by
Peter Bruin
Abstrat. Weprovethat oeients of
q
-expansionsof modularforms anbeomputed in polynomial timeunder ertain assumptions, the most important of whih is the Riemannhypothesisfor
ζ
-funtionsofnumberelds. WegiveappliationstoomputingHekeoperators, ountingpointsonmodularurvesoverniteelds,and omputingthenumberofrepresenta-tionsofanintegerasasumofagivennumberofsquares.
Résumé (Sur lealul desoeients desformes modulaires). Ondémontreque
les oeients des
q
-développements des formes modulaires peuvent être alulés en temps polynomialsousertainesonditions,dontlaplusimportanteestl'hypothèsedeRiemannpourlesfontions
ζ
desorpsdenombres.Ondonnedesappliationsauxproblèmessuivants:aluler desopérateursdeHeke;ompterlenombredepointsd'uneourbemodulairesurunorpsni;alulerlenombredereprésentationsd'unentierommesommed'unnombredonnédearrés.
1. Introdution
Let
n
andk
be positive integers, and letM k (Γ 1 (n))
be theomplex vetor spaeof modularformsof weight
k
for thegroupΓ 1 (n)
. Amodularformf ∈ M k (Γ 1 (n))
isdeterminedbyn
,k
and its
q
-expansion oeientsa m (f)
for0 ≤ m ≤ k · d(Γ 1 (n))
,whered(Γ 1 (n))
isa funtiongrowing roughlyquadratially in
n
.Anaturalquestiontoaskiswhether,given
a m (f )
for0 ≤ m ≤ k · d(Γ 1 (n))
,oneaneientlyompute
a m (f )
for largem
. In the asen = 1
, Couveignes, Edixhoven et al. [2℄ desribed a deterministi algorithm that aomplishes this in time polynomial inlog m
for xedk
.Under the generalised Riemann hypothesis, their algorithm runs in time polynomial in
k
and
log m
. Earlieralgorithms, basedonmodularsymbols,requiretimepolynomialinm
. The2000 MathematisSubjetClassiation. 11E25,11F11,11F30,11F80,11Y16.
Key wordsand phrases. Algorithms,Hekealgebras,modularforms,sumsofsquares.
The results of this artile are based on those of my thesis [1 ℄. I am muh indebted to my advisors Bas
Edixhovenand RobindeJong for theirsupport. I wouldalso like tothanktheorganisers ofthe onferene
Théorie des nombres et appliations for theopportunity to speak about thissubjet, whihhas ledto this
artile.
TheresearhforthisartilewassupportedbytheNetherlandsOrganisationforSientiResearh.
method of [2℄ is, very briey, to ompute two-dimensional Galois representations assoiated
toeigenforms of level
1
overnite elds.Inthisartile,resultsfromtheauthor'sthesis[1℄onomputingGaloisrepresentationsassoi-
atedto eigenforms ofhigherlevels areusedto generalisetheresultofCouveignes,Edixhoven
et al., be it that we an urrently only give a probabilisti algorithm. The preise result
from[1℄ thatweneed isTheorem3.1below. We willusethisto proveour mainresult,whih
readsasfollows.
Theorem 1.1. Let
n 0 be a positive integer. There exists a probabilisti algorithm that, given
a positive integer
k
,a squarefree positive integer
n 1 oprime ton 0,
a number eld
K
,a modular form
f
of weightk
forΓ 1 (n)
overK
,wheren = n 0 n 1, and
a positive integer
m
in fatored form,omputes
a m (f )
, and whose expeted running time is bounded by a polynomial in the lengthof the inputunder the Riemann hypothesis for
ζ
-funtionsof number elds.Letusmakepreisehowthenumbereld
K
andthe formf
shouldbegiventothealgorithmand how it returns
a m (f)
. We representK
by its multipliation table with respet to someQ
-basis(b 1 , . . . , b r )
ofK
. By this we mean the rational numbersc i,j,k with 1 ≤ i, j, k ≤ r
suhthat
b i b j = X r k=1
c i,j,k b k .
We represent elements of
K
asQ
-linear ombinations of(b 1 , . . . , b r )
. We representf
by itsoeients
a 0 (f )
,...,a k · d(Γ 1 (n)) (f )
;these,aswellasthe outputa m (f )
,areelementsofK
.Weshouldalsomakepreisewhattheword`probabilisti'inTheorem1.1means. Theorret
interpretationisthattheresultisguaranteedtobeorret,butthattherunningtimedepends
on random hoies made during exeution. Probabilisti algorithms with this property are
ommonlyalled Las Vegas algorithms. These areto be ontrasted with Monte Carlo algo-
rithms,where therandomnessinuenesthe orretnessoftheoutputinsteadoftherunning
time. It is worth emphasising that the expeted running time is dened by averaging only
over the randomhoies made duringexeution, not over thepossible inputs. For any input
x
, the atual running time of the algorithm given this input an be modelled as a randomvariable
T x. The laim that the expeted running time is polynomial in the length of the
inputmeansthatthereexistsapolynomial
P
suhthatforanyinputx
,theexpetationofT x
isat most
P (
lengthofx)
. We refer to Lenstra andPomerane [10,§
12℄ for an enlightening disussionofprobabilisti algorithms.Remark 1.2. The length of theinput depends not only on
k
,n 0, n 1,log m
and K
,but
log m
andK
,butalsoon theomplexityofthe given oeientsof the modularform
f
. For example,iff
isaprimitiveform
f 0multipliedbyanintegerA
,thenforxedf 0andA
tendingto∞
,thelength
A
tendingto∞
,thelengthof the input inreases approximately by a multiple of
log A
,and the running time inreasesapproximately byapolynomialin
log A
.Remark 1.3. Without the generalised Riemann hypothesis, we are only able to prove
thattherunningtimeofouralgorithm ispolynomialin
exp(n 1 )
,exp(k)
andthelengthoftheinput. Inother words, we arestill ableto prove unonditionally that ifnot only
n 0,but also
n 1 and k
are xed,thentheexpetedrunning timeispolynomialinthelength of theinput.
k
are xed,thentheexpetedrunning timeispolynomialinthelength of theinput.Remark 1.4. Omitting the onditionthat
m
be given in fatored form would be equiv-alent to laiming that integers that are produts of two prime numbers an be fatored in
polynomialtime. Namely,suppose thatthe theoremholds without this ondition. Applying
thehypothetial strongerversionof the theoremwith
k
axedeven integergreaterthan2
,
n 0 = n 1 = 1
,
K = Q
,
f = E k,the lassial Eisensteinseries E k ofweight k
for Γ 1 (1) = SL 2 (Z)
,and
k
forΓ 1 (1) = SL 2 (Z)
,and
m = pq
,wherep
andq
aretwo distint prime numbers,we onlude that there existsa probabilisti algorithm that omputes
a m (E k )
intime poly-nomialin
log m
. From the formulaa m (E k ) = X
d | m
d k−1
= 1 + p k − 1 + q k − 1 + m k − 1 ,
it follows that
{ p k − 1 , q k − 1 }
an be omputed quikly as the set of roots of the polynomialx 2 − (a m (E k ) − m k−1 − 1)x + m k−1 ∈ Z[x]
. Henewe wouldbeabletoompute{ p, q }
fromm
intimepolynomialin
log m
,whih isa laimwe ertainlydo not wishto make.Remark 1.5. The reason why our algorithm is probabilisti is that this is the urrent
state of aairs for the algorithm to whih Theorem 3.1 refers. This algorithm an perhaps
be turned into a deterministi one by replaing the arithmeti over nite elds that is used
in[1℄byapproximatearithmeti overtheomplex numbers. Thelatterapproah istaken by
Couveignes,Edixhovenetal.[2,Chapter12℄formodularformsoflevel1. Thereareurrently
still some diulties with this approah for modular forms of higher level. We refer to [1,
Introdution℄for adisussion ofthese.
Remark 1.6. It wouldbemore satisfatoryifwe ouldprove thetheoremwiththelevel
ranging over all positive integers
n
. We urrently annot do this for the following reason.The modular urve
X 1 (n)
has a regular and semi-stable model over thering of integersZ L
of asuitable numbereld
L
,but ingeneral we do not knowa good bound on thenumberofirreduibleomponentsofthegeometribresofsuhamodelatprimesof
Z Lthatdividen
.
Ifweouldprovethetheoreminthismoregeneralform, thentherestritiontomodularforms
forongruenesubgroupsoftheform
Γ 1 (n)
ouldalsoberemoved. Thereasonforthisisthatthespaeofmodular formsofweight
k
fortheprinipalongruene subgroupΓ(n) ⊆ SL 2 (Z)
anbe embedded into
M k (Γ 1 (n 2 ))
byamapthat onq
-expansions isgivenbyq 7→ q n.
Wenowturnto someappliationsof Theorem1.1. Wewill prove thatthereexistprobabilis-
ti algorithms that solve the following problems in expeted polynomial time in the input,
assumingtheRiemann hypothesisfor
ζ
-funtionsof numberelds:Given a positive integer
k
, a squarefree positive integern
and a positive integerm
infatored form, ompute the matrix of the Heke operator
T m in T(M k (Γ 1 (n)))
with
respetto a xed
Z
-basisofT(M k (Γ 1 (n)))
.Given a squarefree positive integer
n
and a prime numberp ∤ n
, ompute the zetafuntion of the modularurve
X 1 (n)
overF p.
Givenanevenpositive integer
k
andapositiveintegerm
infatored form,ompute thenumber ofways inwhih
m
an be written asa sumofk
squaresof integers.Atually, we do not prove our results in exatly the same order as presented above. We
rst prove Theorem 1.1 in the speial ase where
f
is an Eisenstein series or a primitiveusp form. This sues to solve (a slightly more general version of) the above problem of
omputingHekeoperators. Wethenprove Theorem1.1ingeneral. Finally,weshowhowto
solvetheproblems ofomputing zetafuntionsofmodularurves andndingthenumberof
representationsof aninteger asa sumof squares.
To onlude this introdution, we remark that in order to keep this artile at a reasonable
length, we have omitted, or only briey touhed upon, muh material that an be found in
[1℄and [2℄. Thismeans thatthe ontents of this artile arelargely disjoint from thoseof [1℄
and[2℄.
2. Bakground
We begin by olletingthe neessary preliminaries and introduing our notation. For deni-
tionsand more bakground, we referto the manytextson modular forms, suh asDiamond
andIm [5℄ or Diamondand Shurman [6℄.
2.1. Modular forms. Let
n
andk
be positive integers. LetM k (Γ 1 (n))
denote theC
-vetorspae ofmodular forms ofweightk
for the groupΓ 1 (n) = a
c b d
∈ SL 2 (Z)
a ≡ d ≡ 1 (mod n), c ≡ 0 (mod n)
.
For every
f ∈ M k (Γ 1 (n))
and everym ≥ 0
, we writea m (f)
for the oeient ofq m in the
q
-expansion of f
, so the q
-expansion of f
is the power series P ∞
m=0 a m (f)q m
inK [[q]]
. Foreverydivisor
d
ofn
and every divisore
ofn/d
,thereexistsan injetiveC
-linear mapb d,n e : M k (Γ 1 (d)) M k (Γ 1 (n))
that,on
q
-expansions, hasthe eet ofsendingq
toq e.
We dene
(1)
d(Γ 1 (n)) = 1
12 [SL 2 (Z) : {± 1 } Γ 1 (n)].
This
d(Γ 1 (n))
grows roughlyquadratially inn
. A basi fat that we will need oftenis thefollowing.
Lemma 2.1. Any
f ∈ M k (Γ 1 (n))
is determined byn
,k
and the oeientsa m (f)
for0 ≤ m ≤ k · d(Γ 1 (n))
.Proof. If
n ≥ 5
, thend(Γ 1 (n))
is the degree of the line bundleω
of modular forms ofweight
1
on the modular urveX 1 (n)
. In that ase, we an view modular forms as globalsetions of
ω ⊗k. If f, g ∈ M k (Γ 1 (n))
aresuh that a m (f ) = a m (g)
for 0 ≤ m ≤ k · d(Γ 1 (n))
,
then
f − g
hasazerooforderat leastk · d(Γ 1 (n)) + 1
attheusp∞
ofX 1 (n)
,andweonludethat
f = g
. Onean prove the lemmaingeneral byreduing to theasen ≥ 5
. We refer toSturm [14℄for afull proof.
2.2. Heke algebras. Let
T(M k (Γ 1 (n)))
be the Heke algebra onM k (Γ 1 (n))
. This isa ommutative ring, free of nite rank as a
Z
-module and generated as aZ
-algebra by theHeke operators
T m for m ∈ { 1, 2, . . . }
and thediamond operators h d i
for d ∈ (Z/nZ) ×. It
atson the
C
-vetor spaeM k (Γ 1 (n))
of modularforms.Letus give someuseful formulae. We have
(2)
T m 1 m 2 = T m 1 T m 2 if gcd(m 1 , m 2 ) = 1
and
(3)
T p i+2 = T p T p i+1 − p k − 1 h p i T p i (p
primeandi ≥ 0),
where
h p i
is to be interpretedas0
ifp
dividesn
. For allf ∈ M k (Γ 1 (n))
,we have(4)
a m (T p (f )) = a pm (f ) + p k − 1 a m/p ( h p i f ) (p
primeandm ≥ 1),
where theseond termis
0
ifp
dividesn
or ifp
doesnot dividem
,and(5)
a 1 (T m f ) = a m (f ) (m ≥ 1).
There existsaanonial bilinear map
T(M k (Γ 1 (n))) × M k (Γ 1 (n)) −→ C (t, f ) 7−→ a 1 (tf ),
induing an isomorphism
(6)
M k (Γ 1 (n)) −→ ∼ Hom Z-modules (T(M k (Γ 1 (n))), C)
of
C ⊗ Z T(M k (Γ 1 (n)))
-modules.An eigenform of weight
k
forΓ 1 (n)
is an element ofM k (Γ 1 (n))
spanning a one-dimensional eigenspae for the ation ofT(M k (Γ 1 (n)))
. Letf
be suh a form. Thena 1 (f ) 6 = 0
, and wemaysale
f
suhthata 1 (f ) = 1
. Now (5)impliesthatT m f = a m (f )f
forallm ≥ 1.
Furthermore,there existsaunique grouphomomorphism
ǫ: (Z/nZ) × → C × ,
alledtheharater of
f
,suhthath d i f = ǫ(d)f
for alld ∈ (Z/nZ) × .
Underthe isomorphism (6), the eigenforms
f ∈ M k (Γ 1 (n))
witha 1 (f) = 1
orrespond to thering homomorphisms
T(M k (Γ 1 (n))) → C
.The
C
-vetor spaeM k (Γ 1 (n))
an be written asadiret sumM k (Γ 1 (n)) = E k (Γ 1 (n)) ⊕ S k (Γ 1 (n)).
Here
S k (Γ 1 (n))
denotes the subspae of usp forms andE k (Γ 1 (n))
denotes the subspaeof Eisenstein series. The ation of
T(M k (Γ 1 (n)))
respets these subspaes, and we get aorrespondingdeomposition
T(M k (Γ 1 (n))) = T(E k (Γ 1 (n))) × T(S k (Γ 1 (n)))
of
Z
-algebras.2.3. Eisenstein series. Let
d 1 andd 2 bepositiveintegerssuhthatd 1 d 2 dividesn
,and
d 1 d 2 dividesn
,and
onsiderprimitive haraters
ǫ 1 : (Z/d 1 Z) × → C × , ǫ 2 : (Z/d 2 Z) × → C × .
(Aharater
ǫ : (Z/dZ) × → C ×, with d
a positive integer, is alled primitive ifthere is no
stritdivisor
e | d
suh thatǫ
fatorsthroughthe quotient(Z/dZ) × → (Z/eZ) ×.) We dene
theformal powerseries
(7)
E k ǫ 1 ,ǫ 2 (q) = − δ d 1 ,1 B k ǫ 2 2k +
X ∞ m=1
X
d|m
ǫ 1 (m/d)ǫ 2 (d)d k − 1
!
q m ∈ C[[q]].
Here
B k ǫ 2 is a generalised Bernoullinumber and δ d 1 ,1 is 1
or 0
depending on whether d 1 = 1
1
or0
depending on whetherd 1 = 1
or
d 1 > 1
.If
k 6 = 2
, or ifk = 2
and at least one ofǫ 1 and ǫ 2 is non-trivial, then E k ǫ 1 ,ǫ 2 (q)
is the q
-
E k ǫ 1 ,ǫ 2 (q)
is theq
-expansionofan eigenformin
E k (Γ 1 (d 1 d 2 ))
withharaterǫ 1 ǫ 2. Foranydivisore
ofn/(d 1 d 2 )
,
themap
b d e 1 d 2 ,n : M k (Γ 1 (d 1 d 2 )) → M k (Γ 1 (n))
sendsthisformtoanelementofM k (Γ 1 (n))
withq
-expansionE k ǫ 1 ,ǫ 2 (q e )
. Asfor thease wherek = 2
and bothǫ 1 and ǫ 2 are trivial, for every
divisor
e | n
withe > 1
there is an element ofE 2 (Γ 1 (n))
withq
-expansionE 2 (q) − eE 2 (q e )
,where
E 2 (q)
isthepowerseries(8)
E 2 (q) = − 1
24 + X ∞ m=1
X
d | m
d
! q m .
For
k 6 = 2
,theniteset(9)
F k (Γ 1 (n)) = G
d 1 d 2 |n
G
e|(n/d 1 d 2 )
E k ǫ 1 ,ǫ 2 (q e ) | ǫ i : (Z/d i Z) × → C × primitive
isa
C
-basisofE k (Γ 1 (n))
. Fork = 2
,wetakeallE k ǫ 1 ǫ 2 (q e )
forǫ 1,ǫ 2 not both trivial,together
withthe
E 2 (q) − eE 2 (q e )
for alle | n
withe > 1
.2.4. Cusp forms. We write
S new k (Γ 1 (n))
for the orthogonal omplement, with respetto the Petersson inner produt, of the subspae of
S k (Γ 1 (n))
spanned by the images ofall the
b d,n e with d
stritly dividing n
. The spae S new k (Γ 1 (n))
is preserved by the ation
of
T(S k (Γ 1 (n)))
. Theunique quotient ofT(S k (Γ 1 (n)))
thatats faithfullyonS new k (Γ 1 (n))
isdenotedby
T(S new k (Γ 1 (n)))
.An eigenform
f ∈ S new k (Γ 1 (n))
witha 1 (f ) = 1
isalleda primitive uspform. The niteset(10)
B k (Γ 1 (n)) = G
d|n
G
e|(n/d)
b d,n e
primitive uspforms in
S new k (Γ 1 (n))
isa
C
-basisforS k (Γ 1 (n))
.2.5. Modular forms over other rings. We dene
M int k (Γ 1 (n)) = {
forms inM k (Γ 1 (n))
withq
-expansion inZ[[q]] } .
Thisisa
T(M k (Γ 1 (n)))
-modulethatisfreeofniterankasaZ
-module. ForanyommutativeZ[1/n]
-algebraR
, we dene theR
-module of modular forms of weightk
forΓ 1 (n)
withoeientsin
R
asM k (Γ 1 (n), R) = R ⊗ Z M int k (Γ 1 (n)).
Apartfromthe omplexnumbers, the important examplesforus arenumberelds andnite
elds of harateristi not dividing
n
. IfR
is any eld of harateristi not dividingn
, wedene eigenforms over
R
inthesame wayasin theaseR = C
.If
R
isasub-Z[1/n]
-algebraofC
,weidentifyM k (Γ 1 (n), R)
withthesubmoduleofM k (Γ 1 (n))
onsistingof forms with
q
-expansion inR[[q]]
.3. Modular Galois representations
Let
n
andk
be positive integers, letF
be a nite eldof harateristi not dividingn
, andlet
f ∈ M k (Γ 1 (n), F)
bean eigenformoverF
.Itfollows fromwork ofEihler,Shimura, Igusa,Deligne andSerre thatthere existsa ontin-
uoussemi-simple representation
ρ f : Gal(Q/Q) → Aut F V f ,
where
V f is atwo-dimensional F
-vetor spae,with thefollowing properties:
ρ f isunramied at allprimenumbers p
not dividingnl
;
if
p
issuh a prime number, thentheharateristi polynomialof theFrobenius onju- gay lassatp
equalst 2 − a p (f )t + ǫ(p)p k−1,where ǫ
istheharater of f
.
This
ρ f isunique upto isomorphism.
The end produt of [1℄ is a probabilisti algorithm for omputing representations of the
form
ρ f, where f
is an eigenform over a nite eld F
. This allows us to state thefollowing
theorem.
Theorem 3.1. Let
n 0 be a positive integer. There exists a probabilisti algorithm that, given
a positive integer
k
,a squarefree positive integer
n 1 oprime ton 0,
a niteeld
F
of harateristi greater thank
, andan eigenform
f ∈ M k (Γ 1 (n))
, given by its oeientsa m (f )
for0 ≤ m ≤ k · d(Γ 1 (n))
,omputes
ρ f in the form of the followingdata:
the nite Galois extension
K f of Q
suh that ρ f fators as
Gal(Q/Q) ։ Gal(K f /Q) Aut F V f ,
Gal(Q/Q) ։ Gal(K f /Q) Aut F V f ,
given by the multipliationtable of some
Q
-basis(b 1 , . . . , b r )
ofK f;
for every
σ ∈ Gal(K f /Q)
, the matrix ofσ
with respet to the basis(b 1 , . . . , b r )
andthematrix of
ρ f (σ)
withrespet tosome xedF
-basis ofV f,
andthat runs in expeted time polynomial in
k
,n 1 and #F
.
Moreover, one
ρ f has been omputed, one an ompute ρ f (Frob p )
using a deterministi al-
gorithmin time polynomialin k
, n 1, #F
and log p
.
#F
andlog p
.Remark 3.2. Thisrunningtimeisoptimalfromaertainperspetive,giventhefatthat
thelengthof theinput andoutput ofsuh an algorithm isneessarily at leastpolynomial in
k
,n 1 and#F
(and log p
for theseondpart).
4. Some bounds
Inthis setionwe olletsome boundsthatwe will need in
§
5belowto prove Theorem 1.1.4.1. The disriminant of the new quotient of the Heke algebra. Let
n
andk
be positive integers. The
Z
-algebraT(S new k (Γ 1 (n)))
is redued, beause there is a basis ofeigenformsfor itsation on
S new k (Γ 1 (n))
. Furthermore, itis freeof niterankasaZ
-module.Inpartiular,it hasanon-zero disriminant
disc T(S new k (Γ 1 (n)))
.Lemma 4.1. The logarithm of
| disc T(S new k (Γ 1 (n))) |
is bounded by a polynomial inn
and
k
.Proof. ThemethodofUllmo[15℄,whoonsidereduspformsofweight 2for
Γ 0 (n)
withn
squarefree,extends without diulty to our situation. For ompleteness, let us give a proof
inthis more general setting.
Weabbreviate
T = T(S new k (Γ 1 (n)))
and
r = dim C S new k (Γ 1 (n))
= rank Z T.
ItfollowsfromLemma2.1and(6 )thatthe
Q
-vetorspaeQ ⊗ Z T
isspannedbytheelementsT 1,...,T k · d(Γ 1 (n)),with d(Γ 1 (n))
asin(1). Wean thereforehooseintegers
d(Γ 1 (n))
asin(1). Wean thereforehooseintegers1 ≤ m 1 ≤ · · · ≤ m r ≤ k · d(Γ 1 (n))
suh thattheelements
T m 1,...,T m r of T
areZ
-linearly independent. Welet T ′ denotethe
T
areZ
-linearly independent. WeletT ′ denotethe
subgroup of
T
spannedbyT m 1,...,T m r. ThisT ′ is free of rankr
asa Z
-module, so ithas
T ′ is free of rankr
asa Z
-module, so ithas
nite index
(T : T ′ )
inT
,anddisc T = disc T ′ (T ′ : T) 2 .
Inpartiular, this implies
| disc T | ≤ disc T ′ .
We next usethe denitionof thedisriminant:
disc T ′ = det tr(T m u T m v ) r u,v=1 ,
where
tr(e)
denotes the trae of theZ
-linear mapT ′ → T ′ sending t
to et
. Now the trae
ofan endomorphism
e
ofT ′ equalsthe trae oftheendomorphism dualtoe
on theC
-vetor
spae
Hom Z-modules (T ′ , C) ∼ = S new k (Γ 1 (n)).
We let
f 1,...,f r be the primitive uspforms inS new k (Γ 1 (n))
,andweabbreviate
α t,u = a m t (f u ).
S new k (Γ 1 (n))
,andweabbreviateα t,u = a m t (f u ).
Thenweget
tr(T m u T m v ) = X r t=1
α t,u α t,v .
We thenompute
disc T ′ asfollows:
disc T ′ = det X r
t=1
α t,u α t,v r
u,v=1
= det
α 1,1 α 2,1 . . . α r,1
α 1,2 α 2,2 . . . α r,2
.
.
. .
.
. .
.
. .
.
.
α 1,r α 2,r . . . α r,r
α 1,1 α 1,2 . . . α 1,r
α 2,1 α 2,2 . . . α 2,r
.
.
. .
.
. .
.
. .
.
.
α r,1 α r,2 . . . α r,r
= det (α t,u ) r t,u=1 2
.
Deligne'sboundfortheoeientsofeigenforms,provedin[3℄ and[4℄,impliestheinequality
| α t,u | = | a m t (f u ) |
≤ σ 0 (m t )m (k t − 1)/2 .
Here
σ 0 (m)
denotes the number of positive divisors ofm
. Elementary estimates now showthat
log | disc T |
is bounded byapolynomialinn
andk
.4.2. Primes of small norm in number elds. The Riemann hypothesis for the
ζ
-funtionof anumbereld
K
hasthe following well-knownimpliation offor theexisteneofprimeidealsofsmall normin thering ofintegers of
K
.Lemma 4.2. Let
ǫ
andδ
be positive real numbers. There exist positive real numbersA
andB
suh that the following holds. LetK
be an number eld suh that the Riemannhypothesis is true for the
ζ
-funtion ofK
. LetZ K denote the ring of integers of K
, andfor
everyprimenumber
p
letλ K (p)
denotethe numberof primeidealsofZ K of normequal top
.
Thenfor all real numbers
x ≥ 2
suhthatx δ
(log x) 2 ≥ A[K : Q]
andx δ
log x ≥ B log | disc Z K |
wehave
X
p ≤ x
primeλ K (p) log p − x
≤ ǫx 1/2+δ .
Proof. For every prime number
p
andeverypositiveintegerm
,wedeneΛ K (p m ) = X
t|m
t · # {
prime idealsof normp t inZ K } · log p.
Inpartiular,thisimplies
Λ K (p) = λ K (p) log p
foreveryprimenumberp
. WedeneΛ K (n) = 0
ifn
is not aprimepower. Therelation betweenζ K andΛ K istheDirihlet series
− ζ K ′ (s) ζ K (s) =
X ∞ n=1
Λ K (n)n − s .
Wedene
ψ K : [1, ∞ ) −→ R x 7→ X
n≤x
Λ K (n).
Now there existsa positive real number
c
, independent ofK
, suh thatthe generalisedRie- mannhypothesis forζ K impliestheestimate
| ψ K (x) − x | ≤ c √
x log(x) log x [K:Q] | disc Z K |
for all
x ≥ 2;
seeIwanieand Kowalski [9, Theorem 5.15℄. Byelementary arguments, it follows thatthere
existsapositive realnumber
c ′,alsoindependent ofK
,suh that
X
p ≤ x
primeλ K (p) log p − x ≤ c ′ √
x log(x) log x [K:Q] | disc Z K |
forall
x ≥ 2.
It isnowstraightforward to hekthattaking
A = B = 2c ′ /ǫ
works.5. Proof of Theorem 1.1
Asalready mentioned brieyintheintrodution, Theorem1.1 willbe proved asfollows. We
rstprove thefollowing basi ases:
f
is anelement of theformE k ǫ 1 ,ǫ 2 inE k (Γ 1 (d 1 d 2 ))
,where ǫ 1 : (Z/d 1 Z) × → C × and
ǫ 2 : (Z/d 2 Z) × → C × areprimitiveharaters;
ǫ 2 : (Z/d 2 Z) × → C × areprimitiveharaters;
f
is aprimitive uspforminS k (Γ 1 (n))
.In eah ase, we take
K
to be the number eld generated by the oeients off
, and weassume that
m
is a prime number. After proving these speial ases, we show that we anompute theHeke algebra
T(M k (Γ 1 (n)))
ina sense thatwillbe explainedin§
5.3below. Itisthenstraightforward to dedueTheorem 1.1ingeneral.
5.1. Eisenstein series. We start by onsidering the Eisenstein series
E k ǫ 1 ,ǫ 2, where
ǫ 1 : (Z/d 1 Z) × → C × and ǫ 2 : (Z/d 2 Z) × → C × are primitive haraters and e
is a divisor
ǫ 2 : (Z/d 2 Z) × → C × are primitive haraters and e
is a divisor
of
n/(d 1 d 2 )
. For onveniene, we alsoallow thease of the`pseudo-Eisenstein series'E 2 de-
nedby(8 ). Let
K
betheylotomiextensionofQ
generatedbytheimagesofǫ 1andǫ 2. The
formula(7)showsthatforeveryprimenumber
p
,weanomputetheelementa p (E k ǫ 1 ,ǫ 2 ) ∈ K
intimepolynomialin
n
,k
andlog p
.5.2. Primitive forms. We ontinue with the ase where
f
is a primitive usp formin
S new k (Γ 1 (n))
andK
isthe number eld generated bythe oeientsoff
. LetZ K denote
thering ofintegersof
K
. There existsaunique ringhomomorphisme f : T(S new k (Γ 1 (n))) → Z K
sending eah Heke operator to its eigenvalue on
f
. LetA
denote the image ofe f. It is of
nite index
(Z K : A)
inZ K,and we have
disc A = (Z K : A) 2 disc Z K .
Furthermore, we have
| disc A | ≤ | disc T(S new k (Γ 1 (n))) |
and[K : Q] ≤ rank Z T(S new k (Γ 1 (n))).
Lemma 4.1 now implies that
log | disc A |
, and hene alsolog | disc Z K |
andlog(Z K : A)
, areboundedbya polynomial in
n
andk
. The same learly holdsfor[K : Q]
.Nowlet
p
beaprimenumber. Wehavetoshowthatweanomputea p (f )
intimepolynomialin
n
,k
andlog p
. In Couveignes, Edixhoven et al. [2,§
15.2℄ it is explained indetail how todothis. We only givea sketh.
Wemayassumethat
p
doesnot dividen
;namely,ifp
doesdividen
,thenwe an spendtimepolynomial in
p
,sousingmodular symbolsisfast enough;see§
5.3below.ByLemma 4.2 applied to
K
and thefatthatlog(Z K : A)
isbounded bya polynomial inn
and
k
,wean hoosex
suiently large,butboundedbyapolynomial inn
andk
,suh thatif
M
is the set of maximal ideals ofA
whose norm is a prime number lying in the interval(k, x]
and dierent fromp
,we have(11)
Y
m ∈ M
Norm(m) ≥
2 ([K:Q]+1)/2 · 2p (k − 1)/2 [K:Q]
.
An explanation for the right-hand side will be given below. We ompute
a p (f )
using thefollowing algorithm.
1. Compute a
Z
-basisforA
.2. Compute abound
x
and the setM
ofmaximal idealsofA
suhthat thesetM
denedabove satises(11 ).
3. For all
m ∈ M
,ompute the Galois representationρ f mod m : Gal(Q/Q) → GL 2 (A/m)
using Theorem 3.1 .
4. For all
m ∈ M
,ompute(a p (f ) mod m) = tr(ρ f mod m (Frob p )) ∈ A/m,
again using Theorem3.1.
5. Compute an LLL-redued
Z
-basisfor theideala = Q
m∈M mof A
.
6. Fromthe
a p (f ) mod m
,omputethe image ofa p (f )
inA/a
.7. Using the LLL algorithm, reonstrut
a p (f )
as the shortest representative inA
of theimage of
a p (f )
inA/a
. Thisworksbeause ofthe inequality (11) .5.3. ComputingHeke operators. Werepresent
T(M k (Γ 1 (n)))
inthefollowingform:we speify its multipliation table with respet to a suitable
Z
-basis(b 1 , . . . , b r )
, togetherwith the Heke operators
T m for 1 ≤ m ≤ k · d(Γ 1 (n))
and the diamond operators h d i
for
all
d ∈ (Z/nZ) × as Z
-linear ombinations of (b 1 , . . . , b r )
. These data speify T(M k (Γ 1 (n)))
uniquely beause the above operators generate
T(M k (Γ 1 (n)))
. In other words, if the samedata are given with respet to a dierent basis of
T(M k (Γ 1 (n)))
, there exists exatly onehange of
Z
-basis ompatible withthegivenT m and h d i
.
Theorem 5.1. Let
n 0 be a positive integer. There exists a probabilisti algorithm that, given
a positive integer
k
,a squarefree positive integer
n 1 oprime ton 0, and
a positive integer
m
in fatored form,omputes
the Heke algebra
T(M k (Γ 1 (n)))
as above, wheren = n 0 n 1, and
the element
T m on the basis (b 1 , . . . , b r )
,
and that runs in expeted time polynomial in
k
,n 1 and log m
under the Riemann hypothesis
for
ζ
-funtions of number elds.Proof. We need some more information about the ation of Heke operators on
q
-expansions. Asa basisfor
M k (Γ 1 (n))
we take the union ofthe basisF k (Γ 1 (n))
ofE k (Γ 1 (n))
dened by (9 )and the basis
B k (Γ 1 (n))
ofS k (Γ 1 (n))
dened by(10 ).Let
f
be either an Eisenstein seriesE k ǫ 1 ,ǫ 2 ∈ E k (Γ 1 (d 1 d 2 ))
as above or a primitive formin
S k (Γ 1 (d))
. In the rst ase, we putd = d 1 d 2. Theformula(4) for theation oftheHeke
operator
T p shows thattherelationbetween T p andthe mapsb d,n e : M k (Γ 1 (d)) → M k (Γ 1 (n))
,
b d,n e : M k (Γ 1 (d)) → M k (Γ 1 (n))
,where
e
runsthrough thedivisorsofn/d
,is asfollows:(12)
T p (b d,n e f ) =
a p · b d,n e f
ifp ∤ n;
b d,n e/p f
ifp | e;
a p · b d,n e f − p k − 1 ǫ(p)b d,n pe f
ifp ∤ d
,p ∤ e
andp | n;
a p · b d,n e f
ifp | d
andp ∤ e.
Thisformulagivesthematrix of
T p withrespettothebasisF k (Γ 1 (n))
of E k (Γ 1 (n))
andthe
basis
B k (Γ 1 (n))
ofS k (Γ 1 (n))
.We rst ompute the
q
-expansions of the Eisenstein seriesE k ǫ 1 ,ǫ 2 ∈ E k (Γ 1 (d 1 d 2 ))
, withǫ i : (Z/d i Z) × → C × primitive haraters suh that d 1 d 2 | n
, as in §
2.3 . From these q
-
expansionsand(12 )wethenompute theHekealgebra
T(E k (Γ 1 (n)))
intheform desribedabove intime polynomialin
n
andk
.Given a prime number
p
, we ompute all thea p (E k ǫ 1 ,ǫ 2 )
as in§
5.1 , and we nd the matrixof
T p using (12). WethenexpressT p on thebasisofT(E k (Γ 1 (n)))
thatweomputed earlier.
T(E k (Γ 1 (n)))
thatweomputed earlier.Inthis way,we anompute theHeke operator
T p ∈ T(E k (Γ 1 (n)))
intimepolynomialinn
,k
andlog p
.For uspforms, the
q
-expansions areomputed fromtheHeke algebra insteadofvieversa.We ompute theHekealgebras
T(S k (Γ 1 (d)))
,whered
runsthrough thedivisorsofn
,intheform desribed above. These data an be omputed in time polynomial in
n
andk
usingdeterministialgorithmsbasedonmodularsymbolsandtheLLLlattiebasisredutionalgo-
rithm;seeStein[13,Chapter8℄andtheauthor'sthesis[1,
§
IV.4.1℄. FromeahT(S k (Γ 1 (d)))
,we ompute the
q
-expansionsof the primitiveusp formsinS k (Γ 1 (d))
.Sofar,wehaveonlyusedexistingmethods. ToomputetheHekeoperator
T p ∈ T(S k (Γ 1 (n)))
for a primenumber
p
in timepolynomial inlog p
,we need our new tools. For every divisord
ofn
and every primitive formf ∈ S k (Γ 1 (d))
,weomputea p (f )
as in§
5.2 . Using (12),weobtain the matrix of
T p with respet to the basis B k (Γ 1 (n))
. We nally express T p on the
basisof
T(S k (Γ 1 (n)))
thatweomputed earlier.Nowlet
m
beanarbitrarypositive integer,andsupposethatwe knowthefatorisation ofm
.Thenwe an ompute the element
T m ∈ T(M k (Γ 1 (n))) = T(E k (Γ 1 (n))) × T(S k (Γ 1 (n)))
fromthe
T pforforp | m
primeintimepolynomialinlog m
usingtheidentities(2 )and(3).
5.4. Proof of Theorem 1.1 in general. Given
n
,k
,K
,f
andm
as in thetheorem,weompute
a m (f )
asfollows. We rstomputeT(M k (Γ 1 (n)))
usingmodularsymbols. Fromf
we thendetermine the uniqueZ
-linear mape f : T(M k (Γ 1 (n))) → K
sending
T i to a i (f )
for all i
with1 ≤ i ≤ k · d(Γ 1 (n))
. Using Theorem5.1 , we thenompute
theHeke operator
T m. Finally,we ompute a m (f )
as
a m (f ) = e f (T m ).
Itis straightforward to hekthat allthese omputationsan be done intimepolynomial in
thelength of theinput.
Remark 5.2. The proof shows that the Riemann hypothesis only needs to be assumed
forthe
ζ
-funtionsofnumberelds thatariseaseldsof oeients ofprimitive uspforms.6. Appliations
6.1. Counting points on modular urves. Thease
k = 2
ofTheorem 5.1impliesanewresulton ountingpointson modularurvesovernite elds.
Theorem 6.1. There exists a probabilisti algorithm that, given a squarefree positive in-
teger
n
and a prime numberp ∤ n
, omputes the zeta funtion of the modular urveX 1 (n)
over
F p, andthat runs in time polynomial in n
and log p
under the Riemann hypothesisfor
ζ
-funtionsof number elds.
Proof. Let
J 1 (n) F pdenotetheJaobianofX 1 (n) F p. Letχ
betheharateristipolynomial
of the Frobenius endomorphism of the l
-adi Tate module T l J 1 (n) F p, where l
is any prime
χ
betheharateristipolynomial of the Frobenius endomorphism of thel
-adi Tate moduleT l J 1 (n) F p, where l
is any prime
number dierent from
p
; thenχ
hasintegral oeients and doesnot depend on the hoieof
l
. Beauseof thewell-knownidentityZ X 1 (n)/F p (t) = χ ∗ (t) (1 − t)(1 − pt) ,
where
χ ∗ (t) = t deg χ χ(1/t)
isthe reiproal polynomial ofχ
,itsuesto omputeχ
.Let