Wiles 1995

Annals of Mathematics
Modular Elliptic Curves and Fermat's Last Theorem

Author(s): Andrew Wiles
Source: Annals of Mathematics, Second Series, Vol. 141, No. 3 (May, 1995), pp. 443-551
Published by: Annals of Mathematics
Stable URL: http://www.jstor.org/stable/2118559 .
Accessed: 25/09/2013 07:52
Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .
http://www.jstor.org/page/info/about/policies/terms.jsp
.
JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range of
content in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new forms
of scholarship. For more information about JSTOR, please contact support@jstor.org.
Annals of Mathematics is collaborating with JSTOR to digitize, preserve and extend access to Annals of
Mathematics.
http://www.jstor.org
This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

All use subject to JSTOR Terms and Conditions
Annals of Mathematics, 142 (1995), 443-551
Modular elliptic curves

and
Fermat's Last Theorem
By ANDREW WILES*
For Nada, Clare, Kate and Olivia
Cubumautemin duos cubos,aut quadratoquadratum in duos quadra-

toquadratos,et generaliternullam in infinitumultra quadratum
potestatemin duos ejusdem nominisfas est dividere: cujus rei
demonstrationem mirabilemsane detexi. Hanc marginisexiguitas
non caperet.
Pierre de Fermat
Introduction
An ellipticcurveoverQ is said to be modularifit has a finitecoveringby

a modularcurveof the formXo(N). Any such ellipticcurvehas the property
that its Hasse-Weilzeta functionhas an analyticcontinuationand satisfiesa
functionalequation of the standardtype. If an ellipticcurve over Q with a
givenj-invariantis modularthenit is easy to see that all ellipticcurveswith
the same j-invariantare modular (in whichcase we say that the j-invariant
is modular). A well-known conjecturewhichgrewout of the workof Shimura
and Taniyamain the 1950's and 1960's assertsthat everyellipticcurveoverQ
is modular. However,it onlybecamewidelyknownthroughits publicationin a
paper ofWeil in 1967 [We] (as an exerciseforthe interestedreader!),in which,
moreover,Weil gave conceptualevidenceforthe conjecture.Althoughit had
been numericallyverifiedin manycases, priorto the resultsdescribedin this
paper it had onlybeen knownthat finitely manyj-invariantsweremodular.
In 1985 Freymade the remarkableobservationthat thisconjectureshould
implyFermat'sLast Theorem. The precisemechanismrelatingthe two was
formulatedby Serreas the E-conjectureand this was thenprovedby Ribet in
the summerof 1986. Ribet's resultonly requiresone to provethe conjecture
forsemistableellipticcurvesin orderto deduce Fermat'sLast Theorem.
*The work on this paper was supported by an NSF grant.

444 ANDREW WILES
Our approachto the studyof ellipticcurvesis via theirassociated Galois

representations.Suppose that pp is the representation of Gal(Q/Q) on the
p-divisionpointsof an ellipticcurveoverQ, and suppose forthe momentthat
p3 is irreducible.The choiceof 3 is criticalbecause a crucialtheoremofLang-
lands and Tunnellshowsthat ifp3 is irreduciblethen it is also modular. We
thenproceedby showingthat underthe hypothesisthat p3 is semistableat 3,
togetherwith some milderrestrictions on the ramification of p3 at the other
primes,everysuitableliftingofp3 is modular. To do thiswe linkthe problem,
via some novelargumentsfromcommutativealgebra,to a class numberprob-
lem of a well-known type. This we thensolvewiththe help ofthe paper [TW].
This sufficesto provethe modularityof E as it is knownthat E is modularif
and onlyifthe associated 3-adic representation is modular.
The keydevelopmentin theproofis a newand surprising linkbetweentwo
strongbut distincttraditionsin numbertheory,therelationshipbetweenGalois
representations and modularformson the one hand and the interpretation of
special values of L-functionson the other. The formertraditionis of course
more recent. Followingthe originalresults of Eichler and Shimura in the
1950's and 1960's the othermain theoremswereprovedby Deligne,Serreand
Langlands in the period up to 1980. This includedthe constructionof Galois
representations associatedto modularforms,the refinements ofLanglandsand
Deligne (latercompletedby Carayol),and the crucialapplicationby Langlands
of base changemethodsto give converseresultsin weightone. Howeverwith
the exceptionofthe ratherspecial weightone case, includingthe extensionby
TunnellofLanglands' originaltheorem,therewas no progressin the direction
of associatingmodularformsto Galois representations.Fromthe mid 1980's
the main impetus to the fieldwas given by the conjecturesof Serre which
elaboratedon the -conjecturealludedto before.BesidestheworkofRibet and
otherson thisproblemwe drawon some ofthe morespecializeddevelopments
of the 1980's, notablythose of Hida and Mazur.
The second traditiongoes back to the famousanalyticclass numberfor-
mula of Dirichlet,but owes its modernrevivalto the conjectureof Birch and
Swinnerton-Dyer. In practicehowever,it is the ideas ofIwasawa in thisfieldon
whichwe attemptto draw,and whichto a largeextentwe have to replace. The
principlesof Galois cohomology,and in particularthe fundamentaltheorems
of Poitou and Tate, also play an importantrole here.
The restrictionthat p3 be irreducibleat 3 is bypassed by means of an
intriguingargumentwith familiesof elliptic curves which share a common
P5. Using this, we completethe proofthat all semistableellipticcurves are
modular. In particular,thisfinallyyieldsa proofofFermat'sLast Theorem.In
addition,this methodseemswell suitedto establishingthat all ellipticcurves
over Q are modularand to generalizationto othertotallyreal numberfields.
Now we presentour methodsand resultsin moredetail.

MODULAR ELLIPTIC CURVES AND FERMAT'S LAST THEOREM 445
Let f be an eigenformassociated to the congruencesubgroupr1(N) of

SL2(Z) of weightk > 2 and characterX. Thus if Tn is the Hecke operator
associatedto an integern thereis an algebraicintegerc(n, f) such that Tnf =
c(n, f)f foreach n. We let Kf be the numberfieldgeneratedover Q by the
{c(nr,f)} togetherwith the values of X and let Of be its ring of integers.
For any primeA of Of let Of,> be the completionof Of at A. The following
theoremis due to Eichlerand Shimura (fork = 2) and Deligne (fork > 2).
The analogousresultwhenk = 1 is a celebratedtheoremof Serreand Deligne
but is morenaturallystated in termsof complexrepresentations.The image
in that case is finiteand a converseis knownin manycases.
THEOREM 0.1. For each primep E Z and each primeA I p of Of there

is a continuousrepresentation
pf,\:Gal(QQ
A GL2(Of,X)
outsidetheprimesdividingNp and such thatforall primes
whichis unramified
q t Np,
(FRobq) = c(q, f),
tracepf,A detpf,A
(Frobq) = X(q)qk-l
We willbe concernedwithtryingto proveresultsin the oppositedirection,

that is to say, with establishingcriteriaunderwhicha A-adicrepresentation
arises in this way froma modular form. We have not foundany advantage
in assumingthat the representation is part of a compatiblesystemof A-adic
representations exceptthat the proofmaybe easierforsome A than forothers.
Assume
po: Gal(Q/Q) - GL2 (Fp)
is a continuousrepresentation with values in the algebraicclosureof a finite

fieldof characteristicp and that det po is odd. We say that po is modular
if po and pfA mod A are isomorphicover Fp for some f and A and some
embeddingof Of/A in Fp. Serrehas conjecturedthat everyirreduciblepo of
odd determinant is modular. Verylittleis knownabout thisconjectureexcept
whenthe image of po in PGL2(Fp) is dihedral,A4 or S4. In the dihedralcase
it is trueand due (essentially)to Hecke,and in the A4 and S4 cases it is again
true and due primarilyto Langlands,withone importantcase due to Tunnell
(see Theorem 5.1 fora statement). More preciselythese theoremsactually
associate a formof weightone to the correspondingcomplex representation
but the versionswe need are straightforward deductionsfromthe complex
case. Even in the reduciblecase not much is knownabout the problemin
the formwe have describedit, and in that case it should be observedthat
one must also choose the lattice carefullyas only the semisimplification of
PfA= PfAmodA is independentof the choiceof latticein KA.

446 ANDREW WILES
If 0 is the ringofintegersof a local field(containingQp) we will say that

p: Gal(Q/Q) - GL2(0) is a liftingof po if,fora specifiedembeddingof the
residuefieldof (9 in Fp, 1pand po are isomorphicover FP. Our point of view
will be to assume that po is modularand then to attemptto give conditions
underwhicha representation p liftingpo comes froma modular formin the
sense that p - pfa, over Kf,Aforsome f,A. We will restrictour attentionto
two cases:
(I) po is ordinary(at p) by whichwe mean that thereis a one-dimensional

subspace of P , stable undera decompositiongroupat p and such that
the action on the quotient space is unramifiedand distinctfromthe
action on the subspace.
(II) po is flat (at p), meaningthat as a representationof a decomposition

group at p, po is equivalentto one that arises froma finiteflatgroup
scheme over Zp, and detpo restrictedto an inertiagroup at p is the
cyclotomiccharacter.
We say similarlythat p is ordinary(at p) if,viewedas a representationto QP,

thereis a one-dimensional subspace of Q2 stable undera decompositiongroup
at p and such that the action on the quotientspace is unramified.
Let e: Gal(Q/Q) -* denote the cyclotomiccharacter. Conjectural
conversesto Theorem0.1 have been part of the folkloreformany years but
have hithertolacked any evidence. The criticalidea that one mightdispense
withcompatiblesystemswas alreadyobservedby Drinfeldin the functionfield
case [Dr]. The idea that one onlyneeds to make a geometricconditionon the
restrictionto the decompositiongroupat p was firstsuggestedby Fontaineand
Mazur. The following versionis a naturalextensionofSerre'sconjecturewhich
is convenientforstatingour resultsand is, in a slightlymodifiedform,the one
proposedby Fontaineand Mazur. (In the formstatedthisincorporatesSerre's
conjecture.We could insteadhave made the hypothesisthat po is modular.)
CONJECTURE. Suppose thatp: Gal(Q/Q) GL2(0) is an irreducible

-
liftingof po and thatp is unramified outsideof a finiteset of primes. There

are two cases:
(i) Assume thatpo is ordinary.Then ifp is ordinaryand detp = ek-1X for

some integerk > 2 and some X offiniteorder,p comesfroma modular
form.
(ii) Assume thatpo is flat and thatp is odd. Then if p restrictedto a de-
compositiongroupat p is equivalentto a representationon a p-divisible
group,again p comesfroma modularform.

In case (ii) it is not hard to see that if the formexists it has to be of

weight2; in (i) of course it would have weightk. One can of course enlarge
this conjecturein severalways,by weakeningthe conditionsin (i) and (ii), by
consideringothernumberfieldsin place of Q and by consideringgroupsother
than GL2.
We provetwo resultsconcerningthis conjecture. The firstincludesthe
hypothesisthat PO is modular. Here and forthe rest of the paper we will
assume that p is an odd prime.
THEOREM 0.2. Suppose thatPO is irreducibleand satisfieseither(I) or

(II) above. Suppose also thatPO is modularand that
irreducible
(i) PGis absolutely to Q
whenrestricted ( (-1)Tp).
(ii) If q - modp is ramifiedin PO then eitherPOIDq is reducibleover

the algebraicclosurewhereDq is a decomposition groupat q or polIq is
absolutelyirreduciblewhereIq is an inertiagroupat q.
p as in the conjecturedoes indeedcomefroma mod-

Then any representation
ular form.
The only conditionwhichreallyseems essentialto our methodis the re-

quirementthat PObe modular.
The mostinterestingcase at the momentis whenp = 3 and POcan be de-
finedoverF3. Then sincePGL2(F3) - S4 everysuchrepresentation is modular
by the theoremofLanglands and Tunnellmentionedabove. In particular,ev-
eryrepresentationintoGL2(Z3) whosereductionsatisfiesthe givenconditions
is modular. We deduce:
THEOREM 0.3. Suppose thatE is an ellipticcurve definedover Q and

thatPO is the Galois action on the 3-divisionpoints. Suppose thatE has the
followingproperties:
reductionat 3.
(i) E has goodor multiplicative
to Q (\-).
(ii) P0 is absolutelyirreduciblewhenrestricted
(iii) For anyq -1 mod3 eitherPOIDqis reducibleoverthealgebraicclosure

or POIIqis absolutelyirreducible.
Then E is modular.
We should pointout that whilethe propertiesof the zeta functionfollow

directlyfromTheorem0.2 the strongerversionthat E is coveredby Xo(N)

448 ANDREW WILES
requiresalso the isogenytheoremprovedby Faltings(and earlierby Serrewhen

E has nonintegralj-invariant,a case whichincludesthe semistablecurves).
We note that if E is modular then so is any twist of E, so we could relax
condition(i) somewhat.
The importantclass of semistablecurves,i.e., thosewithsquare-freecon-
ductor,satisfies(i) and (iii) but not necessarily(ii). If (ii) failsthenin factpo
is reducible.Rathersurprisingly, Theorem0.2 can oftenbe applied in thiscase
also by showingthat the representation on the 5-divisionpointsalso occursfor
anotherellipticcurvewhichTheorem0.3 has alreadyprovedmodular. Thus
Theorem0.2 is appliedthistimewithp = 5. This argument,whichis explained
in Chapter 5, is the only part of the paper whichreallyuses deformations of
the ellipticcurveratherthan deformations of the Galois representation.The
argumentworksmoregenerallythan in the semistablecase but in this setting
we obtain the followingtheorem:
THEOREM0.4. Suppose thatE is a semistableellipticcurvedefinedover

Q. Then E is modular.
More generalfamiliesof ellipticcurveswhichare modularare givenin Chap-

ter 5.
In 1986, stimulatedby an ingeniousidea of Frey [Fr],Serre conjectured
and Ribet proved(in [Ril]) a propertyofthe Galois representations
associated
to modularformswhichenabled Ribet to showthat Theorem0.4 implies'Fer-
mat's Last Theorem'. Frey'ssuggestion,in the notationof the followingtheo-
rem,was to showthat the (hypothetical)ellipticcurvey2 = X(X + UP) (X - VP)
could not be modular. Such ellipticcurveshad already been studied in [He]
but withoutthe connectionwithmodularforms.Serremade precisethe idea
ofFreyby proposinga conjectureon modularformswhichmeantthat the rep-
resentationon thep-divisionpointsofthisparticularellipticcurve,ifmodular,
would be associated to a formof conductor2. This, by a simple inspection,
could not exist. Serre's conjecturewas then provedby Ribet in the summer
of 1986. However,one still needed to knowthat the curvein questionwould
have to be modular,and this is accomplishedby Theorem0.4. We have then
(finally!):
THEOREM 0.5. Suppose thatuP+vP+wP = O withu,v,wE Q andp > 3,

then uvw = 0.
The second result we prove about the conjecturedoes not requirethe

assumptionthat po be modular (since it is alreadyknownin this case).

THEOREM 0.6. Suppose thatpo is irreducible

and satisfiesthehypotheses
of the conjecture,including(I) above. Supposefurtherthat
(i) po = IndQ Ko for a characterKo of an imaginaryquadraticextensionL

of Q whichis unramified at p.
(ii) det po I P = .
Then a representation
p as in the conjecturedoes indeedcomefroma modular
form.
This theoremcan also be used to prove that certainfamiliesof elliptic
curves are modular. In this summarywe have only describedthe principal
theoremsassociated to Galois representations and ellipticcurves. Our results
concerninggeneralizedclass groupsare describedin Theorem3.3.
The followingis an account of the originsof this workand of the more
specialized developmentsof the 1980's that affectedit. I began workingon
theseproblemsin the late summerof 1986 immediatelyon learningof Ribet's
result. For several years I had been workingon the Iwasawa conjecturefor
totallyreal fieldsand some applicationsof it. In the process,I had been using
and developingresultson ?-adicrepresentations associatedto Hilbertmodular
forms.It was therefore naturalforme to considerthe problemof modularity
fromthe pointofview of ?-adic representations. I began withthe assumption
that the reductionof a givenordinary?-adic representation was reducibleand
triedto proveunderthis hypothesisthat the representation itselfwould have
to be modular. I hoped rathernaivelythat in this situationI could apply the
techniquesof Iwasawa theory.Even moreoptimistically I hoped that the case
e = 2 would be tractableas thiswould sufficeforthe studyof the curvesused
by Frey. From now on and in the main text,we writep fore because of the
connectionswithIwasawa theory.
Afterseveralmonthsstudyingthe 2-adic representation, I made the first
realbreakthrough in realizingthatI could use the3-adicrepresentation instead:
the Langlands-Tunnell theoremmeantthat p3,the mod3 representation ofany
given ellipticcurve over Q, would necessarilybe modular. This enabled me
to try inductivelyto prove that the GL2 (Z/3n Z) representationwould be
modularforeach n. At thistimeI consideredonlythe ordinarycase. This led
quicklyto the studyof Hz(Gal(Fx/Q), Wf) fori = 1 and 2, whereFx is the
splittingfieldofthem-adictorsionon the Jacobianofa suitablemodularcurve,
mbeingthe maximalideal ofa Heckeringassociatedto p3 and Wf the module
associated to a modular formf describedin Chapter 1. More specifically, I
neededto comparethiscohomologywiththe cohomologyofGal(QE/Q) acting
on the same module.
I triedto apply some ideas fromIwasawa theoryto this problem. In my
solutionto the Iwasawa conjecturefortotallyreal fields[Wi4],I had introduced

450 ANDREW WILES
a new techniquein orderto deal withthe trivialzeroes. It involvedreplacing

the standardIwasawa theorymethodofconsideringthe fieldsin the cyclotomic
Zp-extensionby a similaranalysisbased on a choiceofinfinitely manydistinct
primes qi _ 1 modpfi with ni -* oo as i -* oo. Some aspects of this method
suggestedthat an alternativeto the standard techniqueof Iwasawa theory,
whichseemedproblematicin the studyof Wf, mightbe to make a comparison
betweenthe cohomologygroupsas E varies but with the fieldQ fixed. The
new principlesaid roughlythat the unramified cohomologyclasses are trapped
by the tamelyramifiedones. Afterreadingthe paper [Grel],I realizedthatthe
dualitytheoremsin Galois cohomologyof Poitou and Tate wouldbe usefulfor
this. The crucialextractfromthis lattertheoryis in Section2 of Chapter 1.
In orderto put these ideas into practiceI developedin a naive formthe
techniquesof the firsttwo sectionsof Chapter 2. This drew in particularon
a detailed study of all the congruencesbetweenf and othermodular forms
of differing levels,a theorythat had been initiatedby Hida and Ribet. The
outcomewas that I could estimatethe firstcohomologygroupwell undertwo
assumptions,firstthat a certainsubgroupof the second cohomologygroup
vanishedand second that the formf was chosen at the minimallevel form.
These assumptionsweremuchtoo restrictive to be reallyeffective
but at least
theypointedin the rightdirection.Some of these argumentsare to be found
in the secondsectionofChapter1 and some formthe firstweak approximation
to the argumentin Chapter3. At that time,however,I used auxiliaryprimes
q -1 modp whenvaryingE as the geometrictechniquesI workedwithdid
not apply in generalforprimesq _ 1 modp. (This was formuch the same
reason that the reductionof level argumentin [Ril] is much more difficult
when q -1 modp.) In all this workI used the moregeneralassumptionthat
pp was modularratherthan the assumptionthat p = 3.
In the late 1980's,I translatedtheseideas intoring-theoretic
language. A
fewyearspreviouslyHida had constructedsome explicitone-parameterfam-
ilies of Galois representations.In an attemptto understandthis,Mazur had
been developingthe languageofdeformations ofGalois representations. More-
over,Mazur realizedthat the universaldeformation ringshe foundshould be
givenby Hecke rings,at least in certainspecial cases. This criticalconjecture
refinedthe expectationthat all ordinaryliftingsof modular representations
should be modular. In makingthe translationto this ring-theoretic language
I realizedthat the vanishingassumptionon the subgroupof H2 whichI had
neededshouldbe replacedby the strongerconditionthat the Heckeringswere
completeintersections.This fittedwell with their being deformationrings
whereone could estimatethe numberofgeneratorsand relationsand so made
the originalassumptionmoreplausible.
To be of use, the deformation theoryrequiredsome development.Apart
fromsome special examples examinedby Boston and Mazur therehad been

littleworkon it. I checkedthatone could makethe appropriateadjustmentsto

thetheoryin orderto describedeformation theoriesat theminimallevel. In the
fallof 1989,I set Ramakrishna,thena studentof mineat Princeton,the task
of provingthe existenceof a deformation theoryassociated to representations
arisingfromfiniteflatgroup schemesover Zp. This was needed in orderto
removethe restriction to the ordinarycase. These developmentsare described
in the firstsectionof Chapter 1 althoughthe workof Ramakrishnawas not
completeduntil the fall of 1991. For a long time the ring-theoretic version
of the problem,althoughmore natural,did not look any simpler. The usual
methodsof Iwasawa theorywhen translatedinto the ring-theoretic language
seemedto requireunknownprinciplesofbase change. One neededto knowthe
exact relationsbetweenthe Hecke ringsfordifferent fieldsin the cyclotomic
Zp-extension Q,
of and not just the relations up to torsion.
The turningpoint in this and indeed in the whole proofcame in the
springof 1991. In searchingfora clue fromcommutativealgebra I had been
particularlystrucksome yearsearlierby a paper ofKunz [Ku2]. I had already
neededto verifythat the HeckeringswereGorensteinin orderto computethe
congruencesdevelopedin Chapter 2. This propertyhad firstbeen provedby
Mazur in the case of primelevel and his argumenthad alreadybeen extended
by other authors as the need arose. Kunz's paper suggestedthe use of an
invariant(the 77-invariantof the appendix) whichI saw could be used to test
forisomorphismsbetweenGorensteinrings. A different invariant(the p/p2_
invariantof the appendix) I had already observedcould be used to test for
isomorphisms betweencompleteintersections.It was onlyon readingSection6
of [Ti2] that I learned that it followedfromTate's account of Grothendieck
dualitytheoryforcompleteintersections that these two invariantswereequal
forsuchrings.Not longafterwards I realizedthat,unlikelythoughit seemedat
first,the equalityof these invariantswas actuallya criterionfora Gorenstein
ringto be a completeintersection.These argumentsare givenin the appendix.
The impactofthisresulton the main problemwas enormous.Firstly,the
relationshipbetweenthe Heckeringsand the deformation ringscould be tested
In I
just usingthesetwo invariants. particular could providethe inductivear-
gumentof Section 3 of Chapter 2 to show that if all liftingswith restricted
ramification are modularthenall liftingsare modular. This I had been trying
to do fora longtimebut withoutsuccess untilthe breakthrough in commuta-
tive algebra. Secondly,by means of a calculationof Hida summarizedin [Hi2]
the main problemcould be transformed into a problemabout class numbers
of a type well-knownin Iwasawa theory.In particular,I could checkthis in
the ordinaryCM case usingthe recenttheoremsofRubin and Kolyvagin.This
is the contentof Chapter 4. Thirdly,it meantthat forthe firsttime it could
be verifiedthat infinitelymanyj-invariantswere modular. Finally,it meant
that I could focuson the minimallevelwherethe estimatesgivenby myearlier

452 ANDREW WILES
Galois cohomologycalculationslooked morepromising.Here I was also using

the workof Ribet and otherson Serre's conjecture(the same workof Ribet
that had linkedFermat'sLast Theoremto modularformsin the firstplace) to
knowthat therewas a minimallevel.
The class numberproblemwas of a type well-knownin Iwasawa theory
and in the ordinarycase had alreadybeen conjecturedby Coates and Schmidt.
However,the traditionalmethodsof Iwasawa theorydid not seem quite suf-
ficientin this case and, as explained earlier,when translatedinto the ring-
theoreticlanguage seemed to requireunknownprinciplesof base change. So
instead I developed furtherthe idea of using auxiliaryprimesto replace the
change of fieldthat is used in Iwasawa theory.The Galois cohomologyesti-
mates describedin Chapter 3 werenow muchstronger,althoughat that time
I was still using primesq -1 modp forthe argument.The main difficulty
was that althoughI knew how the q7-invariant changed as one passed to an
auxiliarylevel fromthe resultsof Chapter 2, I did not knowhow to estimate
the change in the p/p2-invariant precisely.However,the methoddid give the
rightbound forthe generalisedclass group,or Selmergroupas it is oftencalled
in this context,underthe additionalassumptionthat the minimalHecke ring
was a completeintersection.
I had earlierrealizedthat ideallywhatI neededin thismethodofauxiliary
primeswas a replacementforthe powerseriesringconstructionone obtainsin
themorenaturalapproachbased on Iwasawa theory.In thismoreusual setting,
the projectivelimitof the Hecke ringsforthe varyingfieldsin a cyclotomic
towerwould be expected to be a power series ring,at least if one assumed
the vanishingofthe a-invariant.However,in the settingwithauxiliaryprimes
whereone wouldchangethe levelbut not the field,the naturallimitingprocess
did not appear to be helpful,withthe exceptionofthe closelyrelatedand very
importantconstruction ofHida [Hil]. This methodofHida oftengave one step
towardsa powerseriesringin the ordinarycase. Therewerealso tenuoushints
of a patchingargumentin Iwasawa theory([Scho], [Wi4, ?10]), but I searched
withoutsuccess forthe key.
Then, in August,1991, I learnedof a new constructionof Flach [Fl] and
quicklybecame convincedthat an extensionof his methodwas more plausi-
ble. Flach's approachseemed to be the firststep towardsthe constructionof
an Euler system,an approach whichwould give the preciseupper bound for
the size of the Selmergroup if it could be completed. By the fall of 1992, I
believed I had achievedthis and began then to considerthe remainingcase
wherethe mod3 representation was assumed reducible. For severalmonthsI
triedsimplyto repeat the methodsusing deformationringsand Hecke rings.
Then unexpectedlyin May 1993,on readingofa constructionoftwistedforms
of modularcurvesin a paper of Mazur [Ma3], I made a crucialand surprising
breakthrough:I foundthe argumentusing familiesof ellipticcurveswith a

commonp5 whichis given in Chapter 5. Believingnow that the proofwas

complete,I sketchedthe wholetheoryin threelecturesin Cambridge,England
on June21-23. However,it became clearto me in the fallof 1993 that the con-
structionof the Euler systemused to extendFlach's methodwas incomplete
and possiblyflawed.
Chapter 3 followsthe originalapproach I had taken to the problemof
boundingthe Selmergroup but had abandoned on learningof Flach's paper.
Darmonencouragedme in February,1994,to explainthe reductionto the com-
plete intersectionproperty,as it gave a quick way to exhibitinfinitefamilies
of modularj-invariants. In presentingit in a lectureat Princeton,I made,
almostunconsciously, a criticalswitchto the special primesused in Chapter3
as auxiliaryprimes.I had onlyobservedthe existenceand importanceofthese
primesin the fallof1992 whiletryingto extendFlach's work.Previously,I had
only used primesq -1 modp as auxiliaryprimes. In hindsightthis change
was crucialbecause of a developmentdue to de Shalit. As explainedbefore,I
had realizedearlierthat Hida's theoryoftenprovidedone step towardsa power
seriesringat least in the ordinarycase. At the Cambridgeconference de Shalit
had explainedto me that forprimesq -1 modp he had obtaineda versionof
Hida's results. But except forexplainingthe completeintersectionargument
in the lectureat Princeton,I still did not give any thoughtto my initialap-
proach,whichI had put aside since the summerof 1991, since I continuedto
believethat the Euler systemapproachwas the correctone.
Meanwhilein January,1994, R. Taylorhad joined me in the attemptto
repairthe Euler systemargument.Then in the springof 1994, frustratedin
the efforts to repairthe Euler systemargument,I began to workwithTaylor
on an attemptto devisea newargumentusingp = 2. The attemptto use p = 2
reachedan impasseat theend ofAugust. As Taylorwas stillnotconvincedthat
the Euler systemargumentwas irreparable,I decidedin Septemberto take one
last look at myattemptto generaliseFlach, ifonlyto formulatemoreprecisely
the obstruction.In doing this I came suddenlyto a marvelousrevelation:I
saw in a flashon September19th,1994,that de Shalit's theory,ifgeneralised,
could be used togetherwithdualityto glue theHeckeringsat suitableauxiliary
levelsintoa powerseriesring.I had unexpectedlyfoundthe missingkeyto my
old abandonedapproach. It was the old idea ofpickingqi's withqi- 1 modpfi
and ni -x oc as i -* oc that I used to achievethe limitingprocess. The switch
to the special primesof Chapter3 had made all this possible.
AfterI communicatedthe argumentto Taylor,we spentthe nextfewdays
makingsure of the details. The fullargument,togetherwiththe deductionof
the completeintersection property, is givenin [TW].
In conclusionthe keybreakthrough in the proofhad been the realization
in the springof 1991 that the two invariantsintroducedin the appendixcould
be used to relate the deformationringsand the Hecke rings. In effectthe 71-

454 ANDREW WILES
invariantcould be used to countGalois representations. The last step afterthe

June,1993, announcement,thoughelusive,was but the conclusionof a long
processwhosepurposewas to replace,in thering-theoretic setting,themethods
based on Iwasawa theoryby methodsbased on the use of auxiliaryprimes.
One improvement that I have not includedbut whichmightbe used to
simplifysome of Chapter2 is the observationof Lenstrathat the criterionfor
Gorensteinringsto be completeintersections can be extendedto moregeneral
ringswhich are finiteand freeas ZP-modules. Faltings has pointed out an
improvement, also not included,whichsimplifiesthe argumentin Chapter 3
and [TW]. This is howeverexplainedin the appendixto [TW].
It is a pleasureto thankthosewhoread carefullya firstdraftofsomeofthis
paper afterthe Cambridgeconferenceand particularlyN. Katz who patiently
answeredmany questions in the course of my work on Euler systems,and
togetherwithIllusieread criticallytheEuler systemargument.Theirquestions
led to mydiscoveryof the problemwithit. Katz also listenedcriticallyto my
firstattemptsto correctit in the fallof 1993. I am gratefulalso to Taylorfor
his assistancein analyzingin depththe Euler systemargument.I am indebted
to F. Diamond forhis generousassistancein thepreparationofthe finalversion
ofthispaper. In additionto his manyvaluable suggestions,severalothersalso
made helpfulcommentsand suggestionsespeciallyConrad,de Shalit,Faltings,
Ribet, Rubin, Skinnerand Taylor. Finally,I am most gratefulto H. Darmon
forhis encouragement to reconsidermyold argument.AlthoughI paid no heed
to his advice at the time,it surelyleftits mark.
Table of Contents
Chapter 1 1. Deformationsof Galois representations

2. Some computationsof cohomologygroups
3. Some resultson subgroupsof GL2(k)
Chapter 2 1. The Gorensteinproperty
2. CongruencesbetweenHecke rings
3. The main conjectures
Chapter 3 Estimatesforthe Selmergroup
Chapter4 1L The ordinaryCM case
2. Calculationof qj
Chapter 5 Applicationto ellipticcurves
Appendix
References

Chapter 1
This chapteris devoted to the study of certain Galois representations.

In the firstsectionwe introduceand study Mazur's deformationtheoryand
discuss various refinements of it. These refinementswill be needed later to
make precisethe correspondence betweenthe universaldeformation ringsand
the Hecke ringsin Chapter 2. The main resultsneeded are Proposition1.2
whichis used to interpretvariousgeneralizedcotangentspaces as Selmergroups
and (1.7) whichlaterwill be used to studythem. At the end ofthe sectionwe
relatethese Selmergroupsto ones used in the Bloch-Kato conjecture,but this
connectionis not needed forthe proofsof our main results.
In the second sectionwe extractfromthe resultsof Poitou and Tate on
Galois cohomologycertaingeneralrelationsbetweenSelmergroupsas E varies,
as well as betweenSelmergroupsand theirduals. The most importantobser-
vationof the thirdsectionis Lemma 1.10(i) whichguaranteesthe existenceof
the special primesused in Chapter3 and [TW].
1. Deformations of Galois representations
Let p be an odd prime. Let E be a finiteset of primesincludingp and

let Qs be the maximal extensionof Q unramifiedoutside this set and oc.
Throughoutwe fixan embeddingof Q, and so also of Qy, in C. We will also
fixa choiceof decompositiongroupDq forall primesq in Z. Suppose that k
is a finitefieldof characteristic
p and that
(1.1) Po: Gal(Q/Q) -* GL2(k)
is an irreduciblerepresentation. In contrastto the introductionwe willassume

in the rest of the paper that po comes with its fieldof definitionk. Suppose
furtherthat detpo is odd. In particularthis impliesthat the smallestfieldof
definition forpo is givenby the fieldkogeneratedby the tracesbut we will not
assume that k = ko. It also impliesthat po is absolutelyirreducible.We con-
siderthe deformations [p]to GL2(A) of po in the sense of Mazur [Mal]. Thus
ifW(k) is the ringofWitt vectorsof k, A is to be a completeNoetherianlocal
W(k)-algebrawithresiduefieldk and maximalideal m, and a deformation[p]
is just a strictequivalenceclass ofhomomorphisms p: Gal(Qr/Q) -* GL2(A)
such that p modm = po, two such homomorphisms beingcalled strictlyequiv-
alent if one can be broughtto the other by conjugationby an elementof
ker: GL2(A) -* GL2(k). We oftensimplywritep instead of [p] forthe
equivalenceclass.

456 ANDREW WILES
We will restrictour choiceof po furtherby assumingthat either:

of po to the decompositiongroupDp
(i) po is ordinary;viz., the restriction
has (fora suitablechoiceof basis) the form
(1.2) PO (? X *)
whereXi and X2 are homomorphisms fromDp to k* withX2 unramified.
Moreoverwe requirethat Xi #4X2. We do allow here that POIDP be
semisimple.(If Xi and X2 are both unramifiedand POIDp is semisimple
thenwe fixour choicesof Xi and X2 once and forall.)
(ii) po is flat at p but not ordinary(cf. [Sel] wherethe terminology finiteis

used); viz., POIDP is the representation associated to a finiteflatgroup
schemeoverZp but is notordinaryin thesenseof(i). (In generalwhenwe
referto the flatcase we will mean that po is assumed not to be ordinary
unless we specifyotherwise.) We will assume also that detpo0Iip= w
where Ip is an inertiagroup at p and w is the Teichmiillercharacter
givingthe action on pth rootsof unity.
In case (ii) it followsfromresultsof Raynaud that POIDP is absolutely
irreducibleand one can describeP0IIPexplicitly.Forextendinga Jordan-Holder
seriesforthe representation space (as an Ip-module)to one forfiniteflatgroup
schemes(cf. [Rayl]) we observefirstthatthetrivialcharacterdoes notoccuron
a subquotient,as otherwise(usingthe classification of Oort-Tateor Raynaud)
the group schemewould be ordinary.So we findby Raynaud's results,that
PoIIP0k k - b1? 02 where b1 and 02 are the two fundamentalcharactersof
degree 2 (cf. Corollary3.4.4 of [Rayl]). Since b1and 02 do not extend to
charactersof Gal(Qp/Qp), PoIDP mustbe absolutelyirreducible.
We will sometimeswish to make one of the followingrestrictions on the
deformations we allow:
(i) (a) Selmerdeformations.In thiscase we assumethat po is ordinary,
withno-
tation as above, and that the deformationhas a representative
p: Gal(Q/Q) -* GL2(A) withthe property that (fora suitablechoice
of basis)
with 52 unramified, X2 X2 mod m, and det PIip = Fw 1X1X2 where

e is the cyclotomiccharacter,E: Gal(Qr/Q) -* Zp, givingthe action
on all p-powerroots of unity,w is of orderprimeto p satisfyingw _ E
modp, and Xi and X2 are the charactersof (i) viewedas takingvalues in

(i) (b) Ordinarydeformations.The same as in (i) (a) but withno conditionon

the determinant.
(i) (c) Strictdeformations. This is a varianton (i) (a) whichwe onlyuse when
POIDP is not semisimpleand not flat (i.e. not associated to a finiteflat
group scheme). We also assume that XiX-1 = w in this case. Then a
strictdeformation is as in (i)(a) exceptthat we assume in additionthat
(X1/X2)IDP = E.
(ii) Flat (at p) deformations.
We assumethat each deformation p to GL2(A)
has the propertythat forany quotientA /a of finiteorderPIDP mod a
is the Galois representation
associated to the Qp-pointsof a finiteflat
groupschemeover Zp.
In each ofthesefourcases, as wellas in the unrestricted

case (in whichwe
imposeno local restriction
at p) one can verify
thatMazur's use ofSchlessinger's
criteria[Sch]provesthe existenceof a universaldeformation
p: Gal(Q/Q) -* GL2(R).
In the ordinaryand unrestrictedcase this was provedby Mazur and in the
flatcase by Ramakrishna[Ram]. The othercases requireminormodifications
of Mazur's argument. We denote the universalring RE in the unrestricted
case and Rs, Rrd, Rstr,Rf in the otherfourcases. We oftenomitthe E ifthe
contextmakesit clear.
There are certaingeneralizationsto all of the above whichwe will also
need. The firstis thatinsteadofconsideringW(k) -algebrasA we mayconsider
0-algebras for0 the ringof integersof any local fieldwithresiduefieldk. If
we need to recordwhich0 we are usingwe will writeRE,0 etc. It is easy to
see that the naturallocal map of local 0-algebras
RE,0- RE ?0 0
W(k)
is an isomorphismbecause forfunctorialreasonsthe map has a naturalsection

whichinducesan isomorphismon Zariskitangentspaces at closed points,and
one can then use Nakayama's lemma. Note, however,that if we change the
residue fieldvia i: k ,-* k' then we have a new deformationproblem associated
to the representationp0 = i o po. There is again a natural map of W(k')-
algebras
R(pO) -* R C W(k')
W(k)
whichis an isomorphismon Zariskitangentspaces. One can checkthat this
is again an isomorphism by consideringthe subringR1 of R(pO) definedas the
subringof all elementswhose reductionmodulo the maximal ideal lies in k.
Since R(pO) is a finiteRl-module,R1 is also a completelocal Noetherianring

458 ANDREW WILES
with residuefieldk. The universalrepresentationassociated to pO is defined

over R1 and the universalpropertyof R then definesa map R -* R1. So we
obtain a sectionto the map R(p') 0-* R 0 W(k') and the map is therefore
W(k)
an isomorphism.(I am gratefulto Faltingsforthis observation.)We will also
need to extendthe considerationof 0-algebras to the restrictedcases. In each
case we can requireA to be an 0-algebra and again it is easy to see that
R'_ E'0 RjE 0 O in each case.
W(k)
The secondgeneralizationconcernsprimesq #&p whichare ramifiedin PO.
We distinguishthreespecial cases (types (A) and (C) need not be disjoint):
(A) POIDq= ( *) fora suitablechoiceofbasis, withXi and X2 unramified,

Xi X2-1= w and the fixedspace of Iq of dimension1,
(B) POIIq= X7# 1, fora suitablechoiceof basis,
?),Xq
(C) H1(Qqi WA)= 0 whereWAis as definedin (1.6).
Then in each case we can definea suitabledeformation
theoryby imposing
additionalrestrictionson those we have alreadyconsidered,namely:
(A) PIDq = ( 1) fora suitable choice of basis of A2 with 41 and 02 un-

ramifiedand b10'41= E;
(B) PIIq = (Xq f)fora suitablechoiceof basis (Xqoforderprimeto p, so the

samecharacter
as above);
(C) detplIq = detpolIq,i.e., of orderprimeto p.
Thus if M is a set of primesin E distinctfromp and each satisfyingone of
(A), (B) or (C) forPo, we will impose the correspondingrestrictionat each
primein M.
Thus to each set of data D = {., E, 0, M} where is Se, str,ord, flator
unrestricted,we can associate a deformation theoryto Po provided
(1.3) po: Gal(Q/Q) -* GL2(k)
is itselfof typeD and 0 is the ringof integersof a totallyramifiedextension
of W(k); Po is ordinaryif . is Se or ord, strictif . is strictand flat if . is fl
(meaningflat); POis of type M, i.e., of type (A), (B) or (C) at each ramified
primeq #&p, q E M. We allow different types at different q's. We will refer
to these as the standarddeformation theoriesand writeRD forthe universal
ringassociated to D and p-Dforthe universaldeformation(or even p if D is
clear fromthe context).
We note here that if D = (ord, E, 0, M) and V' = (Se, E, 0, M) then
thereis a simplerelationbetweenRD and Rn'. Indeed thereis a naturalmap

RD-* R-D by the universalpropertyof RD, and its kernelis a principalideal

generatedby T = 6-1(-y)detpv(-y)- 1 where-yE Gal(Q/Q) is any element
to Gal(QOO/Q) is a generator(whereQO is the Zp-extension
whoserestriction
to Gal(Q((Np)/Q) is trivialforany N primeto p
of Q) and whoserestriction
with(N E Qs, C(Nbeing a primitiveNth root of 1:
(1.4) R-DIT -_R-D.

It turnsout that under the hypothesisthat po is strict,i.e. that POIDP
is not associated to a finiteflat group scheme,the deformationproblemsin
is alreadya strict
(i)(a) and (i)(c) are the same; i.e., everySelmerdeformation
deformation.This was observedby Diamond. The argumentis local, so the
decompositiongroupDp could be replacedby Gal(QP/Qp).
PROPOSITION 1.1 (Diamond). Suppose that7r:Dp-* GL2(A) is a con-

tinuousrepresentation whereA is an Artinianlocal ringwithresiduefieldk, a
p. Suppose ire (OX"F*) withXi and X2 unramified
finitefieldof characteristic
and X1 #4 X2. Then the residualrepresentation-ris associatedto a finiteflat
groupschemeoverZp.
Proof (taken from[Dia, Prop. 6.1]). We may replace 7rby 7r0 X2-1and
we let p = XX2-1. Then 7r-( t determinesa cocyclet: Dp -* M(1) where
1)
M is a freeA-moduleof rank one on whichDp acts via A. Let u denote the
cohomologyclass in H1 (Dp, M(1)) definedby t, and let uo denote its image
in H1(Dp,Mo(1)) whereMO = M/mM. Let G = kerr and let F be the fixed
extensionof Qp). Choose n so that pnA
fieldof G (so F is a finiteunramified
= 0. Since H2(G, ppr) -+ H2(G, lps) is infectivefor r < s, we see that the
natural map of A[Dp/G]-modules H1(G, 1upn)?zp M - H1(G,M(l)) is an
isomorphism. By Kummer theory,we have H1 (G, M(1)) Fx /(FX)Px 0zpM
as Dp-modules.Now considerthe commutativediagram
H1 (G, M(l))DP -+((Fxl(Fx )pn &Zp M)Dp ) MDP
H1(GI MoI()) (Fx/(Fx)P) IFpMO ) MO

wherethe right-handhorizontalmaps are inducedby vip Fx -* Z. If fp#1,
then MDp C mM, so that the elementresuo of H1(G, Mo(1)) is in the image
of (OFx/(OFx)P)?Fp Mo. But thismeansthat 7ris "peu ramifie&in the senseof
[Se] and thereforet comes froma finiteflatgroupscheme. (See [El, (8.2)].)
Remark. Diamond also observesthat essentiallythe same proofshows

that if 7r: Gal(Qq/Qq) -* GL2(A), whereA is a completelocal Noetherian

460 ANDREWWILES
ringwithresiduefieldk, has theformlrIIq - (0 *) with -rramified

then7ris
of type (A).
Globally,Proposition1.1 says that ifpo is strictand ifD = (Se, A, (9, M)

and 7' = (str, A, (9, M) then the natural map RD -* R-D is an isomorphism.
In each case the tangentspace of RD may be computedas in [Mal]. Let
A be a uniformizer for(9 and let UA, k2 be the representationspace forpo.
(The motivationforthe subscriptA willbecomeapparentlater.) Let VAbe the
representationspace of Gal(Qr/Q) on Adpo = Homk(UA,UA) --M2(k). Then
thereis an isomorphismof k-vectorspaces (cf. the proofof Prop. 1.2 below)
(1.5) Homk(mvD/(m2, A), k) Hv HD(Q/QVA)
whereHD,(Q/Q, VA)is a subspace of H1 (Q/Q, VA)whichwe now describe
and my is the maximal ideal of RD. It consistsof the cohomologyclasses
at p and at the primesin M. We call
whichsatisfycertainlocal restrictions
mv/(m%,2A) the reducedcotangentspace of RD.
We begin withp. First we may write (since p 5$2), as k[Gal(Qr/Q)]-
modules,
(1.6) VA=WA k, whereWA = {ff eHomk(UA,UA):tracef=0}
~ (Sym2? det-1)po
and k is the one-dimensionalsubspace of scalar multiplications.Then if po

is ordinarythe action of Dp on UAinduces a filtration
of UA and also on WA
and VA. Suppose we writethese 0 C UCOc UA, 0 c WA?C WA C WA1and
0 C VA?c VA C VA. Thus UA?is definedby the requirementthat Dp act on it
via the characterXi (cf. (1.2)) and on UA/UA?
via X2. For WAthe filtrations
are definedby
WA = {f E WA: f(U?) c UA},
WA = {EfWeWA:f=0onUA?},
and the filtrationsforVAare obtained by replacingW by V. We note that
these filtrations
are oftencharacterizedby the action of Dp. Thus the action
of Dp on WA?is via Xl/x2; on WA/WA? it is trivialand on WA/WAit is via
X2/xl. These determinethe filtrationifeitherX1/x2 is not quadraticor POIDP
is not semisimple.We definethe k-vectorspaces
VArd fE VA :f= in Hom(UA/UAUA/U?)},

Hse(Qp, VA) = ker{H1(Qp,VA,-), Hl(Qunr, V/W?)}
Hord(Qp, VA) = VA) Hl(Q nr VA Vord)}

ker{H1(Qp,
Hstr(Qp,
VA) = ker{H1(Qp,VA) Hl (QpWA/WA?)?DHl (Qunr k)}.

In the Selmercase we make an analogous definitionforHl,(QP, WA) by

replacingVAby WA,and similarlyin the strictcase. In the flatcase we use
the factthat thereis a naturalisomorphismof k-vectorspaces
H1(Qp, VA) -? Extk[D](UA UA)
wheretheextensionsare computedin thecategoryofk-vectorspaces withlocal

Galois action. Then Hi (Qp, VA) is definedas the k-subspaceof H1 (Qp, VA)
whichis the inverseimage of Ext1(G, G), the groupof extensionsin the cate-
goryoffiniteflatcommutativegroupschemesoverZp killedbyp, G beingthe
(unique) finiteflatgroupschemeoverZp associated to UA. By [Rayl] all such
extensionsin the inverseimageevencorrespondto k-vectorspace schemes.For
moredetails and calculationssee [Ram].
For q different fromp and q E M we have threecases (A), (B), (C). In
case (A) thereis a filtrationby Dq entirelyanalogous to the one forp. We
write this 0 c W q cWq c WA and we set
A AWanwee
Iker:
H1(Qq, VA)
H1(Qq WA/W' ) ? H1(Qunr k) in case (A)

Hhq(QqVA>)V=
ker: H1(Qq, VA)
-* Hl(Qunr, VA) in case (B) or (C).
Again we make an analogous definitionfor Hh (Qq, WA) by replacing VA

by WA and deletingthe last term in case (A). We now definethe k-vector
space Hi (QE/Q, VA) as
Hb(Q/Q, VA) = {a e H1(Q/Q, VA): aq E HDq(Qq, VA) forall q E M,
oapE H1 (Qp, VA)}

where* is Se, str,ord,flor unrestrictedaccordingto the typeof D. A similar
definitionapplies to Hi,(Qr/Q, WA) if* is Selmeror strict.
Now and forthe restof the sectionwe are goingto assume that po arises
fromthe reductionof the A-adicrepresentationassociated to an eigenform.
More preciselywe assume that thereis a normalizedeigenformf of weight2
and level N, divisibleonly by the primesin E, and that thereis a prime A
of Of such that po = Pf,AmodA. Here Of is the ringof integersof the field
generatedby the Fouriercoefficientsof f so the fieldsof definitionof the two
representations need not be the same. Howeverwe assume that k D (9f,A/A
and we fixsuch an embeddingso the comparisoncan be made over k. It will
be convenientmoreoverto assume that if we are consideringpo as being of
type D then D is definedusing 0-algebras where(9 D Of,A is an unramified
extensionwhose residuefieldis k. (Althoughthis conditionis unnecessary,it
is convenientto use A as the uniformizerfor0.) Finallywe assume that pf,A,

462 ANDREW WILES
itselfis of typeD. Again this is a slightabuse of terminology

as we are really
consideringthe extensionof scalars Pf,AX0 0 and not PfA itself,but we will
OfA\
do thiswithoutfurther mentionifthe contextmakesit clear. (The analysisof
this sectionactually applies to any characteristiczero liftingof po but in all
our applicationswe will be in the morerestrictivecontextwe have described
here.)
With these hypothesesthereis a unique local homomorphism RD -* (9
of 0-algebras whichtakes the universaldeformation to (the class of) pf,A.Let
pD = ker: RD -* 0. Let K be the fieldof fractionsof 0 and let Uf = (K/0)2
withthe Galois action takenfromPf,A>Similarly,let Vf = Ad Pif,A
(o K/O
(K/O)4 withthe adjoint representation
so that
Vf -- Wf ( K/0
where Wf has Galois action via Sym2pf,A0 detp7enand the action on the
second factoris trivial. Then if po is ordinarythe filtrationof Uf underthe
Ad p action of Dp induces one on Wf whichwe write0 C WC C W1 C Wf.
Oftento simplify the notationwe willdropthe indexf fromWe,Vfetc. There
is also a filtrationon WAn= {ker An: Wf ) Wf} given by Wn = WAnn WI
(compatiblewithour previousdescriptionforn = 1). Likewisewe writeVAn
for{ker An: Vf ) Vf}.
We now explain how to extend the definitionof HD, to give meaningto
H9,(Q/Q, VAn) and H-j,(Q/Q, V) and these are O/An and 0-modules, re-
spectively.In the case wherepo is ordinarythe definitions are the same with
VAnor V replacingVAand O/An or K/O replacingk. One checkseasily that
as 0-modules
(1.7) H-,(QE/Q, VAn) H-,(QE/QV)An1
where as usual the subscriptAn denotes the kernelof multiplicationby An.
This just uses the divisibilityof H0(Qr/Q, V) and HO(Qp, W/W0) in the
strictcase. In the Selmercase one checksthat form > n the kernelof
Hl(Qunr, VAn/WAn
) - Hl(Qunr VAm/WAm)
has onlythezeroelementfixedunderGal(Qunr/Qp)and the ordcase is similar.
Checkingconditionsat q E M is done withsimilararguments.In the Selmer
and strictcases we make analogous definitions with WAnin place of VAnand
W in place of V and the analogue of (1.7) still holds.
We now considerthe case wherepo is flat (but not ordinary).We claim
firstthat thereis a naturalmap of 0-modules
(1.8) H1(Qp, VAn) -* ](U'
Ext,[D m, UAn)
for each m > n where the extensionsare of 0-modules with local Galois
action. To describethis suppose that ar E H1 (Qp, VAn). Then we can asso-
ciate to ar a representation pa: Gal(Qp/Qp) -* GL2(0(n[e]) (where On[-E]=

(9[e]/(An7e, e2)) whichis an 0-algebra deformation

ofpo (see theproofofPropo-
sition 1.1 below). Let E = On0[E]2 wherethe Galois action is via p,. Then
thereis an exact sequence
0 -) E/Am -E/Am - (E/e)/Am ) 0
UAn UAM
and hence an extensionclass in Ext1(UAM, UAn). One checksnow that (1.8)
is a map of 0-modules. We defineHf (Qp,VAn) to be the inverseimage of
Ext1(UAn, UAn) under(1.8), i.e., thoseextensionswhichare alreadyextensions
in the categoryoffiniteflatgroupschemesZp. Observethat Ext' (UAn,UAn) n
ExtO[Dp] (UAn, UAn) is an 0-module, so Hf'(Qp,VAn) is seen to be an 0-sub-
moduleofH1 (Qp,VAn). We observethat our definition is equivalentto requir-
ing that the classes in Hf (Qp,VAn) map under(1.8) to Exti(UAm,UAn) forall
m > n. For ifem is the extensionclass in Ext1(UAMi,UAn) thenem - en E UAm
as Galois-modulesand we can apply resultsof [Rayl] to see that em comes
froma finiteflatgroupschemeover Zp ifen does.
In the flat(non-ordinary) case polipis determinedby Raynaud's resultsas
mentionedat the beginningof the chapter.It followsin particularthat, since
POID is absolutelyirreducible,V(Qp) = Ho (Qp, V) is divisiblein this case
(in factV(Qp) i K/0). Thus H1 (Qp, VAn) _-H1 (Qp, V)An and hencewe can
define 00
Hf (QpV) = U Hf (Qp, VAn)

n_1
and we claimthatHf (Qp,V)An - Hf1(Qp,VAn). To see thiswe have to compare
form > n,
representations
pn,m: Gal(Qp/Qp) -?GL2(On [e]/Am)
f m,n
pm,m: Gal(Qp/Qp) )GL2(Om[e]/Am)

where Pn,m and Pm,mare obtained froman e H1(Qp, VAn) and im(an) E
H'(Qp, yAm) and
V-mn: a+be a+Am2be. By [Ram,Prop 1.1 and Lemma
2.1] ifPn,mcomesfroma finiteflatgroupschemethenso does Pm,m.Conversely
SPm,n is injectiveand so Pn,mcomesfroma finiteflatgroupschemeifPm,mdoes;
cf. [Rayl]. The definitions of HD,(Q/Q, VAn)and H-D(QE/Q, V) now extend
to the flatcase and we note that (1.7) is also valid in the flatcase.
Still in the flat (non-ordinary)case we can again use the determination
of PoI' to see that H1(Qp, V) is divisible. For it is enough to check that
H2(Qp, VA)= 0 and thisfollowsby dualityfromthe factthat H?(Qp, V*) = 0

464 ANDREW WILES
whereV* = Hom(VX,tip) and lp is the group of pth roots of unity. (Again

this followsfromthe explicitformof P?jD ) Much subtleris the fact that
Hf (Qp, V) is divisible. This resultis essentiallydue to Ramakrishna. For,
usinga local versionof Proposition1.1 below we have that
Homrn(PR/Pt, K/O) - Hf (Qp, V)
whereR is the universallocal flatdeformationringforPOIDPand 0-algebras.

(This existsby Theorem1.1 of [Ram] because PoIDp is absolutelyirreducible.)
Since R - Rfl (9 whereRfl is the correspondingringforW(k)-algebras
W(k)
the main theoremof [Ram, Th. 4.2] shows that R is a powerseriesringand
the divisibilityof Hf (Qp, V) thenfollows.We referto [Ram] formoredetails
about RM.
Nextwe need an analogueof(1.5) forV. Againthisis a variantofstandard
resultsin deformation theoryand is given (at least forD = (ord,E, W(k), q)
withsome restriction on X1,X2 in i(a)) in [MT, Prop 25].
PROPOSITION Suppose that pf,A,is a deformationof po of type

1.2.
D = (-,E,O,M) with0 an unramified
extensionof(Of,A,.Then as O-modules
- H.D(QF/QV)
Home)(pDlp2,KIO)
Remark. The isomorphismis functorialin an obviousway ifone changes

D to a largerD'.
Proof. We will just describethe Selmer case with M =4 as the other

cases use similararguments.Suppose that a is a cocyclewhichrepresentsa
cohomologyclass in HSe(Q,/Q, VAn). Let O4e] denotethering0[e]/(AnF, e2).
We can associate to a a representation
p,: Gal(Q/Q) -+ GL2(O0n[E])
as follows:set p, (g) = a(g)pf,A(g)wherepf,A(g),a prioriin GL2 (0), is viewed

in GL2(On [E]) via the natural mapping0 -e On [e]. Here a basis for 02
is chosen so that the representation pf,> on the decompositiongroup Dp C
Gal(QE/Q) has the upper triangularformof (i)(a), and then a(g) E V,\n is
viewedin GL2(046[e])by identifying
V,\n
_{( ZE X t = {ker: GL2 (On[e]) - GL2(O)}.
Then
=? { (1 XE )}

W.n {= ( jey6)}
=n { (1 e
6yZ) : }
and
VXn = tE )
One checksreadilythat pa is a continuoushomomorphism

and that the defor-
mation [p,>]is unchangedifwe add a coboundaryto a.
We need to check that [p,>] is a Selmer deformation. Let 7Y =
Gal(Qp/Qunr)and g = Gal(Qunr/Qp). Considerthe exact sequence of (9[g]-
modules
0 -) (V- 1 ) (VAn/Wn)o -
X - 0
whereX is a submoduleof (VAn/Vln)7H. Since the action of Dp on VXn/V~nis
via a characterwhichis nontrivialmodA (it equals X2X11modA and X1 # X2),
we see that Xg = 0 and H1(g, X) = 0. Then we have an exact diagramof
0-modules
H'(91 (V1
'I
0 "LI)1) - H'(9, (V~n/WA~)t
H1 (Qp, VAn/Wn)
H1(Qpunr,
Vn/Xn )
By hypothesisthe image of a is zero in H1(Qq'TnrhVnf/jW>). Hence it

is in the image of H1(g, (VXn/Wo~)7H). Thus we can assume that it is rep-
resentedin H1(QpV~n/W>?) by a cocycle,whichmaps g to Vn/W~n; i.e.,
f(Dp) C VAn/WAn f(Ip) = 0. The difference betweenf and the image of a is
a coboundary{la i-- - i} forsome u e VAn. By subtractingthe coboundary
{f F-* oU - u} froma globallywe get a new a such that a = f as cocycles
mappingg to VXn/W Thus a(Dp) C V1,n,oa(Ip) C W\nand it is now easy
to checkthat [p,] is a Selmerdeformation of po.
Since [p,] is a Selmer deformationthere is a unique map of local (9-
algebras is,>: Rz -* On [E] inducingit. (If M $ q we must check the

466 ANDREW WILES
otherconditionsalso.) Since p =_pf,Amod E we see that restricting

ip, to PD
givesa homomorphism of 0-modules,
n
mu,:PD -- FO/A
such that ('' (P2) - 0. Thus we have defineda map p: a - ,,

y(p:H~e (Qs/Q, VAn) - Hom e(PD/PD,(9/Afl)
It is straightforward to checkthat this is a map of 0-modules. To checkthe

injectivityof(p supposethat .p,(pD) == 0. Then , factorsthroughR/p-D - (9
and beingan 0-algebra homomorphism thisdetermines ,>. Thus [pfA] = [paC].
If A-'pA = pf,A,thenA mode is seen to be centralby Schur'slemma and so
maybe takento be I. A simplecalculationnow showsthat a is a coboundary.
To see that p is surjectivechoose
' E Homo(pP)/p2,0/An).
Then pT: Gal(QE/Q) -- GL2(Rv/(p2, kerI)) is inducedby a representative

of the universaldeformation(chosento equal pf,Awhenreducedmod PD) and
{
we definea map aog: Gal(Qr/Q) -+ VAn by
1 + pD/(p2, ker I) pD/(p2,ker I) 1

() PT (g) pf, (g)1 E VAn
pv/(p2, ker I) 1 + pD/(p2,ker I) J
wherepfA(g) is viewed in GL2(RD/(pD,kerI)) via the structuralmap 0
RD (RD being an (9-algebraand the structuralmap being local because of
the existenceof a section). The right-handinclusioncomes from
PD/ PD,ker T) )/A n ((D/AnE

1 1 ) E.
Then ca is readilyseen to be a continuouscocyclewhose cohomologyclass

lies in HSe(QE/Ql VA\n). Finally (p(aT) = T. Moreover,the constructionsare
compatiblewithchangeofn, i.e., forVA\n ' + V,+i and A: /D 0/An+l. E
We now relatethe local cohomologygroupswe have definedto the theory
of Fontaineand in particularto the groupsof Bloch-Kato [BK]. We will dis-
tinguishtheseby writingHF forthe cohomologygroupsof Bloch-Kato. None
of the resultsdescribedin the rest of this sectionare used in the rest of the
paper. They serveonlyto relatethe Selmergroupswe have defined(and later
compute)to the morestandardversions.Usingthelatticeassociatedto pf,Awe
obtain also a latticeT - (94 withGalois actionvia Adpf,A. Let V = T Oz QP
be the associated vectorspace and identifyV with V/T. Let pr: V -* V be

the naturalprojectionand definecohomologymodulesby
V) = ker: H1(QpV) 0 Bcrys),

Hk(Qp,
F -+ H1(Qp, V~~~~~~~~~QP
Hk(Qp, V) = pr %H4(Qp, V)) C H1(Qp, V),
HF(Qp, V~n) = (jn) (HF(Qp, V)) C H1(Qp, Vsn),

wherejn: VXn- V is the natural map and the two groupsin the definition
of HF(Qp, V) are definedusingcontinuouscochains. Similardefinitions apply
to V* = HomQP(V, Qp(1)) and indeed to any finite-dimensional continuous
p-adic representationspace. The reader is cautioned that the definitionof
HF(Qp, VAn) is dependenton the lattice T (or equivalentlyon V). Under
certainconditionsBloch and Kato show, using the theoryof Fontaine and
Lafaille, that this is independentof the lattice (see [BK, Lemmas 4.4 and
4.5]). In any case we will considerin what followsa fixedlatticeassociated to
p = PfA, Ad p, etc. Henceforth we willonlyuse the notationH1(QP, -) when
the underlying vectorspace is crystalline.
PROPOSITION 1.3. (i) If po is flat butnot ordinaryand pf,Ais associated
to a p-divisiblegroupthenfor all n
Hf (QpV>,n) = HF (QpVAn)).
(ii) If pf,Ais ordinary,
detpfA |I = Eand pf,Ais associatedto a p-divisible
group,thenfor all n,
HF (QpV,\n) C HSe(QpVn).
Proof. Beginningwith (i), we defineHf(QpV) = {E e H1(QpV)

s(a/An) E Hf (Qp, V) for all n} where K : H1(Qp, V) H1 (Qp, V). Then
we see that in case (i), Hf (Qp,V) is divisible.So it is enoughto show that
Hk(QP, V) = Hfl(Qp IV).
We have to comparetwo constructionsassociated to a nonzeroelementa of
H1 (Qp, V). The firstis to associate an extension
(1.9) 0 -- V -- E -6 K -O 0
of K-vectorspaces with commutingcontinuousGalois action. If we fixan e
with 8(e) = 1 the action on e is definedby oe = e + &(of) with & a cocycle
representinga. The secondconstruction beginswiththe imageofthe subspace
(a) in H1 (Qp, V). By the analogue of Proposition1.2 in the local case, there
is an 0-module isomorphism
H1 (QP, V) c Homo (PR/p2,K/()

468 ANDREW WILES
where R is the universaldeformationring of po viewed as a representation

of Gal(Qp/Qp) on 0-algebras and PR is the ideal of R correspondingto pD
(i.e., its inverseimage in R). Since a $&0, associated to (a) is a quotient
PR/(PR, a) of PR/pi whichis a free0-module of rank one. We then obtain a
homomorphism
Pa: Gal(Qp/Qp) -) GL2 (R/(iR, a))
inducedfromthe universaldeformation(we pick a representation in the uni-

versalclass). This is associatedto an 0-module ofrank4 whichtensoredwith
K gives a K-vectorspace E' (K)4 whichis an extension
(1.10) 0 -? U -E -? U -? 0
whereU - K2 has the Galois representation pf,A(viewedlocally).
In the firstconstruction a E HF(QP, V) ifand onlyiftheextension(1.9) is
crystalline,as the extensiongivenin (1.9) is a sum of copies of the moreusual
extensionwhereQp replacesK in (1.9). On the otherhand (a) C Hf (Qp,, V) if
and onlyifthe secondconstruction can be made throughRfl,or equivalentlyif
and onlyifE' is the representation associatedto a p-divisiblegroup. (A priori,
the representationassociated to p, only has the propertythat on all finite
quotientsit comes froma finiteflat group scheme. Howevera theoremof
Raynaud [Rayl] says that then p, comes froma p-divisiblegroup. For more
details on Rfl,the universalflatdeformationringof the local representation
po, see [Ram].) Now the extensionE' comes froma p-divisiblegroup if and
onlyifit is crystalline;cf. [Fo, ?6]. So we have to showthat (1.9) is crystalline
ifand onlyif (1.10) is crystalline.
One obtains (1.10) from(1.9) as follows.We view V as HomK(U, U) and
let
X = ker:{HomK(UU ) OU - U}
wherethe map is the naturalone f 0 w | - f(w). (All tensorproductsin this

proofwill be as K-vectorspaces.) Then as K[Dp]-modules
E' (E ? U)/X.
To checkthis,one calculatesexplicitlywiththe definitionof the action on E

(givenabove on e) and on E' (givenin the proofofProposition1.1). It follows
fromstandardpropertiesof crystallinerepresentations that ifE is crystalline,
so is E 0 U and also E'. Conversely,we can recover E fromE' as follows.
Consider E' 0 U - (E 0 U 0 U)/(X 0 U). Then there is a natural map
up: E 0 (det) -- E' 0 U inducedby the directsum decompositionU 0 U
(det) EDSym2U. Here det denotes a 1-dimensionalvectorspace over K with
Galois action via detpf,A,.Now we claim that p is injectiveon V 0 (det). For

if f E V then p(f) = f 0 (Wi 0 w2 - W2Owl) where W1,W2 are a basis forU

forwhichW1AW2= 1 in det K. So if p(f) E X 0 U then
f(Wl)08)w2-f(w2)08)W1 =0 in U X U.
But this is false unless f(wi) = f(w2) = 0 whence f = 0. So o is injective

on V 0 det and if 'p itselfwerenot injectivethen E would split contradicting
a 7&0. So 'p is injectiveand we have exhibitedE0 (det) as a subrepresentation
of E' 0 U which is crystalline. We deduce that E is crystalline if E' is. This
completesthe proofof (i).
To prove (ii) we checkfirstthat HSe(QP,(H7e(QP,
VAn) jn1 V)) (this -
was alreadyused in (1.7)). We nexthaveto showthatH (Qp, V) C H~e(QP, V)

wherethe latteris definedby
HSe(QP ,V) = ker: H1 (Qp IV) -) H1 (Qunr,V/VO)
with V0 the subspace of V on whichIp acts via E. But this followsfromthe

computationsin Corollary3.8.4 of [BK]. Finallywe observethat
pr (HSe(QP V)) C HSe(QP 7 V)

althoughthe inclusionmay be strict,and
pr (Hk(QP, V)) = Hk(QP, V)
by definition.This completesthe proof. C1
These groupshave the propertythat fors > r,
(1.11)
HF7 (
H1(Qp, VVr s)) = HF(Qp, VAr)
whereir,s: -) Vs is the natural injection. The same holds forV* and

V,\r
V* in place of V\r and V1\AwhereV)*ris definedby
V*r = Hom(V\r q lpr)
and similarly for V*. Both results are immediate from the definition (and
indeedwerepart of the motivationforthe definition).
We also givea finitelevelversionof a resultofBloch-Katowhichis easily
deduced fromthe vectorspace version.As beforelet T C V be a Galois stable
lattice so that T - 04. Define
HI (Q T) = i (H
1 (Q7 V))
under the natural inclusion i: T --+ V, and likewise for the dual lattice T* =
Homzp(V,(Qp/Zp)(1)) in V*. (Here V* = Hom(V, Qp(l)); throughoutthis
paper we use M* to denote a dual of M with a Cartier twist.) Also write

470 ANDREW WILES
prn: T -- T/A' for the natural projectionmap, and for the mapping it
induceson cohomology.
PROPOSITION 1.4. is associatedto a p-divisiblegroup(the ordi-

If pf,A,
nary case is allowed) then
(i) prr (H1 (Qp, T)) -

HF1(Qp, T/An) and similarlyforT*, T*/An.
(ii) HF(Qp, VAn)is the orthogonalcomplementof HF(Qp, VAhn) under Tate

H1 (Qp, VAn)and H1 (Qp, VAin)
local dualitybetween and similarlyforWAn
and WA*nreplacing Vxn and VA*n.
More generallytheseresultsholdfor any crystallinerepresentation

V' in
in K' whereK' is anyfiniteextensionof Qp
place of V and A' a uniformizer
withK' C EndGal(V/Q)V.
Proof. We first observe that prn (HF(Qp, T)) C HF(Qp, T/An). Now
fromthe constructionwe may identifyT/An with VAn. A result of Bloch-
Kato ([BK, Prop. 3.8]) says that HF (Qp, V) and HF(Qp, V*) are orthogonal
complementsunderTate local duality. It followsformallythat HF(QP, V*n)
and prn(HFk(Qp,T)) are orthogonalcomplements, so to provethe proposition
it is enoughto showthat
(1.12) # HF(Qp, VA'n) # HFp(Qp, VAn)= #H1(Qp VAn).
Now if r = dimK Hk(Qp, V) and s = dimK HF(Qp, V*) then
(1.13) r + s = dimK HO(Qp, V) + dimKHO(Qp, V*) + dimK V.
Fromthe definition,
(1.14) # HF (Qpa VAn)=# (Q/)Any) #ker{H1 (Qp, VAn) >H1 (Qp, V)}.
The secondfactoris equal to # {V (Qp)/AnV (Qp)}. Whenwe writeV (Qp)div

forthe maximaldivisiblesubgroupof V (Qp) this is the same as
# (V(Q )/V(Q )div)/An = # (V(QP)/V(Qp)diV)An

= # V(Qp)An/# (V(Qp)div )An.
Combiningthis with (1.14) gives
(1.15) # HF(Qp, VAn) = # (O/Afn)y
# H0(Q VAn) / # (Q/An)dimKHO(Qp, V)
This, togetherwithan analogousformulafor# HF(Qp, Vi*n)and (1.13), gives
# H (QP, VAn)# H4(Q Van) = # (O/A )4.#H0(Qp, # H0(Qp, VAn).

VAin)

As #Ho (Qp, V*) = # H2 (Qp, VAn)the assertionof (1.12) now followsfrom

the formulaforthe Euler characteristic
of Vain.
The proofforWarn,or indeedmoregenerallyforany crystallinerepresen-
tation,is the same. R
We also give a characterizationof the orthogonal complementsof

Hie(Qp, WAn)and H"e(QpS I/n), underTate's local duality. We writethese
duals as H e*(Qp Wan) and H e* (Qp, Vn) respectively.Let
:p H1 (Qp W~n) H'(Qp, W~n/(W~n) )
be the natural map where (W~n)i is the orthogonal complement of W%-T in

WAn,and let Xni be definedas the image underthe compositemap
i = im: Z /(ZX)pn (- O/An H1(Q tpn (9/An)
H (Qp7 WA~n/(WA~n)
)
where in the middle term tLpn0 O/Anis to be identifiedwith (W~n)1/(W~n)0.

Similarlyifwe replace W*n by Vn we let be the image Ofo ZX/(ZX)Pn0
(O/An)2 in H1 (Qp, V*'n/(W~n)0), and we replace ,owby the analogous map (pv.
PROPOSITION 1.5.
HA = -1 (Xni)
*(QpWX*n)
Hse*(QpV>*n) = f01 (yni)
Proof. This can be checkedby dualizingthe sequence
o Hstr(Qp, WArn)-4 HSe(QP7 W~n)

ker: {H1(Qp, W\n/(W\n) ) - H(Qunr, WAn/(WAn)},
where Hltr(Qp, WAn) = ker: H1(Qp, H1(Qp, WAn/(W~n)0) The

W\n) --
firstterm is orthogonal to ker: H1 (Qp, Wn) H1 (Qp, Win /(W*n)1). By the
-
naturalityof the cup productpairing withrespectto quotientsand subgroups

the claim then reduces to the well knownfact that under the cup product
pairing
H1(Qp, ipn) x H1(Qp, Z/pn) _+Z/pn
the orthogonalcomplementofthe unramifiedhomomorphisms is the image of

the units ZpX/(ZX)pn H1 (Qp, Ippn). The proof for VAnis essentially the
4
same. 0

472 ANDREW WILES
2. Some computations of cohomology groups
We now make some comparisonsof ordersof cohomologygroups using

the theoremsof Poitou and Tate. We retainthe notationand conventionsof
Section 1 thoughit will be convenientto state the firsttwo propositionsin a
moregeneralcontext.Suppose that
L =JJLq C I7 H1(QqX)
qes
is a subgroup,whereX is a finitemodule forGal(Qr/Q) of p-powerorder.
We defineL* to be the orthogonalcomplementof L underthe perfectpairing
(local Tate duality)
7 H1(QqX) X 7 Hl(QqX*) QP/ZP
qEF2 qEF2
where X* = Hom(X, pp.). Let
Ax: H1(Q/QX) -- JJ
H1(QqX)
qEF,
be the localizationmap and similarlyAx* forX*. Then we set
HL(Q/QX) = A-1(L), HL*(Q/QX*) =AX1*(L*).
The followingresultwas suggestedby a resultof Greenberg(cf. [Grel]) and
is a simpleconsequenceof the theoremsof Poitou and Tate. Recall that p is
alwaysassumedodd and that p E E.
PROPOSITION 1.6.
#HL(Q/QX) /#HL*(QE/QX*) = hoo fJhq

qEF
where
hq = #HO(Qq,X*)/[Hl(Qq,X):Lq]
{ho= #H0(R, X*) #HO(QX)/#HO(QX*).
Proof. Adaptingtheexact sequenceofPoitou and Tate (cf. [Mi2,Th. 4.20])
we get a seventermexact sequence
0 HLj(Q/QX) , H1(Q/QX) Hj H1(QqX)/Lq
qEF,
H H2(Qq X) H2(Q/QX) HL*(QE/QX*))

qEF
L+ HO(Q/Q, X*)A - 0,

whereMA = Hom(M, Qp/Zp).Now usinglocal dualityand global Euler char-

acteristics(cf. [Mi2,Cor. 2.3 and Th. 5.1]) we easilyobtain the formulain the
proposition.We repeat that in the above propositionX can be arbitraryof
p-powerorder. []
We wishto applythe propositionto investigateHDi. Let D = ( I, , (9, M)

be a standarddeformationtheoryas in Section 1 and definea corresponding
groupLn = LDn by setting
H1(Qq, Vn) forq p and qf M
Ln~q= HDq(Qq, van) for q 5 p and q EM
H.1(Qp , 1Vn) for q = p.
Then HID(QE/Q, Vain)= HLn (Q /Q,Vain)and we also define

HD*(Q/Q = HL.(QE/Q,
V*,Vn) V*)
We will adopt the conventionimplicitin the above that ifwe considerA' D E
thenHE,(QE//Q, VAn)places no local restrictionon the cohomologyclasses at
primesq E A'-S. Thus in HD* (Q//Q,V\n)we willrequire(by duality)that
the cohomologyclass be locallytrivialat q E -E.
We need now some estimatesforthe local cohomologygroups. First we
consideran arbitraryfiniteGal(QE/Q)-module X:
PROPOSITION 1.7. finiteGal(QE/Q)-

If q V E, and X is an arbitrary
moduleofp-powerorder,
#HL,(QEuq/QX)/#HL(Q/QX) ? #H0(QqX*)
whereL' = Le for? E E and L- = H1 (Qq ,X).
Proof. Considerthe shortexact sequenceof inflation-restriction:
0 - -- Hom(Ga1(QEUq/QE), X)Gal(QF/Q)
HHL(QE/QX) HLj,(QEuq/QX)
H1(Qunr, X)Ga1(Q nr/Qq) H1(Qunr

X)Ga1(Q nr/Qq)
The propositionfollowswhenwe note that
#H0(Qq, X*) = #H1 (Qunr, X)Gal(Qunr/
Now we returnto the studyof Vain

and Wan.
PROPOSITION 1.8. If q E M (q 7&p) and X = VAin

then hq = 1.

474 ANDREW WILES
calculation. For example if q is of type

Proof. This is a straightforward
(A) thenwe have
Lnq = ker{H1 (Qq, Vain) - H1 (Qq, W\n/W2?n)E H1(Qnr, O/A>n)}.
Using the long exact sequenceof cohomologyassociated to

0> Wa n - W rn-> W,) 0/Wn O
one obtains a formulafor the order of Lnq in terms of #Ht(Qq, WAn),

#Hi (QqI WAn /WAC ) etc. Using local Euler characteristics
these are easily re-
duced to ones involvingHO(Qq, WAgn) etc. and the resultfollowseasily. M
The calculation of hp is more delicate. We contentourselveswith an

inequalityin some cases.
PROPOSITION 1.9. (i) If X = VAnthen
#Ho (QV)*n)
hpho = # (O/A)3n#Ho (QpVA*n)/
in the unrestricted
case.
(ii) If X = VAXn
then
hpho < # (0/A)n # H0(Qp, (Vord)*)/#Ho(Q,W;n)

in the ordinarycase.
(iii) If X = VAnor WArnthen hp hoo < #IHI(QpI (W n)*)/#Ho(Q Wan)
in the Selmer case.
(iv) If X = VAnor WAn then hp hco = 1 in the strict case.
(v) If X = V,\nthenhpho, = 1 in theflat case.
(vi) If X = VAn or WArnthen hp hoo = / #Ho (Q, Vv*\n)
if Ln -
HF(Qp,X) and pf,Aarises froman ordinaryp-divisiblegroup.

Proof. Case (i) is trivial.Considerthencase (iii) withX = V n. We have
a long exact sequenceof cohomologyassociated to the exact sequence:
(1.16) 0 -> Wn - Vxn -> V,\n/WOn O-.
In particularthis givesthe map u in the diagram

H 1(Qp, VAn)
\6
1 -) Z=Hl(Qpnr/Qp, (VAn/W~n)H) ) H1(Qp, VXn/WXn)~~Hl(Qpnr, VXn/WXn)g1~
where g = Gal(Qpnr/Qp),7-= Gal(Qp/Qunr) and 6 is definedto make the

triangle commute. Then writing hi(M) for #HI(Qp, M) we have that #Z =

ho(Vn/W0) and #im6 > (#imu)/(#Z). A simple calculation using the

long exact sequenceassociated to (1.16) gives
(1.17) #imu = h1(VW? )h2(V n)

h2(WA~h2(V,\n/IW5~)
Hence
Ln,p]= #im6 ? #(O/A)3nho(V*,\n)/ho(WXn*).

[H1(Qp,VAn):
The inequalityin (iii) followsforX = VAnand the case X = WAnis similar.
Case (ii) is similar. In case (iv) we just need # imu whichis givenby (1.17)
withWAnreplacingV\n.In case (v) we have alreadyobservedin Section1 that
Raynaud's resultsimplythat #Ho(Qp, V*n) = 1 in the flatcase. Moreover
#Hf (Qp, Vn) can be computedto be #(Q/A)2nfrom
Hf(Qv, VAn) Hf (Qp , V)An Homo(PR/PR , K/O)An
ringofpo for0-algebras. Using

whereR is the universallocal flatdeformation
the relationR _RRXl (9 where Rfl is the correspondingring for W(k)-
W(k)
algebras,and the main theoremof [Ram] (Theorem4.2) whichcomputesRfl,
we can deduce the result.
We now prove(vi). Fromthe definitions
#HF4(Qp V) { (#O/Al)r #H0(Qp, WAn) if Pf,AIDp does not split
,\n
(#O/An')r if PfAID~ Splits
wherer = dimKHF(Qp, V). This we can computeusing the calculationsin
[BK, Cor. 3.8.4]. We findthat r = 2 in the non-splitcase and r = 3 in the
split case and (vi) followseasily. E
3. Some results on subgroups of GL2(k)
We now give two group-theoretic results which will not be used until
Chapter 3. Althoughthese could be phrasedin purelygroup-theoretic terms
it will be moreconvenientto continueto workin the settingof Section 1, i.e.,
withpo as in (1.1) so that impo is a subgroupof GL2(k) and detpo is assumed
odd.
LEMMA 1.10. If impo has orderdivisiblebyp then:
(i) It containsan element'Yoof orderm > 3 with(m,p) = 1 and 'Yotrivial
on any abelian quotientofimpo.
(ii) It containsan elementpo(G) withany prescribedimage in the Sylow
2-subgroupof (impo)/(impo)' and withthe ratio of the eigenvaluesnot equal
to w(of). (Here (impo)' denotesthe derivedsubgroupof (impo).)

476 ANDREW WILES
The same resultshold if the image of theprojectiverepresentation

po as-
sociatedto po is isomorphicto A4,S4 or A5.
Proof. (i) Let G = impo and let Z denote the centerof G. Then we
have a surjectionG' - (G/Z)' wherethe ' denotes the derivedgroup. By
Dickson's classificationof the subgroupsof GL2(k) containingan elementof
orderp, (G/Z) is isomorphicto PGL2(k') or PSL2(k') forsome finitefieldk' of
characteristic p or possiblyto A5 whenp = 3, cf.[Di, ?260]. In each case we can
find,and then liftto G', an elementof orderm with (m, p) = 1 and m > 3,
except possiblyin the case p - 3 and PSL2(F3) - A4 or PGL2(F3) S4.
Howeverin thesecases (G/Z)' has orderdivisibleby 4 so the 2-Sylowsubgroup
of G' has ordergreaterthan 2. Since it has at mostone elementofexact order
2 (the eigenvalueswouldboth be -1 sinceit is in the kernelofthe determinant
and hencethe elementwould be -I) it mustalso have an elementof order4.
The argumentin the A4, S4 and A5 cases is similar.
G = impo has no fixedline.
(ii) Since po is assumedabsolutelyirreducible,
We claim that the same then holds forthe derivedgroup G'. For otherwise
sinceG' < G we could obtain a secondfixedline by taking(gv) where(v) is the
originalfixedlineand g is a suitableelementofG. Thus G' wouldbe contained
in the group of diagonal matricesfora suitable basis and eitherit would be
centralin which case G would be abelian or its normalizerin GL2(k), and
hence also G, would have orderprimeto p. Since neitherof thesepossibilities
is allowed,G' has no fixedline.
By Dickson's classificationof the subgroupsof GL2(k) containingan el-
ementof orderp the image of impo in PGL2(k) is isomorphicto PGL2(k')
or PSL2(k') forsome finitefieldk' of characteristic p or possiblyto A5 when
p = 3. The only one of these with a quotientgroup of orderp is PSL2(F3)
whenp = 3. It followsthat p t [G: G'] except in this one case whichwe treat
separately.So assumingnow that p t [G: G'] we see that G' containsa non-
trivialunipotentelementu. Since G' has no fixedline theremustbe another
noncommuting unipotentelementv in G'. Pick a basis forPOIG' consisting
of theirfixedvectors. Then let r be an elementof Gal(Qr/Q) forwhichthe
image of po(r) in G/G' is prescribedand let Po(T) = (a d). Then
(a b )(1sax)(
has det(6) = detpo(r) and trace6 = sa (ra/3+ c) + br/3+ a + d. Since p > 3

we can choose this trace to avoid any two givenvalues (by varyings) unless
ra3 + c = 0 for all r. But ra/3+ c cannot be zero forall r as otherwise
a = c = 0. So we can finda 6 forwhichthe ratio of the eigenvaluesis not
w(T), det(6) being, of course, fixed.

Now suppose that impo does not have orderdivisibleby p but that the
po has image isomorphicto S4 or A5, so
associated projectiverepresentation
necessarilyp 7&3. Pick an element such that the image of po(r) in G/G' is
r
any prescribed class. Since this fixes both det po(r) and w(r) we have to show
that we can avoid at most two particularvalues of the trace for'r. To achieve
thiswe can adapt our firstchoiceof r by multiplying by any elementofG'. So
picka E G' as in (i) whichwe can assume in thesetwo cases has order3. Pick
a basis forpo,by extendingscalars ifnecessary,so that a -+ (a a- ). Then one
checkseasilythat ifPo(r) = (c d) we cannothave the tracesofall ofT, ar and
a2 lyingin a set of the form{Tt} unless a = d = 0. Howeverwe can ensure
that po(r) does not satisfythis by firstmultiplyingr by a suitableelementof
G' since G' is not containedin the diagonal matrices(it is not abelian).
In the A4 case, and in the PSL2(F3) - A4 case when p = 3, we use a
differentargument.In both cases we findthat the 2-Sylowsubgroupof G/G'
is generatedby an elementz in the centreofG. Eithera powerofz is a suitable
candidate forpo(a) or else we mustmultiplythe powerof z by an elementof
G', the ratio of whose eigenvaluesis not equal to 1. Such an elementexists
because in G' the onlypossibleelementswithoutthispropertyare {TI} (such
elementsnecessarilyhave determinant1 and orderprimeto p) and we know
that #G' > 2 as was noted in the proofof part (i). L
Remark. By a well-knownresulton the finitesubgroupsofPGL2 (Fp) this

lemma coversall po whose images are absolutelyirreducibleand forwhich -5
is not dihedral.
Let K1 be the splittingfieldof po. Then we can view WA and W* as
Gal(K1((p)/Q)-modules. We need to analyze theircohomology.Recall that
we are assumingthat po is absolutelyirreducible.Let -5be the associated
to PGL2(k).
projectiverepresentation
The followingpropositionis based on the computationsin [CPS].
PROPOSITION 1.1 1. Suppose thatpo is absolutelyirreducible.Then

= 0.
H1 (K1 ((p)/Q1WA*)
Proof. If the image of po has orderprimeto p the lemma is trivial. The

subgroupsof GL2(k) containingan elementoforderp whichare not contained
in a Borel subgrouphave been classifiedby Dickson [Di, ?260] or [Hu, II.8.27]
Their images inside PGL2(k') where k' is the quadratic extensionof k are
conjugate to PGL2(F) or PSL2(F) for some subfieldF of k', or they are
isomorphicto one of the exceptionalgroupsA4,S4, A5.
groupH1(Ki((p)/Q, WA)$ 0. Then
Assumethenthatthecohomology
by consideringthe inflation-restriction
sequence with respect to the normal

478 ANDREW WILES
subgroupGal(Ki ((p)/KI) we see that (p E K1. Next,since the representation

is (absolutely) irreducible,the center Z of Gal(Ki/Q) is contained in the
diagonalmatricesand so acts triviallyon WA. So by consideringthe inflation-
restrictionsequencewithrespectto Z we see that Z acts triviallyon (p (and
on We). So Gal(Q((p)/Q) is a quotientof Gal(Ki/Q)/Z. This rules out all
cases whenp #3, and whenp = 3 we onlyhave to considerthe case wherethe
image of the projectiverepresentation is isomorphicas a group to PGL2(F)
forsome finitefieldof characteristic3. (Note that S4 - PGL2(F3).)
Extendingscalars commuteswithformationof duals and H1, so we may
assume without loss of generalityF C k. If p = 3 and #F > 3 then
H1(PSL2(F), WA) = 0 by results of [CPS]. Then if p5 is the projective
representation associated to po suppose that g-1 imp g = PGL2(F) and let
H = g PSL2(F)g-1. Then WA WAoverH and
(1.18) H1 (H WA)(FF H (g- Hg, gA1(WA F)) = 0.
F F
We deduce also that H1(impo, WA) = 0.

Finallywe considerthe case whereF = F3. I am gratefulto Taylorforthe
followingargument.Firstwe considerthe actionofPSL2(F3) on WAexplicitly
by consideringthe conjugationactionon matrices{A E M2(F3): traceA = 0}.
One sees that no such matrixis fixedby all the elementsof order2, whence
H1(PSL2(F3), WA) - H1(Z/3, (W,)C2xc2) = 0
whereC2 x C2 denotesthenormalsubgroupoforder4 in PSL2 (F3) - A4. Next

we verifythat thereis a unique copy of A4 in PGL2 (F3) up to conjugation.
For suppose that A, B E GL2(F3) are such that A2 = B2 = I withthe images
ofA, B representing distinctnontrivialcommutingelementsofPGL2(F3). We
can choose A = (O _0)by a suitable choice of basis, i.e., by a suitable conju-
gation. Then B is diagonal or antidiagonalas it commuteswith A up to a
scalar, and as B, A are distinctin PGL2(F3) we have B = (? -') forsome
a. By conjugatingby a diagonal matrix(whichdoes not change A) we can
assume that a = 1. The groupgeneratedby {A, B} in PGL2(F3) is its own
centralizerso it has indexat most6 in its normalizerN. Since N/(A, B) - S3
thereis a unique subgroupofN in which(A, B) has index3 whencethe image
of the embeddingof A4 in PGL2(F3) is indeedunique (up to conjugation). So
arguingas in (1.18) by extendingscalars we see that H1 (impo, WA*)= 0 when
F = F3 also. O
The followinglemma was pointedout to me by Taylor. It permitsmost

dihedralcases to be coveredby the methodsof Chapter3 and [TW].
LEMMA 1.12. Suppose thatpo is absolutelyirreducibleand that
(a) po is dihedral(the case wherethe image is Z/2 x Z/2 is allowed),

(b) PoK is absolutelyirreduciblewhereL = Q (A(-1)(P1)/2P).
Thenfor any positiveintegern and any irreducibleGalois stable subspaceX

of WA0 k thereexistsan elementa E Gal(Q/Q) such that
(i) f3o(a) $ 1,
(ii) a fixesQ((pn),
(iii) a has an eigenvalue1 on X.
Proof. If fiois dihedralthenpo X k = IndH X forsome H of index 2 in G,

whereG = Gal(Ki/Q). (As before,K1 is the splittingfieldof po.) Here H
can be takenas the fullinverseimageof any ofthe normalsubgroupsof index
2 definingthe dihedralgroup. Then WAX k b 80 IndG(X/X') where6 is the
quadraticcharacterG -* C/H and X' is the conjugateof X by any elementof
G - H. Note that X 54X' since H has nontrivialimage in PGL2(k).
To finda a such that 8(cr) = 1 and conditions(i) and (ii) hold, observe
that M((pn) is abelian where M is the quadratic fieldassociated to 8. So
conditions(i) and (ii) can be satisfiediffiois non-abelian.If fiois abelian (i.e.,
the imagehas the formZ/2 x Z/2), thenwe use hypothesis(b). If IndG(X/X')
is reducibleoverk thenWA0 k is a sum ofthreedistinctquadraticcharacters,
none of whichis the quadratic characterassociated to L, and we can repeat
the argumentby changingthe choice of H forthe other two characters. If
X = IndG(x/x') ? k is absolutelyirreduciblethenpick any a E G - H. This
satisfies(i) and can be made to satisfy(ii) if(b) holds. Finally,sincea GC -H
we see that a has trace zero and o2 = 1 in its action on X. Thus it has an
eigenvalueequal to 1.
Chapter 2
In this chapterwe study the Hecke rings. In the firstsectionwe recall

some of the well-knownpropertiesof these ringsand especially the Goren-
stein propertywhose proofis rathertechnical,dependingon a characteristic
p versionof the q-expansionprinciple.In the second sectionwe computethe
relationsbetweenthe Heckeringsas the levelis augmented.The purposeis to
findthe changein the r7-invariant as the level increases.
In the thirdsectionwe state the conjecturerelatingthe deformationrings
of Chapter 1 and the Hecke rings. Finally we end with the criticalstep of
showingthat if the conjectureis true at a minimallevel then it is true at
all levels. By the resultsof the appendix the conjectureis equivalentto the

480 ANDREW WILES
forthe Hecke ringsand the p/p2-invariant

equalityof the 71-invariant forthe
deformationrings. In Chapter 2, Section 2, we compute the change in the
and in Chapter 1, Section 1, we estimatedthe changein the p/p2_
7a-invariant
invariant.
1. The Gorenstein property
For any positiveintegerN let Xj (N) = X1(N)/Q be the modularcurve

over Q corresponding to the grouprl(N) and let J1(N) be its Jacobian. Let
T1 (N) be theringofendomorphisms ofJ1(N) whichis generatedoverZ by the
standardHecke operators{Ti = Tl* for1 t N, Uq = Uq* forq j N, (a) = (a)*
for (a, N) = 1}. For precise definitionsof these see [MW1, Ch. 2, ?5]. In
particularif one identifiesthe cotangentspace of Ji(N) (C) withthe space of
cusp formsofweight2 on rF(N) thenthe actioninducedby T1 (N) is the usual
one on cusp forms. We let A = { (a) : (a, N) = 1}.
The group (Z/NZ)* acts naturallyon X1 (N) via A and for any sub-
group H C (Z/NZ)* we let XH(N) = XH(N)/Q be the quotientX1(N)/H.
Thus forH = (Z/NZ)* we have XH(N) = Xo(N) corresponding to the group
ro (N). In Section2 it will sometimesbe to
convenient assume that L decom-
poses as a productH = HIHq in (Z/NZ)* -_f(Z/qrZ)* wherethe product
is over the distinctprimepowersdividingN. We let JH(N) denote the Ja-
cobian of XH(N) and note that the above Hecke operatorsact naturallyon
JH(N) also. The ringgeneratedby these Hecke operatorsis denotedTH(N)
and sometimes,if H and N are clear fromthe context,we abbreviatethis
to T.
Let p be a prime > 3. Let m be a maximal ideal of T = TH(N) with
p E m. Then associated to m there is a continuousodd semisimpleGalois
representation pm,
(2.1) pm: Gal(Q/Q) -- GL2(T/m)

outsideNp whichsatisfies
unramified
tracepm(Frobq) = Tq, detpm(Frobq) = (q)q
foreach primeq t Np. Here Frobq denotes a Frobeniusat q in Gal(Q/Q).
The representation pmis unique up to isomorphism.If p t N (resp.p I N) we
say that m is ordinaryifTp ? m (resp. Up ? m). This implies(cf.,forexample,
theorem2 of [Wil]) that forour fixeddecompositiongroupDp at p,
P
I D(O X 2)
Dp 0 X2
fora suitable choice of basis, with X2 unramifiedand X2(Frobp) = Tp mod

m (resp. equal to Up). In particularPmis ordinaryin the sense of Chapter 1

providedXi 54X2- We will say that m is Dp-distinguished ifm is ordinaryand

Xi 5 X2. (In practiceXi is usuallyramifiedso thisimposesno extracondition.)
We cautionthe readerthat ifPmis ordinaryin the sense of Chapter 1 thenwe
can onlyconcludethat mis Dp-distinguished ifp t N.
Let Tmdenotethe completionofT at mso that Tmis a directfactorofthe
completesemi-localringTp = T X Zp. Let D be the pointsof the associated
m-divisiblegroup
D = JH(N) (Q)m - JH(N) (Q)poo (0 Tm.

Tp
It is knownthat D = HomzP(D, Qp/Zp) is a rank 2 Tm-module,i.e., that

D 0 Qp - (Tm 0 Qp)2. Brieflyit is enoughto show that H1(XH(N), C) is
zp zp
freeof rank 2 over T 0 C and this reducesto showingthat S2 (rH(N), C),
the space of cusp formsof weight2 on rH(N), is freeof rank 1 over T 0 C.
One showsthen that if {fi, ... , f4} is a completeset of normalizednewforms
in S2(rH(N), C) of levels mj,...,m, then if we set di = N/mi, the form
f = E fi(diz) is a basis vectorof S2(rH(N), C) as a T 0 C-module.
If m is ordinarythen Theorem2 of [Wil], itselfa straightforward gener-
alization of Proposition2 and (11) of [MW2], shows that (forour fixedde-
compositiongroupDp) thereis a filtration of D by Pontrjaginduals of rank 1
Tm-modules(in the sense explainedabove)
(2.2) 0 Do D __DE _0
whereDo is stable underDp and the inducedaction on DE is unramified with

Frobp = Up on it ifp I N and Frobp equal to the unit root ofx2 -Tpx + p(p)
- 0 in Tm if p t N. We can describeDo and DE as follows. Pick a of E
Ip whichinduces a generatorof Gal (Qp((Npoo)/Qp((Np)). 14 Z-
Let e:Dp x
be the cyclotomic character. Then DO = ker (oa _ E(a))diV, the kernel being
taken inside D and 'div' meaningthe maximal divisiblesubgroup. Although
in [Wil] this filtrationis given only for a factorAf of Ji(N) it is easy to
deduce the result for JH(N) itself. We note that this filtrationis defined
withoutreference to characteristic
p and also that ifm is Dp-distinguished,DO
(resp. DE) can be describedas the maximal submoduleon whicho - 1(of)
is topologicallynilpotentfor all of E Gal(Qp/Qp) (resp. quotient on which
f- X2(07) is topologicallynilpotentforall of E Gal(Qp/Qp)), where%i(of)is
any liftingof Xiy(o)to Tm.
The Weil pairing ( , ) on JH(N)(Q)pM satisfiesthe relation (tx,y) =
(x, t*y) forany Hecke operatort. It is more convenientto use an adapted
pairingdefinedas follows. Let w(, forC a primitiveNth root of 1, be the
involutionofXi (N)/Q(() definedin [MW1, p. 235]. This inducesan involution
of XH(N)/Q(() also. Then we can definea new pairing[ , ] by setting(fora

482 ANDREW WILES
fixedchoiceof ()
(2.3) [x,Iy]= (x,wy).

Then [tx, y] = [x,ty] forall Hecke operatorst. In particularwe obtain an
inducedpairingon Dpa.
The followingtheoremis the crucial resultof this section. It was first
provedby Mazur in the case ofprimelevel [Ma2]. It has sincebeen generalized
in [Til], [Ril] [M Ri], [Gro]and [El], but the fundamentalargumentremains
that of [Ma2]. For a summarysee [E1, ?9]. Howeversome ofthe cases we need
are not coveredin these accountsand we will presentthese here.
THEOREM 2.1. (i) If p t N and Pmis irreduciblethen
JH(N) (Q) [m] --(T/M)2.
(ii) If p t N and Pmis irreducibleand m is Dp-distinguished

then
JH(Np) (Q) [m] (T/)2 .

(In case (ii) m is a maximal ideal of T = TH(Np).)
COROLLARY 1. In case (i), JH(N)(Q)m. T2 and Tam (JH(N)(Q)>

TM.
T2m
In case (ii), JH(Np)(Q)m - T. and Tam(JH(Np) (Q)) T. (where
Tm = TH(NP)m)
COROLLARY 2. In eitherof cases (i) or (ii) Tm is a Gorensteinring.
In each case the firstisomorphisms of Corollary1 followfromthe theorem

2
togetherwith the rank result alluded to previously. Corollary2 and the
second isomorphisms of corollary1 then followon applyingduality(2.4). (In
the proofand in all applicationswe will only use the notionof a Gorenstein
Zp-algebraas definedin the appendix. For finiteflat local Zp-algebrasthe
notionsof Gorensteinring and GorensteinZp-algebraare the same.) Here
Tam(JH(N) (Q)) = Tap (JH(N) (Q)) 0 Tm is the m-adic Tate module of
JH(N).
We shouldalso pointout that althoughCorollary1 givesa representation
fromthe m-adicTate module
P = PTm: Gal(Q/Q) -* GL2 (Tm)

thiscan be constructedin a muchmoreelementary way. (See [Ca3] foranother
argument.)For,the representation 0
existswithTm Q replacingTm whenwe
use the factthat Hom(Qp/Zp,D) 0 Q was freeofrank2. A standardargument

using the Eichler-Shimurarelationsimpliesthat this representationp' with

values in GL2(Tm 0 Q) has the propertythat
tracep'(Frobe) = Te, detp'(Frobe) = f(f)
forall e t Np. We can normalizethis representation by pickinga complex
conjugationc and choosinga basis suchthatp'(c) = (' _), and thenbypicking
a r forwhichp'(r) = (a br) with b, c, # 0(m) and by resealingthe basis so
that b, = 1. (Note that the explicitdescriptionof the traces showsthat ifPm
is also normalizedso that pm(c) = (o _) then
b1c1 mod m = brmCrm where
P.m(T)= (lrm dbrn,). The existenceof a r such that b~c, 0 0(m) comes from
the irreducibility
of pm.) With this normalizationone checksthat p' actually
takes values in the (closed) subringof Tm generatedover Zp by the traces.
One can even constructthe representation directlyfromthe representations in
Theorem0.1 usingthis ringwhichis reduced. This is the methodof Carayol
whichrequiresalso the characterizationof p by the traces and determinants
(Theorem 1 of [Ca3]). One can also ofteninterpretthe Uq operatorsin terms
of p forq I N usingthe 7rq 7r(gq) theoremof Langlands (cf. [Cal]) and the
Up operatorin case (ii) usingTheorem2.1.4 of [Wil].
Proof(of theorem).The importanttechniqueforprovingsuchmultiplicity-
one resultsis due to Mazur and is based on the q-expansionprinciplein char-
acteristicp. Sincethe kernelof JH(N) (Q) -* Ji (N) (Q) is an abelian groupon
whichGal(Q/Q) acts throughan abelian extensionof Q, the intersection with
kerm is trivialwhen pmis irreducible.So it is enoughto verifythe theorem
forJ1(N) in part (i) (resp. Ji(Np) in part (ii)). The methodforpart (i) was
developedby Mazur in [Ma2, Ch. II, Prop. 14.2]. It was extendedto the case
of Fo(N) in [Ril, Th. 5.2] whichsummarizesMazur's argument.The case of
F1(N) is similar(cf. [El, Th. 9.2]).
Now considercase (ii). Let A(p) = {(a): a l(N)} C A\. Let us first
assumethatA(p) is nontrivialmodm,i.e., that6-1 mforsome6 E A(p). This
case is essentiallycoveredin [Til] (and also in [Gro]). We brieflyreviewthe
argumentforuse later. Let K = Qp((p), (p beinga primitivepth rootofunity,
and let 0 be the ringof integersof the completionof the maximalunramified
extensionof K. Using the fact that A(p) is nontrivialmod m togetherwith
Proposition4, p. 269 of [MW1] we findthat
J (Np)m/tO (Fp) - (Pic?
E0t X Pico E1)m (Fp)
wherethe notationis taken from[MW1] loc. cit. Here E't and El, are the
two smoothirreduciblecomponentsofthe special fibreofthe canonicalmodel
of X1(Np)l0 describedin [MW1, Ch. 2]. (The smoothnessin this case was
provedin [DR].) Also Jl(Np)6t1 denotesthe canonical etale quotientof the
r-divisiblegroupover 0. This makes sense because J1(Np)m does extendto

484 ANDREW WILES
a p-divisiblegroupover(9 (again by a theoremof Deligne and Rapoport [DR]

and because A(p) is nontrivialmodm). It is ordinaryas followsfrom(2.2) when
we use the main theoremof Tate ([Ta]) since 'Do and 'DE clearlycorrespond
to ordinaryp-divisiblegroups.
Now the q-expansionprincipleimpliesthat dimp X [m'] < 1 where
p
X = {Ho (EP, Q1) e Ho (EtI Q1)}

and m' is definedby embeddingT/m - Fp and settingm' = ker:T Fp -Fp
under the map t 0 a - at mod m. Also T acts on Pico EA x Pico Et, the
abelian varietypart of the closed fibreof the Neron model of J1(Np) IO, and
hence also on its cotangentspace X. (For a proofthat X[m'] is at most one-
dimensional,whichis readilyadapted to this case, see Lemma 2.2 below. For
similarversionsin slightlysimplercontextssee [Wi3, ?6] or [Gro,?12].) Then
the Cartiermap inducesan injection(cf. Prop. 6.5 of [Wi3])
6: {Pic0 imx Pico E6t}[] (p) 0 Ip C 4 X.
The composite6 owa can be checkedto be Hecke invariant(cf. Prop. 6.5 of

[Wi3]. In checkingthe compatibilityforUp use the formulasof Theorem5.3
of [Wi3]but note the correctionin [MW1, p. 188].) It followsthat
J,(Np)m/0(Fp) [m] T/ra

as a T-module. This shows that if H is the Pontrjagin dual of
H = Ji(Np)m10(Fp) then H - Tmsince H/m - T/m. Thus
Ji(Np)m1(Fp) [p] Z Hom(Tm/p, Z/pZ)

Now our assumptionthat m is Dp-distinguished
enables us to identify
'Do= J,(Np)m0 (Qp) XDE = j, (Np)6t (
For the groupson the rightare unramifiedand those on the leftare dual to
groupswhereinertiaacts via a characterof finiteorder(dualitywithrespect
to Hom( , Qp/Zp(l))). So
'Do[p] Tm/p, DE[p] Hom (T./p, Z/pZ)
as Tm-modules,the formerfollowing fromthe latterwhenwe use dualityunder
the pairing[,]. In particularas m is Dp-distinguished,
(2.4) 'D[p] - Tm/p G Hom (T./p, Z/pZ).
We now use an argumentof Tilouine [Til]. We pick a complexconjugation
T-. This has distincteigenvalues?1 on Pmso we may decompose D[p] into
eigenspaces for -r:
D[p] = D[p]+ D[p]-.

Since Tm/pand Hom (Tm/p, Z/pZ) are both indecomposableHecke-modules,

by the Krull-Schmidttheoremthis decompositionhas factorswhichare iso-
morphicto those in (2.4) up to order. So in the decomposition
D[m] = D[m]+ GDD[m]
one of the eigenspacesis isomorphicto T/m and the otherto (Tm/p)[m].But

sincePmis irreducibleit is easy to see by consideringD[m]G Hom(D [m],detPm)
that ir has the same numberofeigenvaluesequal to +1 as equal to -1 in D[m],
whence #(Tm/p) [m] = #(T/m). This shows that D[m]+ A D[m]- T/m as
required.
Now we considerthe case where \(p) is trivialmod m. This case was
treated (but only forthe group Fo(Np) and Pm 'new' at p-the crucial re-
strictionbeing the last one) in [M Ri]. Let Xl (N, p)/Q be the modularcurve
corresponding to I1 (N) n Fo(p) and let Ji(N,p) be its Jacobian. Then since
the composite of natural maps J1(N, p) -) Ji (Np) -+ Ji(N, p) is multiplication
by an integerprimeto p and since A(p) is trivialmodm we see that
J,(N. p) m(Q) -_J,(Np) m(Q)

It willbe enoughthento use J1(N, p), and the corresponding
ringT and ideal
m.
The curve X1(N,p) has a canonical model X1(N,p)z which over Fp

consistsof two smooth curves SEt and E2 intersectingtransversallyat the
supersingularpoints (again this is a theoremof Deligne and Rapoport; cf.
[DR, Ch. 6, Th. 6.9], [KM] or [MW1] formoredetails). We willuse the models
describedin [MW1, Ch. II] and in particularthe cusp ox will lie on E?. Let
Q denotethe sheafof regulardifferentials on X1(N,P)/FP (cf. [DR, Ch. 1 ?2],
[M Ri, ?7]). OverFp, sinceX1 (N, p) j; has ordinarydoublepointsingularities,
the differentials
may be identifiedwith the meromorphic differentials
on the
normalizationX1 (N, p)j = SEt U EA whichhave at mostsimplepoles at the
supersingular points(theintersection pointsofthetwocomponents)and satisfy
resxj+ resX2= 0 if x1 and x2 are the two points above such a supersingular
point. We need the followinglemma:
LEMMA 2.2. dimT/.HO(X1(N,p)/FpA) [m] = 1.
Proof. First we remarkthat the action of the Hecke operatorUp here is

most conveniently definedusingan extensionfromcharacteristic zero. This is
explainedbelow. We will first
show that dimT/. HO (Xi (N, P)/Fp iQ) [m]< 1,
this being the essential step. If we embed T/m - Fp and then set
m' = ker: T 0 Fp - Fp (the map given by t 0 a - atmodim) then it is
enoughto showthat dimppHO(X1(N, p)/li, Q) [m']< 1. Firstwe willsuppose

486 ANDREW WILES
in Ho (Xi (N, p)/F Q)[i'],

that thereis no nonzeroholomorphicdifferential
formwhichpulls back to holomorphicdifferentials
i.e., no differential on E'
and EJ. Then if wi and W2 are two differentials in Ho(Xi (N, p)/,-,) [n'],
the q-expansionprincipleshowsthat ,uW1- AW2has zero q-expansionat oo for
some pair (,u,A) $&(0, 0) in F2 and thus is zero on EyL* As ,uwl - AW2= 0 on
EY it is holomorphicon Eet. By our hypothesisit would then be zero which
showsthat wi and W2 are linearlydependent.
This use ofthe q-expansionprinciplein characteristic p is crucialand due
to Mazur [Ma2]. The pointis simplythat all the coefficients in the q-expansion
are determinedby elementaryformulaefromthe coefficient of q providedthat
w is an eigenform forall the Hecke operators. The formulaeforthe action of
these operatorsin characteristicp followfromthe formulaein characteristic
zero. To see this formally(especially forthe Up operator) one checks first
that Ho(Xi (N, p)/zp,Q), whereQ denotesthe sheafofregulardifferentials on
-
Xi(N,p)lzp, behaves well under the base changes Zp -*
Fp and Zp Qp;
cf. [Ma2, ?11.3]or [Wi3, Prop. 6.1]. The action of the Hecke operatorson
J1(N, p) inducesan actionon the connectedcomponentofthe Neronmodelof
J1(N, P)/QP, so also on itstangentspace and cotangentspace. By Grothendieck
duality the cotangentspace is isomorphicto H0(Xj(N,p)1ZPQ); see (2.5)
below. (For a summaryof the dualitystatementsused in this context,see
[Ma2, ?II.3]. For explicitdualityover fieldssee [AK, Ch. VIII].) This then
definesan actionofthe Heckeoperatorson this group. To checkthat overQp
thisgivesthe standardactionone uses the commutativity ofthe diagramafter
Proposition2.2 in [Mil].
Now assume that thereis a nonzeroholomorphicdifferential in
H?(Xi(N.p)1V' 7 [in'].
Q)
We claimthat the space ofholomorphicdifferentialsthenhas dimension1 and

w :$ 0 is actually nonzeroon EY. The dimension
that any such differential
claim followsfromthe second assertionby usingthe q-expansionprinciple.To
provethat w :$ 0 on EA we use the formula
Up*(x,y) = (Fx, y')
for (x, y) E (Pico YE x Pico E)(Fp), where F denotes the Frobeniusendo-

morphism. The value of y' will not be needed. This formulais a variant
on the second part of Theorem 5.3 of [Wi3] where the correspondingre-
sult is proved forXi (Np). (A correctionto the firstpart of Theorem 5.3
was noted in [MW1, p. 188].) One checks then that the action of Up on
Xo = Ho (EI, IQ71)E Ho (YEt,Q1) viewed as a subspace of HO(Xi (N,p)/ F ,
is the same as the action on Xo viewed as the cotangentspace of Pico Ell x
Pic? SEt. Fromthis we see that ifw = 0 on EA then Upw= 0 on E't. But Up

acts as a nonzeroscalar whichgivesa contradictionifw :$ 0. We can thus as-

sumethatthespace ofm'-torsion holomorphic has dimension1 and
differentials
is generatedby w. So ifW2 is now any differential HO(X1 (N, p)/P, Q) [m']
in
thenW2- Awhas zero q-expansionat oX forsome choiceofA. Then W2 - Aw = 0
on EA whenceW2 - Awis holomorphicand so W2 = Aw. We have now shown
in general that dim(HO (Xi(N,p)1F , Q) [m']) < 1.
The singularitiesof X1(N,p)lZP at the supersingularpointsare formally
isomorphic overZunr to Zunr[[X, Y]] /(XY -k) with k = 1, 2 or 3 (cf. [DR,
Ch. 6, Th. 6.9]). If we considera minimalregularresolutionM1(Np)1ZP
then H0(M1(NP)/Fp, 7) HO(X1(N.P)/FP, 7) (see the argumentin [Ma2,
Prop. 3.4]), and a similarisomorphismholds forHO (Ml (N, p)/ZP,Q)
As M1(N,P)/ZP is regular,a theoremof Raynaud [Ray2] says that the
connectedcomponentof the Neron model of J1(N,P)/QP is Ji(N,P)5/z
PicO(M1(N,p)1zP). Takingtangentspaces at the origin,we obtain
(2.5) Tan(Ji(N, p)?Z ) H1 (Ml (N. p)1ZP, OM1 (N,p)))
Reducingboth sides modp and applyingGrothendieckdualitywe get an iso-
morphism
(2.6) Tan(J1(Np)5/F) Hom(H0(Xl(NP)/FP, I Fp).

(To justifythe reductionin detail see the argumentsin [Ma2, ?II. 3]). Since
Tan(J1(N, p)?Z ) is a faithfulT 0 Zp-moduleit followsthat
H? (Xi (N. A)/FPiQ) [ml

is nonzero.This completesthe proofof the lemma. OI
To completethe proofof the theoremwe choose an abelian subvariety

A of J1(N, p) with multiplicativereductionat p. Specificallylet A be the
connectedpart of the kernelof J1(N, p) -* J1(N) x J1(N) underthe natural
map f describedin Section2 (see (2.10)). Then we have an exact sequence
0 *A -* J1(N,P) -*B O*
and J1(N,p) has semistablereductionover Qp and B has good reduction.
By Proposition1.3 of [Ma3] the correspondingsequence of connectedgroup
schemes
0 -' A[p]ZP] J1(N. p) [p]/Z B[IZP -o 0
is also exact, and by Corollary1.1 of the same propositionthe corresponding
sequenceoftangentspaces ofNeronmodelsis exact. Usingthis we may check
that the naturalmap
(2.7) Tan(J1(N,p)[p]t ) 0 Tm Tan(J(Np) ) 0 Tm
TP Ta(l N P) TP

488 ANDREW WILES
is an isomorphism,wheret denotesthe maximalmultiplicative-type subgroup

scheme(cf. [Ma3, ?1]). For it is enoughto checksuch a relationon A and B
separatelyand on B it is true because the r-divisiblegroupis ordinary.This
followsfrom(2.2) by the theoremof Tate [Ta] as before.
Now (2.6) togetherwiththe lemmashowsthat
Tan(Ji(Np))/zP (0 Tm Tm.
Tp
We claim that (2.7) togetherwiththis impliesthat as Tm-modules
V := Ji(N,p) [p]t(Q,)m - (Tm/p).
To see this it is sufficient

to exhibitan isomorphismof Fp-vectorspaces
(2.8) Tan(G/f ) G(Qp) 0 Fp
groupscheme(finiteand flat) GIzp whichis killed

forany multiplicative-type
by p and moreoverto give such an isomorphismthat respectsthe action of
endomorphisms ofGlzp. To obtainsuch an isomorphismobservethat we have
isomorphisms
(2.9) Hom-, (lp, G) X9Fp Hompp(lp, G) (9 Fp
Hom (Tan(tip/. Tan(G/V))

P)
whereHomQ denoteshomomorphisms of the groupschemesviewedover Qp

and similarlyforHomF . The secondisomorphism can be checkedby reducing
p
to the case G = lip. Now pickinga primitivepth root of unitywe can iden-
tifythe left-handtermin (2.9) with G(Qp) 0 Fp. Pickingan isomorphismof
Fp
Tan(,ip/l;) with Fp we can identifythe last termin (2.9) with Tan(G/y ).

Thus afterthese choices are made we have an isomorphismin (2.8) which
respectsthe action of endomorphisms of G.
On the otherhand the action of Gal(Qp/Qp) on V is ramifiedon every
subquotient,so V C 'DQ[p]. (Note that our assumptionthat A(p) is trivial
modm impliesthat the action on V0[p] is ramifiedon everysubquotientand
on DyE[p]is unramifiedon everysubquotient.) By again examiningA and B
separatelywe see that in fact V = ZDo[p].For A we note that A[p]/A[p]tis
unramified because it is dual to A[p]twhereA is the dual abelian variety.We
can now proceedas we did in the case whereA(p) was nontrivialmodm. E

2. Congruences between Hecke rings
Suppose that q is a primenot dividingN. Let IF(N, q) = P1(N) n 1o(q)

and let Xi (N, q) = X1 (N, q)/Q be the corresponding
curve. The two natural
maps X1 (N, q) -- Xi (N) induced by the maps z -- z and z -- qz on the
upper halfplane permitus to definea map Ji(N) x J1(N) -- Ji(N, q). Using
a theoremofIhara, Ribet showsthat thismap is injective(cf. [Ri2,Cor. 4.2]).
Thus we can definep by
(2.10) 0- Ji(N) x J, (N) J1 (N, q)

Dualizing,we defineB by
0 -+B B- Ji(N, q) - Ji(N) x Ji(N) -+0.
Let T1 (N, q) be the ring of endomorphismsof Ji(N, q) generatedby the
standard Hecke operators {T1* for I; Nq, Ul* for I | Nq, (a) = (a)* for
(a, Nq) = 1}. One can checkthat Uq preservesB eitherby an explicitcalcu-
lation or by notingthat B is the maximalabelian subvarietyof J1(N, q) with
multiplicative reductionat q. We set J2= Ji(N) x Ji(N).
More generally,one can considerJH(N) and JH(N, q) in place of J1(N)
and J1(N,q) (whereJH(N, q) correspondsto X1(N, q)/H) and we writeTH(N)
and TH(N, q) forthe associated Hecke rings. In this case the corresponding
map p may have a kernel.Howeversince the kernelof JH(N) -+ J1(N) does
not meet kerm forany maximal ideal m whose associated Pmis irreducible,
the above sequences remainexact if we restrictto M(q)-divisiblegroups,M(q)
being the maximal ideal associated to m of the ringT (q)(N,q) generatedby
the standard Hecke operatorsbut omittingUq. With this minormodifica-
tion the proofsof the resultsbelow forH #& 1 followfromthe cases of full
level. We will use the same notationin the generalcase. Thus 0 is the map
J2 = JH(N)2 --+ JH(N, q) induced by z -- z and z -- qz on the two factors,
and B = ker . (B will not be an abelian varietyin general.)
The followinglemma is a straightforward generalizationof a lemma of
Ribet ([Ri2]). Let nq be an integersatisfyingnq q(N) and nq 1(q), and
write(q) = (nq) E TH(Nq).
LEMMA 2.3 (Ribet). 4'(B) n PO(J2).(q)

= sc(J2) [Uq -
(q)]M(q) for irre-
duciblePm.
Proof. The left-handside is (im o n ker 3), so we compute p1 (im ofn

ker )= ker( o p).
An explicitcalculationshowsthat
Woc~o=[F q+ I -qi1 onJ2

490 ANDREW WILES
whereT= Tq (q)'. The matrixactionhereis on the left.We also findthat

on J2
(2.11) q q Tq ]
[~~~~~
whence
(U2-(q))o = ? -(q) ]?( ? )
Now suppose that m is a maximal ideal of TH(N), p E m and Pmis

irreducible.We will now give a slightlystrongerresultthan that givenin the
lemma in the special case q = p. (The case q $ p we will also strengthenbut
we will do this separately.) Assume then that p 1 N and Tp ? m. Let ap be
the unit root of x2 - Tpx + p(p) = 0 in TH(N)m. We firstdefinea maximal
ideal mpof TH (N, p) withthe same associated representation as m. To do this
considerthe ring
S1 = TH(N)[Ul]/(U2 - TpUl + p(p)) C End(JH(N)2)
whereU1 is the endomorphism

of JH(N)2 givenby the matrix
Tp -(Pj
It is thus compatiblewiththe action of Up on JH(N, p) whencomparedusing

A. Now = (i, Ui-
-p ) is a maximal ideal of S1 whereap is any element
of TH(N) representingthe class aip e TH(N)m/m - TH(N)/m. Moreover
Si,mi - TH(N)m and we let mp be the inverseimage of ml in TH(N,p) under
the naturalmap TH(N, p) -* S1. One checksthat mpis Dp-distinguished.For
any standardHeckeoperatort except Up (i.e., t = TI, Uqiforq' :$p or (a)) the
image of t is t. The image of Up is U1.
We need to checkthat the inducedmap
a: TH(NP)m -+ Silmi TH(N)m
is surjective.The onlyproblemis to showthatTpis in theimage. In thepresent

contextone can provethis using the surjectivityof in (2.12) and usingthe
A
factthat the Tate-modulesin the rangeand domainof are freeof rank2 by

A
Corollary1 to Theorem2.1. The resultthenfollowsfromNakayama'slemmaas

one deduceseasilythatTH(N)m is a cyclicTH(N, p)m-module. This argument
was suggestedby Diamond. A second argumentusingrepresentations can be
foundat the end ofProposition2.15. We willnow givea thirdand moredirect
proofdue to Ribet (cf. [Ri4, Prop. 2]) but foundindependentlyand shownto
us by Diamond.

Forthefollowing lemmawe let TM, foran integerM, denotethesubringof

End (S2(rF (N))) generatedby the Hecke operatorsTn forpositiveintegersn
relativelyprimeto M. Here S2 (I1 (N)) denotesthe vectorspace of weight2
cusp formson IF (N). WriteT forT'. It will be enoughto show that Tp is
a redundantoperatorin T', i.e., that TP = T. The resultforTH(N)m then
follows.
LEMMA (Ribet). Suppose that (M,N) = 1. If M is odd then TM = T.

If M is even thenTM has finiteindexin T equal to a powerof2.
As the ringsare finitely

generatedfreeZ-modules,it suffices
to provethat
TM 0 F1 -- T 0 F1 is surjectiveunless 1 and M are both even. The claim
followsfrom
1. TM OF1 -- TM/POF1 is surjectiveifp I M and p t IN.
2. T' 0 F1 -+ T 0 F1 is surjectiveif 1 t 2N.
Proofof1. Let A denotethe Tate moduleTal(Ji(N)). Then R = TM/PO

on A. Let R' = (R 0 Ql) n Endz, A and choose d so that
Zj acts faithfully
Zl
ldR' C iR. Consider the Gal(Q/Q)-module B = Jl(N)[1d] x uN~d. By
Cebotarevdensity,thereis a primeq not dividingMNl so thatFrobp = Frobq
on B. Using the fact that Tr = Frobr + (r)r(Frobr)'- on A forr = p and
r = q, we see that Tp = Tq on J1(N) [Id]. It followsthat Tp - Tq is in ld Endz, A
and therefore in idR' c iR. a
Proofof2. Let S be theset ofcusp formsin S2 (r1 (N)) whoseq-expansions

in Z. Recall that S2(rF(N)) = SoC and that S is stable
at oo have coefficients
underthe action of T (cf. [Shl, Ch. 3] and [Hi4, ?4]). The pairingT 0 S-+ Z
definedby T 0 f | 4 a, (Tf) is easily checkedto induce an isomorphismof
T-modules
S Homz (T, Z).
The surjectivityof T'/lT' -+ T/lT is equivalentto the injectivityof the dual
map
Hom(T, Fj) -* Hom(T1, Fj).
Now use the isomorphismS/iS Hom(T, Fj) and note that if f is in the
kernel of S -- Hom(T1, Fj), then an(f) = al(Tnf) is divisible by l for all n
primeto 1. But thenthe modl formdefinedby f is in the kernelofthe operator
q d , and is therefore
trivialif I is odd. (See Corollary5 of the main theorem
of Ka].) Thereforef is in IS.
Remark. The argumentdoes not provethat TMd = Td if (d, N) $&1.

492 ANDREW WILES
We now returnto the assumptionsthat Pm is irreducible,p t N and

Tp ? m. Next we definea principalideal (Lxp) of TH(N)m as follows. Since
TH(N, p)mpand TH(N)m are both Gorensteinrings(by Corollary2 of Theo-
rem2.1) we can definean adjoint a^ to
a: TH(NPp)m - TH(N)m
-
Siml
in the mannerdescribedin the appendix and we set ?p = (a o a^)(1). Then
(zAp) is independentof the choice of (Hecke-module)pairingson TH(N, p),,
and TH(N)m. It is equal to the ideal generatedby any compositemap
TH(N)m TH(NP)nvp TH(N)m

providedthat /3is an injectivemap ofTH (N, p)mp
-moduleswithZp torsion-free
cokernel.(The modulestructureon TH(N)m is definedvia a.)
PROPOSITION2.4. Assume thatm is Dp-distinguished
and thatPm is
irreducibleof levelN withp t N. Then
(Ap) = (T2 (p)(1 +p)2) = (a2 -

(p)).
Proof. Considerthe maps on p-adic Tate-modulesinducedby o and A':
Tap(JH (N) 2) Tap(JH (N, p)) Tap (JH (N) 2

These maps commutewith the standardHecke operatorswith the exception
of Tp or Up (whichare not even definedon all the terms). We define
S2 = TH(N)[U2]/(U22 - Tp U2 +p(p)) C End (JH(N)2)
where U2 is the endomorphismof JH(N)2 definedby (? j(P)). It satisfies

(O U2 = Upo. Again m2 = (m, U2 - p) is a maximal ideal of S2 and we have,
on restricting
to the ml, mpand m2-adicTate-modules:
Tam2 (JHN2)
( Tamp (JH(N,p)) -Tam, (JH(N)2)
(2.12) TI V2 T Ivl
Tam(JH(N)) Tam(JH(N))
The verticalisomorphismsare definedby v2: x (-(p)x, apx) and vi: x - -
(apx, px). (Here ap E TH(N)m can be viewed as an elementof TH(N)p

HITH(N)n where the product is taken over the maximal ideals containing
p. So vi and v2 can be viewed as maps to Tap (JH(N)2) whose images are
respectivelyTam,(JH(N)2) and Tam2 (JH(N)2).)
'
Now is surjectiveand p is injectivewithtorsion-free
cokernelby the re-
sult of Ribet mentionedbefore. Also Tam(JH(N)) TH(N)2 and

Ta,,p(JH(N, p p)) TH(N, p)2 by Corollary1 to Theorem2.1. So as I,

are maps of TH(N, p),p-moduleswe can use this diagramto computeLp as
remarkedjust priorto the statementof the proposition. (The compatibility
the completionsS1iml and S2,M2
of the Up actionsrequiresthat,on identifying
withTH(N)m, we get U1 = U2 whichis indeedthe case.) We findthat
VT11 == 1( 2 (p)) (Z).
We nowapplyto J1(N, q2) (but q #p) the same analysisthat we havejust

appliedto J1(N,p). Here Xi(A, B) is thecurvecorresponding to F1(A)nro(B)
and J1(A, B) its Jacobian. First we need the analogue of Ihara's result. It is
convenientto workin a slightlymoregeneralsetting.Let us denotethe maps
Xi(Nqr-lqr) X(Nqr'-1) induced by z -- z and z -- qz by 7lrl, and 7r2,r
respectively. Similarly we denote the maps Xi (Nqr, qr+l) - X1(Nqr) induced
by z -+ z and z -? qz by 7r3,rand 7r4,rrespectively. Also let ir: Xi (Nqr) -
X1(Nqr-l, qr) denote the natural map induced by z -- z.

In the following lemmaifm is a maximalideal of Tl(Nqr-1) or T1(Nqr)
we use M(q) to denote the maximal ideal of Ti (Nqr, qr+l) compatiblewith
m, the ring T(q) (Nqr, qr+l) c T, (Nqr, qr+l) being the subring obtained by
omittingUq fromthe list of generators.
LEMMA2.5. If q $&
p is a prime and r > 1 thenthesequenceof abelian
varieties
0 -O Ji(Nqr-l) 1 J(Nqr) x J1(Nq) ) J1(Nq " qr+l)
where 6i = ((irr o 7r)*, - (7r2,ro 7r)*) and (? = (7r*r, 7r*,r) induces a corre-
spondingsequenceofp-divisiblegroupswhichbecomesexact whenlocalizedat
any M(q)for whichPmis irreducible.
Proof. Let Fl(Nqr) denotethe group ad E P1(N): a = d = l(qr),
C O(qr-1), b O(q)}. Let Bi and B1 be givenby
B1 = Fl(Nqr) jFi(Nqr) n r(q), B1 = f1(Nqr)/F (Nq r) n )
and let Aq = Fi(Nqr-l)/ri(Nqr) n r(q). Thus Aq ~ SL2 (Z/q) if r = 1 and

is of ordera powerof q ifr > 1.
The exact sequencesof inflation-restriction
give:
A1
H1 (Fi (Nqr) Q /Z)-*H1 (r1(Nqr) n F(q), Qp/Zp)B1
withA1replacingA1and B1 replacingB1.
togetherwitha similarisomorphism
We also obtain
H1 (F1(Nq r-l), Q /Z ) -H1 (F1 (Nqr) n F(q), QP/ZP)Aq.

494 ANDREW WILES
The vanishingof H2(SL2(Z/q), Qp/Zp) can be checkedby restricting to the

Sylowp-subgroupwhichis cyclic.Note thatimA1,nimAl C H1 (Fl (N qr)nr(q),
Qp/Zp)~'q since B1 and B1 togethergenerateAq. Now considerthe sequence
(2.13) 0 - H1(Fl(Nqr-l), Q /Z)
Hl (Fl (Nqr ,Qp/Zp)EDHl (rl (Nqr ,Qp/Zp)
>> , Hl (F1(Nqr) n F(q), Qp/Zp).

We claim it is exact. To check this, suppose that A1(x) = -A1(y). Then
Al(X) E Hl(Fi(Nqr) n F(q),QP/ZP)Aq. So Al(x) is the restrictionof an
x' E Hl (IF(Nqr-l), Qp/Zp) whence x - resi(x') E kerAl = 0. It follows
also that y =-res1 (x').
bythematrix(O?) inducesisomorphisms
Nowconjugation
rl(Nqr) -_F(Nqr), r
F(Nqr) nfF(q) - Fi(Nqr, qr+l).
So our sequence (2.13) yieldsthe exact sequenceofthe lemma,exceptthat we
have to change fromgroup cohomologyto the cohomologyof the associated
completecurves. If the groups are torsion-free then the differencebetween
thesecohomologiesis Eisenstein(morepreciselyT1- I-1 for1 _ 1 modNqr+l
is nilpotent)so will vanish when we localize at the preimageof M(q) in the
abstractHecke ringgeneratedas a polynomialringby all the standardHecke
operatorsexcludingTq. If M < 3 then the group r1(M) has torsion. For
M = 1,2,3 we can restrictto F(3), F(4), F(3), respectively, where the co-
homologyis Eisensteinas the corresponding curveshave genus zero and the
groupsare torsion-free. Thus one onlyneeds to checkthe action of the Hecke
operatorson thekernelsoftherestriction maps in thesethreeexceptionalcases.
This can be done explicitlyand again theyare Eisenstein.This completesthe
proofof the lemma. E
Let us denote the maps X1 (N, q) X1 (N) induced by z -* z and z -* qz
-*
by 7r1and 7r2respectively.Similarlywe denotethe maps X1 (N, q2) -* X1 (N, q)

induced by z -* z and z -* qz by 7r3and 7r4respectively.
From the lemma (with r = 1) and Ihara's result (2.10) we deduce that
thereis a sequence
(2.14) 0 -* J1(N) x J1(N) x J1(N) , Ji(Nq2)
where = (7rlo 7r3)*x (7r20 13)* x (7r2 14)* and that the induced map of p-
0
divisiblegroupsbecomesinjectiveafterlocalizationat M(q)'s whichcorrespond

to irreduciblepm's.By dualitywe obtain a sequence
J1(Nq2) - J1(N)3 -* 0
whichis 'surjective'on Tate modules in the same sense. More generallywe
can proveanalogous resultsforJH(N) and JH(N, q2) althoughtheremay be

a kernel of order divisible by p in JH(N) -* Ji (N). However this kernel will

not meetthe M(q)-divisiblegroupforany maximalideal m(q)whose associated
Pm is irreducibleand hence,as in the earliercases, will not affectthe resultsif
afterpassingto p-divisiblegroupswe localize at such an m(q). We use the same
notationin the generalcase whenH $&1 so ( is themap JH(N)3 -* JH(N, q2).
We suppose now that m is a maximalideal of TH(N) (as alwayswithp E
m) associated to an irreduciblerepresentation and that q is a prime,q { Np.
We now definea maximal ideal mq of TH(N, q2) with the same associated
representation as m. To do this considerthe ring
Si = TH(N)[Ui]/Ul(U?2-Tq U1 ? q(q)) C End (JH(N)3)
wherethe action of U1 on JH(N)3 is givenby the matrix

Tq -(q) 0
q 0 0
L q Oj
Then U1 satisfiesthe compatibility
0 Uq = Ul 0
One checksthis using the actions on cotangentspaces. For we may identify

the cotangentspaces withspaces ofcusp formsand withthisidentification
any
Heckeoperatort. inducesthe usual actionon cusp forms.There is a maximal
ideal ml = (U1, m) in Si and Slmi TH(N)m. We let mq denote the reciprocal
image of ml in TH(N, q2) under the natural map TH(N, q2) -* S1.
Next we definea principalideal (A') of TH(N)m using the fact that
TH(N, q2)m,and TH(N)m are both Gorensteinrings(cf. Corollary2 to The-
orem 2.1). Thus we set (A') = (a' o ') where
a': TH(N, q2)mq -
Slm l TH(N)m
is the naturalmap and a" is the adjointwithrespectto selectedHecke-module
pairingson TH(N, q2)mqand TH(N)m. Note that a' is surjective. To show
that the Tq operatoris in the imageone can use the existenceofthe associated
2-dimensionalrepresentation(cf. ?1) in whichTq = trace(Frobq) and apply
the Cebotarevdensitytheorem.
PROPOSITION 2.6. Suppose thatm is a maximalideal ofTH(N) associ-
Pm.Suppose also thatq { Np. Then
ated to an irreducible
(A') = (q - 1) (T2 - (q)(1 + q)2).
Proof. We provethis in the same manneras we provedProposition2.4.

Considerthe maps on p-adic Tate-modulesinducedby ( and (:
(2.15) Tap (JH(N)) Tap (JH (N, q2)) A Tap (JH(N)3)

496 ANDREW WILES
These maps commutewith the standardHecke operatorswith the exception

of Tq and Uq (whichare not even definedon all the terms). We define
S2 = TH(N)[U2]/U2(U22-Tq U2 + q(q)) C End (JH(N)3)
whereU2 is the endomorphism

of JH(N)3 givenby the matrix
q 0 -(q) .
0 q Tq
Then Uq( = ( U2 as one can verifybycheckingthe equality( o0) U2 = U1( o )

because ( o ( is an isogeny. The formulafor ( o ( is given below. Again
m2 = (i, U2) is a maximal ideal of S2 and S2,m2Z TH(N)m. On restricting
(2.15) to the M2,Mq and m1-adicTate moduleswe get
Tam2(JH(N)3) >Tamq(JH(N, q2) > Tam,(JH(N)3)

(2.16) U2 TU1
Tam(JH(N)) Tam(JH(N)).
The vertical isomorphisms are induced by u2: z ((q)z, -Tqz, qz) and -
ui: z > (0, 0, z). Now a calculation shows that on JH(N)3
q(q + 1) Tq *q Tq2-(q)(1 + q)
0of = Tq*- q q(q+ 1) Tq - q
Tq*2- (q)-'l ( + q)
q~~~~~~qq1 Tq**q q(q + 1
where Tq*= (q)-1Tq.
We computethenthat
(U1 o 2)=(q-l)(q 1)(Tq-(q)(1 + q))
Now using the surjectivityof ( and that ( has torsion-free

cokernelin (2.16)
(by Lemma 2.5) and that Tam (JH(N)) and Tam (JH(N, q2)) are each freeof
rank2 overthe respectiveHeckerings(Corollary1 ofTheorem2.1), we deduce
the resultas in Proposition2.4. [1
There is one further(and completelyelementary)generalizationof this
result. We let ir: XH(Nq,q2) -* XH(N,q2) be the map given by z -* z.
Then lr*: JH(N, q2) -* JH(Nq, q2) has kernel a cyclic group and as before
this will vanish when we localize at M(q) if m is associated to an irreducible
representation.(As beforethe superscriptq denotesthe omissionof Uq from
the list of generatorsof TH (Nq, q2) and M(q) denotes the maximal ideal of
T(q) (Nq, q2) compatible with m.)

We thus have a sequence (not necessarilyexact)

-* JH(N)3 -)JH(Nq q) ) Z 0
wherei = lr*o ( whichinducesa corresponding sequenceofp-divisiblegroups

whichbecomesexact whenlocalizedat an m(q)corresponding to an irreducible
Pm.Here Z is the quotientabelian varietyJH(Nq, q2)/ ims. As beforethere
is a naturalsurjectivehomomorphism
a: TH(Nq, q )mq -
Si)me TH(N)m
wheremq is the inverseimage of ml in TH(Nq, q2). (We note that one can
replace TH(Nq, q2) by TH(Nq2) in the definitionof a and Proposition2.7
below would still hold unchanged.) Since both ringsare again Gorensteinwe
can definean adjoint a' and a principalideal
(Aq) = (a o a')
PROPOSITION Suppose thatm is a maximalideal of T = TH(N)
2.7.
associatedto an irreduciblerepresentation.Suppose thatq { Np. Then
(Aq) = (q -
1)2 (T2 - (q)(1 + q)2))
The proofis a trivialgeneralizationof that of Proposition2.6.

Remark2.8. We have includedthe operatorUq in the definition of Tmq=
TH(Nq, q2)mqas in the applicationofthe q-expansionprincipleit is important
to have all the Heckeoperators.HoweverUq = 0 in Tmq. To see this we recall
that the absolute values of the eigenvaluesc(q, f ) of Uq on newformsof level
Nq with q { N are known(cf. [Li]). They satisfyc(q, f )2 = (q) in Of (the
ringof integersgeneratedby the Fouriercoefficients of f) if f is on Fl (N, q),
and c(q, f)I = q1/2 if f is on rF(Nq) but not on r1(N, q). Also when f is
a newformof level dividingN the roots of x2 - c(q, f )x + q xf (q) = 0 have
absolutevalue q1/2wherec(q, f) is the eigenvalueofTq and Xf(q) of (q). Since
forf on r1(Nqq2), Uqf is a formon Fr(Nq) we see that
Uq(Uq2 - (q)) JJ(Uq - c(qf)) f (u - c(q, f)Uq + q(q)) = 0

fESl fES2
in TH(Nq, q2) 0 C whereSi is the set of newforms on Ii (Nq) whichare not

on IF (N, q) and S2 is the set of newforms
of level dividingN. In particularas
Uq is in mqit mustbe zero in Tmq
situationarisesifmis a maximalideal ofT = TH(N, q)

A slightlydifferent
(q $ p) whichis not associatedto any maximalideal oflevelN (in the senseof
havingthe same associatedPm). In thiscase we mayuse the map 3 = (r4, 7r)
to give
(2.17) JH(N, q) x JH(N, q) JH(Nfq) -
JH(N,q) x JH(N,q).

498 ANDREW WILES
Then (3 o (3 is givenby the matrix

q Uq]
LUq qj
on JH(N, q)2, where Uq*= Uq(q)1- and TT2= (q) on the m-divisiblegroup. The
second of these formulaeis standardas mentionedabove; cf. forexample [Li,
Th. 3], sincePmis not associatedto any maximalideal oflevelN. For the first
considerany newformf of level divisibleby q and observethat the Petersson
innerproduct ((UUq - l)f(rz), f(mz)) is zero forany r,m I (Nq/ levelf)
by [Li, Th. 3]. This shows that Uq*Uqf(rz),a priori a linear combinationof
f(miz), is equal to f(rz). So Uq*Uq= 1 on the space of formson rH(N,q)
which are new at q, i.e. the space spanned by forms{f(sz)} where f runs
throughnewformswith q I levelf. In particularUq*preservesthe m-divisible
groupand satisfiesthe same relationon it, again because Pmis not associated
to any maximalideal of level N.
Remark2.9. Assume that Pmis of type (A) at q in the terminologyof

Chapter 1, ?1 (which ensuresthat Pmdoes not occur at level N). In this
case Tm = TH(N, q)m is already generatedby the standardHecke operators
withthe omissionof Uq. To see this,considerthe GL2 (Tm) representation of
Gal(Q/Q) associatedto the m-adicTate moduleof JH(N, q) (cf.the discussion
followingCorollary2 of Theorem 2.1). Then this representationis already
definedoverthe ZP-subalgebraTt' ofTm generatedby the tracesofFrobenius
elements,i.e. by the Te fore { Nqp. In particular(q) E Tt'. Furthermore, as
Tt' is local and complete,and as TT2= (q), it is enough to solve X2 = (q)
in the residuefieldof Tt'. But we can even do this in ko (the minimalfield
of definitionof Pm) by lettingX be the eigenvalueof Frobq on the unique
unramifiedrank-onefreequotientof ko and invokingthe lrq 7r(aq) theorem
of Langlands (cf. [Cal]). (It is to ensurethat the unramifiedquotientis free
of rankone that we assume Pmto be of type (A).)
We assumenow that Pmis oftype (A) at q. DefineSi thistimeby setting
= TH(N, Uq) C End

Si (JH(N, q)2)
whereU1 is givenby the matrix
q)[Ui]/Ui(Ui-
(2.18) U, =
on JH(N, q)2. The map (3 is not necessarilysurjectiveand to remedythis we

introduceM(q) = mn T(q)(N, q) whereT(q) (N, q) is the subringof TH (N, q)
generatedbythestandardHeckeoperatorsbut omittingUq. We also writeM(q)

forthe correspondingmaximal ideal of T )(Nq, q2). Then on M(q)-divisible

groups, 3 and (3 o 7r*are surjectiveand we get a natural restrictionmap of
localizationsTH(Nq, q2)(m(q)) I Sl(m(q)). (Note that the image of Uq under
thismap is U1 and not Uq.) The ideal ml = (m,U1) is maximalin S1 and so also
in Si,(m(q)) and we let mqdenotethe inverseimage of ml underthis restriction
map. The inverseimageofmqin TH(Nq, q2) is also a maximalideal whichwe
again write mq. Since the completions TH(Nq, q2)mqand Si,m, - TH(N, q)m
are both Gorensteinrings (by Corollary2 of Theorem 2.1) we can definea
principal ideal (Aq) of TH(N, q)m by
(Aq) = (a o )
where a: TH(Nq, q2)mq_*Sl,ml - TH(N, q)m is the restriction map induced
by the restriction
map on M(q)-localizations
describedabove.
PROPOSITION 2.10. Suppose that m is a maximal ideal of TH(N, q)
associatedto an irreducible
m of type(A). Then
(Aq) = (q - 1)2 (q + 1).
Proof. The methodis a straightforwardadaptationofthatused forPropo-

sitions2.4 and 2.6. We let S2 = TH(N, q)[U2]/U2(U2- Uq) be the ringof
endomorphisms of JH(N, q)2 whereU2 is givenby the matrix
[Uq qg]
This satisfies the compatibility 43U2 = Uq 43. We define m2 = (m,U2) in S2

and observethat S2,m2 - TH(N, q)m.
Then we have maps
1ro0 /30-r
Tam2(JH(N, q)2) Tanmq
(JH(Nq, q2)) Tam,(JH(Nt q) )
TI V2 T I vi
Tam(JH(N, q)) Tam(JH(N, q)).

The maps vi and v2 are given by v2: z -* (-qz, aqz) and vi: z -* (z, 0)
where Uq = aq in TH(N, q)m. One checks then that v1 o0(c3o7r*)o(7r*oc3)ov2
is equal to -(q - 1) (q2 1) or -2(q - 1)(q2 _ 1).
-
of$3 o7r*on the completionsis equivalentto the statement

The surjectivity
that
JH(Nq, q2) [p]mq JH(Nj q) 2[P]ml
is surjective.We can replace this conditionby a similarone withM(q) substi-
tuted formqand forml, i.e., the surjectivityof
JH(Nq, q2)[P]m(q) ) JH(N, q)2 [P]m(q)

500 ANDREW WILES
By our hypothesisthat Pmbe oftype (A) at q it is even sufficient

to showthat
the cokernelof JH(Nq, q2)[p](? Fp- JH(N, q)2[p]?&Fp has no subquotientas
a Galois-modulewhichis irreducible,two-dimensional and ramifiedat q. This
statement,or ratherits dual, followsfromLemma 2.5. The injectivity of 7r*
063
on the completionsand the fact that it has torsion-freecokernelalso follows
fromLemma 2.5 and our hypothesisthat Pmbe of type (A) at q. E
The case that correspondsto type (B) is similar.We assume in the anal-
ysis of type (B) (and also of type (C) below) that H decomposesas HIHq as
describedat the beginningof Section 1. We assume that m is a maximalideal
of TH(Nqr) whereH containsthe Sylowp-subgroupSp of (Z/qrZ)* and that
(2.19) Pm 1)
(Xq
fora suitablechoiceof basis withXq $&1 and condXq = qr. Here q { Np and
we assume also that Pmis irreducible.We use the sequence
*
(7r/) 06 r r~~l ~2O07r' xJHN
JH(Nq') x JH(Nqr ) I* JH,(Nqrq r+l) JxJ(
J )
definedanalogouslyto (2.17) where62 was as definedin Lemma 2.5 and where

H' is definedas follows.Using the notationH = rHH1 as at the beginningof
Section1 set Hl = H1 for1$&q and Hq x Sp = Hq. Then defineH'= rl Hl and
let ir' XHi(Nqr, qr+l) -+ XH(Nqr, qr+l) be the natural map z z. Using
Lemma 2.5 we checkthat 62 is injectiveon the M(q)-divisible
group. Again we
set S1 = TH(Nqr) [Ul]/Ul (U1 - Uq) C End(JH(Nqr)2) where U1 is given by
the matrix in (2.18). We define ml = (m, U1) and let mq be the inverse image
of ml in T.H(Nqr, qr+l). The natural map (in which Uq -+ Ui)
a: TH,(Nqr, qr+ )mq iSiml ~ TH(Nqr)m
is surjectiveby the followingremark.
Remark2.11. Whenwe assumethatPmis oftype(B) thenthe Uq operator

is redundantin Tm = TH(Nqr)m. To see this,firstassume that Tm is reduced
and considerthe GL2(Tm) representation of Gal(Q/Q) associated to the m-
adic Tate module. Pick a aq E Iq, the inertiagroupin Dq in Gal(Q/Q), such
that Xq(aq) $&1. Then because the eigenvaluesofUq are distinctmodm we can
diagonalizetherepresentation withrespectto aq. IfFrobq is a Frobeniusin Dq,
thenin the GL2(Tm) representation the image of Frobq normalizesIq and we
can recoverUq as the entryofthe matrixgivingthe value ofFrobq on the unit
eigenvectorforaq. This is by the lrq (iq)X theoremof Langlands as before
(cf. [Cal]) applied to each of the representations obtained frommaps Tm
(9f,A. Since the representation is definedoverthe Zp-algebraTtr generatedby
the traces,the same reasoningapplied to TM showsthat Uq E Tm.

If Tm is not reduced the above argumentshows only that there is an

operatorvq E Ttr such that (Uq - vq) is nilpotent. Now TH(Nqr) can be
viewed as a ringof endomorphisms of S2(FH(Nqr)), the space of cusp forms
of weight2 on FH(Nqr). There is a restriction map TH(Nqr) -? TH(Nqr)new
whereTH(Nqr)new is the image of TH(Nqr) in the ringof endomorphisms of
S2(FH(Nqr))/S2(FH(Nqr))old, the old part being definedas the sum of two
copies of S2(Fr(Nqr-l)) mapped via z -? z and z -? qz. One sees that on
m-completionsTm - (TH(Nqr)new)m since the conductor of Pmis divisible by
qr. It followsthat Uq E Tm satisfiesan equation of the formP(Uq) = 0 where
P(x) is a polynomialwith coefficients in W(km) and with distinctroots. By
extendingscalars to (9 (the integersof a local fieldcontainingW(km)) we can
assume that the rootslie in T -i Tm 0 (9.
W(km)
Since (Uq - vq) is nilpotentit followsthat P(vq)r = 0 forsome r. Then
since vq E Ttr whichis reduced,P(vq) = 0. Now considerthe map T -? HT(p)
wherethe productis taken over the localizationsof T at the minimalprimes
p of T. The map is injectivesince the associated primesof the kernelare all
maximal,whence the kernelis of finitecardinalityand hence zero. Now in
each T(p), Uq = ai and Vq = aj for roots aj, aj of P(x) = 0 because the roots
are distinct. Since Uq - Vq E p foreach p it followsthat ai = aj foreach p
whence Uq = Vq in each T(p). Hence Uq = Vq in T also and this finallyshows
that Uq E Ttr in general.
definea principalideal
We can therefore
(Lq) = (a o
using,as previously, thattheringsTH' (Nqr, qr+l)M and TH(Nqr)m are Goren-
stein. We compute(Aq) in a similarmannerto the type (A) case, but using
this timethat Uq*Uq = q on the space of formson FH(Nqr) whichare new at
q, i.e., the space spanned by forms{f(sz)} where f runs throughnewforms
withqr I levelf . To see this let f be any newformof level divisibleby qr and
observethat the Peterssoninnerproduct ((UUq - q)f (rz), f (mz)) = 0 for
any m I (Nqr/ levelf ) by [Li, Th. 3(ii)]. This shows that (Uq*Uq- q) f (rz),
a priori a linear combinationof {f(miz)}, is zero. We obtain the following
result.
PROPOSITION 2.12. Suppose that m is a maximal ideal of TH(Nqr)
associatedto an irreducible
Pmof type(B) at q, i.e., satisfying
(2.19) including
thehypothesisthatH containsSp. (Again q t Np.) Then
(LAq) = ((q - 1)2 ).
Finally we have the case wherePmis of type (C) at q. We assume then

that mis a maximalideal of TH(Nqr) whereH containsthe Sylowp-subgroup

502 ANDREW WILES
Sp of (Z/q'Z)* and that

(2.20) H1(Qqi WA)= 0
whereWAis definedas in (1.6) but withPmreplacingpO,i.e., WA= ado Pm.
This time we let mq be the inverseimage of m in THI (Nqr) under the
natural restrictionmap THI (Nqr) ) TH(Nqr) with H' definedas in the
case of type B. We set
(Aq) = (a(o&)
where a: TH' (Nq')m TH(Nq') is the induced map on the completions,
whichas beforeare Gorensteinrings. The proofof the followingproposition
is analogous (but simpler)to the proofof Proposition2.10. (Notice that the
propositiondoes not requirethe conditionthat Pmsatisfy(2.20) but thisis the
case in whichwe will use it.)
PROPOSITION 2.13. Supposethatm is a maximalideal ofTH(Nqr) asso-
PmwithH containingtheSylowp-subgroupof(Z/qrZ)*.
ciatedto an irreducible
Then
(Aq) = (q- 1).
Finally,in this sectionwe state Proposition2.4 in the case q ? p as this

will be used in Chapter3. Let q be a prime,q { Np and let Si denotethe ring
(2.21) TH(N)[Ul] /{U2 -

TqUi + (q)q} C End(JH(N)2)
where ): JH(N, q) -- JH(N)2 is the map defined after (2.10) and Ui is the
matrix
Tq - (q)
Lq ? i
Thus, 'Uq = U1. Also (q) is definedas (nq) wherenq _ q(N), nq l(q).
Let ml be a maximal ideal of Si containing the image of m, where m is a
maximalideal of TH(N) with associated irreduciblePm.We will also assume
that pm(Frobq) has distincteigenvalues. (We will only need this case and
it simplifiesthe exposition.) Let mq denote the correspondingmaximal ide-
als of TH(N, q) and TH(Nq) underthe naturalrestrictionmaps TH(Nq)
TH(N, q) -- SI. The corresponding maps on completionsare
(2.22) TH(Nq)m, A TH(N, q)m,
Siml - TH(N)m 0 W(k )

W(km)
wherek+ is the extensionof kmgeneratedby the eigenvaluesof {pm(Frobq)}.

Thus k+ is eitherequal to kmor its quadratic extension. The maps A, a are
surjective,the latterbecause Tq is a trace in the 2-dimensionalrepresentation

to GL2(TH(N)m) given afterTheorem 2.1 and hence is 'redundant'by the

Cebotarevdensitytheorem.The completionsare Gorensteinby Corollary2 to
Theorem2.1 and so we defineinvariantideals of Si,m1
(2.23) (A) = (aUo &),

O,(). (A') = (aUof6) o (C
Let aq be the imageofU1 in TH(N)m 09 W(k+) underthe last isomorphism
W(km)
in (2.22). The proofof Proposition2.4 yields
PROPOSITION Suppose thatPmis irreduciblewherem is a maximal

2.4'.
ideal ofTH(N) and thatpm(iFrobq) has distincteigenvalues.Then
(A) = (agq-(a2
-
)X
(A/) = (c(2 _ (q))(q - 1).
Remark. Note that ifwe suppose also that q 1(p) then (i\) is the unit
ideal and a is an isomorphismin (2.22).
3. The main conjectures
As we suggestedin Chapter 1, in orderto studythe deformation-theory

of po in detail we need to assume that it is modular. That this should always
be so fordetpo odd was conjecturedby Serre. Serrealso made a conjecture
(the 'e'-conjecture)makingprecisewhereone could finda liftingof po once
one assumed it to be modular (cf. [Se]). This has now been provedby the
combinedeffortsof a numberof authors includingRibet, Mazur, Carayol,
Edixhovenand others. The most difficultstep was to show that if po was
unramified at a prime1 thenone could finda liftingin which1 did not divide
the level. This was proved(in slightlyless generality)by Ribet. For a precise
statementand completereferenceswe referto Diamond's paper [Dia] which
removedthe last restrictionsreferredto in Ribet's surveyarticle [Ri3]. The
following is a minoradaptationofthe epsilonconjectureto our situationwhich
can be foundin [Dia, Th. 6.4]. (We wish to use weight2 only.) Let N(po) be
the primeto p part of the conductorof po as definedforexamplein [Se].
THEOREM 2.14. Suppose thatpo is modularand satisfies(1.1) (so in

particularis irreducible)and is of typeVD= (.,2,O,M) with. = Se, str or fl.
Suppose thatat least one of thefollowingconditionsholds(i) p > 3 or (ii) po
is not inducedfroma characterof Q(V'/=). Then thereexistsa newformf
of weight2 and a prime A of Of such thatpf,> is of typeV' = (,EO',M)
for some O', and such that (pf,Amod A) - po over Fp. Moreoverwe can
assume thatf has characterXf of orderprime to p and has levelN(po)p6(PO)

504 ANDREW WILES
whereb(Po) = 0 if POIDp is associated to a finiteflat groupscheme over Zp

and detpo = aw,and 6(po) = 1 otherwise.Furthermore in the Selmer case
we can assume thatap(f) X2(FRobp) mod A in the notationof (1.2) where
ap(f) is the eigenvalueof Up.
For the rest of this chapterwe will assume that po is modular and that
ifp = 3 then po is not inducedfroma characterof Q(j/=S). Here and in the
restof the paper we use the term'induced' to signifythat the representation
is inducedafteran extensionof scalars to the algebraicclosure.
For each D = {., E, (, M} we will now definea Hecke ring TD except
where is unrestricted.Suppose firstthat we are in the flat,Selmeror strict
cases. Recall that when referring to the flat case we assume that po is not
ordinary and that detpolI = W. Suppose that E {qj} and that N(po) =
H qi with si > 0. If UA,- k2 is the representation space of po we set nq =
dimk(U>,)Iq where Iq is the inertiagroup at q. Define MO and M by
(2.24) MO= N(po) 11 qj* 1 M
=Mzp(PO)
qi,
nqi=l nq =2
qj FMU{p}
where r(po) = 1 if po is ordinary and r(po) = 0 otherwise. Let H be the

subgroupof (Z/MZ)* generatedby the Sylowp-subgroupof (Z/qiZ)* foreach
qi E M as well as by all of (Z/qiZ)* foreach qi E M of type (A). Let T/(M)
denotethe ringgeneratedby the standardHeckeoperators{T1 for1 t Mp, (a)
for(a, Mp) = 1}. Let m' denotethe maximalideal ofT'(M) associatedto the
f and A givenin the theoremand let kmibe the residuefieldT'(M)/m'. Note
that m' does not dependon the particularchoiceofpair (f, A) in theorem2.14.
Then km_ '- kowherekois thesmallestpossiblefieldofdefinition forpo because
kmiis generatedby the traces. Henceforth we will identifyko withkm'.There
is one exceptionalcase wherepo is ordinaryand POIDp is isomorphicto a sum
oftwo distinctunramified characters(X1 and X2 in the notationof Chapter 1,
?1). If po is not exceptionalwe define
(2.25(a)) YD = T (M)m' 0 (9.
W(ko)
If po is exceptionalwe let T4(M) denotethe ringgeneratedby the operators

{T1 for1 t Mp, (a) for (a, Mp) = 1, Up}. We choose m" to be a maximal
ideal of TV(M) lyingabove m' forwhichthereis an embeddingkmno e 4 k (over
ko = kmi)satisfyingUp -- X2(Frobp). (Note that X2 is specifiedby D.) Then
in the exceptionalcase km"e is eitherkoor its quadraticextensionand we define
(2.25(b)) Tv= T(M)mi 0 (9.

The o soW(k) f ..
The omissionofthe HeckeoperatorsUq forq I MO ensuresthat TD is reduced.

We need to relateTD to a Hecke ringwithno missingoperatorsin order

to apply the resultsof Section 1.
PROPOSITION 2.15. In the nonexceptionalcase thereis a maximalideal

m for TH(M) withmn T' (M) = m' and ko = km,)and such thatthe natural
map T' (M)m, > TH(M)m is an isomorphism, thus giving
TD2 TH(M)m 0 (9O

W(ko)
In the exceptional case the same statements hold withm" replacing m', TV (M)
replacingT'(M) and kintreplacingko.
Proof. For simplicity

we describethenonexceptionalcase indicatingwhere
appropriatethe slightmodificationsneeded in the exceptionalcase. To con-
structmwe taketheeigenform foobtainedfromthenewform f ofTheorem2.14
by removingthe Euler factorsat all primesq E - {M U p}. If po is ordinary
and f has level prime to p we also remove the Euler factor (1 - .p p-s) where
3p is the non-uniteigenvaluein (Of,,. (By 'removingEuler factors'we mean
take the eigenformwhose L-series is that of f with these Euler factorsre-
moved.) Then fo is an eigenform ofweight2 on FH(M) (thisis ensuredby the
choice of f) with Of,, coefficients. We have a correspondinghomomorphism
lrf0: TH(M) -* O9f,Aand we let m 7r-i(A).
=
Since the Hecke operatorswe have used to generateT'(M) are primeto
the level thereis an inclusionwithfiniteindex
T' (M)c+fl O9g
whereg runsoverrepresentatives of the Galois conjugacyclasses of newforms

associatedto FH(M) and wherewe notethatbymultiplicity one O., can also be
describedas the ringof integersgeneratedby the eigenvaluesof the operators
in T'(M) acting on g. If we considerTH(M) in place of T'(M) we get a
similarmap but we have to replacethe ring0g by the ring
Sg = 09g[Xql ..., Xqr, XP]/{Yi, zp}tj1
where{p, qj,... , qr } are the distinctprimesdividingMp. Here
( Xqr2i f
= (Xqj - qi(g)) (Xqi - qi (9()) if qj t level(g)
Xqrii(Xqi -aqi(g)) if qi I level(g),
where the Euler factor of g at qj (i.e., of its associated L-series) is

(1-Oqi (g)qTS) (1- !q (g)q S) in the firstcase and (1- aqj (g)qi-S) in the second
case, and qijll(M/ level(g)). (We allow aqi(g) to be zero here.) SimilarlyZp is

506 ANDREW WILES
definedby
J Xp2- + pXg(P) ifp M, p t level(g)
ap(g)XXp
Zp =
1
Xp-ap(g) ifp{M
Xp- ap(g) ifplevel(g),
wherethe Euler factorof g at p is (1 - ap(g)p-8+ Xg(p)pl-2.) in the first
two cases and (1 - ap(g)p-8) in the thirdcase. We then have a commutative
diagram
T',(M) C0 H Og
(2.27) f f
TH(M) C HSg =
r 09g[Xql ,***Xqr, Xp]/{Yi, Zp}1
9 9
where the lower map is given on {Uqi, Up or Tp} by Uq% o Xqj, Up or
Tp ) Xp (accordingas p I M or p t M). To verifythe existenceof such
a homomorphism one considersthe action of TH(M) on the space of formsof
weight2 invariantunderrH(M) and uses that I>r=1gj(mjz) is a freegener-
ator as a TH(M) 0 C-module where{gj} runs overthe set of newformsand
mj = M/level(gj).
Now we tensorall the ringsin (2.27) with Zp. Then completingthe top
rowof (2.27) withrespectto m' and the bottomrowwithrespectto m we get
a commutativediagram
T' (M)mi 5
( 0U))m r 9
(2.28) { { I
TH(M)m C ( Sg) ( H(Sg)m.
Here IL runs throughthe primesabove p in each 0g forwhichm' IL under

TH'(M) X)9. Now (Sg)mis givenby
(2.29) (Sg 0 Zp)m - ((Og 0 Zp) [Xq X
I... *Xqr Xp]/{YiX Zp}i )m
\ilp Jm
whereA9,j denotesthe productof the factorsof the completesemi-localring
X***Xqr,Xp]/{Yi, Zp};r1 in whichXqj is topologicallynilpotentfor
Qgqs[Xq1,.

qj V M and in whichXp is a unitifwe are in the ordinary case (i.e., when

p IM). This is because Uqj E m ifqj V M and Up is a unit at m in theordinary
case.
Now if m' -u then in (Ag-,,)mwe claimthatYi is givenup to a unitby
Xqi- biforsomebi E 0g,,,withbi= 0 ifqj V M. Similarly Zp is givenup to a
unitbyXp - ap(g) whereap(g) is theunitrootofx2-ap(g)x + PXg(P)= 0 in
0g,4 ifp t levelg and p I M and byXp - ap(g) ifp levelg orp t M. This
willshowthat(Ag,,)m- (g,ji whenm' --+ILand (Ag,,i)m = 0 otherwise.
Forqi E M andforp, theclaimis straightforward. Forqj V M, it amounts
to the following.Let Uglydenotethe 2-dimensional Kg,'vectorspace with
Galoisactionvia pgq and let nqi(g,,u) = dim(Ug,,,)'qi. We wishto checkthat
Yj = unit.Xqj in (Ag,,i)mand from the definitionof Yi in (2.26) thisreduces
to checking = -
thatri nqi(g,,u)bythelrq ir(oq) theorem(cf.[Cal]). We use
herethatcxq2 (9), /qj(g) andaqi(g) arep-adicunitswhentheyarenonzerosince
theyare eigenvalues ofFrobqi. Nowbydefinition thepowerofqi dividingM
is thesameas thatdividing N(po)qji (cf.(2.21)). By an observation ofLivne
(cf.[Liv],[Ca2,?1]),
(2.30) (levelg) = ordqi(N(po)qini
ordqi (q,)) fqi
As bydefinitionqirII(M/levelg) we deducethatri = nqi(g,I) as required.
We havenowshownthateachAg, - 09g,4(whenm'-* iL) and it follows
from(2.28) and (2.29) thatwe havehomomorphisms
T1 (M)mt C TH(M)m )9gj J
wheretheinclusions are offiniteindex.Moreover we haveseenthatUqi= 0
in TH(M)m forqj V M. We nowconsiderthe primesqi E M. We have
to showthatthe operatorsUq forq E M are redundant in the sensethat
theylie in T' (M)m', i.e., in the ZP-subalgebra of TH(M)m generated by the
{Ti: 1 t Mp, (a): a E (Z/MZ)*}. For q E M of type (A), Uq E T' (M)m'
as explained in Remark 2.9 and forq E M of type (B), Uq E T/(M)m' as
explainedin Remark2.11. Forq E M oftype(C) butnotoftype(A), Uq= 0
by the lrq - ir(aq) theorem(cf. [Cal]). For in thiscase nq = 0 whencealso
nq(g,,) = 0 foreach pair (g,1L)withm' -+ IL. If po is strictor Selmerat p then
Upcanbe recovered from thetwo-dimensional representation p (describedafter
to Theorem2.1) as theeigenvalue
thecorollaries ofFrobpon the(free,ofrank
one) unramified quotient(cf.Theorem2.1.4of [Wil]). As thisrepresentation
is definedoverthe Zp-subalgebrageneratedby the traces,it followsthatUp
is containedin thissubring.In the exceptionalcase Up is in TV4(M)mttby
definition.
Finallywe haveto showthatTpis also redundant in thesenseexplained
{
abovewhenp M. A proofofthishasalreadybeengiveninSection2 (Ribet's

508 ANDREW WILES
lemma). Here we give an alternativeargumentusing the Galois representa-

tions. We knowthat Tp E m and it will be enoughto show that Tp E (m2,p).
Writingkmforthe residuefieldTH(M)m/m we reduceto the followingsitua-
tion. If Tp V (mi2,p) then there is a quotient
TH(M)m/(m , P)+km[E] = TH(M)m/a

wherekm[E]is the ringof dual numbers(so E2 = 0) with the propertythat
Tp As withA $&0 and such that the image of T' (M)m, lies in km.Let G/Q
denotethe four-dimensional space associated to the representation
km-vector
Pe: Gal(Q/Q) ) GL2 (km[6])

in Theorem2.1. It has the form
inducedfromthe representation
G/Qc?GO/Q ED GO/Q
whereGO is the correspondingspace associated to po by our hypothesisthat

the traceslie in km.The semisimplicityofG/Q hereis obtainedfromthe main
theoremof [BLR]. Now G/QP extends to a finiteflat group scheme GIZP.
Explicitlyit is a quotientofthe groupschemeJH(M)m[p]/Zp.Since extensions
to Zp are unique (cf. [Rayl]) we know
G/ZP -G/Z eGo/Z

Now by the Eichler-Shimura
relationwe knowthat in JH(M)/F
Tp = F + (p)F
Since Tp E m it followsthat F + (p)FT = 0 on Go/F and hencethe same holds
on G/FP. But Tp is an endomorphismof G/ZP whichis zero on the special
fibre,so by [Rayl, Cor. 3.3.6], Tp = 0 on G/ZP. It followsthat Tp = 0 in km[E]
whichcontradictsour earlierhypothesis. So Tp E (m2,p) as required. This
completesthe proofof the proposition. Oi
FRomthe proofofthe propositionit is also clear thatmis the uniquemax-
imal ideal of TH(M) extendingm' and satisfyingthe conditionsthat Uq E m
forq e E - {M U p} and Up V m ifpo is ordinary.For the restof this chapter
we will alwaysmakethis choiceof m (givenpo).
Next we defineTD in the case when D = (ord, ,Q,M). If n is any
ordinarymaximalideal (i.e. Up V n) of TH(Np) withN primeto p thenHida
has constructeda 2-dimensionalNoetherianlocal Hecke ring
Too = e TH(Np')n:= lim e TH(Npr)nr
which is a A = ZpJ[T-algebra satisfyingTO/T - TH(Np)n. Here nr is the

map. Also T = lim(1+ Np) -1
inverseimageofn underthe naturalrestriction

and e = limU!. For an irreduciblepo of type D we have definedTpD in

r
(2.25(a)), whereD' = (Se , (IM), by
YD TH(MOP)m 0 (9,
W(km)
the isomorphismcomingfromProposition2.15. We will defineTD by
(2.31) YD = eTH(MOPc?)m 0 (9)

W(km)
In particularwe see that
(2.32) TD YDIT',
i.e., whereD' is the same as D but with 'Selmer' replacing'ord'. Moreoverif
q is a heightone primeideal of TD containing((1 + T)P' - (1 + NP)Pn(k-2))
forany integersn > 0 k > 2, then TD/q is associated to an eigenformin a
naturalway (generalizingthe case n = 0, k = 2). For moredetails about these
ringsas well as about A-adic modularformssee forexample [Wil] or [Hil].
For each n > 1 let Tn = TH(Mopl),n. Then by the argumentgiven
afterthe statementofTheorem2.1 we can constructa Galois representation Pn
unramified outsideMp withvaluesin GL2(Tn) satisfying tracep (Frob 1) = T1,
detp, (Frobl) = 1(1) for(1, Mp) = 1. These representations can be patched
togetherto give a continuousrepresentation
(2.33) p = limp,: Gal(Q/Q) - GL2(TD)
where 3 is the set of primesdividingMp. To see this we need to checkthe

commutativity of the maps
Rs-> Tn
Tn-1
wherethe horizontalmaps are inducedby Pn and Pn-1 and the verticalmap is

the naturalone. Now the commutativity is valid on elementsof RE, whichare
tracesor determinants in the universalrepresentation, sincetrace (Frob1) |-4 T1
underboth horizontalmaps and similarlyfordeterminants.Here RE is the
universaldeformationringdescribedin Chapter 1 with respectto po viewed
withresiduefieldk = km.It suffices thento showthat RE is generated(topo-
logically)by tracesand thisreducesto checkingthat thereare no nonconstant
deformations ofpo to k[E]withtraceslyingin k (cf. [Mal, ?1.8]). For thenifRtr
denotesthe closed W(k)-subalgebraof RE generatedby the traceswe see that
Rtr --(RE/m2) is surjective,m being the maximalideal of RE, fromwhich
we easily conclude that Rtr = RE. To see that the conditionholds, assume

510 ANDREW WILES
that a basis is chosenso that po(c) = (1 _0)fora chosencomplexconjugation

c and poG(o) with b = 1 and c, 54 0 forsome a. (This is possible
because po is irreducible.)Then any deformation [p]to k[e]can be represented
by a representation p such that p(c) and p(or) have the same properties. It
followseasily that if the traces of p lie in k then p takes values in k whence
it is equal to po. (Alternatively one sees that the universalrepresentationcan
be definedoverRtrby diagonalizingcomplexconjugationas before.Since the
two maps Rt-+ Tn_- induced by the triangleare the same, so the associ-
ated representations are equivalent,and the universalpropertythen implies
the commutativity of the triangle.)
The representations (2.33) werefirstexhibitedby Hida and werethe orig-
inal inspirationforMazur's deformation theory.
For each D = {., 3,0, M} where is not unrestrictedthere is then a
canonicalsurjectivemap
Wp : RD YTD
whichinducestherepresentations describedafterthecorollariesto Theorem2.1

and in (2.33). It is enoughto checkthis when (9 = W(ko) (or W(km,,)in the
exceptionalcase). Then one just has to checkthat foreverypair (g,pa)which
is of typeD. For thenwe claim
appears in (2.28) the resultingrepresentation
that the imageofthe canonicalmap RpD TO = II 0g9,,is TD wherehere
denotesthe normalization.(In the case where*is ord thisneeds to be checked
instead forT, 0 (9 foreach n.) For this we just need to see that RD is
W(ko)
generatedby traces. (In the exceptionalcase we have to show also that Up is
usingTheorem2.1.4 of
in the image. This holds because it can be identified,
[Wil], with the image of u E RD whereu is the eigenvalueof Frobp on the
uniquerankone unramified quotientofRD witheigenvalue X2 (Frobp) which
is specifiedin the definitionof D.) But we saw above that this was true for
RE. The same thenholds forRD as RE RD is surjectivebecause the map
-
on reduced cotangentspaces is surjective(cf. (1.5)). To checkthe condition

on the pairs (g,Iu) observefirstthat forq E M we have imposedthe following
conditionson the level and characterof such g's by our choiceof M and H:
q of type (A): qIIlevelg, detpg,|I = 1,
q of type (B): condXqlIlevelg,detpg,| I = Xq,
q of type (C): det pg,tI is the Teichmiiller liftingof det po I.
In the firsttwo cases the desiredformof PgqDi then followsfromthe

1rq 'K(0q) theoremof Langlands (cf. [Cal]). The thirdcase is already of

type (C). For q = p one can use Theorem2.1.4 of [Wil] in the ordinarycase,
the flatcase beingwell-known.
The followingconjecturegeneralizesa fundamentalconjectureof Mazur
and Tilouine forV = (ord,A, W(ko), 0); cf. [MT].
CONJECTURE 2.16. FpD is an isomorphism.
Equivalentlythis conjecturesays that the representation describedafter

thecorollariesto Theorem2.1 (or in (2.33) in theordinarycase) is the universal
one fora suitablechoiceofH, N and m. We remindthe readerthatthroughout
this section we are assumingthat if p = 3 then po is not induced froma
characterof Q(+-/=).
Remark. The case of most interestto us is whenp = 3 and po is a rep-
resentationwith values in GL2(F3). In this case it is a theoremof Tunnell,
extendingresultsof Langlands,that po is alwaysmodular. For GL2(F3) is a
double coverof S4 and can be embeddedin GL2(Z[VA/=])whencein GL2(C);
cf. [Se] and [Tu]. The conjecturewill be provedwitha mild restrictionon po
at the end of Chapter3.
Remark. Our originalrestrictionto the types (A), (B), (C) forpo was
motivatedby the wish that the deformationtype (a) be of minimalconduc-
tor among its twists,(b) retainproperty(a) underunramifiedbase changes.
Withoutthiskindofstabilityit can happenthataftera base changeofQ to an
extensionunramifiedat A, po 0 ' has smaller'conductor'forsome character
A. The typicalexampleof thisis wherePo D = IndQ (x) withq -I(p) and
X is a ramifiedcharacterover K, the unramifiedquadratic extensionof Qq.
What makesthisdifficult forus is that thereare thennontrivialramifiedlocal
deformations(IndQPXf for( a ramifiedcharacterof orderp of K) whichwe
cannotdetectby a changeof level.
For the purposesof Chapter 3 it is convenientto digressnow in orderto
introducea slightvariantof the deformation ringswe have been considering
so far. Suppose that D = ( *, (9, M) is a standard deformationproblem
(associated to po) with * = Se, str or fl and suppose that H, MO, M and m
are definedas in (2.24) and Proposition2.15. We choose a finiteset of primes
Q = {q1, ... , qr} with qi { Mp. Furthermorewe assume that each qi _ 1(p)
and that the eigenvalues{ci, /i} of po(Frobqi) are distinctforeach qi E Q.
This last conditionensuresthatpo does notoccuras theresidualrepresentation
of the A-adic representation associated to any newformon FH(M, q..qr)
whereany qi dividesthe levelofthe form.This can be seen directlyby looking
at (Frobqi) in such a representation or by usingProposition2.4' at the end of
Section 2. It will be convenientto assume that the residuefieldof (9 contains
si, /3ifor each qi.

512 ANDREW WILES
problemassociated to
Pick ai foreach i. We let DQ be the deformation
representationsp of Gal(QsuQ/Q) whichare oftypeD and whichin addition
satisfythe propertythat at each qj E Q
(2.34) P _,( X
X Xlqi )
with X2,qiunramifiedand X2,qi (Frobqj) -i mod m fora suitable choice of
basis. One checksas in Chapter 1 that associated to DQ thereis a universal
deformation ringRQ. (These new conditionsare reallyvariantson type (B).)
We willonlyneed a corresponding Heckeringin a veryspecial case and it
is convenientin thiscase to defineit usingall the Heckeoperators.Let us now
set N = N(po)p6(Po)where6(po) is as definedin Theorem2.14. Let modenote
a maximal ideal of TH(N) given by Theorem 2.14 with the propertythat
pmo PO over Fp relative to a suitable embedding of kM0-+ k over ko. (In the
exceptionalcase we also imposethe same conditionon moabout the reduction
of Up as in the definition
ofTD in the exceptionalcase before(2.25)(b).) Thus
pmo pf,A,mod A over the residue fieldof OfA, forsome choice of f and A
with f of level N. By droppingone of the Euler factorsat each qj as in the
proofof Proposition2.15, we obtain a formand hence a maximalideal mQ of
TH(Nql ... qr) with the property that pmQ PO over Fp relative to a suitable
embeddingkmQ-+ k over kM0.The fieldkmQis the extensionof ko (or kmin in
the exceptionalcase) generatedby the caj,fi. We set
(2.35) TQ = TH(Nql ... qr).Q 0 (9.

W(k-Q)
It is easy to see directly(or by the argumentsof Proposition2.15) that

TQ is reducedand that thereis an inclusionwithfiniteindex
(2.36) TQ >TQ = J09g,1

ofthe Galois conjugacyclasses
wherethe productis takenoverrepresentatives
of eigenformsg of level Nqj ... q. with mQ -4 pi. Now define DQ using the
choices ai for which Uq, -+ ai under the chosen embedding kQ -- k. Then
each of the 2-dimensionalrepresentations associated to each factorOgt is of
type DQ. We can check this for each q E Q using eitherthe Wq - 7r(q)
theorem(cf. [Cal]) as in the case of type (B) or using the Eichler-Shimura
relationifq does not dividethe levelofthe newform associatedto g. So we get
a homomorphism of (9-algebrasRQ -+ TQ and hence also an 0-algebra map
(2.37) (oQ: RQ- TQ

as RQ is generatedby traces. This is not an isomorphismin generalas we
have used N in place of M. Howeverit is surjectiveby the argumentsof
Proposition2.15. Indeed, forq I N(po)p, we checkthat Uq is in the image of

(pQ using the argumentsin the second halfof the proofof Proposition2.15.
For q E Q we use the fact that Uq is the image of the value of X2,q(Frobq)
in the universalrepresentation; cf. (2.34). For q I M, but not of the previous
types,Tq is a trace in PTQ and we can apply the Cebotarev densitytheorem
to show that it is in the image of ,oQ.
0, then set pQ = ker7rand let pp de-
Finally, if there is a section a: TQ -+
note the 2-dimensionalrepresentation

to GL2(0) obtainedfromPTQ mod pQ.
Let V = Adpp0 K/c whereK is the fieldof fractionsof 0. We pick a basis
0
forPP satisfying(2.34) and thenlet
a
(2.38) {( g)}
CAdpOK/O
- P()
= {( c d )
b):a~b~cdE0}?K/0
' ' '
and let V(qi)= V/V(qi). Then as in Proposition1.2 we have an isomorphism
(2.39) Hom0(pRQ/p1Q2K/O) HEQ (QEuQ/QIV)
wherePRQ = ker(iro FQ) and the second termis definedby
(2.40) HDQ (QEUQ/Q,

(2140) V) = H1/(Q/
~~~~~~~~~~~~~~q ker: uQ/Q1 V)
H v
(Qunr, (qi)
i=1
We returnnow to our discussionof Conjecture2.16. We will call a de-

formationtheoryD minimalif E = M U {p} and is Selmer,strictor flat. .
This notionwill be criticalin Chapter 3. (A slightlystrongernotionof mini-

malityis describedin Chapter3 wherethe Selmerconditionis replaced,when
possible,by the conditionthat the representations arise fromfiniteflatgroup
schemes-see the remarkafterthe proofof Theorem3.1.) Unfortunately even
up to twist,not everypo has an associated minimalD even whenpo is flator
ordinaryat p as explainedin the remarksafterConjecture2.16. Howeverthis
could be achievedif one replaced Q by a suitable finiteextensiondepending
on po.
Suppose nowthat f is a (normalized)newform, A is a primeofOf above p
and PfA a deformation ofpo of typeD whereD = (., Z, (Of,A,M) with . = Se,
str or fl. (Strictlyspeakingwe may be changingpo as we wish to choose its
fieldof definition to be k = Of,A/A.)Suppose furtherthat level(f) I M where
M is definedby (2.24).
Now let us set 0 = OfA forthe restof this section. There is a homomor-
phism
(2.41) r= rD,f: T-D 40

514 ANDREW WILES
whose kernelis the primeideal PT,f associated to f and A. Similarlythereis

a homomorphism
Rv - C0
whose kernelis the primeideal PR,fassociated to f and A and whichfactors
throughlrf. Pick perfectpairingsof 0-modules, the second one TD-bilinear,
(2.42) 0 X 0 -0, (, ): Tv xTv O.
In each case we use the termperfectpairingto signifythat the pairsofinduced

maps ( --) Homo (0, 0) and Tv -* Homo (Tv, 0) are isomorphisms. In
additionthesecondone is requiredto be Tv-linear. The existenceofthesecond
pairingis equivalentto the Gorensteinproperty,Corollary2 of Theorem2.1,
as we explain below. Explicitlyif h is a generatorof the freeTV-module
Homo(TD, 0) we set (tl, t2) = h(tit2).
A prioriTH(M)m (occurringin the descriptionofTv in Proposition2.15)
is only Gorensteinas a Zp-algebrabut it followsimmediatelythat it is also a
GorensteinW(km)-algebra.(The notionof Gorenstein0-algebra is explained
in the appendix.) Indeed the map
Homw(k.) (TH(M)m, W(km)) Homzp (TH(M)m, Zp)

givenby 'p F traceof is easily seen to be an isomorphism,as the reduction
modp is injectiveand the ranksare equal. Thus Tv is a Gorenstein0-algebra.
Now let *: 0 -- Tv be the adjoint of ir withrespectto these pairings.
Then definea principalideal (71)of Tv by
(10 = OqDj) = (* (1)).
This is well-defined
independently of the pairingsand moreoverone sees that
TvD/71is torsion-free
(see the appendix). Fromits description(71)is invariant
underextensionsof 0 to 0' in an obviousway. Since Tv is reduced r(r1)+ 0.
One can also verifythat
(2.43)
up to a unit in 0.
We will say that Di D V if we obtain Di by relaxingcertain of the
hypotheseson D, i.e., if19= (., E, 0, M) and Di = (, 0iOi, M1) we allow
that El D E, any O1, M D M1 (but of the same type) and if is Se or str
in D it can be Se, str, ord or unrestrictedin Di, if . is fl in Di it can be fl
or unrestricted in 1i. We use the termrestrictedto signifythat is Se, str,
fl or ord. The followingtheoremreducesconjecture2.16 to a 'class number'
criterion. For an interpretation of the right-handside of the inequalityin
the theoremas the orderof a cohomologygroup,see Proposition1.2. For an
interpretation of the left-handside in termsof the value of an innerproduct,
see Proposition4.4.

THEOREM 2.17. Assume, as above, thatpf,Ais a deformation of po of

type19= (E,A, 0 = O9f,), M) with = Se, str or fl. Suppose that
#0/l(mDf) > #PR,f/IRf
Then
(i) pD1,:RD1 _TM1 is an isomorphism
for all (restricted)Di D 1.
(ii) TD1 is a completeintersection(over 01 if. is Se, str or fl) for all re-
strictedD, D D.
Proof. Let us writeT forTD, PT forPTf, PR forPRf and 71for?D,f.

Then we alwayshave
(2.44) #//p ? #1'T/1T
(Here and in what followswe sometimeswriteq forir(r) ifthe contextmakes
this reasonable.) This is provedas follows.T/rqacts faithfully
on PT. Hence
the Fittingideal of PT as a T/rq-moduleis zero. The same is then true of
/p2 as an X)7 = (T/71)/pT-module.So the Fittingideal of PT/p2 as an
0-module is containedin (r1)and the conclusionis theneasy. So togetherwith
the hypothesisof the theoremwe get inequalities(and henceequalities)
#0e/7r(7) > #PR/PR >- #PT/PT - c/(1
By Proposition2 ofthe appendixT is a completeintersection
over(9. Part (ii)
of the theoremthenfollowsforD. Part (i) followsforD fromProposition1 of
the appendix.
We now proveinductivelythat we can deduce the same inequality
(2.45) # 01/1"1,f > #PRi,f/PRI1,f
forD, D 9D and R1 = RD1. The above argumentwill thenprovethe theorem
forV1. We explainthisfirstin the case D, = 1Dqwhere1DqdiffersfromD only
in replacingE by E U{q}. Let us writeTq forTDq, PR,q forPRJfwithR = RDq
and rq for 7Pqfj We recall that Uq = 0 in Tq.
We choose isomorphisms
(2.46) T Homo(T, (9), Tq Homo(Tq, (9)
comingfromthe fact that each of the rings is a Gorenstein0-algebra. If
aq: Tq -- T is the naturalmap we mayconsiderthe elementAq = aqo &q E T
wherethe adjoint is withrespectto the above isomorphisms.Then it is clear
that
(2.47) (q(q)) = (JAq)
as principalideals of T. In particular7r(rq)= Ir(? Aq) in (.

Now it followsfromProposition2.7 thattheprincipalideal (Aq) is givenby
(2.48) (Aq) = ((q - 1)2(T2 - (q)(1 + q)2))

516 ANDREW WILES
In the statement of Proposition 2.7 we used Zp-pairings
T - -
HomZ, (T, Zp), Tq Homzp (Tq, Zp)
to define (LAq) = (aqo &q). However, using the description of the pairings
as W(km)-algebras derived fromthese Zp-pairings in the paragraph following
(2.42) we see that the ideal (Aq) is unchanged when we use W(km)-algebra
pairings, and hence also when we extend scalars to ( as in (2.42).
On the other hand
#PR~q/PR7q< #PR/R #{O/(q-1)2 (T, - (q)(1 + q2)}
by Propositions1.2 and 1.7. Combiningthiswith(2.47) and (2.48) gives(2.45).

If M $ 0 we use a similar argument to pass fromD to Dq where this time
Dq signifiesthat D is unchanged except for dropping q from M. In each of
types (A), (B), and (C) one checks fromPropositions 1.2 and 1.8 that
#PR~ /p2 < #PR/p2 *#H?(Qqi V*).
This is in agreementwith Propositions2.10, 2.12 and 2.13 which give the

corresponding change in r1by the method described above.
To change from an (9-algebra to an (91-algebra is straightforward(the
complete intersectionpropertycan be checked using [Kul, Cor. 2.8 on p. 209]),
and to change fromSe to ord we use (1.4) and (2.32). The change fromstr
to ord reduces to this since by Proposition 1.1 strict deformations and Selmer
deformationsare the same. Note that forthe ord case if R is a local Noetherian
ring and f E R is not a unit and not a zero divisor, then R is a complete
intersection if and only if R/f is (cf. [BH, Th. 2.3.4]). This completes the
proof of the theorem. -1
Remark 2.18. If we suppose in the Selmer case that f has level N with
p t N we can also consider the ring TH(MO)mo (with Mo as in (2.24) and mo
defined in the same way as for TH(M)). This time set
To = TH(MO)no 0 (9, T = TH(M)m 0 (9.

W(km.0) W(krn)
Define r10,rj,Po and p with respect to these rings, and let (Ap) = oo &p where
OP: T - To and the adjoint is taken with respect to (-pairings on T and To.
We then have by Proposition 2.4
(2.49) (rjp) = (n . Ap) = (T-(p)(l

(22 ?p)2)) = (a2 _ (p))
as principal ideals of T, where ap is the unit root of X2 -Tpx + p(p) = 0.
Remark. For some earlier work on how deformation rings change with E
see [Bo].

Chapter 3
In this chapterwe prove the main resultsabout Conjecture 2.16. We

beginby showingthat the bound forthe Selmergroupto whichit was reduced
in Theorem 2.17 can be checkedif one knowsthat the minimalHecke ring
is a completeintersection.Combiningthis with the main resultof [TW] we
completethe proofof Conjecture2.16 undera hypothesisthat ensuresthat a
minimalHecke ringexists.
Estimates for the Selmer group
Let po: Gal(Q/Q) -- GL2(k) be an odd irreduciblerepresentation which

we willassume is modular.Let V be a deformation theoryoftype (-, A,(9,M)
such that po is type D, where is Selmer,strictor flat. We remindthe
reader that k is assumed to be the residuefieldof (. Then as explained in
Theorem 2.14, we can pick a modular liftingpf,Aof po of type D (altering
k if necessaryand replacing( by a ring containingOf,A,) providedthat po
is not induced froma characterof Q (V'Z3) if p = 3. For the rest of this
chapter,we will make the assumptionthat po is not of this exceptionaltype.
Theorem2.14 also specifiesa certainminimumlevel and characterforf and
in particularensuresthat we can pick f to have level primeto p when POIDP
is associated to a finiteflatgroupschemeoverZp and detpoIP = W.
In Chapter 2, Section3, we defineda ringTo associated to D. Here we
make a slightmodificationof this ring. In the case where is Selmerand
POIDP is associated to a finiteflatgroupschemeand detpo I = Wwe set
(3.1) TDo = T'H (MO)' W/0
with Mo as in (2.24), H definedfollowing(2.24) (it is actually a subgroup

of (Z/Mo Z)*) and m' the maximal ideal of T' (Mo) associated to po. The
same proofas in Proposition2.15 ensuresthat thereis a maximalideal moof
TH(MO) withmo n T1 (Mo) = m' and such that the naturalmap
(3.2) TDo = T'(Mo)ma ?(9-)X TH(MO) (9

HW(ko) W(ko)
is an isomorphism.The maximalideal mowhichwe choose is characterizedby

the propertiesthat pnlo= po and Uq e moforq e - M U {p}. (The value of
Tp or of Uq forq e M is determinedby the otheroperators;see the proofof
Proposition2.15.) We now defineTDO in generalby the following:

518 ANDREW WILES
TDO is givenby (3.1) if. is Se and POIDPis associated

to a finiteflatgroupschemeover Zp and
detpolip = W;
(3.3)
TDO = TD if is str or fl,or POIDP is not associated

to a finiteflatgroupschemeover Zp, or
detPOIIP$wW.
We choose a pair (f,A) of minimumlevel and characteras given by Theo-

rem 2.14 and this givesa homomorphism
of (9-algebras
irf: T0o -> (9 D of"Ad
We set PT,f = ker7rfand similarlywe let PRf denotethe inverseimageofPTJf

in RD. We definea principalideal (rTif) of TDo by takingan adjoint#'fof lrf
withrespectto pairingsas in (2.42) and write
n1T,f= ( f (l)).
NotethatPT f/ is finiteand 7rf(IT, ) $ 0 because Tpo is reduced. We

also write Trffor1rf(irTf) ifthe contextmakesthisusage reasonable. We let
Vf = Ad pp 09K/c) wherepp is the extensionof scalars of pf,Ato (.
0
THEOREM 3.1. Assume thatV is minimal,i.e., a, = M U {p}, and that

PO is absolutelyirreduciblewhenrestrictedtoQ ( ) ). Then
(i) #HD1(QE/QVf) < #(PT~f/PTf) #OT
wherecp = #(9/U?2- (p)) < o0 whenPo is Selmer and POID, is associatedto

a finiteflat groupschemeoverZp and detPOIIJ= w, and cp = 1 otherwise;
(ii) ifTD0 is a completeintersectionover( then(i) is an equality,RV
To and TD is a completeintersection.
In general,for any (not necessarilyminimal) D of Selmer,strictor flat
type,and any pf,Aof typeD, #HD,(Qz/Q,Vf) < oo ifpo is as above.
Remarks. The finiteness was provedby Flach in [Fl] undersome restric-

tions on f, p and D by a different method. In particular,he did not consider
the strictcase. The bound we obtain in (i) is in fact the actual order of
H,(QE/Q, Vf/)as followsfromthe main resultof [TW] whichprovesthe hy-
pothesisof part (ii). Then applyingTheorem2.17 we obtain the orderof this
groupformoregeneralD's associatedto Po underthe conditionthat a minimal
D existsassociated to po. This is stated in Theorem3.3.

The case wherethe projectiverepresentation associated to po is dihedral

does not alwayshave the propertythat a twistofit has an associated minimal
D. In the case wherethe associatedquadraticfieldis imaginarywe will give a
differentargumentin Chapter4.
Proof. We willassumethroughoutthe proofthat D is minimal,indicating
onlyat the end the slightchangesneededforthefinalassertionofthe theorem.
Let Q be a finiteset ofprimesdisjointfromE satisfying q- 1(p) and po(Frobq)
having distincteigenvaluesforeach q E Q. For the minimaldeformation
problem'D = (., A, (9,M), let VQ be the deformation problemdescribedbefore
(2.34); i.e., it is the refinementof (., E U Q, (9,M) obtained by imposingthe
additionalrestriction(2.34) at each q E Q. (We will assume forthe proofthat
(9 is chosenso (9/A= k containsthe eigenvaluesof po(Frobq) foreach q E Q.)
We set
T = TDO, R = RD
and recall the definitionof TQ and RQ fromChapter 2, ?3 (cf. (2.35)). We
writeV forVf and recallthe definition ofV(q)following(2.38). Also remember
that mQ is a maximalideal of TH(Nql ... q,) as in (2.35) forwhichpmQ po
over Fp (recall that this uses the same choice of embeddingkmQ k as in -
the definitionof TQ). We use mQ also to denote the maximalideal of TQ if

the contextmakesthis reasonable.
Considerthe exact and commutativediagram
o -0 H~D(QE/QV) -
HD'Q(QEUQ/QV) -Q
JJ
H(Qunrv(q))Ga (Q nr/Qq)
qEQ
I It
T T
0 - T
(PT/P ) (PTQ/PTQ) QS. KQ -O
the cokernelin the horizontalsequenceand * denotes

whereKQ is by definition
Homo( , K/O) forK the fieldof fractionsof (. The keyresultis:
LEMMA 3.2. The map LQ is injectivefor any finiteset of primes Q
satisfying
q_ 1(p),T (q) (1 +q)2modm for all eQ.
Proof. Note that the hypothesesof the lemmaensurethat po(Frobq) has

distincteigenvaluesforeach q E Q. First,considerthe ideal aQ of RQ defined

520 ANDREW WILES
by
(3.4) aQ={ai-1,bicidi-1: ( i
ac b(i) withvi C Iqqi C Q
Then the universalpropertyof RQ shows that RQ/aQ R. This permitsus

to identify(PR/P2)* a
(PR/P2) = {f E (PRQ/ipQ)* : f(aQ) = 0}.
If we provethe same relationforthe Heckerings,i.e., withT and TQ replacing

R and RQ then we will have the injectivityof LQ. We will writeaQ forthe
image of aQ in TQ under the map CpQ of (2.37).
It willbe enoughto checkthat forany q E Q', Q' a subsetof Q, TQ//-aq
TQ,_{q} where aq is definedas in (3.4) but with Q replaced by q. Let
N' = N (po) p'(PO) * rlqiEQ'-{q}qi where6(po) is as definedin Theorem 2.14.
Then take an elementa E Iq C Gal(Qq/Qq) whichrestrictsto a generator
of Gal(Q((Nfq)/Q((N/)). Then det(a) = (tq) e TQ/ in the representation to
GL2(TQ/) definedafterTheorem2.1. (Thus tq 1(N') and tq is a primitive
root modq.) It is easily checkedthat
(3.5) JH(N' q)mQ (Q) JH(N'q)mQ, (Q) [(tq) - 1]

Here H is stilla subgroupof (Z/MoZ)*. (We use herethat po is not reducible
forthe injectivityand also that po is not inducedfroma characterof Q(Vz/3)
forthe surjectivity whenp = 3. The latteris to avoid the ramification
pointsof
the coveringXH(N'q) -4 XH(N', q) oforder3 whichcan give riseto invariant
divisorsof XH(N'q) whichare not the imagesof divisorson XH(N', q).)
Now by Corollary1 to Theorem2.1 the Pontrjaginduals of the modules
in (3.5) are freeof ranktwo. It followsthat
(3.6) (TH(N'q)mQ,)2/((tq) - 1) -_(TH(N' ,q)mQ,)2.
The hypothesesof the lemmaimplythe conditionthat po(Frobq) has distinct

eigenvalues. So applyingProposition2.4' (at the end of ?2) and the remark
followingit (or using the fact remarkedin Chapter 2, ?3 that this condition
impliesthat po does not occur as the residualrepresentation associatedto any
formwhichhas the special representation at q) we see that aftertensoringover
W(kmQ/) with (, the right-hand side of (3.6) can be replacedby TQ,{q} thus
giving
TQ, aq -T Q'-fq} '
since (tq) -1 E dq. Repeated inductivelythis gives the desired relation

TQ/aQ _T, and completesthe proofof the lemma. F1

Suppose nowthat Q is a finiteset ofprimeschosenas in thelemma. Recall

that fromthe theoryof congruences(Prop. 2.4' at the end of ?2)
71TQf/1Tf = 17(q -1),
qEQ
the factors(a2- (q)) being unitsby our hypotheseson q E Q. (We onlyneed

that the right-handside dividesthe leftwhichis somewhateasier.) Also, from
the theoryof Fittingideals (see the proofof (2.44))
#(PT/PT) ? #(O/1TTf)
#(PTQ/PTQ) > #(0O7/TQf).

We deduce that
#KQ > 0 rJ/ (q -1)
wheret = #(PT/P2)/#(O/Tf). Since the rangeof LQ has ordergivenby
we computethat the indexof the image of LQ is < t as LQ is injective.

Keepingour assumptionon Q fromLemma 3.2, considerthe kernelof AM
applied to the diagramat the beginningof the proofof the theorem. Then
with M chosen large enough so that AM annihilatesPT/P4 (which is finite
because T is reduced)we get:
o )~HD,(QE/QV[,XM])-~ HIQ(QSuQ/QSV[AM]) B J7J H1(Qunr, V(q)[AM])Ga1(Qu"/Qq)
qEQ
T T 4Q T LQ
? (PT /p2) (PTQ/PTQ/p)*[AM] `(PT-/p KQ[AM]
See (1.7) forthe justificationthat AM can be taken inside the parenthesesin

the firsttwo terms. Let XQ = 'bQ((PTQ/pTQ)* [AM]). Thenwe can estimate
the orderof 5Q(XQ)usingthe factthat the image of LQ has index at most t.
We get
(3.7) #6Q(XQ) > (II #O/(AM q - 1)) (t) (1/#(PTPT))

qEQ
Now we choose Q to be a set of primeswiththe propertythat
-
(3.8) EQ : HV1*(QE/Q VAM) H1 (Qq7
VAqM)
qEQ

522 ANDREW WILES
is injective. We also keep the conditionthat LQ is injectiveby only allowing

Q to containprimesof the formgivenin the lemma. In addition,we require
theseq's to satisfy
q =_ 1(pM).
To see that this can be done, suppose that x E kereQ and Ax = 0 but
x 5$0. We have a commutativediagram
H'(QE/Q, V;M) [A] Q H'(Qq, V;M) [A] II
qEQ
HI(Q/QV*) II
fJH'(QqV;)
qEQ
the right-hand isomorphisms comingfromour particularchoicesofq's and the

left-handisomorphism fromour hypothesison po. The same diagramwillhold
ifwe replaceQ by Qo = Q U {qo} and we now need to showthat we can choose
qO so that 6Qo(X) 5#0.
The restrictionmap
H' (Q~/Q,V ) - Hom(Gal(Q/Ko((p)), V)Ga1(KO(7P)/Q)
has kernelH'(Ko((p)/Q, k(1)) by Proposition1.11 wherehereKo is the split-
tingfieldof po. Now if x E H'(Ko((p)/Q, k(1)) and x 5$0 thenp = 3 and x
factorsthroughan abelian extensionL of Q((3) of exponent3 whichis non-
abelian over Q. In this exceptionalcase, L must ramifyat some primeq of
Q((3), and if q lies overthe rationalprimeq 543 thenthe compositemap
HI (K ((3) /QIk(1)) H'(Qunr,k(l)) H'(Qunr,((q/Am)(1))
is nonzeroon x. But thenx is not oftypeD* whichgivesa contradiction.This
onlyleaves the possibilitythat L = Q((3, 3V1) but again this means that x is
not of type'D* as locallyat the primeabove 3, L is not generatedby the cube
root of a unitoverQ3((3). This argumentholdswhetheror not D is minimal.
So x, whichwe view in kereQ, gives a nontrivialGalois-equivariantho-
momorphism f., E Hom(Gal(Q/Ko((p)), V*) whichfactorsthroughan abelian
extensionM. of Ko((p) of exponentp. Specificallywe choose M, to be the
minimalsuch extension. Assume firstthat the projectiverepresentationpj3
associated to po is not dihedralso that Sym2po is absolutelyirreducible.Pick
a a E Gal(Mx ((pM)/Q) satisfying
(3.9) (i) po(a) has orderm > 3 with (m,p) = 1,
(ii) a fixesQ(detpo) (4pM),

(iii) fx(orm)7+ Ok
To show that this is possible, observefirstthat the firsttwo conditionscan
be achievedby Lemma 1.10(i) and the subsequentremark. Let al be an el-

ementsatisfying(i) and (ii) and let a1 denote its image in Gal(Ko((p)/Q).

Then (&1) acts on G = Gal(Mx/Ko((p)) and under this action G decom-
poses as G G1 EDGi whereal acts triviallyon G1 and withoutfixedpoints
on G'. If X is any irreducibleGalois stable k-subspaceof f (G) ?Fp k then
ker(al - 1) Ix 0 since Sym2po is assumed absolutelyirreducible. So also
ker(a - 1)If(G) 5$0 and thus we can find r E G1 such that f,(r) ? 0.
Viewingr as an elementof G we thentake
-Gx
At= r x 1 E Gal(Mx((pM)/Ko((p)) Gal(Ko((pM)/Ko((p))
(This decompositionholds because Mx is minimaland because Sym2po and
4p are distinctfromthe trivialrepresentation.)Now rT commuteswitha1 and
eitherfx((ri al)m) #0 or fx(af') #0. Since po(Tiai) = po(ai) this gives
(3.9) with at least one of a = r1al or a = al. We may then choose q0 so that
Frobq0 = a and we willthenhave eQ0(x) 540. Note that conditions(i) and (ii)
implythat q =_1(p) and also that po(a) has distincteigenvalues,thus giving
both the hypothesesof Lemma 3.2.
If on the otherhand po is dihedralthenwe pick a's satisfying
(i) 3o(a) $ 1,
(ii) a fixes Q(4PM),
(iii) fx(am) 540,

withm the orderofpo(a) (and p t m sincejo is dihedral). The firsttwocondi-
tionscan be achievedusingLemma 1.12 and, in addition,we can assume that
a takes the eigenvalue1 on any given irreducibleGalois stable subspace X
of Wx 0 k. Arguingas above, we finda r E G1 such that fx(r) $ 0 and
we proceed as before. Again, conditions(i) and (ii) implythe hypothesesof
Lemma 3.2. So by successivelyadjoiningq's we can assume that Q is chosen
so that eQ is injective.
We have thus shown that we can choose Q = {q1,... ,qI} to be a finite

set ofprimesqj i-=(pM) satisfying the hypothesesofLemma 3.2 as well as the
injectivityof FQ in (3.8). By Proposition1.6, the injectivityof FQ impliesthat
(3.10) #HD(QEuQ/Q, Vf
[AM]) = ho. IJ hq.
qEEUQ
Here we are usingthe conventionexplainedafterProposition1.6 to defineHD1.

Now, as D was chosen to be minimal,hq = 1 forq E EI -{p} by Proposi-
tion 1.8. Also, hq = #(O/AM)2 forq E Q. If. is str or fl then hoohp= 1
by Proposition1.9 (iv) and (v). If * is Se, hochp< cp by Proposition1.9 (iii).
(To computethis we can assume that Ip acts on WA via w, as otherwisewe

524 ANDREW WILES
get hwhp < 1. Then withthis hypothesis,(Won)* is easily verifiedto be un-

ramifiedwithFrobp actingas U2(p)-l by the descriptionof Pf,A DI in [Wil,
Th. 2.1.4].) On the otherhand,we have constructedclasses whichare ramified
at primesin Q in (3.7). These are of type DQ. We also have classes in
Hom(Gal(QsuQ/Q), 0/AM) = H1(QruQ/Q, 0/AM) H1(Q uQ/Q, VAm)

coming fromthe cyclotomic extension Q(Cql ... (qr). These are of type D and
disjointfromthe classes obtained from(3.7). Combiningthese with (3.10)
gives
#Hv(QE/Ql Vf[Am]) < t # P/p2 *@
as required.This provespart (i) of Theorem3.1.

Now if we assume that T is a completeintersectionwe have that t = 1
by Proposition2 of the appendix. In the strictor flat cases (and indeed in
all cases where cp = 1) this impliesthat RD -_TD by Proposition1 of the
appendixtogetherwithProposition1.2. In the Selmercase we get
(3.11) (PT
#(wr/PT) = #(O/1,TJ)cp= #(01/7T,,f) < # (PT,/T')
wherethe centralequality is by Remark 2.18 and the right-handinequality
is fromthe theoryof Fittingideals. Now applyingpart (i) we see that the
inequalityin (3.11) is an equality. By Proposition2 of the appendix, TD is
also a completeintersection.
The finalassertionof the theoremis provedin exactlythe same way on
notingthat we only used the minimalityto ensurethat the hq's were 1. In
general,theyare bounded independentof M and easily computed. (The only
pointto note is that ifpf,Ais of multiplicative
type at q thenpfAIDq does not
split.) D
Remark. The ringTDO definedin (3.1) and used in this chaptershould

be the deformation ringassociated to the followingdeformationproblemDo.
One altersD onlyby replacingthe Selmerconditionby the conditionthat the
deformations be flatin the sense of Chapter 1, i.e., that each deformationp
of po to GL2(A) has the propertythat forany quotientA/a of finiteorder,
PIDPmoda is the Galois representation associated to the Qp-pointsof a finite
flatgroupschemeover Zp. (Of course,po is ordinaryhere in contrastto our
usual assumptionforflatdeformations.)
FromTheorem3.1 we deduce our main resultsabout representations by

using the main resultof [TW], whichprovesthe hypothesisof Theorem 3.1
(ii), and thenapplyingTheorem2.17. More precisely,the main resultof [TW]
shows that T is a completeintersectionand hence that t = 1 as explained
above. The hypothesisof Theorem 2.17 is then given by Theorem 3.1(i),
togetherwith the equality t = 1 (and the centralequality of (3.11) in the

Selmer case) and Proposition 1.2. Strictlyspeaking, Theorem 1 of [TW] refers

to a slightlysmaller class of D's than those covered by Theorem 3.1 but up to
a twist every such D is covered. It is straightforwardto see that it is enough
to check Theorem 3.3 for po up to a suitable twist.
THEOREM
to Q
whenrestricted
3.3.
( Assume that po is modularand absolutelyirreducible
1 p) . Assumealso thatpo is oftype(A), (B)
or (C) at each q : p in S. Then the map WE): RE) TE of Conjecture2.16
is an isomorphism for all D associatedto po, i.e., whereD = (.,E,Y,M) with
* = Se, str,fl or ord. In particularif. = Se, str or fl and f is any newform
for whichpf,Ais a deformation ofpo of typeD then
#H1(QE/Q, Vf) = #(01/iD,f) < 00

where71D,fis theinvariantdefinedin Chapter2 prior to (2.43).
The condition at q : p in E ensures that there is a minimal D associated

to po. The computation of the Selmer group follows from Theorem 2.17 and
Proposition 1.2. Theorem 0.2 of the introduction follows from Theorem 3.3,
afterit is checked that a twist of a po as in Theorem 0.2 satisfiesthe hypotheses
of Theorem 3.3.
Chapter 4
In this chapter we give a different(and slightlymore general) derivation

of the bound for the Selmer group in the CM case. In the first section we
estimate the Selmer group using the main theorem of [Ru 4] which is based on
Kolyvagin's method. In the second section we use a calculation of Hida to relate
the rj-invariantto special values of an L-function. Some of these computations
are valid in the non-CM case also. They are needed if one wishes to give the
order of the Selmer group in terms of the special value of an L-function.
1. The ordinary CM case
In this section we estimate the order of the Selmer group in the ordinary
CM case. In Section 1 we use the proof of the main conjecture by Rubin to
bound the Selmer group in terms of an L-function. The methods are standard
(cf. [de Sh]) and some special cases have been described elsewhere (cf. [Guo]).
In Section 2 we use a calculation of Hida to relate this to the rj-invariant.
We assume that
(4.1) P = IndQ i,: Gal(Q/Q) -* GL2(0)

526 ANDREW WILES
is the p-adic representationassociated to a characterai: Gal(L/L) Ox of

an imaginaryquadraticfieldL. We assumethatp is unramified in L and that i
factorsthroughan extensionofL whoseGalois grouphas the formA - Zp ET
whereT is a finitegroupoforderprimeto p. The ring( is assumedto be the
ringof integersof a local fieldwithmaximalideal A and we also assume that
p is a Selmerdeformation of po = p modA whichis supposed irreduciblewith
detpolII,= w. In particularit followsthat p splits in L, p = po say, and that
preciselyone of a, a* is ramifiedat p (K* beingthe characterr - 1
forany a representing the nontrivialcoset in Gal(Q/Q)/ Gal(Q/L)). We can
suppose withoutloss of generalitythat i is ramifiedat p.
We considerthe representationmoduleV - (K/O)4 (whereK is the field
of fractionsof 0) and the representationis via Ad p. In this case V splitsas
V - Y ED(KIO) (0) EDKIO

where b is the quadratic characterof Gal(Q/Q) associated to L. We let E
denote a finiteset of primesincludingall those which ramifyin p (and in
particularp). Our aim is to computeHs e(Q/Q, V). The decompositionof
V givesa corresponding decompositionof H1(Qr/Q, V) and we can use it to
defineHs e(Q/Q, Y). Since W0 c Y (see Chapter 1 forthe definitionof Wo)
we can define HSle(QF/Q, Y) by
HSe(Q/Q, Y) = ker{H1 (Q/Q, Y) - H1 (Qunr Y/Wo)}.
Let Y* be the arithmeticdual of Y, i.e., Hom(Y,ppoo) 0 Qp/Zp. Write

v for Ke/K* and let L(v) be the splittingfieldof v. Then we claim that
Gal(L(v)/L) - Zp E T' with T' a finitegroupof orderprimeto p. For this
it is enoughto show that X = KK*/efactorsthrougha group of orderprime
to p since v = 82X-1' Suppose that X has orderm = moprwith (mop) = 1.
Then xmoextendsto a characterof Q whichis thenunramified at p sincethe
same is trueofX. Also it factorsthroughan abelian extensionofL withGalois
groupisomorphicto Z2 since X factorsthroughsuch an extensionwithGalois
groupisomorphicto Z E T1 withT1 oforderprimeto p (the compositeofthe
splittingfieldsof n and a*). It followsthat xm0is also unramifiedoutsidep,
whenceit is trivial.This provesthe claim.
Over L thereis an isomorphismof Galois modules
Y*- (KIO) (v) ED(KIO) (v-162).

In analogyto the above we defineHSe(QF/Ql Y*) by
HSe(QF/Q, Y*) = ker{H1(Qr/Q, Y*) -? H1 (Qunr,(Wo)*)}.
Analogousdefinitionsapply if Y* is replacedby Yv*nL. Also we say informally

that a cohomologyclass is Selmerat p ifit vanishesin H1 (Qunr, (WO)*) (resp.

Hl(Qunr, (WOn)*)). Let Moo be the maximalabelian p-extensionof L(v) un-

ramifiedoutsidep. The followingpropositiongeneralizes[CS, Prop. 5.9].
PROPOSITION 4.1. Thereis an isomorphism
Hunr(QF/Q, Y*) A Hom (Gal(Moo/L(v)), (K/O)(V))Ga1(L(v)/L)
whereHlnr denotesthesubgroupof classes whichare Selmerat p and unram-

else.
ifiedeverywhere
sequenceas
Proof. The sequenceis obtainedfromthe inflation-restriction
follows.First we can replaceH1(Qr/Q, Y*) by
{H1 (QF/L, (K/()(v)) E H1 (Q/L, (K/9)(v-162))}
where A = Gal(L/Q). The unramifiedconditionthen translatesinto the

that the cohomologyclass should lie in
requirement
{HunrinE-p(Qr2/L, (K/c)(v)) E Hunrin p* (Qr/L, (K/O)(v162))}

Since A interchangesthe two groups inside the parenthesesit is enough to
computethe firstof them,i.e.,
(4.2) H jnr in _p (Qr/L, K/O(v)).

sequence applied to this givesan exact sequence
The inflation-restriction
(4.3) 0 - Hlnr in U-p (L(v)/L, (K/O)(v))
Hunr in Up (Qr/L, (K/O)(v))
-? Hom (Gal(Mco/L(v))I (K/Q)(V))Gal(L(v)/L)
The firsttermis zero as one easily checksusingthe divisibilityof (K/O)(v).

Next note that H2 (L(v)/L, (K/Q)(v)) is trivial.If v 0 1(A) this is straight-
forward(cf. Lemma 2.2 of [Rul]). If v =1(A) then Gal (L(v)/L) - Zp and so
it is trivialin this case also. It followsthat any class in the finaltermof (4.3)
liftsto a class c in H1 (QE/L, (K/Q)(v)). Let Lo be the splittingfieldof Y*.
Then MooLo/Lo is unramified outside p and Lo/L has degreeprimeto p. It
followsthat c is unramified outsidep. LI
Now writeHltr(QF/Q, 1n*) (whereYn*= Y\' and similarlyforY,) for

the subgroupof Hunr(QF/Q, Yn*)givenby
Hstr(QE/Q YEn)= E Hlnr(QF/Q, Yn) op = 0 in H1(Qp, Yn*/(Yn)O)}

where(Yn*)0is the firststep in the filtration
underDp, thusequal to (Yn/yn?)*
or equivalently to (Y*)%n where (Y*)O is the divisible submodule of Y* on
whichthe action of Ip is via 9. (If p $&3 one can characterize(Yn)0 as the

528 ANDREW WILES
maximalsubmoduleon whichIp acts via E2.) A similardefinition

applies with
It followsfroman examinationofthe actionof Ip on Y), that
Y, replacingY?n*.
(4.4) Hstr(QE/Ql Yn) = Hunr(QE/Q, Yn).
In the case of Y* we will use the inequality
(4.5) # H1tr(QF/Q, Y*) < # Hunr(QF/Qi Y*).

large the map
We also need the factthat forn sufficiently
(4.6) Hstr(Q/Qi Yn*) Hstr(Q/Qi Y*)

is infective.One can checkthis by replacingthese groupsby the subgroups
of H1(L, (K/O)(V)An) and H1(L, (K/O)(v)) whichare unramifiedoutside p
and trivialat p*,in a mannersimilarto the beginningofthe proofof Proposi-
tion 4.1. The above map is theninjectivewheneverthe connectinghomomor-
phism
Ho (Lp*, (K/O) (v)) -+ H1 (Lp*, (K/Q) (V)An)
largen.
whichholds forsufficiently
is infective,
Now, by Proposition1.6,
7 #Hr(Q/Q, Yn) Ho (Q (Yn)*)#H (Q. Yn)

#Hsltr(QE/Q,Y*) HQ () #HO (Q, Yn*)
Also, HO (Q, Yn) = 0 and a simplecalculationshowsthat
(Q. En) = i{
#H
f
#(C)/l-I/(q)) if v-=I1mod A
1 otherwise
whereq runsthrougha set of primesof OL primeto p cond(v) of densityone.

This can be checkedsince Y* = IndQ(v) 0 K/c. So, setting
(4.8) t infq#(O/(1-v(q))) ifvmodA=1

-~~~ 1 vmodA #,41
we get
(4.9)
#HSe(Q/Q, Y) < q # Hom (Gal (Moo/L(v)), (KI 0)(v
-
Iq()I
where?q = #HO(Qq, Y*) forq :&p, ip = lim # HO(Qp, (Yn?)*).This follows
fromProposition4.1, (4.4)-(4.7) and the elementaryestimate
(4.10) #(HSle(Qr/Ql Y)) <

Y)/Hulnr(QE/Ql 1I fqi
qEE-{p}
which follows fromthe fact that #H1 (Qunr, y)Ga1(Qunr/Qq) = e

Our objectiveis to computeH'e(Qr/Q, V) and the mainproblemis to es-

timateH e(Q-/Q, Y). By (4.5) thisin turnreducesto the problemofestimat-
ing
# Hom(Gal(MOO/L(v)), (K/O)(V))Ga1(L(v)/L)). This order can be computed
usingthe 'main conjecture'establishedby Rubin usingideas ofKolyvagin.(cf.
[Ru2] and especially [Ru4]. In the formerreferenceRubin assumes that the
class numberof L is primeto p.) We could now derivethe resultdirectlyfrom
this by referring to [de Sh, Ch. 3], but we will recall some of the steps here.
Let Wfdenote the numberof roots of unity( of L such that ( _ 1 modJ
(Jan integralideal of OL). We choose an f prime to p such that Wf= 1.
Then thereis a grossencharacter Wof L satisfyings((a)) = a fora 1 modJ
(cf. [de Sh, II.1.4]). Accordingto Weil, afterfixingan embeddingQ Qp we
can associate a p-adic characterWpto s (cf. [de Sh, II.1.1 (5)]). We choose
an embeddingcorresponding to a primeabove p and then we findsp = K *X
forsome X of finiteorder and conductorprime to p. Indeed Wpand i are
both unramifiedat p* and satisfyWpIi,= KII = e wheree is the cyclotomic
characterand I. is an inertiagroupat p. Withoutalteringfwe can evenchoose
s so thatthe orderofX is primeto p. This is byour hypothesisthat i factored
throughan extensionof the formZp E T withT of orderprimeto p. To see
thispickan abelian splittingfieldofWpand i whoseGalois grouphas the form
G E G' with G a pro-p-group and G' of orderprimeto p. Then we see that
WPIG has conductordividingfp'. Also the only primeswhichramifyin a Zp-
extensionlie above p so our hypothesison K ensuresthat KIGhas conductor
dividingfp'. The same is then true of the p-partof X whichthereforehas
conductordividingf. We can thereforeadjust s so that X has orderprime
to p as claimed. We will not howeverchoose s so that X is 1 as this would
requirefp? to be divisibleby condx. Howeverwe will make the assumption,
by alteringf if necessary,but still keepingf primeto p, that both v and Wp
have conductor dividing fp'. Thus we replace fp ? by l.c.m.{f, cond v}.
The grossencharacter y (or more preciselysoo NF/L) is associated to a
(unique) ellipticcurveE definedoverF = L(f), the rayclass fieldofconductor
f, with complex multiplicationby CL and isomorphicover C to CIOL (cf.
[de Sh, II. Lemma 1.4]). We may even fixa Weierstrassmodel of E over OF
whichhas good reductionat all primesabove p. For each primeX3of F above
p we have a formalgroup ET, and this is a relativeLubin-Tate group with
respect to FT3over Lp (cf. [de Sh, Ch. II, ?1.10]). We let A = ASE be the
logarithmof this formalgroup.
Let UOObe the productof the principallocal units at the primesabove p
of L(fpoo);i.e.,
Uo=fJ uCo where Uoo,q=limUn,T,

q3lp

530 ANDREW WILES
each Un't being the principallocal units in L(fpn)qp. (Note that the primes
of L(f) above p are totallyramifiedin L(fpoo)so we still call them {q3}.) We
wishto definecertainhomomorphisms bk on Up. These werefirstintroduced
in [CW] in the case wherethe local fieldFspis Qp.
Assume forthe momentthat Fq3is Qp. In this case Eq is isomorphicto
the Lubin-Tategroupassociated to 7rx+ xP where7r= ~p(p). Then lettingwn
be nontrivialrootsof [7rn] it was shown
(x) = 0 chosenso that [ir](an) = wi~n-1
in [CW] that to each elementu = limun E UOq therecorrespondeda unique
powerseriesfu(T) e Zj[T X such that fu(wn) = Un forn > 1. The definition
of 6ke (k > 1) in this case was then
k,9(U)= A'(T) dT logfu(T)

T=O
It is easy to see that6k,q3givesa homomorphism: UooUoo,-- (9wsatisfying

skq3(e) = 09(0)k k,(6) where9: Gal (FIF) Ox is the charactergiving
the action on E[p'].
The constructionof the powerseriesin [CW] does not extendto the case
wherethe formalgrouphas height> 1 or to the case whereit is definedover
an extensionof Qp. A morenaturalapproachwas developedby Coleman [Co]
whichworksin general. (See also [Iwi].) The corresponding generalizationsof
6k weregivenin somewhatgreatergenerality in [Ru3]and thenin fullgenerality
by de Shalit [de Sh]. We now summarizethese results,thus returningto the
generalcase whereFq is not assumed to be Qp.
To an elementu = limun E UOOwe can associate a powerseriesfuv,(T) E
O,3[[T]]x whereOC3is the ringofintegersofFew;see [de Sh, Ch. II ?4.5]. (More
preciselyfuqp(T)is the q3-component ofthe powerseriesdescribedthere.) For
X3 we will choose the primeabove p corresponding to our chosenembedding
Q c* =
Qp. This powerseriessatisfiesUndo (fu,)(wn) forall n > 0, n= 0(d)
whered = [Fqe:L.] and {wn} is chosen as beforeas an inversesystemof irn
divisionpointsofEm. We definea homomorphism
6k: Uoo -- () by
(4.11) Sk(U) = 6k, p(U) =d logfu'(T)

ET ~~~~T=O
Then
(4.12) 5k(UT) = 9(Q)kSk(U) forr E Gal(F/F)
where0 again denotesthe action on E[p']. Now 0 = pp on Gal(F/F). We

actuallywant a homomorphism on UO,witha transformationpropertycorre-
spondingto v on all of Gal(L/L). Observethat v = 2 on Gal(F/F). Let S

forGal(L/L)/ Gal (LIF) and define

be a set of coset representatives
(4.13) dI2(U) = Z v-1(of)62(uU)E Oq3[v].

dES
by (4.8) and it
Each termis independentof the choiceof coset representative
is easily checkedthat
(2(U') = V(Of)4D2(U).
It takes integralvalues in (9q [v]. Let UO (v) denotethe productof the groups
of local principalunits at the primesabove p of the fieldL(v) (by whichwe
mean projectivelimitsof local principalunits as before). Then 12 factors
throughUO (v) and thus definesa continuoushomomorphism
-+CP I
(D2:UOO,,(VJ)
Let COObe the groupofprojectivelimitsofellipticunitsin L(v) as defined
in [Ru4]. Then we have a crucialtheoremof Rubin (cf. [Ru4], [Ru2]), proved
usingideas of Kolyvagin:
THEOREM 4.2. There is an equalityof characteristicideals as A =
Zp[[Gal(L(v)/L)]] -modules:
charA(Gal (Moo/L(v))) = charA(UO(v)/CJ,).
Let vo = v mod A. For any Zp[Gal(L(vo)/L)]-moduleX we writeX(vO)

forthe maximalquotientofX 0 (9 on whichthe actionof Gal(L(vo)/L) is via
zp
liftofvo. Since Gal(L(v)/L) decomposesintoa directproduct
the Teichmiiller
of a pro-pgroupand a groupof orderprimeto p,
Gal (L(v)IL) Gal (L(v)IL(vo)) x Gal(L(vo)IL),
we can also considerany Zp[[Gal(L(v)/L)]]-modulealso as a Zp[Gal(L(vo)/L)]-
module. In particularX(vO)is a moduleover Zp[Gal(L(vo)/L)](v0)O 0. Also
A(vO)-- [[T]].
Now accordingto resultsof Iwasawa ([1w2, ?12], [Ru2, Theorem 5.1]),
UOO(v)(vO) is a freeA(vO)-module of rank one. We extend (2 0-linearly to
Uoo(v) Ozp (9 and it then factorsthroughUoo(v)(v?). Suppose that u is a
generatorofUO (v) (vO)and ,3an elementofC( ?). Then f (y- 1)u = fiforsome
f (T) E 0[[T]] and -ya topologicalgeneratorof Gal (L(v)/L(vo)). Computing
(2 on both u and ,3 gives
(4.14) f (v(-y)- 1) = (2(0)/P2(U)-

Next we let e(a) be the projectivelimit of ellipticunits in limLpn for
a some ideal prime to 6fp describedin [de Sh, Ch. II, ?4.9]. Then by the
propositionof Chapter II, ?2.7 of [de Sh] this is a 12th powerin limL xn We

532 ANDREW WILES
let 31i= !3(a)1/12be the projectionof e(a)1/12to UOOand take ,3 = Normol3

wherethe normis fromLfpo to L(v). A generalizationof the calculationin
[CW] whichmay be foundin [de Sh, Ch. II, ?4.10] showsthat
(4.15) (D2(3) = (root of unity)Q-2 (Na - v(a)) Lf(2, i) E OT9[v]
whereQ is a basis forthe (9L-moduleofperiodsofour chosenWeierstrassmodel

ofE/F. (Recall that thiswas chosento have good reductionat primesabove p.
The periodsare those of the standardNerondifferential.) Also v here should
be interpretedas the grossencharacter whose associated p-adic character,via
the chosenembeddingQ )-k QP, is v, and vPis the complexconjugateof v.
The only restrictionswe have placed on f are that (i) f is prime to p;
(ii) wf= 1; and (iii) condv
fpv . Now let fop' be the conductorof v withto
primeto p. We show now that we can choose f such that Lf(2,-P)/Lf0 (2, vP)is
a p-adic unit unless vo = 1 in whichcase we can choose it to be t as defined
in (4.4). We can clearlychoose Lf(2,y)/Lf(2,17) to be a unit if vo #41, as
-P(q)v(q)= Normq2forany ideal q primeto fop.Note that if vo = 1 then also
p = 3. Also if vo = 1 then we see that
inf# {O/{JLoq(21P)/Lfo(2P)}} = t
since PF-2 =-
We can compute(D2(u) by choosinga special local unit and showingthat
forus to knowthat it is integral.Then
(D2(U)is a p-adicunit,but it is sufficient
since Gal (Moo/L(v)) has no finiteA-submodule(by a resultof Greenberg;see
[Gre2,end of ?4]) we deduce fromTheorem4.2, (4.14) and (4.15) that
#Hom(Gal (Moo/1L(v)),I (K/c)) (z,))Ga1(L(v)1L)
< #O/Q-2Lfo(2, I) ifvo7&1

?l (#O/Q-2Lfo (2,v)) t if vo = 1.
Combiningthis with (4.9) gives:
#HSe(QE/Ql Y) < # (O/Q-2Lfo(2,v)) . fJq

qEE
where4 # HI(Qq, Y*) (forq 4 p), tp # HO (Qp, (YO)*). -
Since V - Y E (K/O)(f ) EDK/(9 we need also a formulafor
# ker{H1(QE/Q, (K/O)(0) E K/c) -- Hl(Qunr, (K/( )(0) EDK/()}
This is easily computedto be
(4.16) #(O/hL)- J 4q
qEE-{p}

where4 = #H0(Qq, ((K/O)(0) E K/O)*) and hL is the class numberof OL.

Combiningthese gives:
PROPOSITION 4.3.
#HSe(QE/QV) ? #(9/Q2L10(2,v)) #(O(/hL) J4q

qEE
where4 = H
OH(Qq,V*) (forq #4
p), tp = #HO(Qp,(YO)*)-
2. Calculation of rq
We needto calculateexplicitlytheinvariantsrD,f introducedin Chapter2,

?3 in a special case. Let po be an irreduciblerepresentation
as in (1.1). Suppose
that f is a newform ofweight2 and levelN, A a primeofOf above p and pf,A a
deformation ofpo. Let mbe the kernelofthe homomorphism T1 (N) -* Of /A
arisingfromf. We writeT forTl(N)m 0 (9, where(9 = Qf,A and kmis
W(km,)
the residue fieldof m. Assume that p t N. We assume here that k is the
residuefieldof ( and that it is chosento containkm.Then by Corollary1 of
Theorem2.1, T1 (N)m is Gorensteinand it followsthat T is also a Gorenstein
O-algebra (see the discussionfollowing(2.42)). So we can use perfectpairings
(the second one T-bilinear)
OxO -, (,): TxT -O
to definean invariantrj of T. If 7r: T (9 is the natural map, we set

(r7)= (*ir(1))where*- is the adjoint of 7rwith respectto the pairings. It is
well-definedas an ideal of T, dependingonlyon 7r. Furthermore, as we noted
in Chapter2, ?3, 7r(r)= (?7,77)up to a unitin 0 and as notedin the appendix
77 = Ann p = T[[p] where p = ker7r. We now give an explicit formula for 77
developedby Hida (cf. [Hi2]fora surveyof his earlierresults)by interpreting
(, ) in termsof the cup product pairingon the cohomologyof X1 (N), and
then in termsof the Peterssoninnerproductof f with itself. The following
account(whichdoes not requirethe CM hypothesis)is adapted from[Hi2]and
we referthereformoredetails.
Let
(4.17) ( ):H1 (Xi(N), IOf) x H1 (Xi(N), I)O Of

be the cup productpairingwith Of as coefficients.(We sometimesdrop the
C fromX1(N)IC or Ji(N)IC if the contextmakes it clear that we are re-
ferringto the complex manifolds.) In particular(tx,y) = (x,t*y) for all
x, y and foreach standardHecke correspondence t. We use the action of t on
H1 (Xi (N), Of) givenby x - t*x and simplywritetx fort*x. This is the same

534 ANDREW WILES
as the action induced by t* E T1(N) on H1(Jl(N),Of) H'(Xi(N), Of

Let pf be the minimalprimeof T1 (N) 0 Of associated to f (i.e., the kernelof
T1(N) 0 Of -- Of givenby t10 f H-* /ctt(f) wheretf = ct(f)f), and let
Lf = H1 (X1 (N), Of) [pf].
If f = Eanqn let fP = Eanqn. Then fPis again a newformand we define
LfP by replacingf by fP in the definition of Lf. (Note here that Of = OfP
as these ringsare the integersof fieldswhichare eithertotallyreal or CM by
a resultof Shimura. Actuallythis is not essentialas we could replace Of by
any ringof integerscontainingit.) Then the pairing(, ) induces anotherby
restriction
(4.18) (,): Lf X LfP Of.

ReplacingOf (and the Of-modules)bythelocalizationofOf at p (ifnecessary)
we can assume that Lf and Lfp are freeof rank 2 and directsummandsas
Of-modulesof the respectivecohomologygroups. Let 61,62 be a basis of Lf.
Then also 61,62 is a basis of Lfp = Lf. Here complexconjugationacts on
H1(Xi(N), Of) via its action on Of. We can thenverifythat
(6, 6) :=det(bi, Ej)

is an elementof Of (or its localizationat p) whose image in Of,> is givenby
r(q2) (unit). To see this,considera modifiedpairing( , ) definedby
(4.19) (x, y) = (x, way)

wherewC is definedas in (2.4). Then (tx,y) = (x, ty) forall x, y and Hecke
operatorst. Furthermore
det(bi,6j) = det(6i,w 6j) = cdet(6i,6j)

forsome p-adic unit c (in Of). This is because wC(LfP) = Lf and wC(Lf) =
Lfp. (One can check this, for example, using the explicit bases described
below.) Moreover,by Theorem2.1,
H (X1(N), Z) ft1(N) mlf~m ~ T1N2
H1(Xi(N), Of) ?Tl(N)?of T - T2.

Thus (4.18) can be viewed (aftertensoringwith Of,, and modifying it as in
(4.19)) as a perfectpairingof T-modulesand so this servesto computeir(iq2)
as explainedearlier(the square comingfromthe fact that we have a rank 2
module).
To givea moreusefulexpressionfor(6, 6) we observethat f and fP can be
viewedas elementsofH1 (X1(N), C) HN), C) via f(z)dz, fPI
fPdz. Then {If, fP} forma basis forLf Oof C. Similarly{If, fP} forma basis

forLfp ?Of C. Define the vectorswi = (ffP), W2 = (f, fP) and write
W1= C6 and W2= C6 withC E M2(C). Then writingfi = If,f2 = fP we set
(WIa) :=det((fi,If)) = (6,6) det(CC).
Now (w, c) is givenexplicitlyin termsofthe (non-normalized)
Peterssoninner
product(,):
(W,c') = -4(f, f)2
where(f, f) = fs/r1(N)ffdxdy.
To computedet(C) we considerintegralsoverclasses in H1 (Xi(N), Of).
By Poincare duality there exist classes C1,C2in Hl(Xi(N), Of) such that
det(f 60) is a unit in Of. Hence detC generatesthe same Of-module as
is generatedby {det (fc f)} forall such choicesof classes (Cl, C2) and with
{fi, f2} = {f, fI}. Lettinguf be a generator
oftheOf-module{det (fc fi)}
we have the followingformulaof Hida:
PROPOSITION 4.4. 7r(iR2)= (f, f)2/Uf f x (unit in Of,,A).
Now we restrictto the case where po = IndQ sio for some imaginary
quadratic fieldL whichis unramifiedat p and some kX-valuedcharacterso
of Gal(L/L). We assume that Po is irreducible,i.e., that so 5$ ",, where
KO'a(b) = Ko(aT-lb6) for any o representingthe nontrivial coset of
Gal(L/Q)/ Gal(L/L). In additionwe wish to assume that Po is ordinaryand
detPoI p = w. In particularp splitsin L. These conditionsimplythat,ifp is a
prime of L above p, soi(a) = a-1 mod p on U. after possible replacement of svo
by Ivoa Here the U. are the unitsofL. and since svois a character,the restric-
tion of svoto an inertiagroupI. inducesa homomorphism on U.. We assume
now that p is fixedand so chosen to satisfythis congruence. Our choice of
so willimplythat the grossencharacter introducedbelowhas conductorprime
top.
We choose a (primitive)grossencharacterp on L togetherwith an em-
beddingQ c-* Qp corresponding to the primep above p such that the induced
p-adic characterfpphas the properties:
(i) ppmodp = so (p = maximalideal of Qp).
(ii) fppfactorsthroughan abelian extensionisomorphicto Zp e T withT of

finiteorderprimeto p.
(iii) p((a)) = a fora _1 (f) forsome integralideal f primeto p.

To obtain p it is necessaryfirstto define pp. Let Mo denote the maximal
abelian extensionof L whichis unramifiedoutside p. Let 0: Gal(Mx/L) -+
QpX be any characterwhichfactorsthrougha Zp-extensionand induces the

536 ANDREW WILES
a | - a-1 on U,
homomorphism --* Gal(M,,/L) whereUp,1= {u E Up:u
l(p)}. Then set ',p = 'coO,and pick a grossencharacter p such that ((p)p = (np.
Note that our choice of p here is not necessarilyintendedto be the same as
the choiceof grossencharacter in Section 1.
Now let f, be the conductorof p and let F be the ray class fieldof con-
ductor . Then overF thereis an ellipticcurve,unique up to isomorphism,
withcomplexmultiplication by OL and periodlatticefree,ofrankone overOL
and withassociatedgrossencharacter ipo NF/L. The curveE/F is the extension
of scalars of a unique ellipticcurveE/F+ whereF+ is the real subfieldof F of
index 2. (See [Shl, (5.4.3)].) Over F+ this ellipticcurvehas onlythe p-power
isogeniesofthe form?pm form E Z. To see thisobservethat F is unramified
at p and po is ordinaryso that the only isogeniesof degreep over F are the
ones that correspondto divisionby kerp and kerp' wherepp' = (p) in L. Over
F+ thesetwosubgroupsare interchanged by complexconjugation,whichgives
the assertion.We let E/O9F+ (p) denote a Weierstrassmodel over OF+,(p), the
localizationof OF+ at p, withgood reductionat the primesabove p. Let WE
be a Neron differential of E/OF+<( ) Let Q be a basis forthe OL-module of
periods of WE. Then Q = u Q for some p-adic unit in FX.
Accordingto a theoremofHecke, p is associatedto a cusp formfAlin such
a way that the L-seriesL(s, cp) and L(s, fg) are equal (cf. [Sh4, Lemma 3]).
Moreoversince p was assumed primitive,f = fgsis a newform.Thus the
integerN = condf = IAL/QINormL/Q(condcp) is primeto p and thereis a
homomorphism
'Of: Ti(N)-4?Rf C Of C 0cp
satisfyingof (Tl) = (p(c)+ (P(C)if1 = ccin L, (I t N) and o)f((Ti)= 0 if1 is inert

in L (1 t N). Also of ((l(l)) = (p((l))?o(l) where / is the quadraticcharacter
associated to L. Using the embeddingof Q in Qp chosen above we get a
prime A of Of above p, a maximal ideal m of T1 (N) and a homomorphism
Tl(N), )
Of,- , such that the associated representationpf,Areduces to
pomodA.
Let po = kerf f:T1 (N) -+ Of and let
Af = Ji (N)/poJ1 (N)
be the abelian varietyassociatedto f by Shimura.OverF+ thereis an isogeny
Af/F+ (E/F+ )d
whered = [Of: Z] (see [Sh4,Th. 1]). To see thisone checksthat thep-adic Ga-
lois representationsassociatedto the Tate moduleson each side are equivalent
to (Ind ,op)Ozp Kf,pwhereKf,p= Of 0 Qp and where(pr:Gal(F/F) Zx *
p
is the p-adic characterassociated to and restrictedto F. (One compares
trace(Frob?) in the two representations for? t Np and ? split completelyin
F+; cf. the discussionafterTheorem2.1 forthe representation on Af.)

Now pick a nonconstantmap
7r:X1(N) IF+ -+EIF+
whichfactorsthroughAf/F+. Let M be the compositeof F+ and the nor-

mal closureof Kf viewedin C. Let WE be a Neron differential
of E/OF+(p).
Extendingscalars to M we can write
ir*WE = E aawfa a, E M
aEHom(Kf ,C)
00
wherewfu = E an(fo)qfln foreach a. By suitablychoosing7rwe can assume

n=l
that aid =$0. Then thereexist Ai E OM and ti E T1(N) such that
Z Aiti-7r*WE = ClWf for some Cl E M.

We considerthe map
(4.20) 7r:Hi (Xi (N)/C, Z) 0 O9M,(p) - Hi (E/C, Z) 0 O9M,(p)
givenby 7r'= E Ai(7ro i). Even if 7r'is not surjectivewe claim that the image
of 7r'always has the formHi(EIC, Z) 0 aOM,(p) forsome a E OM* This is
because tensoredwith Zp 7r'can be viewedas a Gal(Q/F+)-equivariant map
ofp-adic Tate-modules,and the onlyp-powerisogenieson E/F+ have the form
?pm forsome m E Z. It followsthat we can factor7r'as (1 0 a) o a forsome
othersurjectivea
a: Hi(Xi(N)/C, Z) 0 Om -+ H (E/C, Z) 0) OM,
now allowinga to be in OM,(p) . Now definea* on QE bya* = a-lAiti o r*

where-*QE/C QJ1(N)/C is the map induced by 7rand ti has the usual
action on QJ1(N)C Then a*(WE) = CWffor some c E M and
(4.21) Ja*(wE)= JWE

Y a(y)
forany class y e Hi(Xi(N)/C, OM). We note that a (on homologyas in

(4.20)) also comes froma map of abelian varietiesa: J1(N)/F+ 'z OM
E/F+ ?Z OM althoughwe have not used this to definea**.
We claim now that c E OM,(p). We can computea*(WE) by considering
a**(WE 01) = ,Ztlr* 0 a-'Ai on QE0F+ 0 OM and then mapping the image in
Q0 OM to QJ1(N)/F+ OF+ OM = QJ1(N)/M Now let us writeO1 for
9F+,(p). Then thereare isomorphisms
8f
Q1
Q3
)/c~lecv2
-82 Hom~c)M,Q11(N) )
ii(N)1,1-0
Hom(OM, J1 (N)101) N)1~0 6

538 ANDREW WILES
where6 is the different of M/Q. The firstisomorphismcan be describedas

follows.Let e(-y):J1(N) -+ J1(N) 0 Opmfor-yE OM be themap x | >x 0 'y.
Then ti(w)(-y)= e(-y)*w.Similaridentificationsoccur forE in place of Ji(N).
So to checkthat a* (WE 0 1) E Q1 0 OM it is enoughto observethat by
its constructiona comesfroma homomorphism J1(N)/01 0 Om -k E1010 OM.
It followsthat we can comparethe periodsof f and of WE.
For fP we use the fact that flyfPdz = fCYf dz wherec is the OM-linear
map on homologycomingfromcomplexconjugationon the curve. We deduce:
PROPOSITION 4.5. Uf = 4 Q2.(1/(p-adic integer)).
We now give an expressionfor (f,, f,) in termsof the L-functionof p.
This was firstobservedby Shimura [Sh2] althoughthe preciseformwe want
was givenby Hida.
PROPOSITION 4.6.
(ft2'fib
=16 3 2 I (1-- } 27 X) LN(1,4)
whereX is the characterof figand X its restrictionto L;

b is the quadraticcharacterassociatedto L;
LN( ) denotesthattheEuler factorsforprimesdividingN have been
removed;
S. is the set of primesq I N such thatq = qq' withq t condW and q,q'
primesofL, not necessarilydistinct.
Proof. One beginswith a formulaof Peterssonthat foran eigenform

of
weight2 on F1(N) says
(f,f) = (47)-2
F
(2) (1)7r[SL2(Z): Ii(N) (+1)] Res8=2D(s, f,fP)
.
whereD(s, f, fP) = E Ian 2n-s iff = E anqn (cf. [Hi3,(5.13)]). One checks
n=1 n=1
that, removingthe Euler factorsat primesdividingN,
DN(S, f,fP) = LN(S, p2k)LN(S - 1, 0k)(QN(s - 1)/(QN(2s - 2)
by usingLemma 1 of [Sh3]. For each Euler factorof f at a q I N of the form

(1- aq q-8) we get also an Euler factorin D(s, f, fP) ofthe form(1- qaq q-q8).
When f = fl this can only happen fora split primeq whereq' divides the
conductorof p but q does not,or fora ramifiedprimeq whichdoes not divide
the conductorof Ap.In this case we get a term(1 - ql-8) since I9 (q)12 = q.
Puttingtogetherthe propositionsofthissectionwe nowhave a formulafor
as definedat the beginningofthissection. Actuallyit is moreconvenient
7r(?7)

to give a formulafor 7r(?JM),an invariantdefinedin the same way but with

Ti(M)ml 0 (9 replacingTl(N)m 0 (9 where M = pMo with p t MO
and M/N is of the form
lb IJ q2.
qESW q f N
qIM0
that Pm1= po, Uq E m ifq I M (q $ p)

Here ml is definedby the requirements
and thereis an embedding(whichwe fix) km1 ) k over ko takingUp -ap
whereap is the unit eigenvalueof Frobp in pfLA. So if f is the eigenform
obtained fromf by 'removingthe Euler factors'at q I (M/N) (q $ p) and
removingthe non-unitEuler factorat p we have AM = *(1) where7w: T, =
T1 (M)ml 0 (9 - (9 correspondsto f' and the adjointis takenwithrespect
W(km1 )
to perfectpairingsof T1 and (9 with themselvesas 0-modules, the firstone
assumedTi-bilinear.
Property(ii) of fp ensuresthat M is as in (2.24) withD = (Se, A, 0, q)
whereE is the set of primesdividingM. (Note that S. is preciselythe set of
primesq forwhichnq = 1 in the notationof Chapter2, ?3.) As in Chapter2,
?3 thereis a canonicalmap
(4.22) RV 4 TVZ Ti(M)ml 0 0

W(km1 )
whichis surjectiveby the argumentsin the proofof Proposition2.15. Here

we are consideringa slightlymore generalsituationthan that in Chapter 2,
?3 as we are allowingpo to be inducedfroma characterof Q(i/=3). In this
special case we defineTD to be T, (M)m, 0( 0. The existenceof the map
W(km1,)
in (4.22) is provedas in Chapter2, ?3. For the surjectivity,
note that foreach
q I M (with q $ p) Uq is zero in TV as Uq E m1 foreach such q so that we
can apply Remark 2.8. To see that Up is in the image of Rv we use that it
is the eigenvalueof Frobp on the unique unramifiedquotientwhichis freeof
rankone in the representation p describedafterthe corollariesto Theorem2.1
(cf. Theorem 2.1.4 of [Wil]). To verifythis one checksthat TV is reduced
or alternativelyone can apply the methodof Remark2.11. We deduce that
E T, the W(kmi)-subalgebraof T1(M)mi generatedby the traces,and it
followsthenthat it is in the imageof RV. We also need to give a definitionof
TV whereD = (ord,A, 0), ) and po is induced froma characterof Q(V/=3).
For this we use (2.31).
Now we take
M=Np fJq.

540 ANDREW WILES
The argumentsin the proofof Theorem2.17 show that

ir(,M) is divisibleby ir(r/)(a2- (p)) 171
(q - 1)
qESFp
whereap is the unit eigenvalueof Frobp in Pf,A\.
The factorat p is givenby
remark2.18 and at q it comesfromthe argumentofProposition2.12 but with
H = H' = 1. CombiningthiswithPropositions4.4, 4.5, and 4,6, we have that
(4.23) lr(r7M)is divisiblebyQ2LN 2, X) () -(p)) 171

(q- 1).
qIN
We deduce:
THEOREM 4.7. #(O/r(?JM))= #Hs
Proof. As explainedin Chapter2, ?3 it is sufficient
to provethe inequality
#((9/lr(?JM)) > #HSe(Qr/Q7 V) as the oppositeone is immediate.For thisit
sufficesto compare (4.23) withProposition4.3. Since
LN(2, FI) = LN(2, v) = LN(2, 92X)
(note that the right-handtermis real by Proposition4.6) it suffices
to pair up
the Euler factorsat q forq I N in (4.23) and in the expressionforthe upper
bound of # Hse(Q/Q, V). L
We now deduce the main theoremin the CM case using the methodof
Theorem2.17.
THEOREM 4.8. Suppose thatpo as in (1.1) is an irreduciblerepresen-
tation of odd determinantsuch thatpo = Ind? so for a characterso of an
imaginaryquadraticextensionL of Q whichis unramified at p. Assume also
that:
(i) detpo =W;

'p
(ii) po is ordinary.
Thenfor every7D= (-,E,0,0) such thatpo is of typeVDwith = Se or ord,
Rv - Tvz
and TD is a completeintersection.
COROLLARY. For any po as in the theoremsupposethat
p: Gal(Q/Q) -) GL2(0)
is a continuousrepresentationwithvalues in the ring of integersof a local
outsidea finiteset of primes,satisfying
field,unramified p ~ po whenviewed
as representations
to GL2((Fp). Supposefurtherthat:

ELLIPTICCURVESAND FERMAT'S
MODULAR LASTTHEOREM 541
(i) P D is ordinary;
(ii) detp = Xek- withX offiniteorder,k > 2.
Thenp is associatedto a modularformof weightk.
Chapter 5
In this chapterwe provethe main resultsabout ellipticcurvesand espe-

cially show how to removethe hypothesisthat the representation associated
to the 3-divisionpointsshould be irreducible.
Application to elliptic curves
The key resultused is the followingtheoremof Langlands and Tunnell,

extendingearlierresultsof Hecke in the case wherethe projectiveimage is
dihedral.
THEOREM 5.1 (Langlands-Tunnell). Suppose that p: Gal(Q/Q)
GL2(C) is a continuousirreduciblerepresentation whose image is finiteand
solvable. Suppose furtherthatdetp is odd. Then thereexists a weightone
newformf such thatL(s,f) = L(s,p) up to finitelymanyEuler factors.
Langlands actually proved in [La] a much more general result without
restrictionon the determinantor the numberfield(whichin our case is Q).
Howeverin the crucialcase wherethe image in PGL2(C) is S4, the resultwas
onlyobtainedwithan additionalhypothesis.This was subsequentlyremoved
by Tunnellin [Tu].
Suppose thenthat
po: Gal(Q/Q) -- GL2(F3)
is an irreduciblerepresentationof odd determinant. We now show, using

the theorem,that this representationis modular in the sense that over F3,
po - pg,j,mod gLforsome pair (g,I-) withg some newform ofweight2 (cf. [Se,
?5.3]). There existsa representation
i: GL2(F3) -* GL2 (Z [VZ]) c GL2(C).
By composingi withan automorphism ofGL2(F3) ifnecessarywe can assume
that i induces the identityon reductionmod (1 + vi2). So if we consider

542 ANDREW WILES
i 0PO: Gal(Q/Q) -+ GL2(C) we obtain an irreduciblerepresentation whichis

easily seen to be odd and whose image is solvable. Applyingthe theoremwe
finda newform f ofweightone associatedto thisrepresentation.Its eigenvalues
lie in Z [VA/]. Now picka modularformE ofweightone such that E -1(3).
For example,we can take E = 6 El, whereE1,x is the Eisensteinserieswith
Mellintransform givenby ((s) ((s, X) forX the quadraticcharacterassociated
to Q(VA/=). Then fE f mod3 and using the Deligne-Serrelemma ([DS,
Lemma 6.11]) we can findan eigenform g' ofweight2 withthesame eigenvalues
as f modulo a primepu'above (1 + X/=2). There is a newformg of weight2
whichhas the same eigenvaluesas g' foralmostall T1's,and we replace (9', eu')
by (g,,u) forsome prime1I-above (1 + j/2). Then the pair (g,,u) satisfiesour
requirements fora suitablechoiceof ,u (compatiblewithps').
We can apply this to an ellipticcurve E definedover Q by considering

E[3]. We nowshowhowin studyingellipticcurvesour restriction to irreducible
in the deformation
representations theorycan be circumvented.
THEOREM 5.2. All semistableellipticcurvesoverQ are modular.
Proof. Suppose that E is a semistableelliptic curve over Q. Assume

firstthat the representation-E,3 on E[3] is irreducible. Then if Po = PE,3
restrictedto Gal(Q/Q (vZ/=)) werenotabsolutelyirreducible, theimageofthe
restrictionwouldbe abelian of orderprimeto 3. As the semistablehypothesis
impliesthat all the inertiagroupsoutside 3 in the splittingfieldof Po have
orderdividing3 this means that the splittingfieldof Po is unramified outside
3. However,Q(V/Z3) has no nontrivialabelian extensionsunramified outside3
and oforderprimeto 3. So Po itselfwouldfactorthroughan abelian extension
of Q and this is a contradictionas Po is assumed odd and irreducible. So
Po restrictedto Gal(Q/Q(VE/=)) is absolutelyirreducibleand PE,3 is then
modularby Theorem0.2 (provedat the end of Chapter3). By Serre'sisogeny
theorem,E is also modular(in the sense ofbeinga factorof the Jacobianof a
modularcurve).
So assume now that PE,3 is reducible. Then we claim that the represen-
tationP-E,5on the 5-divisionpointsis irreducible.This is because Xo(15) (Q)
has only fourrationalpointsbesides the cusps and these correspondto non-
semistablecurveswhichin any case are modular;cf. [BiKu, pp. 79-80]-.If we
knewthat -E,5 was modularwe could nowprovethe theoremin the same way
we did knowingthat PE,3 was modularonce we observethat PE,5 restrictedto
Gal(Q/Q(x/5)) is absolutelyirreducible.This irreducibility followsa similar
argumentto the one for -E,3 since the only nontrivialabelian extensionof
outside5 and oforderprimeto 5 is Q((5) whichis abelian
Q (Vs) unramified
over Q. Alternatively,it is enoughto checkthat thereare no ellipticcurves
E forwhichPE,5 is an inducedrepresentationoverQ(v/5) and E is semistable

at 5. This can be checkedin the supersingularcase using the descriptionof

PE,5 ID5(in particularit is inducedfroma characterofthe unramified
quadratic
extensionof Q5 whose restrictionto inertiais the fundamentalcharacterof
level 2) and in the ordinarycase it is straightforward.
Consider the twistedformX(p)lQ of X(5)/Q definedas follows. Let
X(5)/Q be the (geometrically disconnected)curvewhose non-cuspidalpoints
classifyellipticcurveswithfulllevel 5 structureand let the twistedcurvebe
definedby the cohomologyclass (even homomorphism) in
H1(Gal(L/Q), AutX(5)/L)
givenby PE,5: Gal(L/Q) GL2(Z/5Z) C Aut X(5)/L whereL denotesthe
-
of
splittingfield PE,5. Then E definesa rationalpoint on X(p)/Q and hence
also of an irreduciblecomponentof it whichwe denote C. This curve C is
smoothas X(p)lQ = X(5)/Q is smooth. It has genus zero since the same is
trueof the irreduciblecomponentsof X(5)-Q.
A rationalpointon C (necessarilynon-cuspidal)correspondsto an elliptic
curveE' overQ withan isomorphism E'[5] - E[5] as Galois modules(cf. [DR,
VI, Prop. 3.2]). We claim that we can choose such a point with the two
propertiesthat (i) the Galois representation PE',3 is irreducibleand (ii) E' (or
a quadratictwist)has semistablereductionat 5. The curveE' (or a quadratic
twist)willthensatisfyall thepropertiesneededto applyTheorem0.2. (For the
primesq : 5 we just use the factthat E' is semistableat q ==> # E-,5(Iq) I 5.)
So E' will be modularand henceso too will PE',5
To picka rationalpointon C satisfying (i) and (ii) we use the Hilbertirre-
ducibilitytheorem.For,to ensurecondition(i) holds,we onlyhave to eliminate
the possibilitythat the image ofPE',3 is reducible.But this correspondsto E'
being the image of a rationalpoint on an irreduciblecoveringof C of degree
4. Let Q(t) be the functionfieldof C. We have therefore an irreduciblepoly-
nomial f(x, t) E Q (t) [x] of degree> 1 and we need to ensurethat formany
values to in Q, f(x, to) has no rational solution. Hilbert's theoremensures
that thereexists a t1 such that f(x, ti) is irreducible.Then we pick a prime
P1 $ 5 such that f(x, t1) has no rootmod P1. (This is easilyachievedusingthe
Cebotarev densitytheorem;cf. [CF, ex. 6.2, p. 362].) So finallywe pick any
to E Q whichis pl-adicallyclose to t1 and also 5-adicallyclose to the original
value of t givingE. This last conditionensuresthat E' (corresponding to to)
or a quadratic twisthas semistablereductionat 5. To see this,observethat
since JE $A0, 1728, we can finda familyE(j): y2 = X3- 92(j)x - 93(j) with
rationalfunctions92(ij, 93(j) whichare finiteat jE and withthej-invariantof
E(jo) equal to jo wheneverthegi(jo) are finite.Then E is givenby a quadratic
twistof E(jE) and so aftera changeof functionsof the form92(i) F u2g2(j),
93(j) F U3g3(j) with u E Qx we can assume that E(jE) = E and that the
equationE(jE) is minimalat 5. Then forj' E Q close enough5-adicallyto jE

544 ANDREW WILES
theequationE(j') is stillminimaland semistableat 5, sincea criterionforthis,

foran integralmodel,is that eitherord5(A(E(j'))) = 0 or ord5(c4(E(j'))) = 0.
So up to a quadratictwistE' is also semistable.
This kindof argumentcan be applied moregenerally.
THEOREM 5.3. Suppose thatE is an ellipticcurvedefinedover Q with
thefollowingproperties:
(i) E has good or multiplicative
reductionat 3, 5,
(ii) For p = 3,5 and foranyprimeq -1 modp either1E,pIDq is reducible
overF orPEIpIIq is irreducibleoverFP.
Then E is modular.
Proof. The main point to be checkedis that one can carryover condi-
tion (ii) to the new curveE'. For this we use that forany odd primep q$
A
PEXpIDq is absolutelyirreducibleand -EXP
IIq is absolutelyreducible
and 3 t #PE)p(Iq)
E acquiresgood reductionoveran abelian 2-powerextensionof

Qqnrbut not overan abelian extensionof Qq.
Suppose thenthat q _-1(3) and that E' does not satisfycondition(ii) at
q (forp = 3). Then we claim that also 3 t #PE',3(Iq). For otherwisePEj,3(Iq)
has its normalizerin GL2(F3) containedin a Borel, whence -El,3(Dq) would
be reduciblewhichcontradictsour hypothesis.So usingthe above equivalence
we deduce,by passingvia -E',5 PE,5, that E also does not satisfyhypothesis
(ii) at p = 3.
We also need to ensurethat PE',3 is absolutelyirreducibleover Q(VW3).
This we can do by observingthatthepropertythattheimageof -E',3 lies in the
Sylow2-subgroupof GL2(F3) impliesthat E' is the image of a rationalpoint
on a certainirreduciblecoveringof C of nontrivialdegree.We can thenargue
in the same way we did in the previoustheoremto eliminatethe possibility
that -E',3 was reducible,this timeusingtwo separatecoveringsto ensurethat
the image of -E',3 is neitherreduciblenor containedin a Sylow2-subgroup.
Finally one also has to show that if both PE,5 is reducibleand PE,3 is
induced froma characterof Q (v/Z3) then E is modular. (The case where
both were reduciblehas already been considered.) Taylor has pointed out
that curvessatisfying both these conditionsare classifiedby the non-cuspidal
rationalpointson a modularcurve isomorphicto Xo(45)/Wg, and this is an
ellipticcurve isogenousto Xo(15) with rank zero over Q. The non-cuspidal
rationalpointscorrespondto modularellipticcurvesof conductor338. El

Appendix
Gorenstein rings and local complete intersections
PROPOSITION 1. Suppose that0 is a completediscretevaluationring

and thatp: S -+ T is a suriectivelocal0-algebra homomorphism betweencom-
plete local Noetherian0-algebras. SupposefurtherthatPT is a prime ideal of
'
T such that T/PT 0 and let Ps = W-1(PT). Assume that
(i) T - O(X1, ...

,Xr1/(fi, .. ,fr-u) wherer is thesize of a minimalset of
0-generatorsofPT/PT,
(ii) o inducesan isomorphism
ps/p2 PT/P2 and thattheseare finitely
generated0-modules whosefreepart has ranku.
Then o is an isomorphism.
Proof. First we considerthe case whereu = 0. We may assume that the

generators x1,. . .,Xr lie in PT by subtracting their residues in T/PT ) 0. By
(ii) we may also write
-
S (9JX1,. * Xrll(gl,* v9,)
with s > r (by allowingrepetitionsif necessary)and Ps generatedby the

images of {x1,.. .,Xr}. Let p = (X1,... ,Xr) in 0Jx1,. .xr] Writing fi
Eaijxj mod p2 withaij E 0, we see that the Fittingideal as an 0-module of
PT/PT is givenby
F (PT/Pp) = det(aij) E 0
and that this is nonzeroby the hypothesisthat u = 0. Similarly,if each

gi =Ebijxj modp2, then
Fo(ps/p2) = {det(bij): i E I, #I=r, I C {1, ...,
By (ii) again we see that det(aij) = det(bij) as ideals of 0 forsome choice Io

of I. Afterrenumbering we may assume that Io = {1, . , r} Then each gi
(i = 1,... ,r) can be written gi = Erijfi for some rij E 0x1,.. .,Xr]J and we
have
det(bij) _ det(rij) det(aij) mod p.
Hence det(rij) is a unit,whence(rij) is an invertiblematrix.Thus the fi's can
be expressedin termsof the gi's and so S _T.
We can extendthisto the case u $&0 by pickingX1,.. , Xr-u so that they
generate(pT/p2)tors. Then we can writeeach fi = Er-uaIj x modp2 and
likewiseforthe gi's. The argumentis now just as beforebut applied to the
Fittingideals of (pT/p2)tors. L

546 ANDREW WILES
For the next propositionwe continueto assume that 0 is a complete

discretevaluationring.Let T be a local 0-algebra whichas a moduleis finite
and freeover (9. In addition,we assume the existenceof an isomorphismof
T-modulesT --0 Homr(T, 0). We call a local 0-algebra whichis finiteand
freeand satisfiesthis extra conditiona Gorenstein0-algebra (cf. ?5 of [Til]).
Now suppose that p is a primeideal of T such that T/p _X.
Let o3:T -- Tlp - 0 be the naturalmap and definea principalideal of T
by
(AT) = C3(1))
where 3: ( --) T is the adjoint of /3with respectto perfect0-pairings on 0

and T, and wherethe pairingof T with itselfis T-bilinear. (By a perfect
pairingon a free0-module M of finiterankwe mean a pairingM x M -( 0
suchthat boththe inducedmaps M -) Homo(M, 0) are isomorphisms.When
M = T we are thusrequiringthat thisbe an isomorphism ofT-modulesalso.)
The ideal (AT) is independentof the pairings.Also T/TT is torsion-freeas an
0-module, as can be seen by applyingHom ( , 0) to the sequence
0 -+ p -T --+ 0 -+0,
to obtain a homomorphism T/T/T Hom(p,0). This also showsthat (OT) =

Annp.
If we let l(M) denotethe lengthof an 0-module M, then
I (P/P ) > 1(0/t)
(wherewe write 7T-for/3(qT)) because p is a faithfulT/IqT-module.(For a
briefaccount of the relevantpropertiesof Fittingideals see the appendix to
[MW1].) Indeed, writingFR(M) forthe Fittingideal of M as an R-module,
we have
FT/1T (P) = 0 X FT(P) C (r/T)E FT/p,(P/P2)C (AT)
and we then use the fact that the lengthof an 0-module M is equal to the
lengthof 09/FO(M) as 0 is a discretevaluationring. In particularwhenp/p2
is a torsion0-modulethen77T $ 0.
We need a criterionfora Gorenstein0-algebra to be a completeinter-
section. We will say that a local 0-algebra S which is finiteand freeover
0 is a complete intersectionover 0 if there is an 0-algebra isomorphism
S _ Oj[xi,.* ., Xrj/(fi *... , fr) forsome r. Such a ringis necessarilya Goren-
stein 0-algebra and {fi, .. , fir}is necessarilya regularsequence. That (i) =X
(ii) in the followingpropositionis due to Tate (see A.3, conclusion4, in the
appendix in [M Ro].)
PROPOSITION 2. Assume that0 is a completediscretevaluationrmng
and thatT is a local Gorenstein0-algebra whichis finiteand freeover0 and

thatPT is a prime ideal of T such thatT/PT _ 0 and PT/PT is a torsion

0-module. Then thefollowingtwo conditionsare equivalent:
(i) T is a completeintersectionover0.
(ii) l(PT/PT) = 1(0//7T-)as 0-modules.
S over0 (so
Proof. To provethat (ii) =X (i), pick a completeintersection
assumedfiniteand flatover0) suchthat a: S-ET and suchthatps/p2 , PT/P2
wherePs = r-1 (PT). The existenceof such an S seems to be well known
(cf. [Ti2, ?6]) but here is an argumentsuggestedby N. Katz and H. Lenstra
(independently).
Write T = 0[xi,. ..,xr]/(fl,.. .,f) with PT the image in T of p =
(Xi, ... , Xr). Since T is local and finite and free over 0 , it follows that also
T - (90X1 X* X ,fs)). We can pick 91,..., gr such that gi = Eaijf;
,Xr]j/(fi, X* X
withaij E (9 and such that
(fiI. ..- f8,p2) = (9gi.. igrp2).
We thenmodifygi, ... , gr by the additionofelements{&I} of (f, .. ., f8)2and

set (gj = gi + a1,,..., gr = gr+ ar). Since T is finiteover0, thereexistsan N
such that for each i, xjv can be written in T as a polynomial hi(Xi .... Xsr) of
total degreeless than N. We can assume also that N is chosengreaterthan
the total degree of gi for each i. Set ai = (x' - hi(xi, ...,x))2. Then set
S = (9 X ... XXr]/(g1 .(.g . , g). Then S is finiteoverX by construction
and also
dim(S) < 1 sincedim(S/A) = 0 where(A) is the maximalideal of 0. It follows
that {9g,. . . ,g9} is a regularsequenceand hencethat depth(S) = dim(S) = 1.
In particularthe maximal 0-torsionsubmoduleof S is zero since it is also a
finitelengthS-submoduleof S.
Now 0/(i-s) - 0/(ft), since l(01/(7s)) = l(ps/p2) by (i) X, (ii) and
I(0/(ftr)) = l(PT/p2) by hypothesis.Pick isomorphisms
T _Homo(T, (), S - Homo(S, 0)

as T-modules and S-modules, respectively.The existenceof the latter for
over 0 is well known;cf. conclusion1 of TheoremA.3
completeintersections
of [M Ro]. Then we have a sequence of maps, in which& and ,3 denote the
adjointswithrespectto these isomorphisms:
oA T S T2-0.
One checksthat & is a map of S-modules (T being givenan S-action via a)
and in particularthat a o & is multiplicationby an elementt of T. Now
(,3 0,L)= (ft7) in O and (,3 o a) o ( o a) = (vs) in (9. As (as) = (NT) in C, we
have that t is a unitmodPT and hencethat a o & is an isomorphism.It follows

548 ANDREW WILES
that S z- T, as otherwiseS _ kera $ ima is a nontrivialdecompositionas

S-modules,whichcontradictsS being local. [1
Remark. Lenstra has made an importantimprovementto this proposi-

tion by showingthat replacing -ZTby /3(annp) gives a criterionvalid forall
local (9-algebraswhichare finiteand freeover(9, thuswithoutthe Gorenstein
hypothesis.
PRINCETON UNIVERSITY, PRINCETON, NJ
REFERENCES
[AK] A. ALTMAN and S. KLEIMAN, An Introductionto GrothendieckDuality Theory,vol.

146, SpringerLecture Notes in Mathematics, 1970.
[BiKu] B. BIRCH and W. KUYK (eds.), Modular Functions of One Variable IV, vol. 476,
SpringerLecture Notes in Mathematics, 1975.
[Bo] Boston, N., Families of Galois representations-Increasingthe ramification, Duke
Math. J. 66, 357-367.
[BH] W. BRUNS and J. HERZOG, Cohen-Macaulay Rings, Cambridge UniversityPress,
1993.
[BK] S. BLOCH and K. KATO, L-Functions and Tamagawa Numbers of Motives, The
GrothendieckFestschrift, Vol. 1 (P. Cartier et al. eds.), Birkhduser,1990.
[BLR] N. BOSTON, H. LENSTRA, and K. RIBET, Quotients of group ringsarisingfromtwo-
dimensionalrepresentations,C. R. Acad. Sci. Paris t312, Ser. 1 (1991), 323-328.
[CF] J. W. S. CASSELS and A. FR6LICH (eds.), AlgebraicNumberTheory,Academic Press,
1967.
[Cal] H. CARAYOL, Sur les representationsp-adiques associees aux formesmodulairesde
Hilbert,Ann. Sci. Ec. Norm. Sup. IV, Ser. 19 (1986), 409-468.
[Ca2] , Sur les representationsgaloisiennesmodulo e attachees aux formesmod-
ulaires, Duke Math. J. 59 (1989), 785-801.
[Ca3] , Formes modulaireset representationsGaloisiennes a valeurs dans un an-
neau local complet,in p-Adic Monodromyand the Birch-Swinnerton-Dyer Conjec-
ture (eds. B. Mazur and G. Stevens), Contemp. Math., vol. 165, 1994.
[CPS] E. CLINE, B. PARSHALL, and L. SCOTT, Cohomologyof finitegroups of Lie type I,
Publ. Math. IHES 45 (1975), 169-191.
[CS] J. COATES and C. G. SCHMIDT, Iwasawa theoryforthe symmetric square of an elliptic
curve, J. reine und angew. Math. 375/376 (1987), 104-156.
[CW] J. COATES and A. WILES, On p-adic L-functionsand ellipticunits,Ser. A26, J. Aust.
Math. Soc. (1978), 1-25.
[CO] R. COLEMAN, Division values in local fields,Invent.Math. 53 (1979), 91-116.
[DR] P. DELIGNE and M. RAPOPORT, Schemas de modulesde courbeselliptiques,in Springer
Lecture Notes in Mathematics,Vol. 349, 1973.
[DS] P. DELIGNE and J-P. SERRE, Formes modulaires de poids 1, Ann. Sci. Ec. Norm.
Sup. IV, Ser. 7 (1974), 507-530.
[Dia] F. DIAMOND, The refinedconjectureof Serre, to appear in Proc. 1993 Hong Kong
Conf. on Modular Forms and Elliptic Curves.
[Di] L. E. DICKSON, Linear Groups withan Expositionof the Galois Field Theory,Teub-
ner, Leipzig, 1901.

[Dr] V. DRINFELD, Two-dimensionalt-adic representationsof the fundamentalgroup of

a curve over a finitefieldand automorphicformson GL(2), Am. J. Math. 105
(1983), 85-114.
[El] B. EDIXHOVEN, The weightin Serre's conjectureon modular forms,Invent.Math.
109 (1992), 563-594.
[E2] , L'action de l'algebre de Hecke sur les groupes de composantesdes jacobi-
ennes des courbes modulairesest "Eisenstein",in Courbes Modulaires et Courbes
de Shimura,Asterisque196-197 (1991), 159-170.
[Fl] M. FLACH, A finiteness theoremforthe symmetricsquare ofan ellipticcurve,Invent.
Math. 109 (1992), 307-327.
[Fo] J.-M.FONTAINE, Sur certainstypesde representations p-adiques du groupede Galois
d'un corps local; constructiond'un anneau de Barsotti-Tate,Ann. of Math. 115
(1982), 529-577.
[Fr] G. FREY, Links between stable elliptic curves and certain diophantineequations,
Annales UniversitatisSaraviensis 1 (1986), 1-40.
[Grel] R. GREENBERG, Iwasawa theoryforp-adic representations, Adv. St. Pure Math. 17
(1989), 97-137.
[Gre2] , On the structureofcertainGalois groups,Invent.Math. 47 (1978), 85-99.
[Gro] B. H. GROSS, A tamenesscriterionforGalois representationsassociated to modular
formsmodp, Duke Math. J. 61 (1990), 445-517.
[Guo] L. Guo, General Selmer groups and critical values of Hecke L-functions,Math.
Ann. 297 (1993), 221-233.
[He] Y. HELLEGOUARCH, Points d'ordre2ph sur les courbeselliptiques,Acta Arith.XXVI
(1975), 253-263.
[Hil] H. HIDA, Iwasawa modules attached to congruencesof cusp forms,Ann. Sci. Ecole
Norm. Sup. (4) 19 (1986), 231-273.
[Hi2] , Theory of p-adic Hecke algebras and Galois representations, Sugaku Ex-
positions 2-3 (1989), 75-102.
[Hi3] , Congruences of cusp formsand special values of their zeta functions,
Invent.Math. 63 (1981), 225-261.
[Hi4] , On p-adic Hecke algebras forGL2 over totally real fields,Ann. of Math.
128 (1988), 295-384.
[Hu] B. HUPPERT, Endliche GruppenI, Springer-Verlag, 1967.
[Ih] Y. IHARA, On modular curves over finitefields,in Proc. Intern. Coll. on discrete
subgroupsof Lie groups and application to moduli, Bombay, 1973, pp. 161-202.
[Iw1] K. IWASAWA,Local Class Field Theory,OxfordUniversityPress, Oxford,1986.
[Iw2] , On Zi-extensionsof algebraic numberfields,Ann. of Math. 98 (1973),
246-326.
[Ka] N. KATZ, A result on modular formsin characteristicp, in Modular Functions of
One VariableV, SpringerL. N. M. 601 (1976), 53-61.
[Kul] E. KUNZ, Introductionto CommulativeAlgebraand AlgebraicGeometry, Birkhaiiser,
1985.
[Ku2] , Almostcompleteintersections are not Gorenstein,J. Alg. 28 (1974), 111-
115.
[KM] N. KATZ and B. MAZUR, ArithmeticModuli ofElliptic Curves,Ann. ofMath. Studies
108, PrincetonUniversityPress, 1985.
[La] R. LANGLANDS, Base Changefor GL (2), Ann. of Math. Studies, PrincetonUniver-
sity Press 96, 1980.
[Li] W. Li, Newformsand functionalequations, Math. Ann. 212 (1975), 285-315.
[Liv] R. LIVNE, On the conductorsof mod e Galois representationscomingfrommodular
forms,J. of No. Th. 31 (1989), 133-141.
[Mal] B. MAZUR, DeformingGalois representations,in Galois Groups over Q, vol. 16,
MSRI Publications,Springer,New York, 1989.

550 ANDREW WILES
[Ma2] , Modular curves and the Eisensten ideal, Publ. Math. IHES 47 (1977),
133-186.
[Ma3] , Rational isogeniesof primedegree,Invent.Math. 44 (1978), 129-162.
[M Ri] B. MAZUR and K. RIBET, Two-dimensionalrepresentationsin the arithmeticof
modular curves, Courbes Modulaires et Courbes de Shimura,Asterisque 196-197
(1991), 215-255.
[M Ro] B. MAZUR and L. ROBERTS, Local Euler characteristics,Invent. Math. 9 (1970),
201-234.
[MT] B. MAZUR and J. TILOUINE, Representationsgaloisiennes,differentielles de Kihler
et conjecturesprincipales,Publ. Math. IHES 71 (1990), 65-103.
[MW1] B. MAZUR and A. WILES, Class fieldsof abelian extensionsof Q, Invent.Math. 76
(1984), 179-330.
[MW2] , On p-adic analytic familiesof Galois representations,Comp. Math. 59
(1986), 231-264.
[Mil] J. S. MILNE, Jacobian varieties,in Arithmetic Geometry (Cornell and Silverman,
eds.), Springer-Verlag, 1986.
[Mi2] , ArithmeticDuality Theorems,Academic Press, 1986.
[Ram] R. RAMAKRISHNA, On a variationof Mazur's deformationfunctor,Comp. Math. 87
(1993), 269-286.
[Rayl] M. RAYNAUD, Schemas en groupes de type (p,p, ... ,p), Bull. Soc. Math. France
102 (1974), 241-280.
[Ray2] , Specialisationdu foncteurde Picard, Publ. Math. IHES 38 (1970), 27-76.
[Ril] K. A. RIBET, On modularrepresentations ofGal(Q/Q) arisingfrommodularforms,
Invent.Math. 100 (1990), 431-476.
[Ri2] ' Congruencerelationsbetweenmodular forms,Proc. Int. Cong. of Math.
17 (1983), 503-514.
[Ri3] , Report on mod 1 representations of Gal(Q/Q), Proc. of Symp. in Pure
Math. 55 (1994), 639-676.
[Ri4] , Multiplicitiesof p-finitemodp Galois representations in Jo(Np), Boletin
da Sociedade Brasileira de Matematica, Nova Serie 21 (1991), 177-188.
[Rul] K. RUBIN, Tate-Shafarevichgroupsand L-functionsof ellipticcurveswithcomplex
multiplication,Invent.Math. 89 (1987), 527-560.
[Ru2] , The 'main conjectures'of Iwasawa theoryforimaginaryquadratic fields,
Invent.Math. 103 (1991), 25-68.
[Ru3] , Elliptic curves with complex multiplicationand the conjectureof Birch
and Swinnerton-Dyer, Invent.Math. 64 (1981), 455-470.
[Ru4] , More 'main conjectures'forimaginaryquadratic fields,CRM Proceedings
and Lecture Notes, 4, 1994.
[Sch] M. SCHLESSINGER, Functorson Artinrings,Trans. A.M.S. 130 (1968), 208-222.
[Scho] R. SCHOOF, The structureof the minus class groups of abelian number fields,
in Seminairede Theorie des Nombres,Paris (1988-1989), Progress in Math. 91,
Birkhauser(1990), 185-204.
[Se] J-P. SERRE, Sur les representationsmodulaires de degre 2 de Gal(Q/Q), Duke
Math. J. 54 (1987), 179-230.
[de Sh] E. DE SHALIT, Iwasawa Theory of Elliptic Curves with Complex Multiplication,
Persp. in Math., Vol. 3, Academic Press, 1987.
[Shl] G. SHIMURA, Introductionto the ArithmeticTheory of AutomorphicFunctions,
Iwanami Shoten and PrincetonUniversityPress, 1971.
[Sh2] , On the holomorphyof certainDirichletseries,Proc. London Math. Soc.
(3) 31 (1975), 79-98.
[Sh3] , The special values ofthe zeta functionassociated withcusp forms,Comm.
Pure and Appl. Math. 29 (1976), 783-804.

[Sh4] , On ellipticcurveswithcomplex multiplicationas factorsof the Jacobians

of modular functionfields,Nagoya Math. J. 43 (1971), 199-208.
[Ta] J. TATE, p-divisiblegroups,Proc. Conf.on Local Fields, Driebergen,1966, Springer-
Verlag, 1967, pp. 158-183.
[Til] J.TILOUINE, Un sous-groupep-divisiblede la jacobienne de X1 (Npr) commemodule
sur l'algebre de Hecke, Bull. Math. Soc. France 115 (1987), 329-360.
[Ti2] , Theorie d'Iwasawa classique et de l'algebre de Hecke ordinaire,Comp.
Math. 65 (1988), 265-320.
[TU] J. TUNNELL, Artin'sconjectureforrepresentationsof octahedral type,Bull. A.M.S.
5 (1981), 173-175.
[TW] R. TAYLOR and A. WILES, Ring theoreticpropertiesof certainHecke algebras,next
paper, this issue.
[We] A. WEIL, Uber die BestimmungDirichletscherReihendurchFunktionalgleichungen,
Math. Ann. 168 (1967), 149-156.
[Wil] A. WILES, On ordinaryA-adicrepresentations associated to modularforms,Invent.
Math. 94 (1988), 529-573.
[Wi2] , On p-adic representations
fortotallyreal fields,Ann. ofMath. 123 (1986),
407-456.
[Wi3] , Modular curves and the class group of Q((p), Invent. Math. 58 (1980),
1-35.
[Wi4] , The Iwasawa conjecturefortotallyreal fields,Ann. of Math. 131 (1990),
493-540.
[Win] J. P. WINTENBERGER, Structuregaloisiennede limitesprojectivesd'unitees locales,
Comp. Math. 42 (1981), 89-103.
(Received October 14, 1994)


Wiles 1995

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Wiles 1995

Uploaded by

Copyright:

Available Formats

Annals of Mathematics

Modular Elliptic Curves and Fermat's Last Theorem

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

Modular elliptic curves

For Nada, Clare, Kate and Olivia

Cubumautemin duos cubos,aut quadratoquadratum in duos quadra-

An ellipticcurveoverQ is said to be modularifit has a finitecoveringby

*The work on this paper was supported by an NSF grant.

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

Our approachto the studyof ellipticcurvesis via theirassociated Galois

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

Let f be an eigenformassociated to the congruencesubgroupr1(N) of

THEOREM 0.1. For each primep E Z and each primeA I p of Of there

We willbe concernedwithtryingto proveresultsin the oppositedirection,

is a continuousrepresentation with values in the algebraicclosureof a finite

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

If 0 is the ringofintegersof a local field(containingQp) we will say that

(I) po is ordinary(at p) by whichwe mean that thereis a one-dimensional

(II) po is flat (at p), meaningthat as a representationof a decomposition

We say similarlythat p is ordinary(at p) if,viewedas a representationto QP,

CONJECTURE. Suppose thatp: Gal(Q/Q) GL2(0) is an irreducible

liftingof po and thatp is unramified outsideof a finiteset of primes. There

(i) Assume thatpo is ordinary.Then ifp is ordinaryand detp = ek-1X for

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

In case (ii) it is not hard to see that if the formexists it has to be of

THEOREM 0.2. Suppose thatPO is irreducibleand satisfieseither(I) or

(ii) If q - modp is ramifiedin PO then eitherPOIDq is reducibleover

p as in the conjecturedoes indeedcomefroma mod-

The only conditionwhichreallyseems essentialto our methodis the re-

THEOREM 0.3. Suppose thatE is an ellipticcurve definedover Q and

(iii) For anyq -1 mod3 eitherPOIDqis reducibleoverthealgebraicclosure

We should pointout that whilethe propertiesof the zeta functionfollow

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

requiresalso the isogenytheoremprovedby Faltings(and earlierby Serrewhen

THEOREM0.4. Suppose thatE is a semistableellipticcurvedefinedover

More generalfamiliesof ellipticcurveswhichare modularare givenin Chap-

THEOREM 0.5. Suppose thatuP+vP+wP = O withu,v,wE Q andp > 3,

The second result we prove about the conjecturedoes not requirethe

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

THEOREM 0.6. Suppose thatpo is irreducible

(i) po = IndQ Ko for a characterKo of an imaginaryquadraticextensionL

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

a new techniquein orderto deal withthe trivialzeroes. It involvedreplacing

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

littleworkon it. I checkedthatone could makethe appropriateadjustmentsto

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

Galois cohomologycalculationslooked morepromising.Here I was also using

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

commonp5 whichis given in Chapter 5. Believingnow that the proofwas

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

invariantcould be used to countGalois representations. The last step afterthe

Chapter 1 1. Deformationsof Galois representations

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

This chapteris devoted to the study of certain Galois representations.

1. Deformations of Galois representations

Let p be an odd prime. Let E be a finiteset of primesincludingp and

(1.1) Po: Gal(Q/Q) -* GL2(k)

is an irreduciblerepresentation. In contrastto the introductionwe willassume

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

We will restrictour choiceof po furtherby assumingthat either:

(ii) po is flat at p but not ordinary(cf. [Sel] wherethe terminology finiteis

with 52 unramified, X2 X2 mod m, and det PIip = Fw 1X1X2 where

This content downloaded from 130.127.238.233 on Wed, 25 Sep 2013 07:52:30 AM

(i) (b) Ordinarydeformations.The same as in (i) (a) but withno conditionon