Ring Theoretic Properties of Certain Hecke Algebras

Richard Taylor D.P.M.M.S., Cambridge University, 16 Mill Lane, Cambridge, CB2 1SB, U.K. Andrew Wiles Department of Mathematics, Princeton University, Washington Road, Princeton, NJ 08544, U.S.A.

7 October 1994

Introduction
In the work of one of us (A.W.) on the conjecture that all elliptic curves dened over Q are modular, the importance of knowing that certain Hecke algebras are complete intersections was established. The purpose of this article is to provide the missing ingredient in [W2] by establishing that the Hecke algebras considered there are complete intersections. As is recorded in [W2], a method going back to Mazur [M] allows one to show that these algebras are Gorenstein, but this seems to be too weak for the purposes of that paper. The methods of this paper are related to those of chapter 3 of [W2]. We would like to thank Henri Darmon, Fred Diamond and Gerd Faltings for carefully reading the rst version of this article. Gerd Faltings has also suggested a simplication of our argument and we would like to thank him for allowing us to reproduce this in the appendix to this paper. R.T. would like to thank A.W. for his invitation to collaborate on these problems and for sharing his many insights into the questions considered. R.T. would also like to thank Princeton University, Universit e de Paris 7 and Harvard University for their hospitality during some of the work on this paper. A.W. was supported by an NSF grant. 1

Notation

Let p denote an odd prime, let O denote the ring of integers of a nite extension K/Qp , let denote its maximal ideal and let k = O /. If L is a perfect eld GL will denote its absolute Galois group and if the characteristic of L is not p then : GL Z p will denote the p-adic cyclotomic character. If L is a number eld and a prime of its ring of integers then G will denote a decomposition group at and I the corresponding inertia group. We will denote by Frob the arithmetic Frobenius element of G /I. If G is a group and M a G-module we will let M G and MG denote respectively the invariants and coinvariants of G on M . If is a representation of G into the automorphisms of some abelian group we shall let V denote the underlying G-module. If H is a normal subgroup of G then we shall let H and H denote the representation of G/H on respectively VH and V,H . We shall also x a continuous representation : GQ GL2 (k ) with the following properties. is modular in the sense that it is a mod p representation associated to some modular newform of some weight and level. The restriction of to the group Gal (Q/Q( irreducible. (1)(p1)/2 p)) is absolutely

If c denotes complex conjugation then det (c) = 1. The restriction of to the decomposition group at p either has the form 1 0 2 with 1 and 2 distinct characters and with 2 unramied; or is induced from a character of the unramied quadratic extension of Qp whose restriction to the inertia group is the fundamental character of level 2, F Ip p2 . If l = p then either |Il 0 0 1 ,

 or |Il

1 0 1

or |Gl is absolutely irreducible and in the case |Il is absolutely reducible we have l 1 mod p. (This implies that |Gl is either unramied or of type A, B or C as dened in chapter 1 of [W2]. On the other hand if |Gl is of type A, B or C then some twist satises the condition above.) 1 we will x the pair of characters 1 , 2 . 0 2 Note that in some cases this may involve making a choice. We will let Q denote a nite set of primes q with the properties In the case that |Gp is unramied at q , q 1 mod p, (Frob q ) has distinct eigenvalues, which we shall denote q and q . Much of our notation will involve a subscript Q to denote dependence on Q, whenever Q = we may simply drop it from the notation. For q Q we shall let q denote the Sylow p-subgroup of (Z/q Z) . We shall let q denote a generator. We will write Q for the product of the q with q Q. We will let aQ denote the kernel of the map O [Q ] O which sends every element of Q to 1. Let q denote the character GQ Gal (Q(q )/Q) q , = (Z/q Z) and let Q = qQ q . We will denote by NQ the product of the following quantities: the conductor of ; the primes in Q; p, if is not at (i.e. does not arise from the action of Gp on the Qp -points of some nite at group scheme over Zp ) or if det |Ip = . (We remark that if |Gp is at but det |Ip = then |Gp arises from an etale group scheme over Zp . We also note that in [W2] the term at is not used when the group scheme is ordinary.) We will let Q denote the inverse image under 0 (NQ ) (Z/NQ Z) of the product of the following subgroups: 3

 the Sylow p-subgroup of (Z/M Z) , where M denotes the conductor of ; for each q Q the unique maximal subgroup of (Z/q Z) of order prime to p. Let T(Q ) denote the Z-subalgebra of the complex endomorphisms of the space of weight 2 cusp forms on Q which is generated by the Hecke operators Tl and l for l |pNQ , by Uq for q Q and by Up if p|NQ . Let m denote the ideal of T(Q ) Z O generated by , by tr (Frob l ) Tl and det (Frob l ) l l for l |pNQ , by Uq q for q Q and by Up 2 (Frob p ) if p|NQ . Note that if Q = this denition only makes sense if O is suciently large that k contains the eigenvalues of (Frob q ) for all q Q. It is a deep result following from the work of many mathematicians that m is a proper ideal (see [D]), and so maximal. We let TQ denote the localisation of T(Q ) Z O at m. Note that TQ is reduced because the operators Tl for l |NQ act semi-simply on the space of cusp forms for Q and the Uq for q Q act semi-simply on all common eigenspaces for the Tl for which the corresponding p-adic representation is either ramied at q or for which (Frob q ) has distinct eigenvalues. There is a natural map O [Q ] TQ , which sends x Q to y where y Z, y x mod q for all q Q and y 1 mod N . It follows from the discussion after theorem 2.1 of [W2] or from the work of Carayol [C2] that there is a continuous representation : GQ GL2 (TQ ); mod Q is unramied at l and we have tr mod such that if l |NQ p then mod Q Q (Frob l ) = Tl mod and det Q (Frob l ) = l l . In particular the reduction of mod modulo the Q maximal ideal of TQ is . From [C1] we can deduce the following. If q Q then mod Q |Gq = 1 2 where 1 is unramied and 1 (Frob q ) = Uq , and where 2 |Iq = q |Iq . If l = p and |Il is non-trivial but unipotent then mod Q |Il is unipotent. If l Q {p} and either |Il = 1 or |Gl is absolutely irreducible then mod Q (Il ) (Il ). = Q where is a character of order prime to p. det mod Q Moreover if |Gp is at and if det |Ip = then p |NQ so mod Q |Gp is at (i.e. the reduction modulo every ideal of nite index is at). If |Gp is not at 4

1 and Up is a unit in TQ . It 0 2 follows from theorem 2 of [W1] (or more directly in the case 1 |Iq = from 1 proposition 12.9 of [G]) that mod , where 2 is unramied Q |Gp 0 2 and 1 (Ip ) has order prime to p. In the case that 1 is unramied we know further that 1 = 2 (see proposition 1.1 of [W2]) and that this character has 1/2 Q of nite order. It will be convenient to introduce the twist Q = mod Q mod Q . In particular we see that det Q is valued in O . The main theorem of this paper is as follows. Recall that we may write T for T . or if det |Ip = then p|NQ , |Gp Theorem 1 The ring T is a complete intersection. We note that if O is the ring of integers of a nite extension K /K then the ring constructed using O in place of O is just TQ O O . Also T is a complete intersection if and only if T O O is (using for instance corollary 2.8 on page 209 of [K2]). Thus we may and we shall assume that O is suciently large that the eigenvalues of every element of are rational over k and that there is a homomorphism : T O . In particular the denition of TQ makes sense for all Q. There is an induced map Q : TQ T O . The map TQ T takes the operators Tl and l to themselves and the operator Uq to the unique root of U 2 Tq U + q q in T above q . We will let Q denote the kernel of Q and will let Q denote the ideal Q (Ann TQ (Q )). Then it is known that > #Q /2 Q #O /Q with equality if and only if TQ is a complete intersection (see the appendix of [W2] or [L], we are using the fact that TQ is reduced).

Generalisation of a Result of de Shalit

In this section we shall use the methods of de Shalit (see [dS]) to prove the following theorem. Theorem 2 The ring TQ is a free O [Q ] module of O [Q ]-rank equal to the O -rank of T. By lemma 3 of [DT] we may choose a prime R with the following properties: R |6NQ p; R 1 mod p; 5

 (Frob R ) has distinct eigenvalues R and R ; (1 + R)2 det (Frob R ) = R(tr (Frob R ))2 . Let Q be dened in the same way as Q but with (Z/q Z) replacing its maximal subgroup of order prime to p in the denition for each q Q. Let Q = Q 1 (R) and let Q = Q 1 (R). The purpose of introducing the auxiliary prime R is to make these groups act freely on the upper half complex plane. Let T (Q ) denote the Z-subalgebra of the complex endomorphism ring of the space of weight two modular (not necessarily cusp) forms generated by the operators Tl and l for l |NQ R and by Ul for l|NQ R. Let mQ denote the maximal ideal of T (Q ) Z O generated by the following elements: ; Tl tr (Frob l ) and l l det (Frob l ) for l |NQ Rp; Uq q for q Q or q = R; Ul tr Il (Frob l ) if l|N and l = p; Up 2 (Frob p ) if p|N ; Tp tr Ip (Frob p ) if p |N . Let TQ denote the localisation of T (Q ) Z O at mQ . Let YQ denote the quotient of the upper half complex plane by Q and let XQ denote its standard compactication. Complex conjugation c acts continuously on these Riemann surfaces. We let H 1 (YQ , O ) and H 1 (XQ , O ) denote the 1 eigenspaces of c on H 1 (YQ , O ) and H 1 (XQ , O ). All these denitions go over verbatim, but with Q replacing Q. Lemma 1 TQ = TQ and TQ = T. This is a standard argument which we will only sketch. First observe that because is irreducible TQ and TQ can be dened using the ring generated by the Hecke operators on the spaces of weight two cusp forms S2 (Q ) and S2 (Q ) (rather than spaces of all modular forms). The same arguments as in the proof of proposition 2.15 of [W2] show that we can drop the Hecke operators Tp if p |NQ and the Hecke operators Ul for l = p and l|N from the denition without changing the Hecke algebra. Next we will show that we need only consider the algebras generated in the endomorphisms of S2 (Q )2 S2 (Q ) #Q+1 and S2 ()2 S2 (Q ). This follows from the following facts. 6

 As R 1 mod p and det is unramied at R, no component of TQ nor of TQ can correspond to an eigenform with a non-trivial action of (Z/RZ) . As R /R = R1 in k , no component of TQ nor of TQ can correspond to an eigenform which is special at R (i.e. an eigenform which corresponds to a cuspidal automorphic representation of GL2 (A) whose component at R is special). As for each prime q Q, q /q = q 1 in k , no component of TQ can correspond to an eigenform which is special at q . The ring generated by the Hecke operators Tl and l for l |pNQ R, by Uq for q Q {R} and by Up if p|NQ on S2 (Q )2 is isomorphic to T(Q )[uR ]/(u2 R TR uR + R R ). In fact UR acts by the matrix TR R R 1 0

on S2 (Q )2 . Similarly the ring generated by the Hecke operators Tl and l for #Q+1 l |pNQ R, by Uq for q Q {R} and by Up if p|NQ on S2 ()2 is isomorphic T u + q q : q Q { R } ). Tensoring to T()[uq : q Q {R}]/(u2 q q q with O and localising at the appropriate maximal ideal we get the desired isomorphism. We have to use the fact that u2 R TR uR + R R has two roots in TQ with distinct reductions modulo the maximal ideal and the similar facts over T for u2 q Tq uq + q q with q Q {R}. Because is irreducible we see that H 1 (YQ , O )mQ = H 1 (XQ , O )mQ and that H 1 (YQ , O )mQ = H 1 (XQ , O )mQ . By corollary 1 of theorem 2.1 of [W2] we 1 see that H 1 (XQ , O ) mQ are free rank one TQ -modules and that H (XQ , O )mQ are free rank one TQ -modules. Hence it will suce to prove the following proposition. Proposition 1 H 1 (YQ , O ) is a free O [Q ]-module, with O [Q ]-rank equal to the O -rank of H 1 (YQ , O ) . Because H 1 (YQ , K ) = H 1 (YQ , K )Q we need only show that H 1 (YQ , O ) is a free O [Q ]-module. Because R 5, Q acts freely on the upper half complex plane and so may be identied with the fundamental group of YQ . In particular we see that Q is a free group. Similarly Q acts freely on the upper half complex plane and we get identications H 1 ( YQ , O ) = H 1 (Q , O ) = H 1 (Q , O [Q ]), 7

the latter arising from Shapiros lemma. Under these identications complex conjugation goes over to the involution induced by conjugation by = 1 0 and trivial action on the coecients. (This follows because the 0 1 action of c on YQ is induced by the map z z of the upper half complex plane to itself.) Because Q is a free group, the cocycles Z 1 (Q , O [Q ]) are a free O [Q ]module. (If 1 , ..., a are free generators of Q then we have an isomorphism Z 1 (Q , O [Q ]) O[Q ]a ( (1 ), ..., (a )).) On the other hand acts trivially on Q and so the coboundaries are contained in Z 1 (Q , O [Q ])+ . Thus H 1 (YQ , O ) = Z 1 (Q , O [Q ]) is a free O [Q ]module, as desired. Before leaving this section we remark the following corollaries of theorem 2. Corollary 1 If q Q then TQ{q} /(q 1) TQ . Moreover (q 1)TQ{q} # 1 and (1 + q + ... + q q )TQ{q} are annihilators of each other in TQ{q} . Corollary 2 If for some Q the ring TQ is a complete intersection then so is T. The proof is by showing that under the assumption that TQ{q} is a complete intersection, so is TQ . The argument in section 2 of [K1] shows that if a complete local noetherian ring S is a complete intersection, if f S and if HomS (S/(f ), S ) = HomS (S/(f ), Ann S (f )) is a free S/(f ) module then S/(f ) is a complete intersection. We apply this result with S = TQ{q} and f = q 1. The nal condition is met because the annihilator of q 1 in TQ{q} is a free rank one TQ module by the last corollary. Corollary 3 Q = #Q . We remind the reader that Q is dened at the end of section one and that = . The proof of the corollary is by showing that Q{q} = Q #q if TQ . q Q. Write Q = Q {q } and let denote the natural surjection TQ It suces to prove that in TQ we have the equation Ann Q = (1 + q + # 1 ... + q q )1 (Ann TQ (Q )). However 1 (Ann TQ (Q )) is just the set of elements t of TQ for which tQ (q 1)TQ . The inclusion Ann Q # 1 (1 + q + ... + q q )1 (Ann TQ (Q )) is now clear. Conversly if t Ann Q 8

then t annihilates ker and hence t = (1 + q + ... + q q )s. Then we see that sQ (q 1)TQ , i.e. that s 1 (Ann TQ (Q )). The other inclusion now follows. Corollary 4 #(aQ TQ /Q aQ TQ )#(O / ) = #(O /Q ). To prove this corollary note that aQ /a2 Q = qQ O /#q . Thus it follows 2 from theorem 2 that aQ TQ /aQ TQ = TQ O[Q ] aQ /a2 Q = qQ T/#q and so O / # . This corollary now follows we deduce that aQ TQ /Q aQ TQ = qQ q from the last one.

# 1

Some Algebra

In this section we shall establish certain criteria for rings to be complete intersections. We shall rely on the numerical criterion established in the appendix of [W2]. In that appendix there is a Gorenstein hypothesis which can be checked in the cases where we will apply the results of this section. However in order to state the results of this section in somewhat greater generality we shall reference the paper [L], where the Gorenstein hypothesis in the appendix of [W2] is removed, rather than the appendix of [W2] directly. Fix a nite at reduced local O algebra T with a section : T O. We will consider complete local noetherian O algebras R together with maps T , by R the R T . We will denote by JR the kernel of the map R induced map R O , by R the kernel of R and by R the image under R of the annihilator in R of R . We will let R = (2 R JR )/R JR . 2 2 R , JS is the pre-image of JR and so S R . If S R T then S We have an exact sequence
2 (0) R JR /R JR R /2 R T /T (0).

From this we deduce the following facts. #R < . (To see this it suces to show that (R )R = (0). However as T is reduced (JR )R = (R )R and so the map (JR /R JR )R (R /2 R )R is an isomorphism. The result follows.)
2 #R #(R /2 R ) = #(T /T )#(JR /R JR ).

Lemma 2 Suppose we have the inequalities #(T /2 T ) #(O /T )#R and #(O /T )#(JR /R JR ) #(O /R ) < and suppose R is a nite at O algebra. Then R is a complete intersection. 9

To show this rst note that we have the inequalities

Now applying the criterion of [L] we see that R is a complete intersection. Lemma 3 We have the inequality: #(O /R ) #(JR /R JR )#(O /T ). As Fitt R (JR ) Ann R (JR ) we see that Fitt O (JR /R JR ) R Ann R (JR ). On the other hand it is easy to see that Ann R (R ) {s R|sR JR }Ann R (JR ). Applying R we see that R T Fitt O (JR /R JR ), and the lemma follows. Lemma 4 If R is a complete intersection which is nite and at over O and R = (0) then #(T /2 T ) #(O /T )#R . If R is a power series ring the same result is true without the assumptions that it is nite over O and that R = (0). For the rst part we see that, as R is a complete intersection, #(R /2 R) = #(O /R ) (see [L]). Thus we see that
2 #R #(O /R ) = #(T /2 T )#(JR /R JR ) #(T /T )#(O /R )/#(O /T ).

The rst result follows. For the second result note that we can factor R T as R R T with R a complete intersection which is nite and at 2 over O and for which R /2 R T /T (by the proof of lemma 9 of [L]). 2 Then #O /R = #R /R < and so #T /2 T #(O /T )#R . However R , so the result follows. R We now return to the notation of the rst section. We will let Q denote TQ and JQ denote JTQ . Proposition 2 Suppose that for a series of sets Qn we have ideals In in TQn with the following properties. 10

1. In is contained in m2 TQn and TQn /In has nite cardinality. 2. In+1 T In T and
n In T

= (0).

TQn /In such that 3. There is a surjective map of O -algebras TQn+1 /In+1 the diagram TQn+1 /In+1 TQn /In T/In+1 T T/In T commutes. (Note that this map is not assumed to take a given Hecke operator to itself.) 4. lim TQn /In is a power series ring. Then for n suciently large TQn is a complete intersection, and hence T is a complete intersection (by corollary 2 of theorem 2). T and can choose Let P denote lim TQn /In . We get a natural map P T and TQn TQn /In . maps P TQn compatible with the maps TQn 2 Because In mTQn we see that the map P TQn is surjective. We have a sequence Qn ((JQn + In ) (2 P Qn + In ))/(JQn Qn + In ). (Note that although the maps P TQn and P Qn are not compatible as n varies the composite map above is.) Moreover P = lim ((JQn + In ) (2 Qn + In ))/(JQn Qn + In ) (using the fact that TQn /In is nite for all n) and so as P is nite we have that the map P ((JQn + In ) (2 Qn + In ))/(JQn Qn + In ) is injective for n suciently large. Thus for n suciently large P Qn . We deduce the inequality #(/2 ) #(O / )#P = #(O / )#Qn , where the rst inequality follows from lemma 4. The proposition follows on applying corollary 4 of theorem 2 and lemma 2. Corollary 1 Suppose that we have an integer r and a series of sets Qm with the following properties: 1. if q Qm then q 1 mod pm ; 2. is unramied at q and (Frob q ) has distinct eigenvalues; 11

3. #Qm = r; 4. TQm can be generated as an O -algebra by r elements. Then T is a complete intersection. To prove this corollary it is useful to have the following denition. By a level n structure we shall mean a quadruple B = (A, , , ), where A is an O -algebra, : O [[T1 , ..., Tr ]] A, : O [[S1 , ..., Sr ]]/(pn , (S1 + 1)p 1, ..., (Sr + 1)p 1) A makes A a n n free module over O [[S1 , ..., Sr ]]/(pn , (S1 + 1)p 1, ..., (Sr + 1)p 1), and : A/(S1 , ..., Sr ) T/pn . If B is a structure of level n and n n then it induces a structure of level n n n by reducing mod(pn , (S1 + 1)p 1, ..., (Sr + 1)p 1). pm 1|q Qm ). This extends to a level m structure Let Am = TQm /(pm , q that we will denote Bm . For n m we will let Bm,n denote the level n structure induced by Bm . There are only nitely many isomorphism classes of structures of level n and so we may choose recursively integers m(n) with the following two properties. 1. Bm(n),n1 = Bm(n1),n1 . 2. Bm(n),n = Bm,n for innitely many integers m. Let In denote the kernel of the map from TQm(n) to the ring underlying Bm(n),n . We claim that the pairs (Qm(n) , In ) for n 2 satisfy the requirements of proposition 2 (we use n 2 to ensure that In m2 Qn ). We need only check that lim Bm(n),n is a power series ring. On the one hand it is a nite free O [[S1 , ..., Sr ]]-module, and so has Krull dimension r + 1. On the other hand it is a quotient of O [[T1 , ..., Tr ]] and so must in fact equal O [[T1 , ..., Tr ]].

n n

Galois Cohomology

It remains to nd a sequence of sets Qm with the properties of corollary 1 of proposition 2. We must recall some denitions in Galois cohomology. We 1 dene Hf (Ql , ad0 ). 12

1 1. If l = p then Hf (Ql , ad0 ) = H 1 (Fl , (ad0 )Il ) = ker (H 1 (Ql , ad0 ) H 1 (Il , ad0 )). 1 2. If |Gp is at and det |Ip = then we will let Hf (Qp , ad0 ) denote those elements in H 1 (Qp , ad0 ) Ext1 k [Gp ] (V , V ) which correspond to extensions which can be realised as the Qp -points on the generic bre of a nite at group scheme over Zp .

3. If |Gp

1 1 (Qp , ad0 ) denote with 1 |Ip = then we let Hf 0 2 the kernel of H 1 (Qp , ad0 ) H 1 (Ip , (ad0 )/Homk (V /F, F )), where F denotes the line in V where Gp acts by the character 1 .

4. Finally if |Gp

1 with 1 |Ip = but is not at then 0 2 1 (Qp , ad0 ) denote the kernel of the map H 1 (Qp , ad0 ) we will let Hf H 1 (Qp , (ad0 )/Homk (V /F, F )), where F denotes the line in V where Gp acts by the character 1 .

1 (Q, ad0 ) to be the inverse image under We dene HQ

H 1 (Q, ad0 )
lQ

H 1 (Ql , ad0 )

1 Hf (Ql , ad0 ). 1 1 (Ql , ad0 (1)) to be the annihilator of Hf (Ql , ad0 ) under We also dene Hf the pairing of Tate local duality H 1 (Ql , ad0 ) H 1 (Ql , ad0 (1)) k . We then 0 1 dene HQ (Q, ad (1)) to be the inverse image under

of

lQ

H 1 (Q, ad0 (1))

H 1 (Ql , ad0 (1))

of

lQ

1 Hf (Ql , ad0 (1)).

0 1 1 (Q, ad0 ) dimk HQ Lemma 5 dimk HQ (Q, ad (1)) + #Q

To see this we apply proposition 1.6 of [W2]. For l Q and l = p we 1 (Ql , ad0 ) in H 1 (Ql , ad0 ) is equal to see that hl = 1 because the index of Hf #H 1 (Il , ad0 )GFl which in turn equals #((ad0 (1))Il )GFl = #((ad0 (1))Il )GFl = #H 0 (Ql , ad0 (1)). 13

For q Q we have that hq = #H 0 (Qq , ad0 (1)) = #k . It remains to check that hp h 1. In the case that |Gp is not at or det |Ip = this is proved in parts (iii) and (iv) of proposition 1.9 of [W2]. Thus suppose that |Gp is at and 1 (Qp , ad0 ) 1 + dimk H 0 (Qp , ad0 ) det |Ip = . We must show that dimk Hf (= 1 if |Gp is indecomposable and = 2 otherwise). Following [FL] let M denote the abelian category of k vector spaces M with a distinguished subspace M 1 and a k -linear isomorphism : M/M 1 M 1 M . Then there are equivalences of categories between: Mop ; nite at group schemes A/Zp with an action of k ; k [Gp ]-modules which are isomorphic as modules over Fp [Gp ] to the Qp points of some nite at group scheme over Zp . See section 9 of [FL] for details. The only point here is that an action of k on the generic bre of a nite at group scheme over Zp extends uniquely to an action on the whole scheme. Let M () denote the object of M corresponding to . Then dimk M () = 2 and dimk M ()1 = 1 (since det |Ip = ). We get 1 an embedding Ext1 M (M (), M ()) H (Qp , ad). We will show that 1. dimk Ext1 M (M (), M ()) = 2 if |Gp is indecomposable and = 3 otherwise;
1 1 2. the composite map Ext1 M (M (), M ()) H (Qp , ad) H (Qp , k ) is non-trivial, where tr denotes the map induced by the trace. tr

The lemma will then follow. For the rst point it is explained in lemma 4.4 of [R] how to calculate 1 Ext1 M (M (), M ()). Let {e0 , e1 } be a basis of M () with e1 M () . Let (e0 , 0) = e0 + e1 and (0, e1 ) = e0 + e1 . Then Ext1 M (M (), M ()) can be identied as a k -vector space with M2 (k ) modulo the subspace of matrices of the form r 0 s t r 0 0 t = 0 (r t ) s + (t r) s ,

for any r, s, t k . Thus dimk Ext1 M (M (), M ()) = 2 if = 0 and = 3 if = 0. However = 0 if and only if M ()1 is a subobject of M () in M. This is true if and only if has a one dimensional quotient on which inertia acts by which itself is true if and only if |Gp is decomposable. 14

For the second point consider the k [Gp ]-module where is the unramied representation 1 1 . Frob p 0 1 Then is an extension of by itself. Moreover its extension class maps to the element of H 1 (Qp , k) = Hom(Q p , k ) which is trivial on Zp and takes p 2. Finally it is isomorphic to the action of Gp on the Qp points of a nite at group scheme over Zp , because this is true over an unramied extension.
1 (Q, ad0 ) eleLemma 6 TQ can be generated as an O algebra by dimk HQ ments.

Let mQ denote the maximal ideal of TQ . It will suce to show that there is an embedding of k -vector spaces
0 1 : Homk (mQ /(m2 Q , ), k ) HQ (Q, ad ).

We rst dene
1 : Homk (mQ /(m2 Q , ), k ) H (Q, ad).

If is a non-zero element of the left hand group we may extend it uniquely to : TQ . We a map of local O -algebras k [ ] where 2 = 0. Let = Q get an exact sequence (0) V V V (0), 1 and hence a class () in Ext1 k [GQ ] (, ) = H (Q, ad). Because det is valued in k k [ ] we see that () actually lies H 1 (Q, ad0 ). 1 (Ql , ad0 ) for l Q. This computation is We claim that resl () lies in Hf very similar to some in [W2], but is not actually carried out there, so we give an argument here. First suppose that l = p and that either p |#(Il ) or |Gl is absolutely irreducible. In this case Q (Il ) (Il ) and det Q |Il has order prime to l. Because either p |#(Il ) or p = 3 and ad(Il ) = A4 we have that H 1 ((Il ), ad0 ) = (0) and so |Il = |Il k k [ ]. The result follows in this case. Secondly suppose that |Il is unipotent and nontrivial. Then the same is true for Q |Il and then also for . However the Sylow p-subgroup of Il is pro-cyclic and so |Il must also be of the form k k [ ] and resl () H 1 (Il , ad0 ()) must vanish. In the case l = p, is at and det |Ip = the claim is immediate 1 with 1 |Ip = from the denitions. In the case l = p and |Gp 0 2 15

use the fact that Q |Ip

1 0 1

1 denotes the Teichmuller lifting where 1 0 2 with 1 |Ip = but

of 1 |Ip . Finally in the case l = p, |Gp not at use the fact that Q |Gp 0

where is an unramied character of order prime to p. It remains to show that is injective. Suppose it were not. Then we could nd a non-zero such that k k [ ]. Thus tr Q is valued in O + ker and in particular TQ is not generated as an O -algebra by tr Q . We will show this is not the case. If q Q and q then we can nd Gq such that (Uq )2 (tr Q ( ))(Uq ) + det Q ( ) = 0. ( will in fact lie above Frob q .) This polynomial has distinct roots in k and so both its roots in TQ lie in the subO -algebra T generated by the image of tr Q . Thus for all q , Uq T . Hence Uq T , and as Uq is a unit, T . Moreover for all l Q for which is unramied we see that Tl Q (Frob l )1/2 T and hence Tl T . If p|NQ then Up Q (Frob p )1/2 is a root of the polynomial X 2 (tr Q ( ))X +det Q ( ) for any element of Gp which lies above Frob p . For some over Frob p this polynomial has two distinct roots in k and so Up T . Thus T = TQ as we required. Finally we turn to the proof of the main theorem. As in [W2] (after equation (3.8)) we may nd a set of primes Qm with the following properties: 1. if q Qm then q 1 mod pm ; 2. if q Qm then is unramied at q and (Frob q ) has distinct eigenvalues;
0 1 3. H (Q, ad (1)) qQm

H 1 (Fq , ad0 (1)).

As for each such q , H 1 (Fq , ad0 ) = k we see that by shrinking Qm we may suppose that the latter map is an isomorphism. Then we have that #Qm = 0 0 1 1 Also HQ dimk H (Q, ad (1)) is the kernel of the map in 3. (Q, ad (1)). m 1 (Q, ad0 ) above and so is trivial. Thus by lemma 5 we see that dimk HQ m 0 1 #Qm and so TQm can be generated by #Qm = dimk H (Q, ad (1)) elements. The main theorem now follows from corollary 1.

References
[C1] H.Carayol, Sur les repr esentations p-adiques associ ees aux formes modulaires de Hilbert, Ann. Sci. Ec. Norm. Super. 19 (1986) 409-468. 16

[C2] H.Carayol, Formes modulaires et repr esentations Galoisiennes ` a valeurs dans un anneau local complet, in p-adic monodromy and the BirchSwinnerton-Dyer conjecture (eds. B.Mazur and G.Stevens), Contemporary Math. 165 (1994). [D] F.Diamond, The rened conjecture of Serre, to appear in the proceedings of the 1993 Hong Kong conference on modular forms and elliptic curves.

[DT] F.Diamond and R.Taylor, Lifting modular mod l representations, Duke Math. J. 74 (1994) 253-269. [FL] J.-M.Fontaine and G.Lafaille, Construction de repr esentations p-adiques, Ann. Sci. Ec. Norm. Super. 15 (1982) 547-608. [G] B.Gross, A tameness criterion for Galois representations associated to modular forms mod p, Duke Math. J. 61 (1990), 445-517.

[K1] E.Kunz, Almost complete intersections are not Gorenstein, J. of Algebra 28 (1974), 111-115. [K2] E.Kunz, Introduction to commutative algebra and algebraic geometry, Birkh auser 1985. [L] H.Lenstra, Complete intersections and Gorenstein rings, to appear in the proceedings of the 1993 Hong Kong conference on modular forms and elliptic curves. B.Mazur, Modular curves and the Eisenstein ideal, Publ. Math. IHES 47 (1977), 133-186.

[M]

[dS] E. de Shalit, On certain Galois representations related to the modular curve X1 (p), to appear in Compositio Math. [R] R.Ramakrishna, On a variation of Mazurs deformation functor, Comp. Math. 87 (1993), 269-286.

[W1] A.Wiles, On ordinary -adic representations associated to modular forms, Invent. Math. 94 (1988), 529-573. [W2] A.Wiles, Modular elliptic curves and Fermats last theorem, preprint (1994).

17

Appendix
The purpose of this appendix is to explain certain simplications to the arguments of chapter 3 of [W2] and to section 3 of this paper. These simplications were found by G.Faltings and we would like to thank him for allowing us to include them here. We should make it clear that the arguments of this appendix (just as those of chapter 3 of [W2] and section 3 of this paper) apply only to proving conjecture 2.16 of [W2] for the minimal Hecke ring and minimal deformation problem. In order to prove theorem 3.3 of [W2] one needs to invoke theorem 2.17 and the arguments of chapter 2 of [W2]. We will keep the notation and assumptions of the main body of this paper. Let Q denote a nite set of primes as described in section 1 of this paper. By a deformation of of type Q we shall mean a complete noetherian local O algebra A with residue eld k together with an equivalence class of continuous representations : GQ GL2 (A) with the following properties: mod mA = ;
1

det is a character of nite order prime to p;

if l Q {p} and |Il is semi-simple then (Il ) (Il ); if l Q {p} and |Il 1 0 1 then |Il 1 0 1 ;

if is at and det |Ip = then is at; if either is not at or if det |Ip = is unramied and 2 mod mA = 2 . : GQ GL2 (RQ ) of type As in chapter 1 of [W2] there is a universal lift univ Q Q. Recall that the universal property is for lifts up to conjugation. Moreover one checks (c.f. the second paragraph of the proof of lemma 6) that there is a natural isomorphism
0 1 Homk (mRQ /(, m2 RQ ), k ) = HQ (Q, ad ).

then |Gp

1 0 2

where 2

pushes forward to a There is also a natural map RQ TQ so that univ Q conjugate of Q . Recall that if Q = we shall often drop it from the notation. In this appendix we shall reprove the following result. 18

Theorem 3 R T and these rings are complete intersections. We note that if O is the ring of integers of a nite extension K /K then the rings TQ and RQ constructed using O in place of O are just TQ O O and RQ O O . Also TQ is a complete intersection if and only if TQ O O is (using for instance corollary 2.8 on page 209 of [K2]). Thus we may and we shall assume that O is suciently large that the eigenvalues of every element of (GQ ) are rational over k . We recall that in the penultimate paragraph of section 4 of this paper we showed that TQ is generated as an O -algebra by tr Q (GQ ). Thus we see that the map RQ TQ is a surjection. We will need the following result. 1 0 0 2 q . these characters factor through q : Iq Lemma 7 If q Q then univ Q |Gq
1 where 1 |Iq = 2 | Iq and both

It suces to check the rst assertion. As is unramied at q , univ Q |Gq Zp (1), where Z is topologically generated by some lift factors through Z f of Frob q , Zp (1) is topologically generated by some element and where f f 1 = q . As (Frob q ) has distinct eigenvalues it is easy to see that after a 0 where a b mod mRQ . conjugation we may assume that univ Q (f ) = 0 b We will show that univ Q ( ) is a diagonal matrix with entries congruent to 1 mod mRQ . We will in fact prove this mod mn RQ for all n by induction on n. For n = 1 there is nothing to prove. So suppose this is true modulo mn RQ with n > 0. Then 1 0 +1 (12 + N ) mod mn univ Q ( ) RQ , 0 2 where each i 1 mod mRQ and where N 0 mod mn RQ . We see that mod n+1 mRQ we have a 0 a 0 1+ N 0 b 0 b q 1 1 0 (1 + N )q q 1 0 2 q 1 1 0 (1 + qN ) q 1 0 2 q 1 0 1 + N, q 1 0 2 19
1

+1 and as a b mod mRQ we deduce that N is diagonal mod mn RQ as required. We can choose 2 so that 2 (f ) q mod mRQ . Then we can dene a to be 2 |2 map q RQ Iq . This makes RQ into an O [Q ]-algebra. Using the last lemma and the universal properties of RQ and of R it is easy to see that RQ /aQ R. It moreover follows from the discussion preceding theorem 1 of this paper that the map RQ TQ is a map of O [Q ]-algebras. The key observation is the following ring theoretic proposition. Theorem 3 follows on applying it to the rings RQn and TQn for the sets Qn constructed in O [Qn ] with section 4 of this paper. Note that there is a map O [[S1 , ..., Sr ]] kernel the ideal ((1 + S1 )#q1 1, ..., (1 + Sr )#qr 1), where Qn = {q1 , ..., qr }. Note also that by the displayed isomorphism a couple of lines before theorem RQn . 3 there exists a surjection of O -algebras O [[X1 , ..., Xr ]]

Proposition 3 Suppose r is a non-negative integer and that we have a map of O -algebras R T with T nite and at over O . Suppose for each positive Tn and a commutative diagram integer n we have a map of O -algebras Rn of O -algebras R O [[S1 , ..., Sr ]] Rn T, Tn where 1. there is a surjection of O -algebras O [[X1 , ..., Xr ]] Rn , 2. (S1 , ..., Sr )Rn ker (Rn R ), T ), 3. (S1 , ..., Sr )Tn = ker (Tn 4. if bn denotes the kernel of O [[S1 , ..., Sr ]] Tn then bn ((1 + S1 )p n 1, ..., (1 + Sr )p 1) and Tn is a nite free O [[S1 , ..., Sr ]]/bn -module. Then R T and these rings are complete intersections. Reducing mod we see that it suces to prove this result with k replacing O everywhere. In this case we see that the last condition becomes bn pn pn , ..., Sr ). Further we may replace R by its reduction modulo mR ker (R ( S1 T ), and so we may assume that R is nite over k . We may replace Tn pn pn pn pn by Tn /(S1 , ..., Sr ) and so assume that bn = (S1 , ..., Sr ). Finally we may replace Rn by its image in R Tn .

20

Now dene an n-structure to be a pair of k -algebras B A together with a commutative diagram of k -algebras k [[S1 , ..., Sr ]] B R k [[X1 , ..., Xr ]] A T, such that 1. B R A, R), 2. (S1 , ..., Sr )B ker (B T ), 3. (S1 , ..., Sr )A = ker (A
p p 4. A is a nite free k [[S1 , ..., Sr ]]/(S1 , ..., Sr )-module.
n n

Note that #B (#T )p #R and so we see that there are only nitely many isomorphism classes of n-structures. If S is an n-structure and if m n then pm (m) pm by replacing A by A/(S1 , ..., Sr ) and B we may obtain an m-structure S m p pm , ..., Sr )). by its image in R (A/(S1 As explained above, it follows from the hypotheses of the proposition that an n-structure Sn exists for each n. We next claim that we can nd, for each n, n-structures Sn such that for m n we have Sm = (Sn )(m) . To prove this observe that we can nd recursively integers n(m) with the following properties
(m) (m) Sn(m) = Sn for innitely many n (m1) (m1) and for m > 1, Sn(m) = Sn(m1) .

nr

Then set Sm = Sn(m) . Thus we obtain a commutative diagram R1 R ... Rn ... R2 T1 T ... Tn ... T2 of k [[X1 , ..., Xr , S1 , ..., Sr ]]-algebras. Moreover we have that Rn Tn , k [[X1 , ..., Xr ]]
p p , ..., Sr )-module, Tn is a nite free k [[S1 , ..., Sr ]]/(S1
n n

(m)

21

 Rn /(S1 , ..., Sr ) R and Tn /(S1 , ..., Sr ) T . Let R denote the quotient of k [[X1 , ..., Xr ]] by the intersection of the ideals ker (k [[X1 , ..., Xr ]] Rn ) and let T denote the quotient of k [[X1 , ..., Xr ]] Tn ). Then we have a by the intersection of the ideals ker (k [[X1 , ..., Xr ]] commutative diagram k [[S1 , ..., Sr ]] R R k [[X1 , ..., Xr ]] T, T such that T , R T is a nite free k [[S1 , ..., Sr ]]-module, R /(S1 , ..., Sr ) R and T /(S1 , ..., Sr ) T . We deduce that T has Krull dimension r and hence we deduce that the T has trivial kernel. That is we have isomorphisms map k [[X1 , ..., Xr ]] k [[X1 , ..., Xr ]] R T . Thus R T . As T has Krull dimension 0 and T = k [[X1 , ..., Xr ]]/(S1 , ..., Sr ) we see that T is a complete intersection, and the proposition is proved.

22