) and

3

4

3
arising from the canonical generic bre descent data induce the identity
on the special bres. This is easy.
Lemma 8.1.2. Let Mbe an object of
1
mod
F
corresponding to a nite at group
scheme G and let [g] be descent data on G

= GO
F
relative to Q
3
. Assume that
(G

, g)
Q3
can be ltered so that each graded piece is isomorphic to Z/3Z or
3
and so that the corresponding ltration of (M

, g) in
1
DD
F

/Q3
has successive
quotients of the form M

rj,j
with r
j
2, 6, 10 and
j
1, . Then
2
4
= 1
on M

/u M

and there exists a basis e

j
of M over F
3
[u]/u
36
so that for all j
e
j

1
(M
1
),
e
j
is an eigenvector of the F
9
-linear map
4
on M

,
e
j
lies in the part of the ltration of M

which surjects onto M

rj,j
and this
surjection sends e
j
onto the standard basis vector e of M

rj,j
over F
9
[u]/u
36
.
Proof. Since
2
4
acts linearly on M

/u M

and (
2
4
)
2
= 1, the action of
2
4
must be
semi-simple. The eigenvalues of
2
4
are all equal to 1, so necessarily
2
4
= 1 on
M

/u M

.
We now argue by induction on the number of Jordan-H older factors in the generic
bre, the case of length 1 being clear. Thus, we can assume we have a short exact
sequence in
1
DD
F

/Q3
,
0 N

r,
0,
so the lemma is known for N

. We just have to nd e
0

1
(M
1
) mapping onto
the standard basis vector e in M

r,
such that e
0
is an eigenvector of
4
. Since

1
(M

1
)

1
((M

r,
)
1
) is a surjective map of F
9
-vector spaces which is compatible
with the semi-simple F
9
-linear endomorphism
4
on each side, we can nd e

1
(M

1
) mapping onto e with e

0
an eigenvector of
4
, say
4
(e

0
) = (

1)
1
e

0
.
Since

4

2
(e

0
) =
2

3
4
(e

0
) =
2
(

1
1
e

0
) =

1
1

2
(e

0
),
the element e
0
= (1/2)(e

0
+
2
(e

0
)) maps to e and is an an eigenvector for
4
.
Also, e
0

1
(M

1
) is
2
-invariant and
2
commutes with

1
, so e
0

1
(M
1
).
8.2. Models for .
Proposition 8.2.1. There exists a unique object (G

, [g]) of FD
F

/Q3,I
such
that (G

, [g])
Q3
corresponds to . If we set (M()

, g) = M

(G

, [g]), then
(M()

, g) is an extension of M

2,
by M

10,1
in
1
DD
F

/Q3
. Moreover Frobenius
is not identically zero on D(G

).
Proof. Let (G

, [g]) be an object of FD
F

/Q3,I
such that (G

, [g])
Q3
corresponds
to , and set (M

, g) = M

(G

, [g]). As in the discussion following Theorem

5.6.1, we have canonically M

F
9

F3
M for a Breuil module M over O
F
, with
2
acting as
2
1. By Lemma 8.1.1, there is a short exact sequence of Breuil modules
over O
F
,
0 M(s, ) MM(r, 1) 0,
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 913
with r, s 2, 6, 10 and this is compatible with descent data after base change to
O
F
in the sense that we obtain an exact sequence
0 M

s,1
M

r,
0,
compatible with descent data. Because is tr`es ramie, it follows that [
G
F

is
non-split, so the sequence
0 M(s, ) MM(r, 1) 0
is non-split.
We rst show that we must have (r, s) = (2, 10). Since is self-dual, in order
to prove (r, s) = (2, 10) we may use Cartier duality (and Lemma 5.2.1) in order to
reduce to the case where r +s e = 12. We will rst rule out cases with r s and
then the case (r, s) = (2, 6).
By Lemmas 8.1.1 and 8.1.2, we can write
M= (F
3
[u]/u
36
)e
1
(F
3
[u]/u
36
)e

, M
1
= u
s
e
1
, u
r
e

+he
1
)
for some h F
3
[u]/u
36
so that

1
(u
s
e
1
) = e
1
,
1
(u
r
e

+he
1
) = e

and

4
(e
1
) = (

1)
s/2
e
1
,
4
(e

) = (

1)
r/2
e

.
Recall from Lemma 5.2.2 that the parameter h gives an isomorphism of abstract
groups
(F
3
[u]/u
36
)/u
s
t u
r
t
3
[t F
3
[u]/u
36
Ext
1
1mod
F
(M(r, 1), M(s, )).
It is easy to see that

4
(M

1
) M

1
,
4

1
=

1

4
on M

1
if and only if
4
(u
r
e

+he
1
) M

1
and
4
(e

) =

4
(u
r
e

+he
1
), or equivalently
(

1)
r/2
h(u) (

1)
s/2
h(

1u) mod u
12+s
.
This says exactly that
j 2 (r +s)/2 mod 4 (8.2.1)
for any j < 12 +s with a non-zero u
j
term appearing in h.
If (r, s) = (6, 2) this would force h 0 mod u
2
, yet u
2
t u
6
t
3
[t F
3
[u]/u
36

contains all multiples of u

2
, so
0 M(2, ) MM(6, 1) 0
is split, a contradiction.
When (r, s) = (10, 2) or (r, s) = (2, 2) we see that h h(0) mod u
4
, yet
u
4
(F
3
[u]/u
36
) u
s
t u
r
t
3
[t F
3
[u]/u
36
,
so the choice of e

may be changed in order to arrange that

h F
3
(though making this change of basis of M may destroy the diagonal form of
4
).
Since
0 M(s, ) MM(r, 1) 0
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
914 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
is non-split, necessarily h ,= 0, so by rescaling e

it can be assumed that h = 1.

Then V
M
(e

) e
1
mod u M (by Theorem 5.1.3) and
(e

) (/c

)u
3(12rs)
e
1
u
3(12rs)
e
1
mod u M.
This forces r +s = 12. In particular, (r, s) = (2, 2) is ruled out.
For (r, s) = (10, 2), a splitting of the generic bre [
F
is induced by the Breuil
module map
M(0, 1) M
dened by
e u
15
e

+ u
3
fe
1
= u
5
(u
10
e

+e
1
) + (uf u
3
)u
2
e
1
,
where f F
3
[u]/u
36
satises f
3
f = u
6
(i.e. f = u
6
u
18
, and a constant
c F
3
can even be added to this if = 1). But [
G
F

must be non-split, so this

rules out (r, s) = (10, 2).
The remaining case with r s is (r, s) = (6, 6). In this case u
s
t u
r
t
3
[t
F
3
[u]/u
36
contains all multiples of u
8
. But we have j 0 mod 4 for all j < 12+s =
18 such that a non-zero u
j
term appears in h, so again (at the expense of possibly
making the
4
-action non-diagonal) we may assume
h = c +c

u
4
for some c, c

F
3
. Writing
4
(e

) = (

1)e

+h
4
(u)e
1
, the commutativity of
4
and

1
amounts to h
4
= h
3
4
, so h
4
(u) = b

for some b F
3
. The condition

4
4
(e

) = e

forces b = 0, so
4
still has diagonal action. This analysis shows that
the map of F
3
-vector spaces
Ext
1
1DD
F

/Q
3
(M

6,
, M

6,1
) Ext
1
1mod
F
(M(6, 1), M(6, ))
has at most a 2-dimensional image. If c

+c = 0, then the Breuil module map

F
9
M(0, 1) M

dened by
e cu
7
e
1
+u
3
(u
6
e

+ (c +c

u
4
)e
1
)
gives a splitting of the corresponding representation of G
F
. Thus the image of
Ext
1
1DD
F

/Q
3
(M

6,
, M

6,1
) Ext
1
F3[G
F
]
(1, ) (8.2.2)
is at most one dimensional and, because [
G
F

is non-split, the pair (c, c

) corre-
sponding to a model of satises c

+c ,= 0.
At this point, we treat the cases = 1 separately. Consider rst the case = 1.
We must have

3
(e
1
) = H
3
(u)
9
e
1
,
3
(e

) = H
3
(u)
9
e

+h
3
(u)e
1
,
where h
3
(u) F
9
[u]/u
36
lies in F
3
[u]/u
36
because
3
commutes with
2
. Evaluat-
ing
3

1

3
mod u M

on u
6
e

+(c+c

u
4
)e
1
M

1
and using our knowledge
of H
3
(u) mod 3, we arrive at
h
3
(0) = (h
3
(0)
3
+ (c +c

)),
which is impossible for h
3
(0) F
3
with = 1 because c +c

= c +c

3
.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 915
Now let us turn to the case = 1, still in the case (r, s) = (6, 6). In this case
Ext
1
F3[G3]
(1, ) Ext
1
F3[G
F
]
(1, ) is injective and so by (8.2.2) we see that the
image of
Ext
1
1DD
F

/Q
3
(M

6,
, M

6,1
) Ext
1
F3[G3]
(1, )
is at most one dimensional. Thus to exclude the case (r, s) = (6, 6) and = 1, it
suces to show that this image contains the peu ramie line (as is tr`es ramie). By
Proposition 5.2.1 of [Man], there is an elliptic curve E

/Q3
which has supersingular
reduction over Q
3
(

1, ), with
E

,3
a non-split, peu ramie extension of 1 by .
The representation
E

,3
[
F
is non-split (again because H
1
(G
3
, ) H
1
(G
F
, ) is
injective in the = 1 case). Let N

be the Breuil module corresponding to the

3-torsion on the Neron model of E

Q3
F

, so N

admits descent data g

via
the universal property of Neron models. The ltration of induces a short exact
sequence in
1
DD
F

/Q3
0 M

a,1
(N

, g

) M

b,
0
for some even a, b with 2 a, b 10. The Neron model of E

Q3
Q
3
(

1, )
has local-local 3-torsion, and the induced local-local integral models G

and G
1
of
the diagonal characters [
Q3(

1,)
and 1[
Q3(

1,)
must be the unique local-local
models (uniqueness follows from Corollary 1.5.1 of [Ra]). Moreover, Corollary 1.5.1
of [Ra] makes it clear that base change to O
F
takes the order 3 group schemes
G

and G
1
to the integral models that lie in the middle of the well-ordered sets of
integral models of [
F
and 1[
F
. It follows that a = b = 6, so the map
Ext
1
1DD
F

/Q
3
(M

6,
, M

6,1
) Ext
1
F3[G3]
(1, )
indeed hits the peu ramie line.
We next exclude the case (r, s) = (2, 6). As a rst step, we check that there is
at most one possibility for the underlying Breuil module M (ignoring the extension
class structure) if (r, s) = (2, 6). We can write
M= (F
3
[u]/u
36
)e
1
(F
3
[u]/u
36
)e

, M
1
= u
6
e
1
, u
2
e

+he
1
)
for some necessarily non-zero h F
3
[u]/u
36
with

1
(u
6
e
1
) = e
1
,
1
(u
2
e

+he
1
) = e

and

4
(e
1
) =

1e
1
,
4
(e

) =

1e

.
The combined conditions
4
(M

1
) M

1
and

1

4
=
4

1
on M

1
are equivalent
to
h(u) h(

1u) mod u
18
.
Since u
6
t u
2
t
3
[t F
3
[u]/u
36
contains u
6
u
2
and all multiples of u
9
, we
may change e

(at the expense of possibly losing the diagonal form for

4
) so that
h = cu
2
for some c F
3
. Since h is necessarily non-zero, we may rescale to get
h = u
2
, so there is indeed at most one possibility for the underlying Breuil module
M when (r, s) = (2, 6).
Again we treat the cases = 1 separately. Consider rst the case = 1. We
have seen above that there is an extension E
6,6
= (N

, g

) of M

6,
by M

6,1
in
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
916 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR

1
DD
F

/Q3
corresponding to a non-split, peu ramie extension of 1 by . Pulling
back E
6,6
by a non-zero map
M

2,
M

6,
in
1
DD
F

/Q3
given by e u
6
e, we get an extension E
2,6
of M

2,
by M

6,1
in

1
DD
F

/Q3
corresponding to a non-split, peu ramie extension of 1 by . The
underlying Breuil module of E
2,6
must be isomorphic to F
9

F3
M for our uniquely
determined M (with h = u
2
). By the injectivity of H
1
(G
3
, ) H
1
(G
F
, ) in the
= 1 case, we conclude that F
9

F3
M cannot admit descent data giving rise to
a tr`es ramie element in Ext
1
F3[G3]
(1, ). This rules out the case (r, s) = (2, 6) and
= 1.
Now turn to the case (r, s) = (2, 6) and = 1. We will show that with the Breuil
module M constructed above (with h = u
2
), the Breuil module M

= F
9

F3
M
does not admit descent data relative to F

/Q
3
(with
2
=
2
1, without loss of
generality). One checks that N(e
1
) = N(e

) = 0, so
N
1
= 0.
We must have

3
(e
1
) = H
3
(u)
9
e
1
,
3
(e

) = H
3
(u)
3
e

+h
3
(u)e
1
for some h
3
F
9
[u]/u
36
. As usual, since
3
and
2
must commute, we have
h
3
F
3
[u]/u
36
. The condition
3
(M

1
) M

1
is equivalent to

3
(u
2
e

+u
2
e
1
) M

1
,
which amounts to
h
3
(u) H
3
3
H
9
3
0 mod u
4
,
so

3
(u
2
e

+u
2
e
1
) = H
3
(u)
1
(u
2
e

+u
2
e
1
) +
_
h
3
H
3
3
+H
9
3
u
4
_
H
2
3
u
6
e
1
.
As N
1
= 0, we have

3

1
=

1

3
on M

1
. Evaluating this identity on u
2
e

+u
2
e
1
M
1
gives
h
3
= H
6
3

_
h
3
H
3
3
+H
9
3
u
4
_
3
,
so h
3
is a cube. Thus, h
3
= u
6
g
3
for some g
3
F
3
[u]/u
36
.
Since H
3
3
1 +u
6
mod u
12
, we compute
H
9
3
H
3
3
u
6
mod u
12
,
so
h
3
H
6
3

_
_
h
3
u
4
_
3
+u
6
_
mod u
7
and
g
3
(0) = g
3
(0)
3
+ 1
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 917
in F
3
. This is absurd. This rules out all possibilities for (r, s) aside from (r, s) =
(2, 10). Uniqueness now follows from Corollary 4.1.5.
From Theorem 5.4.2 of [Man] and Proposition B.4.2 of [CDT] we see that there
is an elliptic curve E
/Q3
such that E[3](Q
3
)

= and
E,3
has type
3
. Let E
denote the Neron model of E
Q3
F

over O
F
. By the Neron property of E
/O
F

we
see that E[3

] has descent data over Q

3
. As in 4.5 we see that I annihilates the
Dieudonne module of E[3

] F
9
. Thus M()

= M

(E[3]) in
1
DD
F

/Q3
and it
follows that Frobenius is non-zero on D(G

).
8.3. Completion of the proof of Theorem 4.5.1.
Lemma 8.3.1. Let (G

, [g]) be the unique object of FD

F

/Q3,I
such that
(G

, [g])
Q3
corresponds to . Set (M()

, g) = M

(G

, [g]). The natural map

of groups
Ext
1
1DD
F

/Q
3
((M()

, g), (M()

, g)) Ext
1
1DD
F

/Q
3
(M

10,1
, M

2,
),
using pushout by (M()

, g) M

2,
and pullback by M

10,1
(M()

, g), is
zero.
Proof. Let (

, g) represent a class in Ext

1
1DD
F

/Q
3
((M()

, g), (M()

, g))
and let (M

, g) be its image in Ext

1
1DD
F

/Q
3
(M

10,1
, M

2,
). By Lemma 5.2.2,
M

= F
9
M with
2
=
2
1 and
M= (F
3
[u]/u
36
)e

(F
3
[u]/u
36
)e

1
, M
1
= u
2
e

, u
10
e

1
+ (c +c

u)e

),
with c, c

F
3
. Also,

1
(u
2
e

) = e

,
1
(u
10
e

1
+ (c +c

u)e

) = e

1
,
and

4
(e

) =

1e

,
4
(e

1
) =

1e

1
+h
4
(u)e

for some h
4
F
9
[u]/u
36
.
The properties
4
(M

1
) M

1
and
4

1
=

1

4
on M

1
amount to
c

= 0, h
4
= h
3
4
u
24
,
so h
4
= 0. If c = 0, then N
1
= 0, so
3

1
=

1

3
on M

1
. From this we
readily see that (M

, g) is split in
1
DD
F

/Q3
, as desired.
Now assume c ,= 0; we will deduce a contradiction. Consider the rank three
Breuil module with descent data
(N, g) = (

, g)/(M

10,1
, g),
where M

10,1
M()

. Then N has an ordered basis e

, e

1
, e

with
respect to which
N
1
= u
2
e

, u
10
e

1
+e

, u
2
e

+he

1
+ (b +b

u)e

)
for some b, b

F
9
and h = a + a

u
4
+a

u
8
F
9
[u]/(u
36
) dened modulo u
10
t
u
2
t
3
(see 8.2.1). Since our base eld F

has absolute ramication degree 12, N

1
contains
u
12
e

= u
10
(u
2
e

+he

1
+ (b +b

u)e

) h(u
10
e

1
+e

) + (h u
10
(b +b

u))e

.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
918 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
From the list of generators of N
1
, it is not dicult to check that in the above
expression for u
12
e

N
1
, u
2
must divide the coecient of e

. Thus a = 0.
We must have N/e

= M()

. Since a = 0, M() has basis e

1
, e

and
M()
1
= u
10
e

1
, u
2
e

+ (a

u
4
+a

u
8
)e

1
).
Since
1
for M() satises

1
(u
10
e

1
) = e

1
,
1
(u
2
e

+ (a

u
4
+a

u
8
)e

1
) = e

,
it follows immediately that 0 mod uM(), which (using Theorem 5.1.3) con-
tradicts Proposition 8.2.1.
Corollary 8.3.2. The natural map

0
: Ext
1
S3
(, ) H
1
(G
3
, )
is zero.
Theorem 4.7.4, and hence Theorem 4.4.1, now follow from the rst case of the
following lemma. We include the second case to simplify the proof.
Lemma 8.3.3. The maps of groups
Ext
1
1DD
F

/Q
3
(M

10,1
, M

10,1
) Ext
1
F3[G3]
(1, 1),
Ext
1
1DD
F

/Q
3
(M

2,
, M

2,
) Ext
1
F3[G3]
(, )
have images inside the line of extension classes that split over an unramied exten-
sion of Q
3
.
Proof. Since
H
1
(G
3
, Z/3) H
1
(G
F
, Z/3)
is injective and induces an isomorphismbetween the subgroups of unramied classes,
it suces to check that
Ext
1
1DD
F

/Q
3
(M

10,1
, M

10,1
) Ext
1
1mod
F
(1, 1),
Ext
1
1DD
F

/Q
3
(M

2,
, M

2,
) Ext
1
1mod
F
(, )
have images consisting of elements split over an unramied extension of F. By
Cartier duality it suces to consider only the second map.
Consider a representative (M

, g) of an element in Ext
1
1DD
F

/Q
3
(M

2,
, M

2,
).
Lemma 5.2.2 ensures that we can write
M= (F
3
[u]/u
36
)e

(F
3
[u]/u
36
)e

, M
1
= u
2
e

, u
2
e

+he

)
for some h = c +c

u +c

u
2
with c, c

, c

F
3
and

1
(u
2
e

) = e

,
1
(u
2
e

+he

) = e

.
We have

4
(e

) =

1e

,
4
(e

) =

1e

+h
4
(u)e

for some h
4
F
9
[u]/u
36
, and the condition
4
(M

1
) M

1
is equivalent to
h(u) h(

1u) mod u
2
,
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 919
so c = c

= 0. The Breuil module extension class M over O

F
(ignoring descent
data) therefore only depends on the parameter c

F
3
. We then have a splitting
F
3

F3
M(2, 1) F
3

F3
M determined by
e ae

+e

,
where a F
3
satises a
3
= a +c

.
9. Proof of Theorems 4.6.1, 4.6.2 and 4.6.3
In this section we will keep the notation of 4.6 and 6.5. We will write F for
F
i
, F

for F

i
and I for I
i
. If G (resp. M) is a nite at O
F
-group scheme (resp.
Breuil module over O
F
) we will write G

(resp. M

) for the base change to O

F
.
9.1. Rank one calculations. We remark that with our choice of polynomials
H
g
(u) in 6.5, any object M in
1
DD
F

/Q3
has an action of
2
,
4
) via
2
and
4
.
(The action of
2
is Frob
3
-semi-linear.) Since
3
and
2
commute and H

1
3
(u)
Z
3
[u] we see that
2
must commute with

1
3
(see Corollary 5.6.2).
By Lemma 5.2.1, the only models for (Z/3Z)
/F
over O
F
are G(r, 1) for r =
0, 2, 4, 6, 8, 10, 12 with G(12, 1)

= (Z/3Z)
/OF
, and the only models for (
3
)
/F
over
O
F
are G(r, 1) for r = 0, 2, 4, 6, 8, 10, 12 with G(0, 1)

= (
3
)
/OF
. Lemma 5.7.1
ensures that the base changes to O
F
admit unique descent data over Q
3
such that
descent of the generic bre to Q
3
is Z/3Z (resp.
3
). We will write G

r,1
(resp. G

r,
)
for the corresponding pair (G(r, 1)
OF
O
F
, [g]) (resp. (G(r, 1)
OF
O
F
, [g])).
We will also let M

r,1
(resp. M

r,
) denote the corresponding object of
1
DD
F

/Q3
.
We have the following useful lemmas, for which the proofs are identical to the
proofs of Lemmas 8.1.1 and 8.1.2.
Lemma 9.1.1. Let 0 r e = 12 be an even integer. The descent data on M
r,1
is determined by

2
(e) = e,
4
(e) = (

1)
r/2
e,

1
3
(e) = H

1
3
(u)
3r/2
e,
and the descent data on M
r,
is determined by

2
(e) = e,
4
(e) = (

1)
r/2
e,

1
3
(e) = H

1
3
(u)
3r/2
e.
In particular,
2
4
= 1 on D(G
r,1
) if and only if
2
4
= 1 on D(G
r,
) if and only
if r = 2, 6 or 10.
Lemma 9.1.2. Let Mbe an object of
1
mod
F
corresponding to a nite at group
scheme G and let [g] be descent data on G

= G
OF
O
F
over Q
3
. Assume that
(G

, g)
Q3
can be ltered so that each graded piece is isomorphic to Z/3Z or
3
and so that the corresponding ltration of (M

, g) in
1
DD
F

/Q3
has successive
quotients of the form M

rj,j
with r
j
2, 6, 10 and
j
1, . Then
2
4
= 1
on M

/u M

and there exists a basis e

j
of M over F
3
[u]/u
36
so that for all j
e
j

1
(M
1
),
e
j
is an eigenvector of the F
9
-linear map
4
on M

,
e
j
lies in the part of the ltration of M

which surjects onto M

rj,j
and this
surjection sends e
j
onto the standard basis vector e of M

rj,j
over F
9
[u]/u
36
.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
920 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
9.2. Models for . Recall that we are assuming that has the tr`es ramie form
_

0 1
_
,
and is not split over F

. We will let
1
DD
F

/Q3,I
denote the full subcategory
of
1
DD
F

/Q3
consisting of objects M

for which the ideal I acts trivially on

(M

/uM

)
F9,Frob3
F
9
.
Proposition 9.2.1. Suppose that (M

, g) is an object of
1
DD
F

/Q3,I
such that
(M

, g)
Q3
is an extension of Z/3Z by
3
. Then we have an exact sequence
(0) M

s,1
M

r,
(0)
with (r, s) = (2, 6), (6, 10), (2, 10) or (6, 6). Moreover we can write M

= M
F3
F
9
with
2
= 1Frob
3
, where Mhas an F
3
[u]/(u
36
)-basis e
1
, e

with e
1
the standard
basis element of M(s, 1) and e

mapping to the standard basis element of M(r, 1).

More precisely we have the following exhaustive list of extension class possibilities,
all of which are well dened. (N denotes the monodromy operator described in
Lemma 5.1.2.)
(1) (r, s) = (2, 6): The natural map
Ext
1
1DD
F

/Q
3
,I
(M

2,
, M

6,1
) Ext
1
F3[G3]
(1, )
is an isomorphism, with elements parametrised by pairs (c, c
1
) F
2
3
corre-
sponding to
M
1
= u
6
e
1
, u
2
e

+cu
2
e
1
),
1
(u
6
e
1
) = e
1
,
1
(u
2
e

+cu
2
e
1
) = e

(so N
1
= 0) with

4
(e
1
) =

1e
1
,
4
(e

) =

1e

1
3
(e
1
) = e
1
,

1
3
(e

) = (1 u
18
)(e

c
1
u
6
e
1
).
The pairs with c = 0 are the ones which generically split over F

. In all
cases 0 mod uM.
(2) (r, s) = (6, 10): The natural map
Ext
1
1DD
F

/Q
3
,I
(M

6,
, M

10,1
) Ext
1
F3[G3]
(1, )
is an isomorphism, with elements parametrised by pairs (c, c
1
) F
2
3
corre-
sponding to
M
1
= u
10
e
1
, u
6
e

+cu
6
e
1
),
1
(u
10
e
1
) = e
1
,
1
(u
6
e

+cu
6
e
1
) = e

(so N
1
= 0) with

4
(e
1
) =

1e
1
,
4
(e

) =

1e

1
3
(e
1
) = (1 u
18
)e
1
,

1
3
(e

) = e

c
1
u
6
e
1
.
The pairs with c = 0 are the ones which generically split over F

. In all
cases 0 mod uM. These cases are Cartier dual to the (2, 6) cases above.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 921
(3) (r, s) = (2, 10): The natural map
Ext
1
1DD
F

/Q
3
,I
(M

2,
, M

10,1
) Ext
1
F3[G3]
(1, )
is an isomorphism, with elements parametrised by pairs (c, c
1
) F
2
3
corre-
sponding to
M
1
= u
10
e
1
, u
2
e

+cu
8
e
1
),
1
(u
10
e
1
) = e
1
,
1
(u
2
e

+cu
8
e
1
) = e

(so N
1
= 0) with

4
(e
1
) =

1e
1
,
4
(e

) =

1e

1
3
(e
1
) = (1 u
18
)e
1
,

1
3
(e

) = (1 u
18
)(e

c
1
u
12
e
1
).
The pairs with c = 0 are the ones which generically split over F

. In all
cases 0 mod uM.
(4) (r, s) = (6, 6): The natural map
Ext
1
1DD
F

/Q
3
,I
(M

6,
, M

6,1
) Ext
1
F3[G3]
(1, )
is an isomorphism, with elements parametrised by pairs (c, c

) F
2
3
corre-
sponding to
M
1
= u
6
e
1
, u
6
e

+ (c +c

u
4
)e
1
),
1
(u
6
e
1
) = e
1
,
1
(u
6
e

+ (c +c

u
4
)e
1
) = e

(it is easily checked that N(e

1
) = 0 and N(e

) = c

u
30
e
1
) and

4
(e
1
) =

1e
1
,
4
(e

) =

1e

1
3
(e
1
) = e
1
,

1
3
(e

) = e

+ (c c

u
12
c

u
30
)e
1
.
In particular, 0 mod uM if and only if c = 0.
In the rst three cases, the peu ramie condition on a class in Ext
1
F3[G3]
(1, ) is
equivalent to the vanishing of c
1
. In the fourth case it is equivalent to the vanishing
of c.
Proof. By Lemma 9.1.1 we have an exact sequence
(0) M

s,1
M

r,
(0)
with r, s 2, 6, 10. As usual
M
1
= u
s
e
1
, u
r
e

+he
1
).
In the cases (r, s) = (2, 2) and (6, 2) as in the proof of Proposition 8.2.1 we may
take h = 0. We will show that in the case (r, s) = (10, 2) we also have h = 0.
Following the proof of Proposition 8.2.1 we may suppose that h F
3
. Without loss
of generality we can take h = 1 and look for a contradiction. Again following the
proof of Proposition 8.2.1 and using
M
1
= u
2
e
1
, u
10
e

+e
1
)
we nd that e

e
1
mod uM. Also

1
3
e
1
= (1 u
18
)e
1
,

1
3
e

= (1 u
18
)e

+h
1
(u)e
1
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
922 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
for some h
1
(u) F
9
[u]/(u
36
), which must actually lie in F
3
[u]/(u
36
) (using, as
usual, the fact that
2
and
3
commute). Thus
e
1
e

(
3

1
3

2
)(e

)
(
3

1
3
)(e

)
(h
1
(0) h
1
(0))e
1
mod uM

.
The inverse linear maps

1
3
on M

/uM

have matrices
_
1 h
1
(0)
0 1
_
with respect to the basis e
1
, e

, so that h
1
(0) = h
1
(0). Thus h
1
(0) = 1. On
the other hand evaluating
3

1

3
mod uM

on u
10
e

+ e
1
and comparing
coecients of e
1
gives h
1
(0) = 0, a contradiction.
Thus if any case (r, 2) arises, the underlying Breuil module must be a split
extension
M= (F
3
[u]/(u
36
))e
1
(F
3
[u]/(u
36
))e

, M
1
= u
2
e
1
, u
r
e

),

1
(u
2
e
1
) = e
1
,
1
(u
r
e

) = e

(so N
1
= 0), with

2
e
1
= e
1
,
2
e

= e

,

4
e
1
=

1e
1
,
4
e

= (

1)
r/2
e

.
We also have

1
3
e
1
= H

1
3
(u)
3
e
1

1
3
e

= H

1
3
(u)
3r/2
e

+h
1
(u)e
1
for some h
1
F
3
[u]/(u
36
). Since N
1
= 0, we have

1
3

1
=

1
3
on
M

1
. Evaluating this on u
r
e

and comparing coecients of e

1
gives h
1
(u) =
u
3(r2)
h
1
(u)
3
H

1
3
(u)
3r
. This forces h
1
(u) = 0 if r ,= 2. If r = 2 it forces
h
1
(u) = c
1
(1 u
18
) for some c
1
F
3
. We will show c
1
= c
1
= 0. Indeed,
evaluating the congruence
(
3

2

1
3

2
) mod uM

on e

gives
0 = (e

) (c
1
c
1
)e
1
mod uM

so that c
1
= c
1
. On the other hand the congruence

3

4

4

1
3
mod uM

gives
_
1 c
1
0 1
__
1 0
0

1
_
=
_
1 0
0

1
__
1 c
1
0 1
_
in M
2
(F
3
), so c
1
= c
1
. Thus c
1
= c
1
= 0 and h
1
= 0 for r = 2 as well. Thus
for r = 2, 6 and 10 the Breuil module with descent data M

is split, so is split, a
contradiction.
This rules out the possibilities (2, 2), (6, 2) and (10, 2). Using Cartier duality we
can also rule out (10, 10) and (10, 6). We are left with the four possible pairs (r, s)
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 923
as asserted in the proposition and must determine which possibilities arise in each
case.
Next consider the case (r, s) = (2, 6). Using the same analysis as in the (r, s) =
(2, 6) case in Proposition 8.2.1, we nd that the possibilities for the Breuil module
M are the ones in the statement of the proposition (and N
1
= 0 is easy to
check), though we only know that

4
(e
1
) =

1e
1
,
4
(e

) =

1e

+h
4
(u)e
1
for some h
4
(u) F
9
[u]/u
36
. The conditions

4
(M

1
) M

1
,
4

1
=

1

4
on M

1
are equivalent to
h
4
0 mod u
4
, h
4
=
_
h
4
u
4
_
3
.
The solutions to this are h
4
= a

1u
6
for a F
3
. Replacing e

by e

+ au
6
e
1
preserves our standardized form but makes h
4
= 0:

4
(e
1
) =

1e
1
,
4
(e

) =

1e

.
The wild descent data must have the form

1
3
(e
1
) = e
1
,

1
3
(e

) = (1 u
18
)e

+h
1
(u)e
1
for some h
1
F
9
[u]/u
36
. The conditions

1
3
(M

1
) M

1
,

1
3

1
=

1
3
on M

1
(recall N
1
= 0) are equivalent to
h
1
0 mod u
4
, h
1
= (1 u
18
)
_
h
1
u
4
_
3
,
whose solutions are
h
1
= c
1
u
6
(1 u
18
)
for some c
1
F
3
. Since N
1
= 0, we have

1
3

1
3

1
=

1
,
so
c
1
= c
1
.
Using Lemma 5.2.2 and Corollary 5.6.2, we see that all of these possibilities are
well dened. We also see that I annihilates M/uM F
9
. It is straightforward to
check that generic splitting over F

(which is equivalent to generic splitting over

F) is equivalent to c = 0, and that such splitting is compatible with descent data
(i.e. descends to Q
3
) if and only if c = c
1
= 0. For dimension reasons, the map on
Ext
1
s is therefore an isomorphism.
Now consider the case (r, s) = (2, 10). Here we have
M
1
= u
10
e
1
, u
2
e

+he
1
)
for some h F
3
[u]/u
36
, with

1
(u
10
e
1
) = e
1
,
1
(u
2
e

+he
1
) = e

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

924 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
and

4
(e
1
) =

1e
1
,
4
(e

) =

1e

.
In order that

4
(M

1
) M

1
,
4

1
=

1

4
on M

1
,
it is necessary and sucient that
h h(

1u) mod u
22
.
But u
10
tu
2
t
3
[t F
3
[u]/u
36
is spanned by u
13
u
11
, u
12
u
8
, u
11
u
5
, u
10
u
2
,
and all multiples of u
15
, so we may suppose
h = c

+c

u
4
+cu
8
, (9.2.1)
for some c

, c

, c F
3
, at the expense of possibly losing the diagonal form of
4
.
The monodromy operator satises
N(e
1
) = 0, N(e

) = (c

u
6
c

u
18
+c

u
30
)e
1
.
Since the wild descent data must take the form

1
3
(e
1
) = (1 u
18
)e
1
,

1
3
(e

) = (1 u
18
)e

+h
1
e
1
for some h
1
F
9
[u]/u
36
, we compute

1
3
(u
2
e

+he
1
) = H
2

1
3
(u)(1 u
18
)(u
2
e

+he
1
) +f
1
(u)e
1
,
where
f
1
(u) = hH
2

1
3
(u)(1 u
18
) +u
2
H
2

1
3
h
1
+ (1 u
18
)h(uH

1
3
). (9.2.2)
Thus, in order that

1
3
(M

1
) M

1
, it is necessary and sucient that f
1
satises
f
1
0 mod u
10
.
Using (9.2.1) and H

1
3
1 u
6
mod 3, this amounts to
h
1
c

u
4
mod u
8
. (9.2.3)
However, N
1
(M
1
) u
6
M, so

1
3

1
3
mod u
6
M

when evaluated on M

1
. This gives
h
1

_
f
1
u
10
_
3
mod u
6
.
Since h
1
is a cube modulo u
6
, by (9.2.3) we must have c

= 0, and so N
1

0 mod u
18
M. Thus,

1
3
and

1
commute modulo u
18
M

when evaluated on M

1
,
so we get
h
1

_
f
1
u
10
_
3
mod u
18
, (9.2.4)
and h
1
is a cube modulo u
18
.
On the other hand, with c

= 0, we see from (9.2.3) that

h
1
0 mod u
8
.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 925
Because h
1
is a cube modulo u
18
, we get the slight improvement
h
1
0 mod u
9
.
Combining this with the vanishing of c

, we deduce from (9.2.2) that f

1

c

u
10
mod u
11
, so by (9.2.4)
h
1
c

mod u.
This forces c

= 0, so N
1
= 0. Thus,

1
3
and

1
commute on M

1
, so
h
1
=
_
f
1
u
10
_
3
in F
9
[u]/u
36
. Using h = cu
8
this becomes (via (9.2.2))
h
1
= (1 u
18
)
_
h
1
u
8
_
3
,
so
h
1
= c
1
u
12
(1 u
18
)
for some c
1
F
3
. As before, we get c
1
= c
1
.
Now we diagonalise
4
. Since we have

4
(e
1
) =

1e
1
,
4
(e

) =

1e

+h
4
(u)e
1
for some h
4
F
9
[u]/u
36
, the conditions

4
(M

1
) M

1
,
4

1
=

1

4
on M

1
are equivalent to
h
4
0 mod u
8
, h
4
=
_
h
4
u
8
_
3
,
which is to say
h
4
= a

1u
12
for some a F
3
. Replacing e

by e

+au
12
e
1
then puts us in a setting with a = 0.
Thus all extensions have the form asserted in the proposition. It is easy to check
that in each case I annihilates (M/uM) F
9
.
Pushout by the non-zero map M

6,1
M

10,1
in
1
DD
F

/Q3
induced by e u
6
e
takes our (2,6) examples to our (2,10) examples (compatibly with the labelling of
parameters c, c
1
as in the statement of the proposition). Thus all 9 possibilities for
(c, c
1
) do occur and we get an isomorphism of Ext
1
s as asserted. Moreover, generic
splitting over F

(which is equivalent to generic splitting over F) is equivalent to

c = 0, and such splitting is compatible with descent data (i.e. descends to Q
3
) if
and only if c = c
1
= 0.
Using Cartier duality and the case (r, s) = (2, 6), we see that in the case (r, s) =
(6, 10) the map of Ext
1
s is an isomorphism. It is easy to check that the objects
in our asserted list of 9 possibilities for (r, s) = (6, 10) are well dened and that
pullback by the non-zero map M

2,
M

6,
induced by e u
6
e takes these to our
(2, 10) examples (compatibly with the labelling of parameters c, c
1
).
Finally, we turn to the case (r, s) = (6, 6). Choosing a basis with respect to
which
4
has a diagonal action, the conditions

4
(M

1
) M

1
,
4

1
=

1

4
on M

1
(9.2.5)
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
926 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
are equivalent to
h(u) h(

1u) mod u
18
.
Since u
6
t u
6
t
3
[t F
3
[u]/u
36
consists of multiples of u
7
, we may change e

to
get
h = c +c

u
4
for some c, c

F
3
, with

4
(e
1
) =

1e
1
,
4
(e

) =

1e

+h
4
(u)e
1
for some h
4
F
9
[u]/u
36
. Feeding this into (9.2.5) we get h
4
= h
3
4
, so h
4
=

1a for some a F
3
. Replacing e

by e

ae
1
returns us to the setting with
diagonal
4
-action and preserves the standardizations we have made so far.
It is easy to compute N(e

) = c

u
30
e
1
(and we know N(e
1
) = 0). The wild
descent data is

1
3
(e
1
) = e
1
,

1
3
(e

) = e

+h
1
e
1
for some h
1
F
9
[u]/u
36
. Using the congruence for t

1
3
in 6.5, the identity

1
3

1
= (1 +t

1
3
N)

1
3
on M

1
amounts to the condition
h
1
= h
3
1
c

u
12
c

u
30
,
whose solutions are
h
1
= c
1
c

u
12
c

u
30
for some c
1
F
3
. The identity

1
3

1
3
1 mod u M

implies c
1
= c
1
. Thus
V
M
(e

) (e

) ce
1
mod u M

, (
3

2

1
3

2
)(e

) c
1
e
1
mod u M

.
Thus I annihilates (M/uM)
F3
F
9
if and only if c
1
= c.
By Lemma 5.2.2 and Corollary 5.6.2, it is easy to see that all of these objects
are well dened. The kernel of
Ext
1
1DD
F

/Q
3
,I
(M

6,
, M

6,1
) Ext
1
F3[GF ]
(1, ) (9.2.6)
consists of pairs (c, c), where generic splittings are induced by any of the (non-
zero) Breuil module maps
M(0, 1) M
dened by
e u
9
e

+ ( cu
9
+cu
3
)e
1
= u(c +u
2
c)u
6
e
1
+u
3
(u
6
e

+ (c cu
4
)e
1
)
with c F
3
. Thus, the pairs (c, c

) corresponding to the which are split over F

(or equivalently, split over F

) are exactly those for which c +c

= 0. The map
Ext
1
1DD
F

/Q
3
,I
(M

6,
, M

6,1
) Ext
1
F3[G3]
(1, )
is therefore injective, because the splitting given above respects descent data if and
only if c = c
1
= 0.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 927
It remains to establish which of the given extensions of Breuil modules corre-
spond to peu ramie extensions of Z/3Z by
3
over Q
3
. We noted above that
the maps among the Ext
1
1DD
F

/Q
3
s in the (2,6), (6,10), (2,10) cases induced by
pushout/pullback along e u
6
e are compatible with the parametrisation by pairs
(c, c
1
). With a little more care, one checks that the maps
Ext
1
1DD
F

/Q
3
,I
(M

2,
, M

6,1
) Ext
1
1DD
F

/Q
3
,I
(M

6,
, M

6,1
)
Ext
1
1DD
F

/Q
3
,I
(M

6,
, M

10,1
)
induced by e u
6
e send the pair (c, c

) in the middle to the pair (c+c

, c) on either
end (to construct the necessary commutative diagrams of short exact sequences in
the two cases, use the maps
(e

, e
1
) (u
6
e

e
1
, e
1
), (e

, e
1
) (e

+c

e
1
, u
6
e
1
)
respectively). This reduces us to checking the (6, 6) case.
By Corollary 2.3.2, the two tr`es ramie extensions,
1
and
2
, of 1 by which
are non-split over F arise from elliptic curves, E
1
and E
2
, over Q
3
for which
Ej,3
is potentially Barsotti-Tate with extended type

i
(see 6.5). Let G
j
denote the
3-torsion in the Neron model of E
j
over O
F
. From the universal property of Neron
models we see that G

j
= G
j

OF
O
F
inherits descent data [g] over Q
3
. By
the same argument used at the end of 4.6 we see that (G

j
, [g]) is an object of

1
DD
F

/Q3,I
. Moreover we see that F ,= 0 on D(G
j
). Since all non-(6, 6) cases
above have 0 mod uM, by the parts of Proposition 9.2.1 which we have already
proved we see that M

(G

j
, [g]) is an extension of M

6,
by M

6,1
and correspond to
a pair (c, c

) with c ,= 0 (since F ,= 0) and c+c

,= 0 (by our analysis of (9.2.6), since

i
is non-split over F). Hence M

(G

1
, [g]) and M

(G

2
, [g]) must correspond in
some order to the lines c

= 0 and c = c

in F
2
3
.
As a non-split peu ramie extension of 1 by remains non-split over F

, we see
that the peu ramie line in
Ext
1
1DD
F

/Q
3
,I
(M

6,
, M

6,1
)

= Ext
1
F3[G3]
(1, )
cannot correspond to c + c

= 0. By the above analysis it cannot correspond to

c

= 0 or c c

= 0. Thus it must correspond to the remaining line c = 0.

The properties of in the cases listed in Proposition 9.2.1 make it clear that the
(6, 6) case there is dierent. We will see further manifestations of this dierence
later.
9.3. Further rank two calculations.
Lemma 9.3.1. For (r, s) = (2, 6), (6, 10) and (2, 10) we have
Ext
1
1DD
F

/Q
3
,I
(M

s,1
, M

r,
) = (0).
Proof. The (6,10) case follows from the (2,6) case by Cartier duality. Thus, we
assume r = 2, s 6, 10. Let (M

, g) be such an extension. By Lemma 8.1.2,

(M

, g) arises from a Breuil module over O

F
of the form
M= (F
3
[u]/u
36
)e

(F
3
[u]/u
36
)e

1
, M
1
= u
2
e

, u
s
e

1
+he

)
with

1
(u
2
e

) = e

,
1
(u
s
e

1
+he

) = e

1
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
928 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
and

4
(e

) =

1e

,
4
(e

1
) = (

1)
s/2
e

1
,
where h F
3
[u]/u
36
.
The combined conditions

4
(M

1
) M

1
,
4

1
=

1

4
on M

1
are equivalent to
(

1)
s/2
h(u) (

1)h(

1u) mod u
14
.
Treating the cases s = 6 and s = 10 separately, we conclude from Lemma 5.2.2
that we may change e

1
so that h = 0 when s = 6 and h F
3
when s = 10. As a
result of this change, we only have

4
(e

) =

1e

,
4
(e

1
) = (

1)
s/2
e

1
+h
4
(u)e

.
However, with h F
3
when s = 10 and h = 0 when s = 6, the condition

1
=

1

4
on M

1
forces h
4
= u
3(s2)
h
3
4
, so that in fact h
4
= 0 after all.
When h = 0, so M is split in
1
mod
F
(compatibly with
4
on M

), and it
is easy to check (using N = 0) that the wild descent data

1
3
must also be
diagonal, so we have the desired splitting in
1
DD
F

/Q3
.
It remains to consider the case (r, s) = (2, 10) with h = c F
3
. It is easy to
compute
N(e

) = 0, N(e

1
) = cu
30
e

.
The wild descent data must have the form

1
3
(e

) = (1 u
18
)e

1
3
(e

1
) = (1 u
18
)e

1
+h
1
e

with h
1
F
9
[u]/u
36
.
It is straightforward to check that

1
3
(M

1
) M

1
, and then the condition

1
3

1
= (1 +t

1
3
N)

1
3
on M

1
gives
h
1
= cu
12
+h
3
1
u
24
(1 u
18
) +cu
30
.
The unique solution to this is
h
1
= c(u
12
+u
30
).
Thus
3

1
3

2
0 mod u, while (e

1
) ce

mod u. This forces c = 0. With

c = 0 we obviously have only the split extension class.
Lemma 9.3.2. The natural map
Ext
1
1DD
F

/Q
3
,I
(M

6,1
, M

6,
) Ext
1
F3[G3]
(, 1)
is an isomorphism, with elements parametrised by pairs (c, c

) F
2
3
corresponding
to
M= (F
3
[u]/u
36
)e

(F
3
[u]/u
36
)e

1
, M
1
= u
6
e

, u
6
e

1
+ (c +c

u
4
)e

),
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 929
where

1
(u
6
e

) = e

,
1
(u
6
e

1
+ (c +c

u
4
)e

) = e

1
, N(e

) = 0, N(e

1
) = c

u
30
e

and the descent data is

4
(e

) =

1e

,
4
(e

1
) =

1e

1
,

1
3
(e

) = e

1
3
(e

1
) = e

1
+ (c c

u
12
c

u
30
)e

.
Proof. The proof is identical to the proof of the case (r, s) = (6, 6) in Proposition
9.2.1, except

1 is everywhere replaced by

1 and when we study splitting

we give M(0, 1) the descent data for the trivial mod 3 character (which amounts to
using
4
(e) = e rather than
4
(e) = e).
Lemma 9.3.3. For r 2, 10, the maps
Ext
1
1DD
F

/Q
3
(M

r,
, M

r,
) Ext
1
F3[G3]
(, )
and
Ext
1
1DD
F

/Q
3
(M

r,1
, M

r,1
) Ext
1
F3[G3]
(1, 1)
are injective and have image consisting of the 1-dimensional space of classes which
split over an unramied extension of Q
3
.
Proof. The cases r = 10 follow from the cases r = 2 using Cartier duality. Thus
we suppose r = 2. We treat only the case of M

2,
, the case M

2,1
being exactly the
same except that

1 replaces

1 everywhere.
Let (M

, g) represent an element in Ext

1
1DD
F

/Q
3
(M

2,
, M

2,
). Lemma 8.1.2
ensures the existence of an ordered F
3
[u]/u
36
-basis e

, e

of M such that
M
1
= u
2
e

, u
2
e

+he

),
1
(u
2
e

) = e

,
1
(u
2
e

+he

) = e

with

4
(e

) =

1e

,
4
(e

) =

1e

.
Carrying out the usual calculation,

4
(M

1
) M

1
,

1

4
=
4

1
on M

1
(9.3.1)
if and only if
h h(

1u) mod u
14
.
Combining this with Lemma 5.2.2, we may change e

so that h = cu
2
, with c F
3
,
at the expense of possibly losing the diagonal form of
4
. But with h = cu
2
and

4
(e

) =

1e

+ h
4
(u)e

, the conditions (9.3.1) imply h

4
= h
3
4
, and so
h
4
(u) = (

1)a for some a F

3
. Then
4
4
= 1 forces a = 0, so
4
still has
diagonal form.
It is easy to check that N(e

) = 0, so N
1
= 0. Thus, we must have

1
3

1
=

1
3
(9.3.2)
on M

1
. Since the wild descent data has to be of the form

1
3
(e

) = (1 u
18
)e

1
3
(e

) = (1 u
18
)e

+h
1
e

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

930 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
for some h
1
F
3
[u]/u
36
, evaluation of (9.3.2) on u
2
e

+cu
2
e

M
1
gives h
1
=
(1 u
18
)h
3
1
, so h
1
= c
1
(1 u
18
) for some c
1
F
3
. The relation

1
3

4

1
3

3
4
(e

) = e

forces c
1
= 0.
We now have described all possibilities in terms of the single parameter c F
3
,
and it is straightforward to use Corollary 5.6.2 to check that all of these examples are
in fact well dened. Generic splittings over unramied extensions of Q
3
correspond
to the maps
F
3

F3
M(0, 1) F
3

F3
M
given by
e au
3
e

+u(u
2
e

+cu
2
e

),
where a F
3
satises a
3
= a + c. Such generic splittings can be dened over Q
3
(i.e. without extending the residue eld) if and only if c = 0.
Lemma 9.3.4. (1) The map of groups
Ext
1
1DD
F

/Q
3
(M

6,1
, M

6,1
) Ext
1
F3[G3]
(1, 1)
is an isomorphism.
Explicitly, the group Ext
1
1DD
F

/Q
3
(M

6,1
, M

6,1
) is parametrised by pairs
(c, c

) F
2
3
corresponding to
M= (F
3
[u]/u
36
)e
1
(F
3
[u]/u
36
)e

1
, M
1
= u
6
e
1
, u
6
e

1
+ (cu
2
+c

u
6
)e
1
),
with

1
(u
6
e
1
) = e
1
,
1
(u
6
e

1
+ (cu
2
+c

u
6
)e
1
) = e

1
,
N(e
1
) = 0, N(e

1
) = cu
24
e
1
and descent data

4
(e
1
) =

1e
1
,
4
(e

1
) =

1e

1
,

1
3
(e
1
) = e
1
,

1
3
(e

1
) = e

1
+c(u
6
u
18
u
24
u
30
)e
1
.
The classes in Ext
1
F3[G3]
(1, 1) which split over an unramied extension of Q
3
correspond to the pairs with c = 0.
(2) The map of groups
Ext
1
1DD
F

/Q
3
(M

6,
, M

6,
) Ext
1
F3[G3]
(, )
is an isomorphism.
Explicitly, the group Ext
1
1DD
F

/Q
3
(M

6,
, M

6,
) is parametrised by pairs
(c, c

) F
2
3
corresponding to
M= (F
3
[u]/u
36
)e

(F
3
[u]/u
36
)e

, M
1
= u
6
e

, u
6
e

+ (cu
2
+c

u
6
)e

),
with

1
(u
6
e

) = e

,
1
(u
6
e

+ (cu
2
+c

u
6
)e

) = e

,
N(e

) = 0, N(e

) = cu
24
e

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use

THE MODULARITY OF ELLIPTIC CURVES OVER Q 931
and descent data

4
(e

) =

1e

,
4
(e

) =

1e

1
3
(e

) = e

1
3
(e

) = e

+c(u
6
u
18
u
24
u
30
)e

.
The classes in Ext
1
F3[G3]
(, ) which split over an unramied extension of
Q
3
correspond to the pairs with c = 0.
Proof. We treat the rst part of the lemma; replacing

1 with

1 throughout
gives the proof of the second part.
As usual, we can nd an ordered F
3
[u]/u
36
-basis e
1
, e

1
of M so that
M
1
= u
6
e
1
, u
6
e

1
+he
1
),
1
(u
6
e
1
) = e
1
,
1
(u
6
e

1
+he
1
) = e

1
,
and
4
(e
1
) =

1e
1
,
4
(e

1
) =

1e

1
. The conditions
4
(M

1
) M

1
and

1
=

1

4
on M

1
amount to
h(u) h(

1u) mod u
18
.
Since u
6
t u
6
t
3
[t F
3
[u]/u
36
consists of multiples of u
7
, we can change the
choice of e

1
so that
h = cu
2
+c

u
6
for some c, c

F
3
, where we may a priori lose the diagonal form of
4
. But the
same kind of calculation as in Lemma 9.3.3 shows
4
(e

1
) =

1e

1
+a

1e
1
for
some a F
3
, so the condition
4
4
= 1 forces a = 0 (i.e.
4
still has diagonal action).
It is straightforward to compute the asserted formula for N, and then the wild
descent data can be computed exactly as in our previous computations of wild
descent data; this yields the formulas

1
3
(e
1
) = e
1
,

1
3
(e

1
) = e

1
+ (c
1
+c(u
6
u
18
u
24
u
30
))e
1
,
where c
1
F
3
. Modulo u, the linear action of
1
3

4

1
3

3
4
sends e

1
to e

1
c
1
e
1
,
but
1
3

4

1
3

3
4
= 1, so c

= 0 for = 1. Thus, we obtain the asserted list of

possibilities. The well-denedness of these examples follows from Lemma 5.2.2 and
Corollary 5.6.2.
It is easy to see that there is a non-zero map F
3

F3
M(0, 1) F
3

F3
M if and
only if c = 0, in which case such non-zero maps are precisely those induced by
e au
9
e
1
+u
3
(u
6
e

1
+c

u
6
e
1
),
where a F
3
satises a
3
= a + c

. The verication that c = c

= 0 corresponds
to being in the kernel of our map of Ext
1
s is now clear, since X
3
= X + c

has a
solution in F
3
if and only if c

= 0.
9.4. Completion of the proof of Theorem 4.6.1. Everything in Theorem 4.6.1
is now clear except for the third assertion, which we now prove. Let (G

, [g]) be
as in the third part of that theorem. We may suppose that G

= G
OF
O
F
for
some G
/OF
. The ltration on
F3
k gives a ltration
(0) G

G G
1
(0),
which is compatible with the descent data over Q
3
. According to Lemma 5.2.3 we
have M

(G

)

= M(k; r

, f

) and M

(G
1
)

= M(k; r
1
, f
1
) for some 0 r
1
, r

12
and some f
1
, f

k[u]/u
36
. We will let denote either 1 or . In particular
M

(G

)
1
= u
r
M

(G

) for = 1, . From this one can conclude that if H is

License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
932 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
a subquotient of G

, then M

(H)
1
= u
r
M

(H). Quite generally, for any Breuil

module Mover O
F
with M
1
= u
r
Mand any short exact sequence of Breuil modules
0 M

MM

0,
we must also have
M

1
= u
r
M

, M

1
= u
r
M

.
Indeed, M M

is a surjection taking M
1
onto M

1
, so the assertion for M

is
clear. Since M

is an F
3
[u]/u
36
-module direct summand of M and
M

1
= M

M
1
= M

u
r
M,
the assertion for M

is likewise clear. We conclude that (M

(G

, g) admits a
ltration with successive quotients M

r,
. Thus r

2, 6, 10.
Consider a xed surjection of F
3
[G
3
]-modules
k .
This gives rise to a nite at O
F
-group scheme Hwith descent data on H

= HO
F

over Q
3
corresponding to and an epimorphism
G H
compatible with descent data. Consider the commutative diagram
0 M

(H
1
) M

(H) M

(H

) 0

0 M

(G
1
) M

(G) M

(G

) 0
where the top row corresponds to the non-split ltration of . The middle vertical
map is an isomorphism of the source onto an F
3
[u]/u
36
-module direct summand
of the target, so the left vertical map is as well, because an injection of F
3
[u]/u
36
into a free F
3
[u]/u
36
-module must be an identication with such a direct summand
(consider torsion). This forces M

(H
1
)

= M

r1,1
and so, by Proposition 9.2.1,
we see that r
1
,= 2. Repeating the analogous argument applied to a submodule
k one sees that r

,= 10.
Thus (G

, [g]) is weakly ltered by G

s,1
, G
r,
for (r, s) = (2, 6), (6, 10), (2, 10)
or (6, 6), as desired.
9.5. Completion of the proof of Theorem 4.6.3. Write A
N
for F
3
[[T]]/(T
N
).
For (r, s) = (2, 6), (6, 10) and (2, 10), we will dene a Breuil module M
N,(r,s)
over
O
F
and descent data g for Gal(F

/Q
3
) on M

N,(r,s)
= M
N,(r,s)

F3
F
9
such that
M
N,(r,s)
and (M

N,(r,s)
, g) have compatible actions of A
N
(and
2
= 1 Frob
3
).
More specically set t = 2, 6 or 8 according as (r, s) = (2, 6), (6, 10) or (2, 10).
Viewing as an extension class, it corresponds to a particular pair (c, c
1
) F
2
3
in Proposition 9.2.1. Fix these values. Motivated by the idea of deforming the
formulae in Proposition 9.2.1, we are led to dene
M
N,(r,s)
= (A
N
[u]/u
36
)e
1
(A
N
[u]/u
36
)e

,
(M
N,(r,s)
)
1
= u
s
e
1
, u
r
e

+ (c +T)u
t
e
1
)
with

1
(u
s
e
1
) = e
1
,
1
(u
r
e

+ (c +T)u
t
e
1
) = e

.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 933
It is straightforward to check that N
1
= 0 on M
N,(r,s)
. We may dene A
N
-linear
descent data on M

N,(r,s)
by setting
2
= 1Frob
3
and using the following formulae.
(1) When (r, s) = (2, 6), set

4
(e
1
) =

1e
1
,
4
(e

) =

1e

1
3
(e
1
) = e
1
,

1
3
(e

) = (1 u
18
)(e

c
1
u
6
e
1
).
(2) When (r, s) = (6, 10), set

4
(e
1
) =

1e
1
,
4
(e

) =

1e

1
3
(e
1
) = (1 u
18
)e
1
,

1
3
(e

) = (e

c
1
u
6
e
1
).
(3) When (r, s) = (2, 10), set

4
(e
1
) =

1e
1
,
4
(e

) =

1e

1
3
(e
1
) = (1 u
18
)e
1
,

1
3
(e

) = (1 u
18
)(e

c
1
u
12
e
1
).
It is readily checked that this denes an object of
1
DD
F

/Q3,I
with an action
of A
N
. Let G
N,(r,s)
and (G

N,(r,s)
, [g]) be the corresponding nite at O
F
-group
scheme and nite at O
F
-group scheme with descent data.
If 1 M < N, then we have a short exact sequence in
1
DD
F

/Q3,I
(0) M

M,(r,s)
M

N,(r,s)
M

NM,(r,s)
(0),
where the rst map is induced by multiplication by T
NM
. The case M = 1 shows
that
(G
N,(r,s)
, [g])
Q3
/T(G
N,(r,s)
, g)
Q3
corresponds to . Thus we get a surjection of A
N
[G
3
]-modules A
2
N
G
N,(r,s)
(Q
3
),
which must in fact be an isomorphism (count orders). Thus (G
N,(r,s)
, [g])
Q3
denes a deformation
N,(r,s)
of to A
2
N
. For N 2 we have
N,(r,s)
mod T
2
=

2,(r,s)
.
We also have an exact sequence
(0) M

s,1

F3
A
N
(M

N,(r,s)
, g) M

r,

F3
A
N
(0)
in
1
DD
F

/Q3
, from which we obtain an exact sequence of A
N
[G
3
]-modules
(0) X

N
X
1
(0).
Note that X
1

= F
N
3
and X

= F
3
()
N
as F
3
[G
3
]-modules. Moreover, this sequence
must split as a sequence of A
N
-modules. (Use, for instance, the kernel of
N
() 1
for any G
3
G
Q3(

3)
.) Thus X
1

= A
N
and X

= A
N
() as A
N
[G
3
]-modules,
so det
N
= .
Finally, we must check that the exact sequence
(0)
2
(0)
is not split. We have maps of Breuil modules
f
1
: M
s,1
M
2,(r,s)
e e
1
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
934 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
and
f
2
: M
2,(r,s)
M
r,
e
1
0
Te
1
0
e

0
Te

e
compatible with descent data. These give rise to maps
f

1
:
2
1
and
f

2
:
2
,
such that the composites

2
f

1
1
and

2

are non-zero.
To check that
(0)
2
(0)
is non-split, it suces to check that
(0) ker f

1
/ Imf

2
1 (0) (9.5.1)
is non-split. However, ker f

1
/ Imf

2
corresponds to an object (N

, g) of

1
DD
F

/Q3,I
satisfying
N

= (F
9
[u]/u
36
)(Te
1
) (F
9
[u]/u
36
)e

, N

1
= u
s
(Te
1
), u
r
e

+u
t
(Te
1
))
with

1
(u
s
(Te
1
)) = (Te
1
),
1
(u
r
e

+u
t
(Te
1
)) = e

.
By Lemma 5.2.2, the sequence of Breuil modules with descent data
(0) M

s,1
N

r,
(0)
is not split. This sequence recovers (9.5.1) under generic bre descent, so by Propo-
sition 9.2.1
(0) ker f

1
/ Imf

2
1 (0)
is not split.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 935
9.6. Completion of the proof of Theorem 4.6.2. Suppose rst that (r, s) =
(2, 6), (6, 10) or (2, 10). By Lemma 9.3.1

0
: Ext
1
S
i,(r,s)
(, ) H
1
(G
3
, )
is the zero map. Lemma 9.3.3 then tells us that if r ,= 6, then

1
: Ext
1
S
i,(r,s)
(, ) H
1
(I
3
, F
3
)
is the zero map; while if s ,= 6, then

: Ext
1
S
i,(r,s)
(, ) H
1
(I
3
, F
3
)
is the zero map. Thus Theorem 4.7.5, and hence Theorem 4.6.2, follows in these
cases.
Now consider the case (r, s) = (6, 6). Choose x H
1
S
i,(6,6)
(G
3
, ad
0
). Let G
denote the corresponding rank 81 nite at O
F
-group scheme with descent data
[g] on G

= G
OF
O
F
. Set M= M

(G). Let H G denote the closed subgroup

scheme (with descent data) corresponding to the kernel of the map (G

, [g])
Q3

F
3
and let N = M

(H). Then N has F

3
[u]/u
36
-basis e

, e

1
, e

with respect
to which
N
1
= u
6
e

, u
6
e

1
+ (b +b

u
4
)e

, u
6
e

+ (c +c

u
4
)e

1
+fe

),
where b, b

, c, c

F
3
, f F
3
[u]/u
36
and
1
sends the indicated generators of N
1
to
e

, e

1
, e

respectively. Also, the descent data has the form

4
(e

) =

1e

,
4
(e

1
) =

1e

1
,
4
(e

) =

1e

+h
4
(u)e

for some h
4
F
9
[u]/u
36
, and

1
3
(e

) = e

1
3
(e

1
) = e

1
+ (b b

(u
12
+u
30
))e

1
3
(e

) = e

+ (c c

(u
12
+u
30
))e

1
+h
1
e

,
where h
1
F
9
[u]/u
36
. Also, as is tr`es ramie, we see that c ,= 0 by Proposition
9.2.1 and Lemma 9.3.4.
The requirement that u
12
N N
1
forces N
1
to contain
u
12
e

= u
6
(u
6
e

+ (c +c

u
4
)e

1
+fe

) (c +c

u
4
)(u
6
e

1
+ (b +b

u
4
)e

)
+ ((b +b

u
4
)(c +c

u
4
) fu
6
)e

,
so N
1
must contain (b +b

u
4
)(c +c

u
4
)e

. As c ,= 0 we get (b +b

u
4
)e

N
1
, and
since e

, u
4
e

, N
1
, we must have b = b

= 0. We conclude that the natural map

0
: Ext
1
S
i,(6,6)
(, ) H
1
(G
3
, )
is the zero map.
Let us further analyse N. Replacing e

by e

+ t
3
e

for t F
3
[u]/u
36
causes
f to be replaced by f u
6
t
3
+ u
6
t and otherwise leaves our standardized form
unchanged (except that h
4
and h
1
may change). Using a suitable choice of such
t, we may assume f has degree at most 6. On the other hand,

4
(u
6
e

+ (c +c

u
4
)e

1
+fe

) =

1(u
6
e

+ (c +c

u
4
)e

1
+fe

)
+ (

1(f(u) +f(

1u)) u
6
h
4
(u))e

,
so
4
(N

1
) N

1
if and only if
f(u) +f(

1u) 0 mod u
6
,
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
936 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
which forces
f = a
2
u
2
+a
6
u
6
for some a
2
, a
6
F
3
. From the wild descent data formulae derived in the proof of
Lemma 9.3.4 we also see that h
1
0 mod u
6
.
Now M has an ordered basis e
1
, e

, e

1
, e

with respect to which

(9.6.1) M
1
= u
6
e
1
, u
6
e

+ (c +c

u
4
)e
1
, u
6
e

1
+he
1
,
u
6
e

+ (c +c

u
4
)e

1
+ (a
2
u
2
+a
6
u
6
)e

+ge
1
),
where g, h F
3
[u]/u
36
and
1
sends the indicated generators of M
1
to e
1
, e

, e

1
, e

.
If we try to expand out u
12
e

as a linear combination of the indicated generators

of M
1
, we nd that
u
12
e

((c +c

u
4
)h +ca
2
u
2
)e
1
mod M
1
.
It follows that u
12
e

M
1
if and only if
(c +c

u
4
)h +ca
2
u
2
0 mod u
6
.
Since is tr`es ramie, the last part of Proposition 9.2.1 tells us that c ,= 0. Thus,
u
12
e

M
1
if and only if h a
2
u
2
mod u
6
. We can now use Lemma 9.3.4 to see
that the wild descent data action is determined by

1
3
(e
1
) = e
1
,

1
3
(e

) = e

+ (c c

(u
12
+u
30
))e
1
,

1
3
(e

1
) = e

1
+f
1
e
1
(with f
1
0 mod u
6
), and

1
3
(e

) = e

+ (c c

(u
12
+u
30
))e

1
+h
1
e

+g
1
e
1
,
where g
1
F
9
[u]/u
36
and h
1
0 mod u
6
.
We must have

1
3
(u
6
e

+ (c +c

u
4
)e

1
+ (a
2
u
2
+a
6
u
6
)e

+g(u)e
1
) M

1
, (9.6.2)
and this expression is easily computed to equal
u
6
H
6

1
3
(e

+ (c c

(u
12
+u
30
))e

1
+h
1
e

+g
1
e
1
)
+ (c +c

u
4
H
4

1
3
)(e

1
+f
1
e
1
)
+ (a
2
u
2
H
2

1
3
+a
6
u
6
H
6

1
3
)(e

+ (c c

(u
12
+u
30
))e
1
) +g(uH

1
3
)e
1
.
Remembering that u
6
e
1
, u
12
M

) M

1
, (9.6.2) becomes
u
6
(e

ce

1
)+(c +c

u
4
H
4

1
3
)e

1
+(a
2
u
2
H
2

1
3
+a
6
u
6
)e

a
2
cu
2
e
1
+g(u)e
1
M

1
.
Using the explicit generators of M
1
given in (9.6.1) and recalling that ha
2
u
2
mod
u
6
, this simplies to
a
2
cu
2
e
1
M

1
.
Thus a
2
cu
2
is divisible by u
6
, so a
2
= 0.
The image of the class x in Ext
1
F3[G3]
(, ) under

corresponds to a nite at
O
F
-group scheme with Breuil module M
x
free of rank two over F
3
[u]/u
36
with basis
e

, e

, and with
(M
x
)
1
= u
6
e

, u
6
e

+a
6
u
6
e

),
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 937
where
1
sends the indicated generators of (M
x
)
1
to e

and e

respectively. Ac-
cording to the proof of Lemma 9.3.4 this implies that the image of the class x in
Ext
1
F3[G3]
(, ) is split over an unramied extension of Q
3
. Thus,

: Ext
1
S
i,(6,6)
(, ) H
1
(I
3
, F
3
)
is the zero map. This completes the proof of Theorem 4.7.5, and hence of Theorem
4.6.2.
10. Corrigenda for [CDT]
We would like to take this opportunity to record a few corrections to [CDT].
Page 523, line 10: Insert K-rational after For each.
Page 532, line 6: The semisimplicity of
n
follows from that of
1
is false
and should be deleted. This assertion was not used anywhere in the rest of
the paper.
Page 536, line 7: Replace GL
2
(C) by GL
2
(R).
Page 538, line 10: Replace of type (S, ) by such that [
G

is of type
and is of type (S, ).
Page 539, lines 1820: Replace each
1
by
1
and each
2
by
2
.
Page 541, line 14: Replace each of the three occurrences of A by A

.
Page 544, line 6: the discrete topology on V
p
should read the -adic
topology on M
p
.
Page 545, part 4 of Lemma 6.1.2: V

should be assumed to be a normal

subgroup of V .
Page 546, line 1: We should have noted that the key component of this
argument is very similar to the main idea of [Kh].
6.2: There are two signicant errors in this section. The assertion =
SL
2
(Z)U
S
satises the hypotheses of Theorem 6.1.1 is false and Hom(L
n
, k)
should be L
n
k. The argument of this section can be repaired by making
the following changes.
Page 546, lines 5 and 6: Replace Setting S = T() r, we nd
that the group = SL
2
(Z) U
S
satises the hypotheses of Theo-
rem 6.1.1. by Set S = T() r; U

S
=

p
U

S,p
where U

S,p
=
U
1
(p
cp
) if p T() and U

S,p
= U
S,p
otherwise; V

S
=

p
V

S,p
where
V

S,p
= U
1
(p
cp
) if p T() and V

S,p
= V
S,p
otherwise; and L

S
=
Hom
O[U

S
/V

S
]
(M

, H
1
(X
V

S
, O))[I

S
]. Then = SL
2
(Z) U

S
satises the
hypotheses of Theorem 6.1.1.
Page 546, lines 713: Replace Y
S
by Y
U

S
, Hom(L
n
, k) by L
n
k, M
S
by M

, F
S
by F
Hom
O
(M

,O)
and L
S
by L

S
.
Page 546, line 13: Replace and N
S
is non-empty. by . Using the fact
that Lemma 5.1.1 holds with U

S
replacing U
S
and

replacing
S
and
the discussion on page 541 we conclude that N
S
is non-empty.
Page 549, line 15: Replace U
{r,r

},p
by U
{r},p
.
Page 549, line 11: Replace U

S
/U

0,S
by V
0
/V
1
.
Page 552, line 4: The assertion is false in the case 5. It can be corrected
by adding and j(E) , 1728 mod (which is true if, for instance, E has
potentially supersingular reduction and 1 mod 4) after if 5.
Page 554, line 11: Replace j
E
by E is isogenous to an elliptic curve
with j-invariant in the set.
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
938 C. BREUIL, B. CONRAD, F. DIAMOND, AND R. TAYLOR
Page 554, line 11: Replace 5(29)
3
/2
5
by 5(29)
3
/2
5
.
Page 554, line 17: Replace the parenthetical comment (and j = 5(29)
3
/2
5
)
by (and isogenous to one with j-invariant 5(29)
3
/2
5
).
Page 554, line 5: Replace p by q and q by p.
Acknowledgements
We would like to thank Bas Edixhoven, Rene Schoof, Tom Weston and the referee
for corrections and improvements to preliminary versions of this paper; Barry Mazur
for helpful conversations; and David Pollack for help with computer calculations
which we made in an earlier attempt to compute some of these local deformation
rings. We are grateful to the Harvard University Clay fund for supporting the rst
author during a key visit paid to Harvard. The second author is grateful to the
Institute for Advanced Study for its stimulating environment and the University of
M unster for its hospitality. The third author is grateful to Harvard University and
the Universite de Paris Sud for their hospitality. The fourth author is grateful to
the University of California at Berkeley for its hospitality and to the Miller Institute
for Basic Science for its support.
References
[BBM] P. Berthelot, L. Breen, W. Messing, Theorie de Dieudonne cristalline II, Lecture Notes
in Mathematics 930, Springer-Verlag, Berlin, 1982. MR 85k:14023
[Br1] C. Breuil, Schemas en groupes et modules ltres, C. R. Acad. Sci. Paris 328 (1999), 9397.
MR 99m:14086
[Br2] C. Breuil, Groupes p-divisibles, groupes nis et modules ltres, Annals of Math. 152
(2000), 489549. CMP 2001:06
[Ca1] H. Carayol, Sur les representations -adiques associees aux formes modulaires de Hilbert,
Ann. Sci.

Ec. Norm. Sup. 19 (1986), 409468. MR 89c:11083
[Ca2] H. Carayol, Sur les representations galoisiennes modulo attachees aux formes modu-
laires, Duke Math. J. 59 (1989), 785801. MR 91b:11058
[Co] B. Conrad, Ramied deformation problems, Duke Math. J. 97 (1999), 439514. MR
2000h:11055
[CDT] B. Conrad, F. Diamond, R. Taylor, Modularity of certain potentially Barsotti-Tate Galois
representations, J. Amer. Math. Soc. 12 (1999), 521567. MR 99i:11037
[deJ] A.J. de Jong, Finite locally free group schemes in characteristic p and Dieudonne modules,
Inv. Math. 114 (1993), 89138. MR 94j:14043
[De] P. Deligne, Formes modulaires et representations -adiques, in: Lecture Notes in Math.
179, Springer-Verlag, 1971, pp. 139172.
[DS] P. Deligne, J.-P. Serre, Formes modulaires de poids 1, Ann. Sci. Ec. Norm. Sup. 7 (1974),
507530. MR 52:284
[Di1] F. Diamond, The rened conjecture of Serre, in Elliptic Curves, Modular Forms and
Fermats Last Theorem (Hong Kong, 1993), International Press, 1995, pp. 2237. MR
97b:11065
[Di2] F. Diamond, On deformation rings and Hecke rings, Ann. Math. 144 (1996), 137166.
MR 97d:11172
[DT] F. Diamond, R. Taylor, Lifting modular mod representations, Duke Math. J. 74 (1994),
253269. MR 95e:11052
[E] T. Ekedahl, An eective version of Hilberts irreducibility theorem, in Seminaire de
Theorie des Nombres, Paris 1988-89, Birkhauser, 1990. MR 92f:14018
[EGA] A. Grothendieck,

Elements de Geometrie Algebrique IV
4
, Publ. Math. IHES 32, 1966-7.
MR 39:220
[Fa] G. Faltings, Endlichkeitss atze f ur abelsche Variet aten uber Zahlk orpern, Inv. Math. 73
(1983), 349366. MR 85g:11026a; erratum MR 85g:11026b
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use
THE MODULARITY OF ELLIPTIC CURVES OVER Q 939
[FM] J.-M. Fontaine, B. Mazur, Geometric Galois representations, in Elliptic Curves, Modular
Forms and Fermats Last Theorem (Hong Kong, 1993), International Press, 1995, pp.
4178. MR 96h:11049
[G] P. Gerardin, Facteurs locaux des alg`ebres simples de rang 4. I, in Groupes Reductifs et
Formes Automorphes, I (Paris, 197677) Univ. Paris VII, 1978, pp. 3777. MR 84f:22023
[Kh] C. Khare, A local analysis of congruences in the (p, p) case: Part II, Invent. Math. 143
(2001), no. 1, 129155. CMP 2001:06
[Kl] F. Klein, Vorlesungen uber das Ikosaeder und die Au osung der Gleichungen vom f unften
Grade, Teubner, 1884.
[L] R.P. Langlands, Base Change for GL(2), Annals of Math. Studies 96, Princeton Univ.
Press, Princeton, 1980. MR 82a:10032
[Man] J. Manoharmayum, Pairs of mod 3 and mod 5 representations arising from elliptic curves,
Math. Res. Lett. 6 (1999), 735754. MR 2000m:11045
[Maz] B. Mazur, Number theory as gady, Amer. Math. Monthly 98 (1991), 593610. MR
92f:11077
[Ra] M. Raynaud, Schemas en groupes de type (p, p, . . . , p), Bull. Soc. Math. France 102
(1974), 241280. MR 54:7488
[Ri] K. Ribet, The -adic representations attached to an eigenform with Nebentypus: a survey,
in: Lecture Notes in Math. 601, Springer-Verlag, 1977, pp. 1752. MR 56:11907
[SBT] N. Shepherd-Barron, R. Taylor, Mod 2 and mod 5 icosahedral representations, J. Amer.
Math. Soc. 10 (1997), 283298. MR 97h:11060
[Se1] J.-P. Serre, Local Fields, Springer-Verlag, 1979. MR 82e:12016
[Se2] J.-P. Serre, Sur les representations modulaires de degre 2 de Gal(Q/Q), Duke Math. J.
54 (1987), 179230. MR 88g:11022
[Sh1] G. Shimura, On elliptic curves with complex multiplication as factors of the Jacobians of
modular function elds, Nagoya Math. J. 43 (1971), 199208. MR 45:5111
[Sh2] G. Shimura, Introduction to the Arithmetic Theory of Automorphic Functions, Princeton
Univ. Press, Princeton, 1971. MR 47:3318
[Sh3] G. Shimura, Response to 1996 Steele Prize, Notices of the AMS 43 (1996), 13441347.
[T] J. Tate, p-divisible groups, in Proceedings of a conference on local elds (Driebergen,
1966), Springer, 1967, 158183. MR 38:155
[TW] R. Taylor, A. Wiles, Ring-theoretic properties of certain Hecke algebras, Ann. Math. 141
(1995), 553572. MR 96d:11072
[T] J. Tunnell, Artins conjecture for representations of octahedral type, Bull. Amer. Math.
Soc. 5 (1981), 173175. MR 82j:12015
[We1] A. Weil,

Uber die Bestimmung Dirichletscher Reihen durch Funktionalgleichungen, Math.
Ann. 168 (1968), 149156. MR 34:7473
[We2] A. Weil, Scientic works. Collected papers. III (19641978), Springer-Verlag, 1979. MR
80k:01067c
[Wi] A. Wiles, Modular elliptic curves and Fermats Last Theorem, Ann. Math. 141 (1995),
443551. MR 96d:11071
Departement de Mathematiques, CNRS, Universite Paris-Sud, 91405 Orsay cedex,
France
E-mail address: [email protected]
Department of Mathematics, Harvard University, Cambridge, Massachusetts 02138
E-mail address: [email protected]
Current address: Department of Mathematics, University of Michigan, Ann Arbor, Michigan
48109
E-mail address: [email protected]
Department of Mathematics, Brandeis University, Waltham, Massachusetts 02454
E-mail address: [email protected]
Department of Mathematics, Harvard University, Cambridge, Massachusetts 02138
E-mail address: [email protected]
License or copyright restrictions may apply to redistribution; see http://www.ams.org/journal-terms-of-use