Analyti c Numbe r Theor y

This page intentionally left blank

 http://dx.doi.org/10.1090/coll/053

America n Mathematica l Societ y

Colloquiu m Publication s
Volum e 53

Analyti c Numbe r Theor y

Henry k Iwanie c
Emmanue l Kowalsk i

America n Mathematica l Societ y

Providence , Rhod e Islan d
 Editorial Boar d
Susan J . Priedlander , Chai r
Yuri Mani n
Peter Sarna k

2000 Mathematics Subject Classification. P r i m a r y H F x x , H L x x , H M x x , H N x x , 1 1 T23 ,

11T24, 1 1 R42 .

For a d d i t i o n a l informatio n a n d u p d a t e s o n t h i s book , visi t

www.ams.org/bookpages/coll-53

Library o f Congres s Cataloging-in-Publicatio n D a t a

Iwaniec, Henryk .
Analytic numbe r theor y / Henry k Iwaniec , Emmanue l Kowalski .
p. cm . — (Colloquiu m publications , ISS N 0065-925 8 ; v. 53 )
Includes bibliographica l reference s an d index .
ISBN 0-821 8-3633- 1 (acid-fre e paper )
1. Numbe r theory . I . Kowalski , Emmanuel , 1 969 - II . Title . III . Colloquiu m publication s
(American Mathematica l Society ) ; v. 53 .
QA241.I85 200 4
512.7'"3-dc22 2004045081

C o p y i n g an d reprinting . Individua l reader s o f thi s publication , an d nonprofi t librarie s

acting fo r them , ar e permitte d t o mak e fai r us e o f th e material , suc h a s t o cop y a chapte r fo r us e
in teachin g o r research . Permissio n i s grante d t o quot e brie f passage s fro m thi s publicatio n i n
reviews, provide d th e customar y acknowledgmen t o f th e sourc e i s given .
Republication, systemati c copying , o r multipl e reproductio n o f an y materia l i n thi s publicatio n
is permitte d onl y unde r licens e fro m th e America n Mathematica l Society . Request s fo r suc h
permission shoul d b e addresse d t o th e Acquisition s Department , America n Mathematica l Society ,
201 Charle s Street , Providence , Rhod e Islan d 02904-2294 , USA . Request s ca n als o b e mad e b y
e-mail t o reprint-permissionQams.org .
© 200 4 b y th e America n Mathematica l Society . Al l right s reserved .
The America n Mathematica l Societ y retain s al l right s
except thos e grante d t o th e Unite d State s Government .
Printed i n th e Unite d State s o f America .
@ Th e pape r use d i n thi s boo k i s acid-fre e an d fall s withi n th e guideline s
established t o ensur e permanenc e an d durability .
Visit th e AM S hom e pag e a t h t t p : //www. ams. o r g/
10 9 8 7 6 5 4 1
3 2 51 41 31 21 1 1 0
 Contents

Preface x i
Introduction
Chapter 1 . Arithmeti c Function s 9
§1.1. Notatio n an d definition s 9
§1.2. Generatin g serie s 0
§1.3. Dirichle t convolutio n 2
§1.4. Example s 3
§1.5.
1 Arithmeti c function s o n averag e 9
§1.6. Sum s o f multiplicativ e function s 2 3
§1.7. Distributio n o f additiv e function s 2 8
Chapter 2 . Elementar y Theor y o f Prim e Number s 3 1
§2.1. Th e Prim e Numbe r Theore m 3 1
§2.2. Tchebyshe v metho d 3 2
§2.3. Prime s i n arithmeti c progression s 3 4
§2.4. Reflection s o n elementar y proof s o f th e Prim e Numbe r Theore m 3 8

Chapter 3 . Character s 4 3
§3.1. Introductio n 4 3
§3.2. Dirichle t character s 4 4
§3.3. Primitiv e character s 4 5
§3.4. Gaus s sum s 4 7
§3.5. Rea l character s 4 9
§3.6. Th e quarti c residu e symbo l 5 3
§3.7. Th e Jacobi-Dirichle t an d th e Jacobi-Kubot a symbol s 5 5
§3.8. Heck e character s 5 6
Chapter 4 . Summatio n Formula s 6 5
§4.1. Introductio n 6 5
§4.2. Th e Euler-Maclauri n formul a 6 6
§4.3. Th e Poisso n summatio n formul a 6 9
§4.4. Summatio n formula s fo r th e bal l 7 1
§4.5. Summatio n formula s fo r th e hyperbol a 7 4
§4.6. Functiona l equation s o f Dirichle t L-function s 8 4
§4.A. Appendix : Fourie r integral s an d serie s 86

Chapter 5 . Classica l Analyti c Theor y o f L- functions 9 3

§5.1. Definition s an d preliminarie s 9 3

v
 vi CONTENT S

§5.2. Approximation s t o L-function s 9 7

1§5.3. Countin g zero s o f L-function s 0 1
§5.4. Th e zero-fre e regio n 0 5
§5.5. Explici t formul a 0 8
1
1§5.6. Th e prim e numbe r theore m 0
§5.7.1
1Th e Gran d Rieman n Hypothesi s 3
1
1§5.8. Simpl e consequence s o f GR H 7
§5.9. Th e Rieman n zet a functio n an d Dirichle1
1
t L-function s 9
§5.10. L-function s o f numbe r fields 1 2 5
§5.11. Classica
1 l automorphi c L-function s 3 1
§5.12.1Genera l automorphi c L-function s 3 6
§5.13. Arti n L-function s 4 1
1
§5.14. L-function s o f varietie s 4 5
1§5.A. Appendix : comple x analysi s 4 9

Chapter
1 6 . Elementar y Siev e Method s 5 3
§6.1. Siev e problem s 5 3
1 §6.2. Exclusion-inclusio n schem e 5 4
1 §6.3. Estimation s o f V+(z), V~(z) 5 7
§6.4. Fundamenta
1 l Lemm a o f siev e theor y 5 8
§6.5. Th e A 2-Sieve 6 0
§6.6. Estimat e fo r th e mai
1 n ter m o f th e A 2-sieve 6 4
§6.7. Estimate s fo r th e remainde r ter1 m i n th e A 2-sieve 6 5
2
§6.8.
1 Selecte d application s o f A -sieve 6 6
Chapter 7 . Bilinea 1r Form s an d th e Larg e Siev e 6 9
§7.1. Genera l principle s o f estimatin
1 g doubl e sum s 6 9
§7.2.
1 Bilinea r form s wit h exponential s 7 1
1§7.3. Introductio n t o th e larg e siev e 7 4
§7.4.
1 Additiv e larg e siev e inequalitie s 7 5
§7.5. Multiplicativ
1 e larg e siev e inequalit y 7 9
§7.4. Application s o f th e larg e siev e t o 1
sievin g problem s 8 0
§7.6. Panoram a o f1 th e larg e siev e inequalitie s 8 3
§7.7. Larg e siev1e inequalitie s fo r cus p form s 8 6
§7.8.
1 Orthogonalit y o f ellipti c curve s 9 2
1§7.9. Powe r moment s o f L-function s 9 4

Chapter 8 . Exponentia l Sum s 9 7

§8.1. Introductio n 9 7
§8.2. Weyl' s metho d 9 8
§8.3. Va n de r Corpu t metho d 20 4
1 §8.4. Discussio n o f exponen t pair s 2 3
1 §8.5. Vinogradov' s metho d 2 6

Chapter 9 . Th e Dirichle t Polynomial s 22 9

§9.1. Introductio n 22 9
§9.2. Th e integra l mean-valu e estimate s 23 0
§9.3. Th e discret e mean-valu e estimate s 23 2
§9.4. Larg e value s o f Dirichle t polynomial s 23 5
§9.5. Dirichle t polynomial s wit h character s 23 8
 TABLE O F CONTENT S vi i

§9.6. Th e reflectio n metho d 24 3

§9.7. Larg e value s o f D(s, x) 24 6
Chapter 1 0 . Zer o Densit y Estimate s 24 9
§10.1. Introductio n 24 9
§10.2. Zero-detectin g polynomial s 25 0
§10.3. Breakin g th e zero-densit y conjectur e 25 4
§10.4. Gran d zero-densit y theore m 25 6
§10.5. Th e gap s betwee n prime s 26 4
Chapter 1 1 . Sum s ove r Finit e Field s 26 9
§11.1. Introductio n 26 9
§11.2. Finit e field s 26 9
§11.3. Exponentia l sum s 27 2
§11.4. Th e Hasse-Davenpor t relatio n 27 4
§11.5. Th e zet a functio n fo r Kloosterma n sum s 27 8
§11.6. Stepanov' s metho d fo r hyperellipti c curve s 28 1
§11.7. Proo f o f Weil' s boun d fo r Kloosterma n sum s 28 7
§11.8. Th e Rieman n Hypothesi s fo r ellipti c curve s ove r finit e field s 29 0
§11.9. Geometr y o f ellipti c curve s 29 1
§11.10. Th e loca l zet a functio n o f ellipti c curve s 29 7
§11.11. Surve y o f furthe r results : a cohomologica l prime r 30 0
§11.12. Comment s 3 3

Chapter 1 2 . Characte r Sum s 3 7

§12.1. Introductio n 3 7
§12.2. Completin g method s 3 8
1 §12.3. Complet e characte r sum s 3 9
§12.4. Shor t characte r sum s 32 4
§12.5. Ver y shor t characte r sum s t o highl y composit e modulu s 330
§12.6. Character s t o powerfu l modulu s 33 5
Chapter 1 3 . Sum s ove r Prime s 33 7
§13.1. Genera l principle s 33 7
§13.2. A varian t o f Vinogradov' s metho d 340
§13.3. Linnik' s identit y 34 2
§13.4. Vaughan' s identit y 34 4
§13.5. Exponentia l sum s ove r prime s 34 5
§13.6. Bac k t o th e siev e 34 8
Chapter 1 4 . Holomorphi c Modula r Form s 35 3
§14.1. Quotient s o f th e uppe r half-plan e an d modula r form s 35 3
§14.2. Eisenstei n an d Poincar e serie s 35 7
§14.3. Thet a function s 36 1
§14.4. Modula r form s associate d t o ellipti c curve s 36 3
§14.5. Heck e L-function s 36 8
§14.6. Heck e operator s an d automorphi c L- functions 37 0
§14.7. Primitiv e form s an d specia l basi s 37 2
§14.8. Twistin g modula r form s 37 6
§14.9. Estimate s fo r th e Fourie r coefficient s o f cus p form s 37 8
viii C O N T E N T S

§14.10. Average s o f Fourie r coefficient s 38 0

Chapter 1 5 . Spectra l Theor y o f Automorphi c Form s 38 3
§15.1. Motivatio n an d geometri c preliminarie s 38 3
§15.2. Th e laplacia n o n H 38 5
§15.3. Automorphi c function s an d form s 38 6
§15.4. Th e continuou s spectru m 38 7
§15.5. Th e discret e spectru m 38 9
§15.6. Spectra l decompositio n an d automorphi c kernel s 39 1
§15.7. Th e Selber g trac e formul a 39 3
§15.8. Hyperboli c lattic e poin t problem s 39 8
§15.9. Distributio n o f lengt h o f close d geodesie s an d clas s number s 40 1
Chapter 1 6 . Sum s o f Kloosterma n Sum s 40 3
§16.1. Introductio n 40 3
§16.2. Fourie r expansio n o f Poincar e serie s 40 4
§16.3. Th e projectio n o f Poincar e serie s o n Maas s form s 40 6
§16.4. Kuznetsov' s formula s 40 6
§16.5. Estimate1 s fo r th e Fourie r coefficient s 4 3
§16.6. Estimate s fo r sum1 s o f Kloosterman sum s 4 5

Chapter 1 7 . Prime s1 i n Arithmeti c Progression s 4 9

§17.1. Introductio n 4 9
§17.2. Bilinea r form s i n arithmeti c progression s 42 1
§17.3. Proo f o f th e Bombieri-Vinogrado v Theore m 42 3
§17.4. Proo f o f th e Barban-Davenport-Halbersta m Theore m 42 4
Chapter 1 8 . Th e Leas t Prim e i n a n Arithmeti c Progressio n 42 7
§18.1. Introductio n 42 7
§18.2. Th e log-fre e zero-densit y theore m 42 9
§18.3. Th e exceptiona l zer o repulsio n 43 4
§18.4. Proo f o f Linnik' s Theore m 43 9
Chapter 1 9 . Th e Goldbac h Proble m 44 3
§19.1. Introductio n 44 3
§19.2. Incomplet e A-function s 44 5
§19.3. A ternar y additiv e proble m wit h A b 44 6
§19.4. Proo f o f Vinogradov' s thre e prime s theore m 44 7
Chapter 20 . Th e Circl e Metho d 44 9
§20.1. Th e partitio n numbe r 44 9
§20.2. Diophantin e equation s 45 6
§20.3. Th e circl e metho d afte r Kloosterma n 46 7
§20.4. Representation s b y quadrati c form s 47 2
§20.5. Anothe r decompositio n o f the delta-symbo l 48 1
Chapter 21 . Equidistributio n 48 7
§21.1. Weyl' s criterio n 48 7
§21.2. Selecte d equidistributio n result s 48 8
§21.3. Root s o f quadrati c congruence s 49 4
§21.4. Linea r an d bilinea r form s i n quadrati c root s 49 6
 TABLE O F CONTENT S i x

§21.5. A Poincar e serie s fo r quadrati c root s 49 8

§21.6. Estimatio n o f th e Poincar e serie s 50 1
Chapter 22 . Imaginar y Quadrati c Field s 50 3
§22.1. Binar y quadrati c form s 50 3
§22.2. Th e clas s grou p 50 8
1 §22.3. Th e clas s grou p L-function s 5 1
1 §22.4. Th e clas s numbe r problem s 5 7
§22.5. Splittin g prime s i n Q(y/D) 52 0
§22.6. Estimation s fo r derivative s L ^ ( 1 , X D ) 52 3
Chapter 23 . Effectiv e Bound s fo r th e Clas s Numbe r 52 9
§23.1. Landau' s plo t o f automorphi c L-function s 52 9
§23.2. A partitio n o f A ^ ( \ ) 53 1
§23.3. Estimatio n o f S 3 an d S 2 53 3
§23.4. Evaluatio n o f 5 i 53 4
§23.5. A n asymptoti c formul a fo r A^)(I ) 53 6
§23.6. A lowe r boun d fo r th e clas s numbe r 53 8
§23.7. Concludin g note s 54 0
§23.A Th e Gross-Zagie r L-functio n vanishe s t o orde r 3 54 1
Chapter 24 . Th e Critica l Zero s o f th e Rieman n Ze t a Functio n 54 7
§24.1. A lowe r boun d fo r N 0(T) 54 7
§24.2. A positiv e proportio n o f critica l zero s 55 0
Chapter 25 . Th e Spacin g o f th e Zero s o f th e Rieman n Ze t a-Function 56 3
§25.1. Introductio n 56 3
§25.2. Th e pai r correlatio n o f zero s 56 4
§25.3. Th e n-leve l correlatio n functio n fo r consecutiv e spacin g 57 0
§25.4. Low-lyin g zero s o f L-function s 57 2

Chapter 26 . Centra l Value s o f L-function s 57 7

§26.1. Introductio n 57 7
§26.2. Principl e o f th e proo f o f Theore m 26. 2 58 0
§26.3. Formula s fo r th e firs t an d th e secon d momen t 58 2
§26.4. Optimizin g th e mollifie r 58 9
§26.5. Proo f o f Theore m 26. 2 59 5
Bibliography 59 9

Index 611
This page intentionally left blank
 PREFACE

This boo k show s th e scop e o f analyti c numbe r theor y bot h i n classica l an d

modern directions . Ther e ar e no division lines ; in fact ou r inten t i s to demonstrate ,
particularly fo r newcomers , th e fascinatin g countles s interrelations . O f course , ou r
picture o f analyti c numbe r theor y i s by n o mean s complete , bu t w e tried t o fram e
the materia l int o a portrai t o f a reasonabl e size , ye t providin g a self-containe d
presentation.
We were writing thi s book i n a period o f time durin g an d afte r teachin g course s
and working with graduate students in Rutgers University, Bordeau x University an d
Courant Institute . W e than k thes e institution s fo r providin g condition s fo r bot h
of u s t o wor k together . W e share d idea s o n wha t thi s boo k shoul d b e abou t wit h
many of our colleagues, who gave us critical suggestions. Amon g them we would like
to mentio n Etienn e Fouvry , Joh n Friedlander , Philipp e Miche l an d Pete r Sarnak .
During a lon g proces s o f typin g an d preparatio n o f thi s boo k fo r publication , w e
received stimulatin g encouragemen t an d technica l advic e fro m Serge i Gelfand , fo r
all o f hi s hel p w e express ou r gratitude . Caro l Hame r helpe d t o polis h som e o f ou r
English phrase s whil e he r littl e boy s trie d t o destro y th e Te X files withou t success .
We thank the m al l fo r th e output .
Henryk Iwanie c
Emmanuel Kowalsk i
15 December, 200 3

xi
This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/01

INTRODUCTION

Analytic Numbe r Theor y distinguishe s itsel f b y th e variet y o f tool s i t employ s

to establis h results , man y o f whic h belon g t o th e mai n stream s o f arithmetic . I t
is no t par t o f analysi s no r o f an y particula r disciplin e o f mathematics , howeve r i t
does interac t indee d wit h variou s fields. Therefor e everybod y seem s t o vie w th e
subject differently . Thi s vas t diversit y o f concept s o f analyti c numbe r theor y i s it s
great attraction . Ou r desir e i n thi s boo k i s t o exhibi t th e wealt h an d prospect s
of th e theory , it s charmin g theorem s an d powerfu l techniques . Howeve r i t i s no t
our primar y objectiv e t o giv e proof s o f th e stronges t results , althoug h i n man y
cases w e com e quit e clos e t o th e bes t possibl e ones . Rathe r w e favo r a reasonabl e
balance betwee n clarity , completenes s an d generality . Th e boo k wa s conceive d
with graduat e student s i n min d s o th e reade r will ofte n find tha t ou r emphasi s i s
on reasonin g throughou t th e arguments . O f cours e ou r presentatio n i s subjective ,
and i n retrospec t ma y los e it s meaning . Certainl y w e d o no t alway s follo w th e
original line s of discovery, bu t occasionall y w e do draw brie f historica l perspectives .
Leonard Eule r mus t ge t credi t fo r th e first us e o f analytica l argument s fo r th e
purpose o f studyin g propertie s o f integers , specificall y b y constructin g generatin g
power series . Euler' s proo f o f th e infinit y o f prim e number s make s us e o f th e
divergence o f th e zet a functio n an d th e correspondin g produc t ove r primes , whic h
is name d afte r him . Thi s wa s th e beginnin g o f analyti c numbe r theory . Nex t
came P . G . L . Dirichle t whos e creatio n o f the theor y o f L- functions fo r characters ,
resulting i n th e proo f o f th e infinit y o f prime s i n arithmeti c progressions , make s
him th e tru e fathe r o f analyti c numbe r theory . Pro m thes e earl y day s t o moder n
times th e distributio n o f prim e number s constitute s th e cor e o f th e subject . Thi s
will b e apparen t i n th e cours e o f ou r book . Th e first tw o chapter s cove r question s
of prime s u p t o th e elementar y method s o f P . Tchebychev .
Chapter 3 provides definition s an d basi c propertie s o f Dirichle t character s an d
the Gaus s sums . Character s o n ideal s o f imaginar y quadrati c fields ar e als o intro -
duced, no t onl y becaus e the y pla y a supportin g rol e i n subsequen t chapter s bu t t o
show a bi t o f analytic numbe r theor y beyon d th e traditiona l domai n o f the rationa l
integers a s well ; ther e wil l b e othe r example s throughou t th e book , fo r instance ,
elliptic curves .
Poisson summatio n fo r numbe r theor y i s wha t a ca r i s fo r peopl e i n moder n
communities - i t transport s thing s t o othe r place s an d i t take s yo u bac k hom e
when applie d nex t tim e - on e canno t liv e without it . Chapte r 4 presents a classica l
account o f this basic technology. Man y reader s do realize now, others will figure out
later, tha t w e ar e alread y talkin g abou t idea s o f modula r forms . Bu t w e continu e
our consideration s alon g traditiona l line s (bot h classica l an d mor e recen t ones )
before th e concep t o f modularit y take s th e leadin g position .
1
2 INTRODUCTION

The celebrate d memoi r o f B . Rieman n o n th e zet a functio n i s embedde d i n

the contex t o f abstrac t L-function s i n Chapte r 5 . I t i s no t ou r styl e t o conside r
things i n term s mor e genera l tha n necessary , s o definin g a clas s o f L-function s
which suits minimum requirement s o f our forthcomin g application s wa s not withou t
difficulty an d hesitation . I n thi s wa y we could conve y to dedicate d researcher s tha t
generalizations ar e no t alway s straightforward . Fo r instance , t o establis h th e zero -
free region for L-functions o f degree > 1 one cannot rel y on the same principles as for
the Dirichlet L-functions . Th e key ingredient i s the Rankin-Selberg convolution. O n
the othe r hand , th e proble m o f exceptiona l zer o i s resolve d fo r man y automorphi c
L-functions o f degree > 1 (not withou t cleve r constructions ) whil e i t remain s ope n
for th e L-function s wit h rea l characters . Furthermor e i n Chapte r 5 a messag e i s
sent tha t a bette r lif e exist s i n th e worl d o f automorphi c form s tha n i n th e zo o of
degree on e L-functions .
Analytic number theor y does not mea n non-elementary. Th e first autho r recall s
that hi s first seriou s encounte r wit h analyti c numbe r theor y starte d b y readin g th e
lovely boo k o f A.O . Gelfon d an d Yu.V . Linnik , "Elementar y Method s o f Analyti c
Number Theory" . Whe n a n ambitiou s beginne r start s fro m ther e her/hi s lov e o f
the subjec t i s seale d forever . Tr y i t yourself ! On e i s instantl y capture d b y siev e
methods. I n thi s boo k w e do no t hav e spac e t o giv e justice t o thi s marvelou s idea ,
nevertheless Chapte r 6 should suffic e fo r basi c applications .
Next come s the "Larg e Sieve" , which is not a sieve but a name fo r othe r things .
Yes, i t di d originat e fro m a shor t pape r b y Linni k o n a siev e problem , bu t i t too k
time t o recogniz e th e tru e natur e o f thes e ideas . I n Chapte r 7 we revea l ou r view -
point an d th e crucia l attribute s (spectra l completeness , orthogonality) , the n w e
demonstrate th e amazin g powe r o f th e larg e siev e o n selecte d ol d an d ne w prob -
lems. Othe r feature s o f the large sieve are scrutinized showin g the good an d th e ba d
sides. Fo r example , th e approac h usin g th e dualit y principl e i s fruitful fo r harmon -
ics o f degre e on e (characters ) whil e producin g poo r result s fo r harmonic s o f large r
degree (lik e fo r exampl e th e eigenvalue s o f Heck e operators). Th e controvers y ove r
the prope r plac e o f th e larg e siev e i s academic . Simpl y speakin g th e larg e siev e
inequalities ar e part s o f bilinea r form s theory .
Estimates fo r exponentia l sum s ar e th e first tool s whic h deepl y penetrat e th e
problems o f analyti c numbe r theor y beyon d natura l structures . Thes e canno t b e
grasped b y harmoni c analysi s alone . Se e what cleve r us e peopl e mad e o f th e prop -
erty tha t a shifte d interva l i s anothe r interval , tha t addin g a n intege r t o a se t o f
integers yield s agai n a se t o f integer s (sorry , prime s ar e no t preserved!) . W e chal -
lenge algebraist s t o fin d a structura l explanatio n o f th e powe r o f suc h arguments !
They shoul d rea d Chapte r 8 t o fin d wha t H . Wey l buil t ou t o f thes e observa -
tions. Va n de r Corpu t an d Vinogrado v ar e als o th e mai n figures fro m th e earl y
stages o f tha t discipline . A lo t o f wor k an d talkin g wen t int o ou r presentatio n o f
Vinogradov's method , becaus e i t i s not quit e correctl y explaine d i n numerou s pub -
lications. A t som e poin t Vinogrado v depart s fro m th e Wey l differencin g proces s
and treat s multi-dimensiona l exponentia l sum s a s bilinea r form s (thi s i s th e wa y
we think o f i t anyway) .
The nex t tw o chapters sho w more recent technolog y which was developed t o re-
place the unprove n Rieman n hypothesi s i n applications t o th e distributio n o f prime
numbers. W e ar e talkin g abou t estimate s fo r th e numbe r o f zero s o f L-function s i n
vertical strip s whic h ar e positivel y distance d fro m th e critica l line . Hopefull y i n a
future on e wil l sa y w e wer e wastin g tim e o n studyin g th e empt y set . Grea t idea s
 INTRODUCTION 3

are camouflage d ther e i n argument s o f enormou s complexity , s o thi s migh t no t b e

enjoyable fo r everyon e a t first . Howeve r i f you thin k th e Rieman n hypothesi s i s not
provable i n you r lifetime , pleas e rea d an d admir e thes e unconditiona l substitutes .
Special mentio n goe s t o Hug h Montgomery , Marti n Huxle y an d Matt i Jutil a fo r
the mos t origina l contributions .
Although w e are primar y intereste d i n rationa l integer s on e ca n lear n an d ben -
efit a lo t fro m arithmeti c o f other fields . No t onl y fro m th e numbe r field s o r p-adi c
fields, bu t indirectl y from th e fields of finite characteristi c a s well. Particularl y fruit -
ful ar e th e method s o f exponentia l sum s ove r th e finit e fields . I n Chapte r 1 1 we
prove (amon g othe r things ) th e Rieman n hypothesi s fo r specia l curve s whic h yield s
the celebrated estimat e of Weil for Kloosterma n sums . Th e Kloosterma n sum s hav e
been employe d t o solve various problems o f analytic numbe r theor y fro m th e begin -
ning o f thei r creatio n i n 1 926 . W e als o mentio n briefl y th e stat e o f th e knowledg e
of exponentia l an d characte r sum s ove r algebrai c varieties . Application s o f thes e
are harde r t o make , ye t ther e i s a handfu l o f example s i n th e literature . I t wa s
a painfu l decisio n t o exclud e al l bu t th e simples t fro m presentatio n i n thi s book .
Otherwise t o d o ful l justic e fo r thes e highl y sophisticate d idea s w e woul d hav e t o
choose th e mos t complicate d applicatio n fo r whic h w e hav e n o room . I t suffice s t o
say that a preparation o f a given problem o f analytic numbe r theor y t o a n estimat e
for characte r su m ove r varietie s ca n b e th e state-of-the-ar t i n it s ow n right , neve r
mind tha t th e final argumen t i s powered by the outside forces o f algebraic geometry .
Dirichlet character s ar e alread y discusse d i n Chapte r 3 and w e return t o the m
in Chapte r 1 2 to trea t ver y shor t characte r sums . Agai n on e mus t b e inventiv e t o
break limit s o f natura l structures . Burges s theore m i s a fin e example .
Sums over primes ar e treated i n the next chapter . Whe n Vinogrado v succeede d
in estimating sums over primes of additive characters, which he needed for a solution
to th e ternar y Goldbac h proble m i n conjunctio n wit h th e circl e method , i t wa s a
shocking result . Befor e hi m th e Gran d Rieman n Hypothesi s coul d d o th e job , bu t
keep in mind that th e Riemann hypothesi s is still not established . Th e original ideas
of Vinogradov were borrowed fro m combinatoria l sieve and were rather complicated .
Recently develope d identitie s offe r muc h simple r treatment s o f mor e genera l sum s
over primes . A s the y shar e th e sam e fundamenta l principl e (reducin g th e su m t o
bilinear forms ) th e result s ar e prett y muc h th e same , s o th e choic e o f th e metho d
is a matte r o f tast e an d technica l convenience . T o captur e th e ke y element s i n
Chapter 1 3 we develop mor e tha n on e identity .
A popular criterio n for analyti c number theor y is that comple x variable analysi s
is bein g used . Perhap s i t i s bette r t o sa y harmoni c analysis , sinc e th e actio n o f
the latte r i s mor e profound . Fo r a lon g time , analyti c numbe r theor y flourishe d
exclusively from abelia n harmonic analysis, that i s to say from th e Fourier transfor m
in IR n. Ther e i s stil l a grea t potentia l i n thi s classica l analysi s t o b e explored .
However muc h stronger fertilizer s bega n to act o n the soi l of analytic number theor y
in recen t times . Thes e ar e automorphi c functions . O f course , modula r form s hav e
been driving algebraic aspect s of number theor y muc h longer , bu t i n a limited scop e
(confined t o holomorphic forms) . Ne w resources of automorphic theor y are found i n
the spectra l analysis , th e foundatio n o f which wa s le d b y H . Maas s an d A . Selber g
at th e turn o f the 1 940' s (real-analyti c cus p forms, Eisenstei n series , trace formula) .
In simpl e term s a non-abelia n harmoni c analysi s foun d it s rol e i n analyti c numbe r
theory. Trul y effectiv e expansio n o f spectra l method s int o analyti c numbe r theor y
began abou t twent y fiv e year s ago , changin g th e fac e o f eithe r subjec t irrevocably .
4 INTRODUCTION

This boo k barel y addresse s th e fascinatin g issue s o f th e ne w directio n throughou t

Chapters 1 4 , 1 5 , an d 1 6 . Ou r feature d applicatio n i s t o estimatio n fo r sum s o f
Kloosterman sums . Thi s i s a goo d choic e (i f n o mor e ca n b e accommodated) ,
because th e reade r ca n appreciat e th e ne w tool s b y comparin g wit h th e earlie r
results derive d i n Chapte r 1 1 b y algebrai c considerations . Anothe r applicatio n o f
the spectra l theor y o f automorphi c form s t o a n arithmetica l questio n i s presente d
in Chapte r 21 , tha t i s t o th e equidistributio n o f root s o f quadrati c congruence s
of prim e moduli . Th e spectra l theor y continue s t o gro w extensively , s o i t woul d
be prematur e t o wra p i t u p her e o r i n an y othe r book . Fo r furthe r readin g w e
recommend [1 3] , [Sa3] .
Although th e spectra l method s o f automorphi c form s predominat e curren t re -
search i n analyti c numbe r theory , th e traditiona l problem s continu e gettin g ou r
attention wit h respectfu l intensit y throughou t th e remainin g chapters . Grea t trea -
sures o f th e subjec t mustn' t b e burie d i n th e past . Firs t o f al l a newcome r shoul d
learn th e storie s o f prime s i n arithmeti c progression s t o larg e moduli . I n Chapte r
17 she/h e wil l find ho w E . Bombier i an d A.I . Vinogrado v bypasse d th e Rieman n
hypothesis t o establis h (b y th e larg e siev e an d othe r means ) unconditiona l result s
with application s a s goo d a s on e ca n ge t fro m th e R H itself . O f cours e ou r argu -
ments ar e no t identica l wit h th e origina l one s (o f 1 965 ) sinc e w e take advantag e o f
later simplification , i n grea t measur e du e t o P.X . Gallagher .
Chapter 1 8 goes further bac k to 1 94 4 when Linnik gave an extraordinary boun d
for th e leas t prim e i n a n arithmeti c progression . Fo r a lon g tim e thi s boun d wa s
considered a s th e mos t difficul t theore m i n analyti c numbe r theory . An d yes , i t i s
still har d b y today' s standards , an d on e ca n stil l lear n a lo t fro m th e technolog y
applied! Face d with the obstacle of the exceptional zero, Linnik brings the repulsio n
effect (h e calls it Deuring-Heilbronn phenomenon ) t o a new level; amazingly enoug h
he turn s th e proble m t o hi s advantage ! Thi s i s a fascinatin g developmen t i n th e
history o f analyti c numbe r theor y whic h w e recommen d on e shoul d maste r fo r a
better understandin g o f the statu s o f th e exceptiona l zer o today .
Once upo n a tim e th e famou s Goldbac h proble m wa s wort h a millio n dolla r
prize award . Fo r application s th e proble m (representation s o f eve n integer s b y th e
sums o f two primes ) ha s n o grea t merit , bu t a s a n intellectua l challeng e on e woul d
be prou d t o crac k it . Probabl y somethin g ne w abou t prim e number s woul d b e
revealed then . Rea d Chapte r 1 9 to improv e you r chances .
Chapter 2 0 i s serious . Her e analyti c method s stor m th e domai n o f diophan -
tine equations , whic h fro m ancien t Greek s wa s exclusively a busines s o f arithmetic .
Started b y Hard y - Ramanujan , continue d b y Hard y - Littlewoo d an d develope d
substantially furthe r b y Kloosterman, th e circl e method use s orthogonality o f addi -
tive characters to detect equations , no t onl y to solve algebraic equation s bu t a larg e
class of additive problem s ove r special integers a s well. Th e toughes t ar e the binar y
additive problems. The y are not completely solved by the Kloosterman method, bu t
at leas t w e get a very reliable picture o f what th e true asymptoti c fo r th e numbe r o f
solutions shoul d be . Kloosterma n sum s whic h w e covered i n the precedin g chapter s
are instrumenta l i n th e circl e methods . After classica l idea s w e propos e a mor e
direct varian t whic h i n principl e shoul d produc e th e sam e results , howeve r withou t
employing Kloosterma n sums . On e shoul d rea d Chapte r 2 0 wit h a n ope n mind ,
separate technica l (stil l attractive ) element s fro m conceptua l device s t o see clearl y
 INTRODUCTION 5

the connection s wit h modular forms . Certainl y Kloosterma n an d Rademache r wer e

aware o f thes e intrinsi c connections , whil e the y ar e overlooke d b y som e specialist s
in th e circl e method .
Equidistribution problem s fo r sequence s o f specia l integers , lattic e point s i n
various domains , solutio n t o diophantin e equations , etc , constitut e a heav y indus -
try over the analyti c number theory. W e regret ther e is no space to run this industr y
in full capacit y i n the book . Th e boo k of M.N. Huxle y [Hu4 ] treat s onl y the lattic e
point problems , howeve r quit e deeply . I n Chapte r 2 1 we are dealin g wit h th e prob -
lem o f distributio n o f roots o f a quadrati c equatio n reduce d modul o prime . A s th e
prime modulu s tend s t o infinit y w e sho w tha t th e root s ar e uniforml y distributed .
The argument s includ e almos t everythin g tha t w e develope d i n th e boo k s o far ,
thus showin g tha t th e industr y i s robust .
Because o f failur e o f th e uniqu e factorizatio n o f algebrai c integers , th e arith -
metic of number fields is not a s easy as for rationa l numbers an d sometime s perplex -
ing. Th e complexity i s measured b y the order o f the ideal class group. Naturall y th e
case of imaginary quadrati c fields receive d th e first an d th e mos t attentio n becaus e
units do not interfere . W e do know that th e class number grow s to infinity (s o there
is only a finite numbe r o f imaginary quadrati c fields with a fixed class number), bu t
the seriou s issu e i s t o estimat e th e clas s numbe r effectively . Chapte r 2 2 describe s
the proble m thoroughl y an d prepare s th e groun d fo r th e advance s i n Chapte r 23 .
The effective lowe r bound fo r the class number (du e to D. Goldfeld) ma y not appea r
strong fo r demandin g researchers , ye t i t i s deep wit h respec t t o result s take n fro m
other sources . Firs t o f all it use s the Gross-Zagie r formul a fo r //-function s o f ellipti c
curves a t th e centra l point . W e d o provid e a substantia l overvie w o f th e involve d
arguments fro m ellipti c curves , althoug h thes e ar e mor e geometri c tha n analytic .
The analyti c argument s themselve s ar e quit e delicate . Actuall y the y cam e first,
the L-function s o f ellipti c curve s bein g supplementary . Indee d w e worke d ou t a n
effective lowe r boun d fo r th e clas s number whic h depend s o n th e orde r o f vanishin g
of general L-function s o f degre e two , suspectin g tha t th e requirement s ar e satisfie d
by quit e a fe w o f them .
In Chapte r 2 4 we prove a very classical result o f Selberg that a positive propor -
tion o f zero s o f th e Rieman n zet a functio n lie s o n th e critica l line . Thi s i s a goo d
place t o lear n abou t th e mollificatio n technique s ( a kin d o f smoothing) , whic h i s
used i n man y work s toda y an d wil l reappea r i n Chapte r 26 .
Assuming th e Rieman n hypothesis , H.L . Montgomery reveale d i n 1 97 4 that th e
distribution o f zero s o f £(s ) follow s th e behavio r o f eigenvalue s o f certai n "ensem -
bles" o f unitar y matrices . Mor e recentl y physicist s joine d th e tea m o f worker s i n
number theory , creatin g a ne w excitemen t an d hop e fo r finding a pat h t o a proo f
of the Rieman n hypothesis . Thi s i s the mai n objectiv e o f the so-calle d rando m ma -
trix theory , on e o f th e mos t popula r subjec t an d drivin g force s o f curren t analyti c
number theory . I t offer s reliabl e model s fo r predictin g th e behavio r o f arithmetica l
quantities whic h fo r a lon g tim e wer e shroude d i n mystery . Th e consistenc y o f
the rando m matri x theor y wit h th e harmon y o f integer s stil l seem s quit e surpris -
ing. Whateve r th e futur e o f this enterpris e wil l be , du e t o th e curren t cooperation ,
analysis i s closer t o arithmeti c tha n eve r before . A subject o f suc h magnitud e can -
not b e full y presente d i n a shor t space . Therefor e i n Chapte r 2 5 w e stic k t o th e
original them e o f th e correlatio n o f zero s o f £(s ) an d it s variation s o n th e zero s o f
6 INTRODUCTION

families o f automorphi c L-function s whic h ar e nea r th e centra l point . W e leave i t

for th e reader t o judge whethe r th e idea s o f random matri x theor y ar e realistic fo r
launching a n attac k o n the Rieman n Hypothesis .

In recen t investigation s th e centra l value s o f L-function s appea r i n a variet y

of formula s wit h vanishin g o r non-vanishin g assumptions . Tak e fo r exampl e [IS2 ]
where a n effectiv e lowe r boun d fo r th e clas s numbe r o f imaginar y quadrati c fields
is derive d essentiall y fro m th e non-vanishin g o f centra l value s o f familie s o f L -
functions, t o th e contrar y o f the vanishin g requirement s i n the previou s investiga -
tions. Anothe r exampl e i s the formul a o f T. Watso n [Wa ] by mean s o f which th e
quantum-ergodicity conjectur e (tha t i s the equidistributio n o f Maas s cus p forms )
is reduced t o a subconvexity boun d fo r certai n L-function s o f degree four . W e con-
sider i n detai l on e non-vanishin g statemen t whic h ha s application s t o arithmeti c
geometry.

We hope thi s boo k will show the picture o f analytic numbe r theor y i n plenty of
colors. Howeve r we must sa y that a lot of significant topic s are left out . Missin g are
the dispersio n method , th e amplificatio n metho d (se e [M2]) , som e analyti c tech -
niques fro m diophantin e approximation s an d transcendence . Moreove r probabilit y
arguments ar e barel y exposed , an d w e didn' t touc h ergodi c theor y either , whos e
impact o n number theor y ha s been fel t strongl y i n the last years .
We als o try to show som e detail s o f the powerful theorie s whic h ar e developin g
as th e mos t usefu l ne w tools fo r analyti c numbe r theory , i n particula r th e theor y
of higher-degree automorphi c form s an d their L-functions , an d algebraic geometry ;
young researcher s i n particular shoul d b e encouraged t o develop expertis e i n thes e
subjects. I t i s certain tha t spectacula r application s hav e onl y begu n an d more wil l
be ope n t o thos e wh o understan d bot h sides . Dually , arithmeti c geometr y an d
algebraic numbe r theor y als o giv e an d promis e a wealt h o f ne w questions, o r new
aspects o f old ones, wher e th e skills an d technique s o f analytic numbe r theor y wil l
be tested to the utmost. Hopefull y the y will bring rich rewards to those who will try
to com e to these ope n fields... W e barely mentio n som e questions relate d t o ellipti c
curves bu t w e believe tha t ther e i s muc h mor e t o discover . Th e dee p conjecture s
of Lan g an d Trotte r [LT ] ar e alread y quit e popular , an d a fe w othe r challengin g
problems ma y be foun d i n [Kol] .

The exercise s insid e eac h sectio n serv e a dua l purpose , som e ar e t o improv e
the reader' s skill , th e other s serv e a s additiona l informatio n abou t th e subject .
Historical remark s ar e brief , t o giv e som e orientatio n i n th e developmen t o f th e
matter, rathe r tha n t o credi t exhaustivel y th e inventors . Th e only advic e w e offe r
to new researchers is read! read ! read ! man y papers with complete proofs. Knowin g
a resul t i n analytic numbe r theor y i s only the first ste p to liking it; more importan t
and rewardin g i s to understan d th e argument s o f its proof . Ou r viewpoin t i s tha t
making mathematic s shoul d no t b e rate d lik e breakin g spor t records . Sometime s
the stronges t resul t i s boring whil e a slightl y weake r on e generates grea t pleasure .

Formal prerequisite s fo r muc h o f the boo k ar e rather slight , no t goin g beyon d

differential calculus , comple x analysi s an d integration, especiall y Fourie r serie s and
integrals. I t i s more importan t fo r the reader t o have or acquire a good understand -
ing o f how to manipulat e inequalitie s an d not simpl e identities .
 INTRODUCTION 7

In late r chapter s automorphi c form s becom e important . W e have include d tw o

survey chapters, ye t w e expect tha t man y readers will have already som e knowledg e
of this importan t topic , o r wil l stud y i t independently .
In som e section s (fo r instanc e Section s 5.1 3 an d 5.1 4) , whic h ar e intende d a s
convenient reference s fo r certai n fact s an d result s which ar e hard t o locate in prope r
form i n th e literature , w e assum e tha t th e reade r ha s som e familiarit y wit h othe r
topics, suc h a s representation s o f group s an d algebrai c geometry .
Sections o f this boo k wer e writte n ove r a perio d o f time , therefor e reader s wil l
notice a sligh t chang e o f styl e an d repetition . W e thin k tha t a smal l redundanc y
is helpfu l fo r readin g lon g arguments . Occasionall y th e sam e objec t i s introduce d
again i n a differen t chapte r i n loca l terminolog y whic h shoul d b e mor e familia r i n
a particula r context . W e believ e thi s flexibilit y i s justified fo r comfort , eve n a t th e
expense o f losin g uniqueness .
Our notation s ar e mostl y standard . Bu t sinc e inequalitie s wit h unspecifie d
constants ar e the lifeblood o f analytic number theory , an d sinc e there ar e sometime s
controversies o n thi s subject , w e spel l ou t th e meanin g o f th e variou s compariso n
symbols 0 ( ) , o() , ~ , x o r <C . Mos t important , w e us e Landau' s / = 0(g) an d
Vinogradov's / < C g a s synonyms ; thu s f(x) < C g(x) fo r x G X (wher e X mus t b e
specified eithe r explicitl y o r implicitly ) mean s tha t \f(x)\ < Cg(x) fo r al l x G X
and som e constan t C ^ 0 . An y valu e o f C fo r whic h thi s hold s i s called a n implie d
constant. Sinc e a constan t i s mos t ofte n a functio n lookin g fo r a variable , th e
"implied constant " wil l sometimes depen d o n other parameters , whic h we explicitly
mention a t th e mos t importan t point s (bu t sometime s i t i s clea r fro m context) . I f
there i s no othe r dependency , w e speak o f "absolut e constants" . Thi s usag e mean s
that ou r <9() , for instance , i s not th e sam e a s tha t o f Landa u o r Bourbaki . W e us e
/ x g t o mea n tha t bot h relation s / < C g an d g <C / hold , o f cours e wit h possibl y
different implie d constants .
However / = o(g) fo r x —> XQ means tha t fo r an y e > 0 ther e exist s som e
(unspecified) neighborhoo d U e o f x$ suc h tha t \f(x)\ ^ eg(x) fo r x G U £. The n
h ~ g mean s h = g + o(g). Thos e ar e th e sam e a s i n Landa u o r Bourbaki .
Among th e fe w notatio n whic h ma y b e unfamilia r t o a beginner , w e mentio n
that p k\\m, wher e p i s prim e an d k a n integer , mean s tha t p k divide s m exactl y
(i.e. p kJrl doe s no t divid e m) . Th e integra l par t [x] is th e intege r n suc h tha t
n< x < n+ 1 .
We sometime s us e th e notatio n Yl* t o denot e sum s restricte d t o a subse t o f
"primitive" objects , whic h wil l b e indicate d i n eac h case , an d ^ t o denot e a su m
restricted t o squarefre e numbers .
This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/02

CHAPTER 1

ARITHMETIC FUNCTION S

1.1. Notatio n an d definitions .

Throughout thi s boo k Z , Q, R, C denot e th e set s o f integers, rationals , rea l an d
complex number s respectively . Th e additio n o f complex number s make s C a group ,
and Z c Q c R c C i s th e natura l inclusio n o f subgroups . Th e Lebesqu e measur e
dx o n th e additiv e grou p R i s th e invarian t measur e (wit h respec t t o additiv e
translations). Th e subset s Q * C R* C C * of the non-zer o number s ar e closed unde r
multiplication, the y ar e considere d a s multiplicativ e groups . I n R * w e distinguis h
the subgrou p o f positiv e rea l number s R + whic h i s endowe d wit h th e invarian t
measure x~ ldx (wit h respec t t o multiplicativ e translations) .
The positiv e integer s ar e calle d natura l numbers . Th e se t o f natura l number s
N = { 1 , 2 , 3 , . . . } contain s th e se t o f prime s P = {2,3 , 5, 7 , . .. }. Thes e ar e funda -
mental element s o f arithmeti c sinc e ever y natura l numbe r factor s uniquel y (u p t o
a permutation ) int o power s o f distinc t primes . Th e distributio n o f prime s i s one of
basic problem s o f analyti c numbe r theory . Prim e number s ar e ofte n denote d b y p.
First o f all, to cultivate analyti c number theor y on e must acquir e a considerabl e
skill fo r operatin g wit h arithmeti c functions . W e begi n wit h a fe w elementar y
considerations.
A comple x value d functio n / define d o n N is calle d a n arithmeti c function . I n
some context s a n arithmeti c functio n / : N - > C i s bette r see n a s th e sequenc e
A = (a n ) wit h a n = f(n). Ver y ofte n i n suc h contex t A i s supporte d o n number s
of primar y interest , an d a n i s jus t a multiplicity , o r som e sor t o f weight , whic h
is introduce d whe n countin g suc h numbers . A differen t jo b i s assigne d t o certai n
functions / : N — > C (suc h a s Dirichle t character s o r Heck e eigenvalues ) whic h
we cal l "arithmeti c harmonics. " Thes e pla y a n instrumenta l rol e i n analyzin g th e
primary sequenc e A = (a n). Essentiall y th e arithmeti c harmonic s ar e employe d
with th e primar y sequenc e A = (a n) t o produc e a famil y o f twiste d sequence s
Af = (a nf(n)). Th e twiste d sequence s wit h appropriat e harmonic s ar e capabl e o f
selecting a special subsequence of A whic h is our target (thin k of Dirichlet character s
being employe d t o detec t prime s i n arithmeti c progressions) .
Thanks t o th e additiv e an d multiplicativ e structure s o f integers th e tw o impor -
tant classe s o f arithmeti c function s ar e distinguished . A functio n / : N— > C i s a n
additive functio n i f i t satisfie s

(1.1) f(mn) = f(m) + f(n)

for m,n relativel y prime . I f thi s propert y hold s fo r al l ra, n , the n / i s sai d t o b e
completely additive . Fo r exampl e f{n) = log n i s completel y additive . Similarly , a
9
10 1. ARITHMETI C FUNCTION S

function / : N—» C i s multiplicative i f i t satisfie s

(1.2) f(mn) = f(m)f(n)
for ra,n relativel y prime , an d / i s completel y multiplicativ e i f (1 .2 ) hold s fo r al l
ra,n. Fo r exampl e f(n) = n~ s wit h 5 G C i s completel y multiplicative . Obviously ,
the additiv e an d th e multiplicativ e function s ar e determine d b y thei r value s a t
prime powers . W e hav e / ( l ) = 0 i f / i s additive , / ( l ) = l i f / i s multiplicativ e no t
identically zero .

1.2. Generatin g series .

To an arithmeti c functio n / w e shall attac h th e tw o most natura l infinit e serie s
(1.3) E f(z) = £/(n)z" ,
n

(1.4) D f(s) = ^/(n)n" s,

n
where z, s ar e comple x variable s (i n th e cas e o f th e powe r serie s Ef(z) w e ma y
also includ e n = 0 i f /(0 ) i s defined) . Thes e ar e calle d th e generatin g series , o r
functions, o f / . Othe r kind s o f generatin g function s ca n b e attractiv e a s wel l t o
capture distinc t propertie s o f / . Fo r man y arithmeti c function s th e correspondin g
generating serie s converge s absolutel y i n a smal l domain , an d i t i s a n interestin g
question ho w fa r th e serie s ca n b e analyticall y continued ? I f th e generatin g se -
ries extend s beyon d th e rang e o f absolut e convergence , the n thi s propert y usuall y
manifests som e grou p structur e i n th e coefficient s f(n) ("random " serie s canno t
be continued) . Fo r exampl e th e analyti c continuatio n o f Arti n L- functions i s inti -
mately relate d t o th e reciprocit y law s i n numbe r fields (i n th e abelia n cas e a t an y
rate), th e analyti c continuatio n o f Hasse-Wei l L- functions i s a majo r ste p toward s
the modularit y o f th e correspondin g ellipti c curve s (ove r Q) , an d s o on .
The generating power series were introduced b y L. Euler (1 707-1 783 ) fo r study -
ing specia l additiv e problem s (whe n doin g s o h e wa s th e first t o mi x analysi s wit h
arithmetic). Le t Ef(z) an d E g(z) b e th e serie s fo r / an d g. The n th e produc t
n
(1.5) E f(z)Eg(z) = J2 Hn)z
n
yields th e powe r serie s fo r th e functio n h give n b y (th e additiv e convolution )

(1.6) h(n) = J2 fW9(™)-

£-\-m=n

Applying Cauchy' s theore m on e expresse s th e coefficien t h(n) b y th e contou r inte -

gral
1 n
(1.7) h(n) = ^- f E f(z)Eg(z)z' ' dz.
J\z\=r

Hence, give n reasonabl e analyti c propertie s o f Ef(z)E g(z) on e ca n deduc e a goo d

estimate fo r h(n), o r even a n asymptoti c formul a a s n tend s t o oo. O f course, Eule r
did no t kno w Cauchy' s theorem , s o hi s idea s wer e limite d t o th e powe r serie s fo r
which on e has a secondary expression , givin g a n exac t formul a fo r f{n) rathe r tha n
 1. A R I T H M E T I C F U N C T I O N 11

an approximatio n (stil l no t a tautology). W e give three cut e examples . Firs t i s th e

identity
oo oo

$>" = II(1 + *2m) = ( 1 -*r 1

10
which i s nothin g bu t a n analyti c statemen t o f th e uniquenes s o f binar y expansio n
of natura l numbers . Her e i s anothe r identity :
oo oo

X>(n)* = II( 1 -* ro )~ 1
n
10

where p(ri) i s the partition function . Les s obvious is the following formul a o f Jacobi:
oo oo

(1.8) £ V * = JJ(1 - z m){\ - z m+1 /2f.

1
— oo

About ninet y year s ag o th e integra l representatio n (1 .7 ) fo r th e additiv e equatio n

(1.6) gav e birth t o th e circl e method. Thes e idea s wil l be develope d i n Chapte r 20 .
By changin g th e variabl e z int o
2niz
(1.9) e(z) = e

the powe r serie s (1 .3 ) i s see n a s a Fourie r series , whic h i s a commo n practic e i n

modern analyti c numbe r theory .
The generatin g serie s Df(s) i s calle d th e Dirichle t serie s afte r bein g use d i n
Dirichlet's fundamenta l work s o n prime s i n arithmeti c progressions , thoug h th e
special cas e
oo
S
(1.10) C(*)=£n-
1

was alread y considere d b y Euler .

The Dirichle t serie s i s particularl y attractiv e fo r multiplicativ e function s (bu t
not limite d t o thi s case) . For , i f / i s multiplicative, the n

(l.n) D f(s) = n(i+ f(p)p~s+f(p2)p~2s+•••)

p

provided th e involve d serie s i n prim e power s an d th e produc t ove r prime s (calle d

the Eule r product ) converg e absolutely . I n particular , w e hav e

(i.i2) C(* ) = II(i-p-T 1

V

if R e (s) > 1 . Thi s representatio n o f £(s ) a s a n Eule r produc t i s a n expressio n i n

analytic languag e o f the uniqu e factorization o f natural number s int o distinct prim e
powers.
12 1. ARITHMETI C FUNCTION S

1.3. Dirichle t convolution .

Assuming tha t th e Dirichle t serie s Df(s),D g(s) converg e absolutel y i t follow s
that th e product Df(s)D g(s) i s also give n b y a Dirichle t series , namel y w e have
oo
s
Df(s)Dg(s) = J2 Kn)n~
1

with th e coefficient s

(1.13) Mn ) = £/(d)ffQ) -
d\n

The arithmeti c functio n h defined b y (1 .1 3 ) (neve r min d th e convergenc e o f gener -

ating series ) i s called th e Dirichle t (o r multiplicative) convolutio n o f / an d g, an d
it i s denoted b y f *g.
The se t of all arithmetic function s wit h th e usual additio n + an d the operatio n
• i s a commutative ring . Th e function 5 : N -> C whose Dirichlet serie s is Ds(s) = 1
is the uni t elemen t o f this ring , i.e.,
J 1 i f n= 1 ,
(1.14) S(n)
~ \ 0 ifn>l .

The functio n ((s) define d i n R e (s) > 1 by (1 .1 0 ) i s calle d th e Rieman n zet a

function (afte r hi s semina l memoi r [Rie]) , it s Dirichle t coefficient s mak e th e con -
stant functio n l(n ) = 1 for al l n ^ 1 . B y virtu e o f (1 .1 2 ) th e invers e o f £(s ) ha s
also a Dirichle t serie s expansion , namel y

(1.15) ' -TT 1^ \_^M» s

0
C(s) 11V psj ^m '
P

say, wit h coefficient s

(1.16) n(m) = |
(-l)r if m = pi . .. pr wit h pi ,.. . , p r distinc t
0 otherwise.
The functio n fi(m) wa s introduced i n 1 83 2 by A. F. Mobius , an d it bear s hi s nam e
ever since . Moreover , b y th e Eule r produc t (1 .1 2 ) th e logarith m o f ((s) ha s th e
Dirichlet serie s expansio n

(i.i7) iogC0 0= £ £ r y •is

e=i P

Recall tha t th e Dirichle t convolutio n * correspond s t o th e multiplicatio n o f th e

generating Dirichle t series , therefor e th e identit y £(s ) • (~ 1 (s) = 1 reads a s
1i fm— 1 ,
(1.18) a(m ) = 5>(<9 = {
0 ifra>l .

Using thi s formul a on e obtains th e followin g

 1. A R I T H M E T I C F U N C T I O N 13

MOBIUS INVERSION . For any / , g : N — > C the following two relations are
equivalent:

(1-19) g(n) = J2f(d),

d\n

(1.20) /(» ) = 5>(d)sQ)-

d\n

REMARK. T O be accurate wit h th e history th e inversion formula s i n the above

form wer e state d onl y i n 1 85 7 by R . Dedekind . Th e origina l versio n state d b y
Mobius wa s somewhat different , i t amount s t o sayin g tha t fo r an y rea l variabl e
functions F , G : [1, x]— > C the following tw o relations ar e equivalent:

(1.21) G{x) = Y. F (xh),

n^x

(1.22) F(x) = ^2 v(rn)G(x/m).

m^x

The Mobiu s functio n i s multiplicative . I f / , g ar e multiplicative, the n s o are

/ • g and / • g. I f g is multiplicative, the n
d
(1.23) E ^ ) ^ d ) = ll(1-5(P))-
d\n p\n

This produc t admit s a probabilisti c interpretation . Viewin g g(p) as a probabilit y

of som e independen t event s whic h ma y occur a t p on e can think o f (1 .23 ) a s the
probability tha t non e o f the events associate d wit h prim e divisor s o f n occur .

1.4. Examples .
Now w e give a larg e sampl e o f arithmeti c function s whic h on e encounters i n
analytic numbe r theory . Man y othe r function s wil l be introduced i n due course.
The diviso r functio n r(n) i s the number o f positive divisor s o f n, so we have
oo
y
(1.24) e(s) = £r(n)n-°.
1

More generall y Tk(n) denote s th e number o f representations o f n a s the product of

k natura l numbers , s o its Dirichlet serie s i s ( k(s). Explicitl y

1
ai
(1-25) ^)=( t-l )--(art-ll) tt'-tf-rf'-
For an y v £ C we define o v{n) b y
oo
s
(1.26) C ( s ) C ( s - ^ ) = 5 > » n - ,
1

so it is the sum of powers o f divisors

(1.27) <7,(n )= ^ < r.
d\n
14 1. ARITHMETI C FUNCTION S

We hav e th e Ramanuja n formul a

oo

(1.28) t(s)C(s - a)C(s - P)t(s - a - P)C 1 (2s _ a - /3 ) = ] T ^(n)ap(n)n'\

l

In particular ,
oo
T n 2n S
(1.29) (\s)C\2s) = J2 () ~-
1

One ca n recognize th e abov e Ramanuja n serie s a s a special cas e of the Rankin -

Selberg convolutio n L- functions attache d t o modula r form s (se e Sectio n 5.1 ) . Di -
viding (1 .29 ) b y £(s ) on e get s
oo
13
(1.30) C (s)C (2s)^J£T(n>~^
1

hence multiplyin g b y ((2s)

oo

<3(s) = £ ( £ T{m>))n-°.
2
1d m=n

This i s th e symmetri c squar e L-functio n associate d wit h ( 2(s). Dividin g agai n b y

£(s) on e get s
oo
12 - 2Uj(n)n S
(1.31) C (s)C (2s) = J2 ~
l

where u(ri) denote s th e numbe r o f distinc t prim e factor s o f n. Nex t w e ge t

oo
n n S
(1.32) ((s)CH2s) = J2\^ )\ ~ -
1

Note tha t |/i(n) | = ^ 2(n) i s th e characteristi c functio n o f squarefre e numbers . B y

(1.32) w e hav e

2
(1.33) „ (n) = £>(<*) •
d2\n

Inverting (1 .32 ) w e obtain anothe r Dirichle t series ,

oo

C\s)C('2s)='£Hn)n-s
1

with coefficient s

(1.34) A(n ) = (-l) n < n >

where fi(n ) denote s th e tota l numbe r o f prime factor s o f n (counte d wit h multiplic -
ity). Thu s A(n) , which i s called th e Liouvill e function , i s completel y multiplicativ e
 1. A R I T H M E T I C F U N C T I O N 15

and i t agrees wit h ji(n) on squarefree numbers . Th e Euler functio n <p(n) i s defined
by the series

(1-35) fcl) =gv,(n)n-.,

so it is also give n by

(i.36) * > ( » ) = » n ( i - £ ) = » E ^ -
p\n d\n

One ma y think o f ip(ri)/n a s th e probabilit y tha t a randoml y chose n intege r m

is coprim e wit h n. I n other words , (p(n) represent s th e number o f residue classe s
(a mo d n) wit h (a,n) — 1 .
Very importan t arithmeti c function s emerg e b y differentiation. W e begin wit h
oo

-C'(«) = ^ ( l o g n ) n - s ,
1

which give s us the logarithm functio n

(1.37) L ( n ) = lo g n.
Since L(n) i s additive, i t follow s tha t L • (/ * g) = (L • / ) * g -h / * (L • g). Thi s
says tha t th e multiplication b y L is a derivation i n the Dirichlet rin g o f arithmeti c
functions.
Next th e function A(n ) ca n be defined a s coefficients i n the Dirichlet serie s for
the logarithmi c derivativ e — (log£(s)y = — C'(s)£(s)_1, so

(1.38) - ^ ( S ) = ^A(n)n" s.
1^

By th e Euler produc t (1 .1 2 ) on e computes tha t

/ 1 o n , K( v f log p i f n = pa, a ^ 1 ,
(1-39) A(n ) =< ^
(^ 0 otherwise .
Thus A(n ) i s supported o n prime powers . On e can read (1 .38 ) a s A = / i * L . Mor e
explicitly,

(1.40) A(n ) = J2 ^ lo ^=-T, V{d) log d.

d\n d\n

By Mobiu s inversio n w e obtain L — 1 • A , or equivalently,

(1.41) log n = 5^A(d) .

The functio n A(n ) wa s know n an d effectively use d b y P. Tchebyshev, nevertheles s

it i s called th e von Mangoldt function . Her e i s another identit y whic h i s useful i n
elementary prim e numbe r theory :
(1.42) A - = / X * ( L - T ) .
16 1. A R I T H M E T I C FUNCTION S

The vo n Mangold t functio n A ha s bee n generalize d t o th e highe r vo n Mangold t

function o f an y degre e k ^ 0 by A k = \i * Lk, i.e. ,

(1.43) A fc(n) = ][>(<* ) (log ^

d\n

(notice tha t A o = 6). Thi s satisfie s th e recurrenc e

(1.44) Afc+ i = LA k + A • A k
from whic h i t follow s b y inductio n tha t A k(n) i s supporte d o n number s havin g a t
most k distinc t prim e factors , i.e. , A k(n) = 0 i f uo{n) > k. Moreover , w e hav e
k
(1.45) 0^A k(n)^(logn) ,
the lowe r boun d followin g b y inductio n fro m (1 .44 ) an d th e uppe r boun d followin g
from th e formul a L k = 1 * A k, o r explicitly ,
fc
(1.46) ( l o g n ) =^Afe(d),
d\n

which come s fro m (1 .43 ) b y Mobiu s inversion . O f course , A ^ i s no t multiplicative ,

but w e hav e th e followin g hand y formul a (se e (1 .44) )

Ak(mn) = 5 Z ( * ) ^i(m )^-i( n )' i f

( m > n ) = 1-

EXERCISE 1 . Fo r an y intege r k ^ 0 an d a rea l numbe r x > 0 , defin e

(1.47) A fc(n,aO = $>(d)(log^)* .

d\n

Note tha t A k(n,x) depend s onl y o n th e squarefre e kerne l o f n. Prov e tha t

k
'k\ A A ,_ nN )/ , l_0x\g ~i
Ak(n,x) = J2 ( ! ) ^ ( D
Then, usin g A . < LJ~ lAk, deriv e tha t fo r n ^ x

(1.48) A fc(n, x) f ^ J < Afc(n ) < A fc(n, x).

Therefore A k(n,x) i s supporte d o n integer s n havin g a t mos t k distinc t prim e

divisors, an d i t satisfie s 0 < A k(n,x) ^ (logx) fe i f n ^ x .
Next sho w tha t i f (ra , n) = 1 , the n

Ak(mn,x) = ^2 [ •) Aj(n)A k-j\jn, - J .

Hence deriv e tha t fo r n — p^1 . . . p% r wit h p i , . . . , P2 distinc t prime s

k r
(1.49) A k(n,x)^r\ ( ) (logx)*- (logpi)... (logp r ).
 1. A R I T H M E T I C F U N C T I O N 17

One ca n associat e a Mobiu s an d vo n Mangold t functio n t o an y multiplicativ e

function. I f Df(s) i s the generatin g Dirichle t serie s fo r / whic h ha s Eule r product ,
then /jbf and A / ar e define d a s coefficient s i n

— = £„,(„,)„,-,
(L50)
w. )^
Hence f*A.f = f-L an d A j = jjLf*(f-L). Clearl y A j i s supported o n prime powers .
If / i s completel y multiplicative , the n fjtf(n) — fJ>(n)f(ri) an d A/(n ) = A(n)/(n) .
There i s a variet y o f trul y interestin g arithmeti c functions . Th e theor y o f
modular form s i s a basi c sourc e fo r multiplicativ e functions . A s a simpl e mode l w e
pick up the functio n r(n) whic h is the number o f solutions t o a 2 + b 2 — n i n integer s
a, b. Th e generatin g Dirichle t serie s fo r r(n) i s equa l t o 4(#(s) , wher e

C*(s) = J2(Na)-°
a

is the zet a functio n o f th e imaginar y quadrati c field K = Q ( \ / ~ l ) whil e th e facto r

4 i s th e numbe r o f unit s i n K. Her e a run s ove r non-zer o integra l ideal s o f K. Al l
these ar e principa l ideal s generate d b y th e Gaussia n integer s a = a + bi G Z[>/—T] ,
a ^ O , an d i f a = (a) , the n Na = a 2 + b 2. Hence , indee d

4 s
Cx(s) = ^ r ( n ) n - .

Clearly \r{n) i s multiplicative . Th e prim e number s whic h ar e represente d a s th e

sum o f tw o square s ar e characterize d i n a beautifu l theore m o f Fermat . I t i s eas y
to se e tha t n o prim e p = — l(mod 4 ) ca n b e s o writte n an d Ferma t prove d tha t al l
other prime s ca n be (uniquel y u p to the sig n an d orde r o f a, b). I n modern languag e
this theorem o f Fermat i s just th e factorization la w in the Gaussia n domai n Z[\/—l] ;
it assert s tha t p — 2 is squar e o f a prim e ideal , p = l(mo d 4 ) split s int o a produc t
of tw o distinc t (comple x conjugate ) prim e ideals , p = — l(mod 4 ) remain s prime .
Using th e factorizatio n la w on e ca n sho w b y verificatio n o f loca l factor s tha t

where % 4 is th e non-trivia l characte r t o modulu s 4 (w e hav e X4( n) = si n ^) an d

L(s, X4 ) i s the associate d Dirichle t L- function

L{S,XA) = ^2x4(n)n~

Hence

(1.51) r(n) = 4^ X4 (d).

d\n
18 1. A R I T H M E T I C FUNCTION S

Besides th e Dirichle t serie s there i s great interes t i n studying th e Fourie r serie s

for r(n) du e t o th e additiv e natur e o f th e equatio n a 2 + b2 = n. W e hav e
oo

^2r(n)e(nz) = 0 2(z)
o
(recall (1 .9 ) tha t e(z) = e27rlz), wher e 0(z) i s the thet a functio n
oo
2
(1.52) 6{z) = ^e{n z).
— OO

This serie s converge s absolutel y fo r I m (z) > 0. Mor e generally , lettin g rk(n) b e
the numbe r o f representation s o f n a s th e su m o f k square s w e hav e
oo

J2rk(n)e(nz) = 0 k(z).
o

On one hand, th e theta function 0(z) i s recognized in analytic number theor y fo r

its abilit y t o selec t squares . O n th e othe r hand , 0{z) i s usabl e du e t o th e followin g
transformation rule :

(1.53) ef^±l\ =v( Cld)(cz + d)h(z)

\ (sZ I (J/ J

which hold s fo r an y z with I m (z) > 0 an d an y integer s a, 6, c, d wit h ad — be = 1,

c = 0 (mod 4) , wher e i/(c,d) depend s onl y o n c , d. W e hav e v 2(c,d) = x±{d), so
z/(c, d) takes onl y fou r value s ± 1 , ±z (se e (3.42 ) fo r th e exac t description) . I n othe r
words, 0(z) i s a modular for m o f weight \ wit h multiplie r v(c,d) whil e 0 2(z) i s a
modular for m o f weight on e an d characte r % 4 on r 0 (4) . I n additio n t o th e modula r
relations (1 .53 ) th e thet a functio n satisfie s th e followin g involutio n equation :

<-) *(S) - G)4»(§)-

Note tha t 0(z) doe s no t vanis h b y virtu e o f th e Jacob i produc t (1 .8) .
More generall y on e associate s a theta functio n wit h an y positiv e definit e qua -
dratic for m i n severa l variable s an d correspondin g harmoni c polynomial s (se e Sec -
tion 1 4.3) .
Yet othe r type s o f arithmetic function s ar e give n b y exponentia l sum s ove r
Z/raZ o r (Z/mZ)* . Th e mos t importan t o f thes e fo r analyti c numbe r theor y ar e
Gauss sums , Ramanuja n sum s an d Kloosterma n sums . W e mentio n her e onl y th e
quadratic Gaus s su m (fo r genera l Gaus s sums , se e Sectio n 3.4 )

G
<™»= E • ( £ ) •
a;(mod m)

They wer e evaluate d exactl y b y Gaus s a s a consequence o f th e Quadrati c Reci -

procity La w (se e Theore m 3.3) :
1 4- ?m
(1.55) G(m) = y -y V ^ -
 1. A R I T H M E T I C F U N C T I O N 19

The Kloosterman sums 5(a, b; c) encompass Ramanujan sums . The y are define d
by
/, -„• > „ ,, N v ^ * /ax + bx\
(1.56) S(a,b;c)= ^ e (— J
x (mo d c)

for integer s a, b and c ^ 1 . Th e symbo l ^ * restrict s th e su m t o x coprim e wit h c

and x denotes th e multiplicativ e invers e o f x modul o c , i.e . xx = 1 (mo d c) . Not e
that S(a, b] c) i s real, an d ha s th e followin g symmetries :

(1.57) S(a,6;c ) = S(6,a;c) ,

(1.58) S(aa' , 6; c) = S(a, 6a'; c) if (a', c) = 1 .

It als o ha s multiplicativ e propertie s (b y Chines e Remainde r Theorem )

(1.59) S(a, b; cd) = S(ac, be; d)S(ad, bd; c) if (c, d) - 1 .

Kloosterman sum s wil l mak e man y appearance s i n this book . The y ar e crucia l to
modern analyti c numbe r theor y throug h link s with spectra l theor y o f automorphi c
forms. Thoug h n o "simpler" formul a exist s (comparabl e t o (1.55) fo r quadratic
Gauss sums) , a sharp boun d fo r individua l Kloosterma n sum s is known, du e to A.
Weil:

(1.60) |5(a,6;c) | ^r(c)(a,6,c) 1 /2Vc.

We will prove thi s in Chapter 1 1 (se e Corollar y 1 1 .1 2 ) an d us e it many times .

When b = 0 (or n — 0 , se e (1 .57)) , th e Kloosterma n su m S(n, 0 ; m) i s called a
Ramanujan sum . I t is also sometimes denote d c^n). Fo r thes e ther e i s an explici t
formula, se e (3.2 ) an d (3.3) .
1.5. Arithmeti c function s o n average.
In th e cours e o f analytic numbe r theor y th e average s o f arithmeti c function s o f
various kind s appea r al l ove r th e place . Th e firs t tas k i s to establish estimate s fo r
finite sum s o f typ e

(1.61) M f(x) = "£ /(») •

In thi s sectio n w e sho w a few elementar y ideas .

The first commo n too l i s summation b y parts , whic h i s the formul a

Mfg{x) = Y. /( n M n ) = Mf(xM*) ~ f Mf(t)g'(t)dt

n^.x

and it s obviou s variants . Thi s formul a ofte n allow s u s to evaluate A4f g(x) i f Mf
is known an d g is smoot h an d doesn' t oscillate .
If / i s defined o n [1 , x], monotonic an d continuous , the n Mf(x) i s well approx -
imated b y a corresponding integral , precisel y

(1.62) M f(x) = j* f(y)dy + 0(\f(x)\ + |/(l)|).

20 1. ARITHMETI C FUNCTION S

For example , w e ge t
fe
(1.63) ^T(logn) = xP fe (logx) + 0((loga;) fc )

where Pk(X) i s a polynomia l o f degre e k

fe
Pk(X) = £ ( - l ) -^,

and
lo
(1.64) 5Z( § ~)k = *'* + 0((logx)*).

If /(y ) i s continuousl y decreasin g t o zero , on e ca n d o bette r tha n (1 .62) . Indeed ,

putting

(1.65) [x] = max{ ^ G Z : < ( x }

(the integra l par t o f x) an d

(1.66) {x} = x - [x]

(the fractiona l par t o f x) th e averag e o f / i s given b y th e Stieltje s integra l

Mf(x) = JX f{y)d[y] = [ f(y)dy - / * f(y)d{y}.

Hence

(1.67) M f(x) = J'f(y)dy + lf + 0(f(x))

where th e constan t j t i s given b y

OO /"O O

/
f(y)d{y} = /(l) + J {y}df(y).
For example , w e ge t

(»•«) E ^ - ^T<'»^)' +1 + <-U*»» + ° ( ! = £)

where 7 & is called th e Stieltje s constant . Fo r k = 0 we get th e Eule r constan t
OO
-2,
(1.69) 7 = l - / {y}y-Ms/"zdy = 0.57721 5 . . .
/

Similarly on e ca n evaluat e average s o f smooth function s i n severa l variable s b y

repeated application s o f th e integra l approximatio n method . I n thi s wa y on e ca n
derive Gaus s formul a fo r th e lattic e point s i n a circl e

(1.70) ] T r(n) = 7TX + 0(yfx).

 1. ARITHMETI C FUNCTIO N 21

Actually, Gauss ' argumen t wa s geometrical i n nature, h e derive d (1 .70 ) b y packin g

the circl e wit h a unit squar e (whic h amount s to the sam e thin g a s the integration) .
In th e sam e wa y on e obtain s
rfc n
(1.71) Yl ( ) = ^+ OO^)

where pk = 7r / C / / 2 /r(| + 1 ) is the volum e of the ^-dimensiona l uni t ball .

Most interestin g arithmeti c function s / ar e no t define d o n all rea l number s in
which cas e it makes n o sens e to approximate Mf(x) b y th e integra l (1 .67) . Yet , if
/ i s given by a convolution o f two functions on e of which i s smooth o n real number s
and th e othe r vanishe s rapidl y o n natural numbers , i t is still possibl e t o evaluate
Mf(x) b y mean s o f integrals an d convergen t series . Specifically , i f / = g*h, wher e
h is monotonic an d bounded , w e arrang e M.f(x) a s

Mf(x)=^29(m)Mh(~),
m^x
py
and, afte r replacin g Mhiy) b y / h(t)dt + 0(1), w e ge t
Jo

m^.x m^.x

Estimating th e partia l su m ove r y < m ^ x triviall y w e ge t

(1.72) M f{x) = [ X F(y)dy + 0(M ]gl(x))

Jo
where

(-3) ™ _ £*£>*(£).
If w e assume tha t th e serie s Y^Qi 1 71 )1 71 '1 converge s absolutely , the n thi s formul a is
a sligh t modificatio n o f a theorem o f A. Wintner . A s a n example w e appl y (1 .72 )
for f(m) — (/?(n)n -1 whic h i s given b y (1 .36 ) gettin g

(••74 ) £»fi!UJL + <,<*„.

n^x

For th e divisor functio n r(n) th e sum M T{x) ha s geometric interpretation ,

namely it represents the number o f lattice points (m , n) G NxN unde r th e hyperbol a
ran ^ x. I n this cas e ou r genera l formul a (1 .72 ) yield s

n^.x n^x

showing tha t th e averag e valu e o f r(n) is log n. I n order t o improve th e erro r ter m
in th e diviso r proble m Dirichle t applie d a simple, ye t powerfu l tric k wit h switchin g
divisors. Firs t b y symmetry o f the equatio n m\rri2 = n we figure tha t

MT(x) = 2 £ g ] - [ ^ ] 2 .
m^y/x
22 1. ARITHMETI C FUNCTION S

Here w e hav e fewe r term s o f summatio n tha n i n th e straightforwar d arrangement .

This tric k i s called th e hyperbol a method . No w dropping th e integra l par t symbol s
we mak e a n erro r 0(y/x), an d i t follow s b y (1 .69 ) tha t

T
(1.75) Yl (n) = x lo
§ x + ( 2 ^ ~ l )x + 0{y/x).

EXERCISE 2 . Prov e b y Dirichlet' s metho d tha t

l
] T r k(n) = xP k(logx) + 0{x ~^)
n^.x

where P k i s a polynomia l o f degre e fc — 1 (Pk(x) i s th e residu e o f ((s) kxss~1 a t

8 = 1 ).

Dirichlet's hyperbol a metho d work s nicel y fo r th e lattic e point s i n a bal l o f

dimension fc ^ 4 . Lagrang e prove d tha t ever y natura l numbe r ca n b e represente d
as the su m o f four squares , i.e. , r±{n) > 0, and Jacob i establishe d th e exac t formul a
for th e numbe r o f representation s

r 4 (n) = 8( 2 + ( - l ) » ) £ d.
d\n,d od d

Hence w e deriv e

^r 4 (n) = 8 ^ (2 + ( - I D£ d
n^.x m^.x dm^.x,d od d
^2
« I > + <-ir>(^ + < >)
m^.x
oo
= 2x 2 ] T ( 2 + ( - l ) m ) m " 2 + 0 ( x logs) .
l

= 3((2)x 2 + 0(x\ogx) = -(irx) 2 + 0(x\ogx).

This resul t extend s easil y fo r an y fc ^ 4 (writ e r ^ a s th e additiv e convolutio n o f

r4 an d r&_4 , appl y th e abov e resul t fo r r^ an d execut e th e summatio n ove r th e
remaining fc — 4 squares b y integration )

(1.76) ] T r fc (n) = ^ 0-+ O f r * " 1 logs) .

Notice tha t thi s improve s th e formul a (1 .71 ) whic h wa s obtaine d b y th e metho d o f

packing wit h a uni t square . Th e exponen t | — 1 in (1 .76 ) i s best possibl e becaus e
the individua l term s o f summatio n ca n b e a s larg e a s th e erro r ter m (apar t fro m
logrr), indeed fo r fc = 4 we have r±(n) ^ 1 6 n if n i s odd b y th e Jacob i formula . Th e
only case s o f th e lattic e poin t proble m fo r a bal l whic h ar e no t ye t solve d (i.e. , th e
best possibl e erro r term s ar e no t ye t established ) ar e fo r th e circl e (f c = 2 ) an d th e
sphere (f c = 3) . W e shal l addres s thes e fascinatin g problem s o n othe r occasions .
 1. ARITHMETI C FUNCTIO N 23

EXERCISE 3 . Prov e b y th e hyperbol a metho d tha t

Y^ r(n 2 + 1) = - z l o gx + 0(x).
n^x

1.6. Sum s o f multiplicativ e functions .

Throughout / i s a multiplicative function . Sinc e / i s determine d b y it s value s
at prim e power s it is possible t o estimat e Mf(x) i n term s o f th e loca l sum s
oo

(1-77) * P (/) = £/&>•>"" •

First w e give simple uppe r bound s whe n / i s non-negative. Th e followin g estimate s

need n o explanatio n

(1.78) M f(x) ^xJ2 fWn- 1 ^x]J a p(f).

One ca n d o bette r i f /i s non-decreasing o n prim e powers . I n thi s cas e h = \x * / i s

non-negative, indee d h(p u) = f(jp v) — / ( p ^ - 1 ) ^ 0 if v > 1 . Writin g / = 1 • h w e
get

(1.79) M f(x) = Y, h(m) [—1 < x J2 fe M™1_

m^x m^x

P^X P^z%

Here i s a crude bu t usefu l applicatio n o f (1.79) fo r f(m) = Tfc(ra)^ . W e hav e

f(p) = k l an d

where c = c(k, £) is a positive constant . W e hav e als o th e elementar y boun d

n(i+^)«io g x.
p^x

for an y x ^ 2 (se e (2.1 5)) . Henc e (1 .79 ) yield s

1
(1.80) Y^ninY^xilogxf-
n^x

where th e implie d constan t depend s o n k,£. Thi s i s a crude bound , bu t o f the

correct orde r o f magnitude . Notic e tha t i t implies tha t
£
(1.81) r k(n)^n

for an y e > 0 , th e implie d constan t dependin g o n e and k. We shal l b e usin g thes e

bounds fo r th e diviso r functio n r^(n ) ofte n withou t mention .
24 1. ARITHMETI C FUNCTION S

If on e take s th e averag e of f(n) ove r squar e fre e numbers , o r what amount s to

the sam e thin g on e assume s / i s supported o n squarefree numbers , the n th e abov e
arguments yiel d

(1.82) M f{x) ^ n l 1 /(P)-I^

P'
p^x
provided f(p) > 1 for all p. W e ma y requir e f(p) ^ 1 to hold tru e fo r primes onl y
in a certain set . The n th e resul t is

(i.83) *,<*>< * n (»+*?) n ( ' + ^ ) -

p^x P^x
/(P)<1 /(P)> 1

Here i s another elementar y approach . W e onl y assum e tha t / i s non-negative

and completel y sub-multiplicative , i.e. , f(mri) ^ / ( r a ) / ( n ) fo r all ra,n> 1 . More-
over, suppos e tha t f(p) i s bounded o n average,

(1.84) ^2 f(m)A(m) < ex

m^x

for an y x ^ 2 , where c ^ 1 is a constant (se e the Tchebyshe v boun d (2.1 2)) . The n
we ge t
^ / ( n ) l o g n = ^2 f{nm)A(m) ^ e x ^ / ( n ) n - 1 .
n^a: nm^ix n^x
Hence by partial summatio n
3CX y^M f(n) ^ 3C(JX yr / , /(p ) ^
M
(1-85) A*,(* )^£ ^n <^ lor^
g x IAI1'(! + )
^ lo gx ^
n^n
x P ^ lo g x - 1 •^x \ p /
where
-f(ny

To illustrat e thi s resul t w e tak e th e multiplicative functio n f(n) = 6(n ) whic h

is the characteristi c functio n o f numbers representabl e a s a sum of two squares. We
have f(p) = 0 if p = — 1 (mod 4 ) and f(p) = 1 otherwise. I n this cas e (se e (2.30) )

n(i+ M )«(io g ^
p^x

for an y x ^ 2 . Hence on e deduce s b y (1.85) tha t

(1.86) B(x) = \{n^x :n = a2 + b2}\ < C x ( l o g x ) " i

Here ever y numbe r whic h admit s a representatio n a s the sum of two square s i s
counted once . Fo r thi s reaso n th e upper boun d (1 .86 ) i s slightly smalle r tha n the
area of the circl e a 2 -ffr 2 ^ x (se e (1 .70)) . Usin g analyti c method s E . Landa u [Lai ]
established th e asymptotic formul a

(1.87) B(x) = Y^ b (n) ~ CxQogx)-* ,

 1. ARITHMETI C FUNCTIO N 25

as x —> co , wher e C is a positiv e constant . Landau' s formul a als o follow s b y a n

elementary metho d wit h C given b y (1 .1 02) .
Very differently , stil l elementarily , th e formul a (1 .87 ) i s proved i n [18] by ap-
plying a half-dimensional sieve . Th e siev e metho d ha s th e advantag e o f producing
results o f great uniformity . Fo r example , on e get s
B{x + y)-B{x)<(l + e)B(y)
for an y e > 0 provided y i s sufficiently larg e i n terms o f e only , wher e x i s an y
positive number . I f x is very larg e compared t o ?/ , then analyti c argument s produc e
rather poo r result s fo r B(x -\-y) — B{x).
Next w e present a n elementar y metho d whic h yield s an asymptoti c formul a fo r
Mf(x) fo r a large clas s o f multiplicative functions . Thi s require s som e regularit y
in th e distributio n o f / a t prime powers . Lettin g A / b e th e vo n Mangold t functio n
associated wit h / (se e (1 .50) ) w e assum e tha t

(1.88) ] T A/(n) = /dog x + O(l )

n^x

where K is a constant. Thi s conditio n mean s f(p) i s about np~ x o n average wherea s
it wa s K previously . Fo r many multiplicativ e function s i n practice (1 .88 ) ca n be
established b y elementary method s (thi s conditio n i s an analogue o f the Merten s
formula (2.1 4) , an d it is weaker tha n th e Prim e Numbe r Theorem) . Wit h som e
applications i n mind w e allo w f(p) t o be slightly negative . Precisel y th e metho d
works if (1.88) hold s wit h K, > — \. I n addition t o this w e nee d a crude estimat e

(1.89) £ | / ( n ) | « ( l o ga; )K
n^x

The metho d begin s wit h evaluatio n o f the logarithmicall y smoothe d su m

(1.90) Y, fW l0 § - = f M f{y)y-ldy.
n^x

Since / • L — f * Ay w e derive b y (1 .88 ) tha t

X ) / ( n ) l o g n = £ / ( d) £ A/(m )
n^.x d^.x m^.x/d

= £/(d)(«log^+0(l)).
d^.x

Estimating th e su m o f error term s b y mean s o f (1.89) w e ge t

lo
(K + 1 ) Y, f( n) § n = *>Mf{x) logz + 0 ( ( l o g x ) N ) .
n^x

Inserting thi s int o (1 .90 ) w e show tha t th e differenc e

1
(1.91) A(x) = M f(x) log x - (K + 1 ) / M f(y)y~ dy

is relativel y small , namel y

(1.92) A ( z ) < ( l o g x ) K
26 . ARITHMETI C FUNCTION S

Next w e divide (1 .91 ) b y x(\ogx)~ K~2 an d integrat e gettin g

[X A(y)y-\logy)- K'2dy = f M f{y)y-\\ogy)-^dy

J2 J2

- ( « + 1) J* y-^logy)-*- 2^ Mf^u-'duyy.
Changing th e orde r o f integratio n i n th e multipl e integra l w e ge t

J' M f(y)y-^hgy)-K1- dy + J 1
X
[(logx)-K1 1
- y-(logu)- K
- ]Mf(u)u- du.
The first integra l cancel s ou t wit h th e secon d par t o f th e las t one , s o w e ar e lef t
with
JX A(y)y- 1 (\ogy)-K-2dy=(\ogx)-^-1 J X
Mfi^u-'du.

Combining thi s wit h (1 .91 ) w e arriv e a t th e followin g identity :

Mf(x) = (logs) * jX -AiyMlogy)-"- 1 + A^Xlogx)" 1 .

Since th e abov e integra l converge s absolutely , w e ca n exten d it s rang e t o infinit y

getting th e identit y
K
(1.93) M f{x) = {c f+rf(x)}(logx)

where cj i s a constan t give n b y th e definit e integra l

/•OO

A
(1.94) Cf =
~J ( ^ ( l o g 2 / ) ~1
"-

and rf(x) i s a functio n give n b y

/•OO

(1.95) r f(x) = / (A(y ) - A(x))d(logy) — \

This tend s t o zer o a s x —> oo , indee d b y (1 .92 )

w K 1
(1.96) r / ( a : ) < (logx) - " .
However, th e constan t (1 .94 ) doe s no t loo k natural .
Now havin g establishe d th e asymptoti c formul a (1 .93 ) wit h th e erro r ter m
(1.96) w e shall us e these result s t o deriv e anothe r expressio n fo r cj b y a n appea l t o
the zet a functio n fo r / ,
oo

1
The series converges absolutely fo r s > 0 by virtue of (1 .89) . W e compute b y partia l
summation tha t
oo />o o

/
y-sdMf(y) = -J M f(y)dy-
S
/•OO pOO
£
= - / Mfie^de-** = - {c f + 0(t- )}tKde~st
Jo Jo
= {c f + 0(s £)}s-KT(K + l)
 1. ARITHMETI C FUNCTIO N 27

as s — » 0+. Comparin g thi s wit h th e ^-t h powe r o f th e Rieman n zet a functio n w e

get
C(8 + l)-"D f{8)~Cfr{K+l).
On th e othe r hand , w e have th e Eule r produc t

as+iyKDf(S) = rj(i - p-'-'r ( £ fW)p- va)

p \i/= 0/

which converge s absolutel y fo r s > 0 and i t ha s limi t a s s — > 0 by virtu e o f (1 .88) .

Hence th e constan t c / i s equal t o

(1^) C /=^nH)Wb) + /(P 2 )+---).

We hav e establishe d th e followin g formul a (i f tt ^ 0 , thi s i s du e t o E . Wirsin g
[Wil]).
THEOREM 1 .1 . Suppose f is a multiplicative function which satisfies (1 .88)
and (1 .89) with n > — \. Then
1
(1.98) ] T f(n) = C/(log*r + Odlogx^- )
n^.x

where Cf is a constant given by (1 .97).

We illustrate th e applicabilit y o f this formula b y solving a simple sieve problem .
Let P b e a se t o f primes. Ou r questio n i s how man y number s n ^ x hav e n o prim e
divisors i n VI Le t / b e th e completel y multiplicativ e functio n wit h

/ ( p ) = f i if peP,
\ 0 otherwise .
Then
9(n) = U(l-f(p))
p\n
is th e characteristi c functio n o f number s i n question . Therefor e w e ar e askin g t o
evaluate
S(V,x) = ^g(n).

Opening th e convolutio n g = fif * 1 we obtai n

S(V,x) = Y,Kd)f(d)[^}=xY,Kd)^-+0(M f(x)).

d^x d^x

To procee d furthe r w e ar e goin g t o assum e tha t th e proportio n o f prime s i n V i s

less tha n th e proportio n o f thos e no t i n V. Mor e precisel y w e nee d th e conditio n

(1.99) V J f(p)
-^- log p - Adog x + O(l )
*-? P
p^x
with a constan t K < \. The n w e derive Mf(x) < C x(\ogx)K~1 b y (1 .85 ) an d
1
J2n(d)^-=c(V)(\ogxr" + 0((\ogxy- )
d^.x
28 1. ARITHMETI C FUNCTION S

by (1 .98) , wher e c(V) i s a constant dependin g o n the set V,

Gathering th e above result s w e conclude

COROLLARY 1 .2 . Let V be a set of primes of density K < \ (that is (1 .99)
holds). Then the number of integers 1 ^ n < x having no prime divisors in V
satisfies
(1.101) S(V,x) = c(V)x(logx)-K{l + 0((logx) 2 "" 1 )}
where c(V) is the constant given by (1 .1 00).

EXERCISE 4 . Appl y (1 .98 ) t o derive Landau's formul a (1 .87 ) wit h the constant

v
p=-l(mo d4 ) ^

1.7. Distributio n o f additiv e functions .

Throughout / i s an additive function , s o / i s determined b y its values at prim e
powers, precisel y
f(n) = ]T /(p").
pa\\n
Hence th e average o f / i s given by
X X
Mf(x) = £ f(P a) ( - +1
pa-^.x i^ _
Dropping th e integral par t symbol s w e arrive at

(1.103) M f(x) = xE(x) + 0 ( x b ( x ) ) ,

where
1
a a
(1.104) E(x)=Y,f(p )p- (l-p- )
a
p ^.x

and
a 2
(1.105) D\x) = £ \f( P )\ p~a.
a
p ^x

Actually th e error ter m i n (1 .1 03 ) i s ^ |/(p Q )| whic h w e estimated b y x*D(x) b y

applying Cauchy' s inequalit y an d tha t p a ^ x. Th e resul t reveal s tha t E(x) i s
approximately equa l t o the mean valu e o f /.
Similarly w e evaluate th e average o f f2. W e have

pa q$ n^x
 1. ARITHMETI C FUNCTIO N 29

where p, q run ove r primes . I f p ^ q, the inne r su m i s equal t o

"X 1 X X X
a
.p qPl
-
.pa+1 3
qf .
-
.p^q^1.
+ pa+lq(3+l pcxqt • H ) ( ' - ; ) + <*>•
and fo r p = q we mus t hav e a = / ? gettin g

,CK+1
FHW
Hence w e derive tha t

M2;(x) = x £ J2 f(P a)f(lP)P~aQ~0^ ~ P_1)(l " T 1 )

+ xp£^ r r/
a
2
( P V ( I - P _1 ) + O (E
a
El/(P a)/(^)l)-
P ?<? ^ ^ ^

Here w e dro p th e conditio n p ^ q an d estimat e th e additiona l term s wit h p — q

using \f{p a)f{pP)\ ^ \f(p«)\ 2 + \f{jf)\ 2. W e arriv e a t
(1.106) MP(x) = xE 2(x) + 0(xD 2
(x)).
This resul t say s tha t E (x) i s approximatel y equa l t o th e mea n valu e o f f 2.
2

Both approximation s (1 .1 03 ) an d (1 .1 06 ) sugges t tha t th e individua l value s

f(n) fo r n ^ x shoul d b e reasonabl y wel l approximate d b y E(x). Indee d thi s i s
true i n th e followin g sense :

THEOREM 1 .3 . For any additive function f : N—» C we have

(1.107) Yl l/( n) - E ^\2 ^ CXD2

^
n^x
where c is an absolute constant.

P R O O F . W e ca n assum e tha t / i s rea l sinc e th e rea l an d th e imaginar y part s

of / ar e additiv e an d the y ca n b e treate d separately . W e can als o assum e tha t x i s
a positiv e integer . Firs t w e evaluate th e varianc e

V{x) = E ( / ( n ) " x- xMf[x))2 = M P(x) - x^Mfe).

n^x

By (1 .1 03 ) an d (1 .1 06 ) w e ge t
2
V{x) < x*\E{x)\D{x) + xD 2(x) « x D (s).
l
Then replacin g x~ Mf(x) b y E(x) w e mak e a n admissibl e error , an d (1 .1 07 ) i s
established.
•
Theorem 1 . 3 i s du e t o P . Tura n [Till ] an d J . Kubiliu s [Kubl] . Th e expecte d
value E(x), i f on e prefers , ca n b e reduce d t o

(1.108) A(x) = J2f^P~ 1

p^x
30 1. ARITHMETI C FUNCTION S

at th e expens e o f th e constan t c i n (1 .1 07) . Ou r argument s abov e yiel d c = 4 i f

x i s sufficientl y large , however , J . Kubiliu s [Kub2 ] showe d (afte r havin g receive d
suggestions from H.L . Montgomery) tha t (1 .1 07 ) holds with c = c(x) ~ | a s x— > oo.
He and other s als o showe d tha t th e | canno t b e replace d b y a smaller constan t (cf .
A. Hildebran d [Hil]) .
As a n exampl e w e choose th e additiv e functio n cj(n) whic h count s th e numbe r
of distinc t prim e divisor s o f n. Therefor e uu(p a) = 1 , E{x) = lo g log x + O(l ) an d
D2(x) — log log x - f 0(1 ) b y (2.1 5) , s o th e Turan-Kubiliu s theore m give s
(1.109) Y^(u;(ri ) — log log x)2 < C x log log x.

Hence i t follow s tha t u(n) ~ lo g log n fo r almos t al l n ( a theore m o f Hard y an d

Ramanujan). Precisel y w e hav e
COROLLARY 1 .4 . For any x ^ 3 and z ^ l ,
2
(1.110) \{n ^ x : \u;(n) - l og lo g a; | > z(loglogx)^} | < z~ x
where the implied constant is absolute.

Furthermore on e ca n as k i f a give n additiv e functio n / ca n b e renormalize d b y

A(x) an d som e B(x) s o that th e frequencie s

have a limi t fo r an y z G R a s x tend s t o oo. Thi s questio n i s addresse d i n th e

probabilistic theor y o f numbers . W e shal l onl y giv e a glimps e o f centra l results .
The ver y firs t on e wa s establishe d b y P . Erdo s an d M . Ka c [EK] .
THEOREM 1 .5 . Let f(n) be an additive function with — 1 ^ f(p a) — f(p) ^ 1
such that
1
(1.112) B{x) = Y,f\p)P~ ^™

as x —> > oo. Then v(x\z) ~ G(z) for any z G R , where G(z) is the Gaussian
distribution function

(1.118) G ^'WJ-, -<'"dt.

In particular , i t follow s fro m th e Erdos-Ka c theore m tha t fo r an y rea l z

\{n ^ x : uj(n) — log log x < z(loglog#)2} | ~ G(z)x
as x tend s t o oo . Man y othe r arithmeti c function s (no t necessaril y additive ) follo w
the Gaussia n limitin g distributio n law . Fo r example , H . Halbersta m [Hal ] prove d
that th e numbe r o f prim e divisor s o f shifte d prime s satisf y
(1.114) \{p^ x :u(p+l) - l o g lo g x < 2:(loglogx)i} | ~ G(Z)TT(X)

for an y fixed z G R a s x —> oc .

 http://dx.doi.org/10.1090/coll/053/03

CHAPTER 2

ELEMENTARY THEOR Y
OF PRIM E NUMBER S

2.1. Th e Prim e Numbe r Theorem .

It wa s observe d nearl y 20 0 year s ag o b y Legendr e an d Gaus s (independently )
that th e densit y o f primes p ^ x i s (logo:) -1 , precisel y the y postulate d th e followin g
P R I M E NUMBE R T H E O R E M . Let ir(x) denote the number of primes p < x. As
x —> oc ; we have
x
(2.1) ir(x)
log x

Gauss observe d tha t a n eve n bette r approximatio n t o TT(X) i s give n b y th e

singular integra l

(2.2) Li(x) = f^ — r(l--) f ^ + l o g l o g x+ 7 .

Jo log y J ± V yJ logy
For x > 1 this satisfie s th e asymptoti c expansio n

f m
(2.3) Li(a O= j ^ - | ] T &(\ogx)- + 0((logxr )\-

Some people use £i(x) — Li(x ) —Li(2) in place of Li(x) fo r x ^ 2 . Later , Tchebyshe v
realized i t i s simpler t o coun t prime s p wit h th e weigh t logp , s o he investigate d th e
sum

p^x
rather tha n n(x). I t i s stil l mor e convenien t t o evaluat e th e averag e o f th e vo n
Mangoldt functio n
(2.4) ip{x) = ] T A(n) .
n^x

Note tha t (2.1 ) i s equivalen t t o eac h o f th e asymptoti c formula s

(2.5) 9{x) ~ x , ip(x) ~ x
by partial summatio n an d trivia l estimatio n o f the contributio n o f n = p a ^ x wit h
a^ 2.
Although A is give n b y th e convolutio n (1 .40 ) th e formul a (1 .72 ) fall s shor t o f
yielding (2.5 ) becaus e th e erro r ter m exceed s th e mai n term . Applyin g Dirichlet' s
trick o f switchin g divisor s (th e hyperbol a method ) on e ca n reduc e th e numbe r o f
31
32 2. ELEMENTAR Y THEOR Y O F PRIM E NUMBER S

error terms , bu t i t doe s not complet e th e work . Thi s onl y lead s t o anothe r proble m
of estimating sum s o f the Mobiu s function . Indeed , on e ca n sho w b y th e hyperbol a
method tha t th e Prim e Numbe r Theore m i s equivalen t t o th e estimat e
m =
(2.6) M(x) = 2\. M ) °{ x)-> a s x ^ oo .

To deriv e (2.5 ) fro m (2.6 ) w e conside r

A(x) = ^ ( l o g r c - r(n) + 27) .

n^x

Subtracting (1 .75 ) fro m (1 .63 ) fo r k = 1 we ge t

(2.7) A(x) < y/x.

Now applyin g (1 .42 ) w e ge t

il>(x) - x + 2 7 = Y, Y, v(d)(logk- r{k) + 2 7)

dk^.x

(2.8) = ^ ( l o g f c _ r ( f c ) + 2 7 ) M ( | ) + J2 l^(d)(A(~)-A(K))
1
k^K d^xK-

for an y 1 ^ K ^ x. Th e last su m is 0(xK~ 1 ^2) b y virtue of (2.7 ) while for an y fixed

K th e precedin g su m i s o(x) b y th e hypothesi s (2.6) . Sinc e K ca n b e arbitraril y
large, thi s show s tha t (2.6 ) implie s (2.5) . T o establis h th e convers e w e us e th e
identity
V^ /i(n ) log — = M(x) lo g x - f Y ^ Y ^ fjb(m)A(n)
n^x mn^x
(1 \ ' C 1
which follow s fro m f - J = — — -. Th e lef t sid e i s triviall y 0(x) (se e (1 .64 ) fo r
k = 1 ) , so
M(x)logx = - ^ /i(ra)V>(— J +0(x).
Applying (2.1 8 ) w e arriv e a t

M(x)logx = Y, M H ( ^ -V>(^) ) +0(X).

m^.x

Now usin g th e hypothesi s (2.5 ) thi s equatio n implie s (2.6) .

2.2. Tchebyshe v method .
The firs t significan t attemp t fo r provin g th e Prim e Numbe r Theore m wa s mad e
by Tchebyshe v i n 1 848-1 852 . Retrospectivel y on e ma y sa y tha t hi s approac h i s
based essentiall y o n th e produc t formul a

N = l[p
P

meaning tha t th e archimedea n valuatio n equal s th e produc t o f p-adi c valuations .

This obviou s formula i s particularly fruitfu l fo r th e factoria l number s J V = 1 • 2 • • • n.
 2. E L E M E N T A R Y T H E O R Y O F P R I M E N U M B E R S 33

Like Tchebyshev w e present th e arguments i n a more friendly analyti c format. Firs t

we hav e
(2.9) L(x) = 2_] log^ = xlog x — x + O(logx) .
n^.x

On th e othe r hand , usin g (1 .41 ) w e ge t

(2.io) ^ ) = EA wS = E ^ ©-
d^x m^.x

Hence evaluatin g L(x) — 2 L ( | ) i n tw o way s on e get s

(2.11) ^ ) - v { f ) + ^ ( | ) - ^ ( | ) + - - -= zlog 2 + 0(logz) .

Using th e monotonicit y o f ip(y) on e derive s fro m (2.1 1 ) th e firs t Tchebyshe v esti -

mates
(2.12) xlog 2 + <3(logx ) < \l>(x) < xlog4 + <3(logx) .

Tchebyshev improve d thes e estimate s severa l time s usin g mor e involve d linea r
combinations o f L ( ^ ) . Fo r exampl e le t u s conside r ij)f(x) = L(x ) — L ( |) — L ( |) —
L(f) + L ( § ) . B y (2.9 )
ipf(x) = ax + O (log x)
where a = \ lo g 2 + § log 3 + \ lo g 5 - ^ lo g 30 = 0.921 2.. . O n th e othe r hand ,

</>/(*) = £A(n)/Q
n^x

where
x~ X1 "X X
ran ra n ra n
L2 J L3 - L5 30 J
+
Since 1 — ^ — | — | + ^ = 0 , th e functio n f(x) i s periodi c o f perio d 30 . On e ca n
check tha t f(x) take s onl y tw o value s 0, 1 an d f(x) = 1 if 1 ^ x < 6. Therefor e

Hence on e derive s
(2.13) ax + O(logx ) < ip(x) < f a x + O(logx) .
Since th e differenc e betwee n th e uppe r an d th e lowe r bound s i n (2.1 3 ) i s quit e
small, on e ca n deduc e fro m (2.1 3 ) tha t an y dyadi c interva l [x , 2x] contain s prime s
provided x i s sufficientl y larg e (actuall y Tchebyshe v prove d thi s fo r al l x ^ 1 ,
which wa s postulated b y Bertrand) . Tchebyshe v faile d t o prov e th e Prim e Numbe r
Theorem, however , h e succeede d t o sho w tha t i f li m ip(x)/x exists , the n thi s limi t
x—>oo
is equal to one (thi s follows instantl y fro m (2.1 1 ) because 1 — § +1_ \~\ = log2) .
Interesting result s wer e derive d fro m Tchebyshe v work s b y Mertens . Firs t
approximating [x/d] i n (2.1 0 ) b y x/d on e get s

(2.14) £ ^ M = i o g x+ 0 (l).
n^.x
34 2. E L E M E N T A R Y T H E O R Y O F P R I M E N U M B E R S

Next applyin g partia l summatio n on e derive s fro m (2.1 4 ) tha t

(2.15) V - = lo g log x + p + 0 ( - 4
^—' p \ log x J
p^x
where ft i s a constant . Finall y usin g log( l — p _ 1 ) = — p~l + 0(p~ 2), w e hav e th e
Mertens formul a

< 216 » n(i-; /

) = iS{i + o(i)}-
p^x x v
'
Mertens showe d tha t th e exponen t 7 i n (2.1 6 ) coincide s wit h th e Eule r constan t
(1.69).
Similarly, usin g S = 1 */x i n plac e o f L = 1 * A, on e derive s a handfu l o f result s
for th e Mobiu s function . Firs t fo r an y x ^ 1 one get s

in plac e o f (2.1 0) . Henc e approximatin g [x/m] b y x/m on e deduce s tha t

/i(m)|
(Z-iS) E 771 o.
2.3. Prime s i n arithmeti c progressions .
Every natura l numbe r n = — l(mod 3 ) ha s a prim e diviso r p = — 1 (mo d 3) .
Hence there ar e infinitel y man y prime s p = — l(mod 3) . I n th e sam e wa y one show s
that fo r q — 4 , 5 there ar e infinitel y man y prime s p = — l(mod q).
Let q be an y intege r > 1 . Conside r th e cyclotomi c polynomia l o f degre e <p(q),

*,(*)= I I (x-eQ)eZ[x]
a ( m o d q)
(a,q) = l

One ca n sho w (quit e easil y b y usin g elementar y propertie s o f th e multiplicativ e

group (Z/gZ)* ) tha t ever y prim e diviso r o f $ q(n) coprim e wit h q satisfie s p = 1
(mod q). I f q is prime thi s propert y i s very simple . Indee d i n thi s cas e w e hav e
xq — 1 q l
<&q(x) - _ = 1 + x + .. . + x ~.

Hence i f p divide s $ q(x), the n x q = 1 (mo d p) , an d comparin g wit h th e littl e

Fermat theore m x p~1 = 1 (mo d p) w e deduc e tha t q divide s p — 1 . No w arguin g
like Eucli d w e ca n deduc e tha t ther e ar e infinitel y man y prime s p = 1 (mo d q).
Naturally ever y clas s a (mod q) wit h (a,q) = 1 shoul d hav e infinitel y man y
primes. I t wa s fo r th e proo f o f thi s theore m tha t L . Dirichle t (1 837 ) invente d
multiplicative character s x(mo d q) (se e Chapte r 3) . T o ever y suc h characte r h e
associated th e Dirichle t serie s

(2.19) L(s, X) = J2x(n)r

 2. E L E M E N T A R Y T H E O R Y O F P R I M E N U M B E R S 35

As i n th e cas e o f th e zet a functio n th e Dirichle t L-functio n ha s th e Eule r produc t

1a
(2.20) L(s,x) = Y[(l-x(p)p- )-
V

which converge s absolutel y fo r s > 1 . Henc e L(s,x ) doe s no t vanis h i f s > 1 .

Taking logarithm s w e get

(2.21) logL(s, X) = 5 > ( p ) p - s + 0(l) .

P

Hence b y th e orthogonalit y o f character s (se e Sectio n 3.2 )

P
(2.22) Yl ~S = Jq) ^ x(a)logL(s, X) + 0(l),
m
p=a(mod q) x( ° d q)

where th e erro r ter m 0(1 ) i s bounde d i n term s o f s uniforml y fo r s > 1 . Fo r th e

principal characte r w e ca n relat e L(s,xo ) t o th e zet a functio n

L(3,xo)=as)H(l-p-s).
p\q

Since

c(s) = ^ > -s = y x -dx + o(i) = j -i + o(i),

we infer tha t th e principa l characte r contribute s t o (2.22 )

(2.23) _ ^ logL(a>Xo) = _^ l o g _l_ + 0 ( 1 ).

Next w e conside r x 7 ^ Xo- Sinc e x i s periodi c o f perio d q an d th e sum s ove r an y

complete se t o f residu e classe s vanish , i t follow s tha t th e characte r su m ove r an y
segment i s bounded, precisel y

| E X(n) <q.

Therefore th e serie s (2.1 9 ) converge s fo r s > 0 an d |L(s,x) | < q, b y partia l sum -

mation. I f w e kno w tha t

(2.24) £ ( , X ) ^ 0 ,

then logL(s , x) ha s a finite limi t a s s — > 1 + . Henc e w e can deduc e tha t fo r s > 1 ,

(2.25) £ P-'-^Th+OW-
p=a(mod q)

In this way Dirichlet prove d there are infinitely man y primes in any primitive residu e
class. Actuall y (2.25 ) show s tha t prim e number s ar e equidistribute d (i n a certai n
sense) amon g th e primitiv e residu e classe s a (mod q).
It remain s t o sho w th e non-vanishin g o f L(l , x) fo r ever y x 7 ^ Xo5 which i s th e
heart o f th e matter . W e d o no t follo w exactl y th e origina l line s o f Dirichle t bu t
36 2. ELEMENTAR Y THEOR Y O F PRIM E NUMBER S

rather mi x his ideas wit h mor e familia r transformation s o f Tchebyshev. W e begin

with

] T Xi^n' 1 lo g n = Y, X(n)n- 1 ^ A(d )

n^.x n^.x d\n

= J2x(d)A(d)d-1 ^ x H " " 1

d<x ra<^§

Y/x(d)A(d)d-1(L(l,x)+0(d/x))
d^.x

L(l,X)^x(d)A(d)d-1+0(l)
d^x

where th e error ter m 0(1 ) is bounded i n terms o f x (but not of q). Th e left sid e is
also bounded (b y Abel's criterion ) becaus e th e average o f x(n) ove r an y segment i s
bounded. Therefor e w e proved tha t

(2-26) I > ( ^ « 1
d^.x

provided L ( l , x ) ^ 0. Similarl y w e obtain

A(n)_^xW_E/i((i)log,
log,+ X : x ( n ) ^ = E ^ E ^ ) l o g J
n ^ x n^.x d\n d\n

d<x m^^

L(l, X )^Mrf)^log^+0(l).

Hence, i f L ( l , x) = 0, we deduce tha t

(2.27) ^ x ( n ) ^= _ i o g x + o(l).
n-Cx

In view of the Mertens formul a (2.1 4 ) th e above conditiona l formul a ca n be written

as
£ ( i + x(p))^f«i.
p^x
This shows that \(p) = —1 for almost al l primes. I n other words, if L(l, x) vanishes,
then th e characte r x ( n ) on squarefree number s mimic s th e Mobiu s functio n /x(n) ,
which i s no t likel y t o b e tru e (bot h function s ar e multiplicativ e bu t fi(n) i s not
periodic). I n any case, w e have show n rigorousl y tha t

(2.28) J2x(n)^=6 A(«)

x\ogx + 0(l)
n^.x
 2. ELEMENTAR Y THEOR Y O F PRIM E NUMBER S 37

an
where 6 X =— 1,0 accordin g t o L ( l , x) = 0 or L ( l , x) / 0 if x ¥" Xo > d Sx = 1 ii
X = ^0 . Summin g (2.28 ) ove r al l characters w e ge t

V
n^x ^ ^ y x(mo d g )
n=l(mod q)

Hence ^2 y5x ^ 0 . Thi s show s tha t ther e i s at most on e character X ¥" Xo with
L ( l , x ) = 0 . I f suc h a n exceptional characte r exists , i t mus t b e real becaus e
L ( l , x ) = 0 implies L(l,x ) = 0 (th e complex conjugatio n i s a continuou s auto -
morphism o f C).
The non-vanishin g o f L ( l , x) fo r a real characte r \ ¥" Xo requires a differen t
argument. T o this en d we consider

where

(2.29) T(n )X) = $>(d).

d\n

Note tha t
T(^X) = n ( x +xw+• • •+x(pn) > o
p«||n
2
for al l n, and r(ra ,x) ^ 1 - Therefor e

On th e other hand , openin g th e convolution (2.29 ) w e evaluate T(x) in the same

fashion a s Dirichlet deal t wit h th e divisor proble m (1 .75 ) (th e hyperbola method) .
We obtai n

mn^x

= ] T xO)m~ ^ ] T n"^ + ^ n" ^ ^ x H ^ ^

J+ +0
- E ^>m-*{j(s) ' ((7) J )} +0 (i:»-**-')
raSC-y/a; n<y/x
= ±L(1 , X)^+0(1 ).

Comparing bot h estimate s w e obtain log x < L(l, x )^ + 0(1), whenc e L(l , x) ^ 0
by lettin g x tend t o oo.
T H E O R E M 2. 1 (DIRICHLET) . For every non-principal character \ (mo d q) we
have L ( l , x ) ^ 0.
By Theore m 2. 1 we complete th e proof o f (2.25). Moreover , w e proved tha t
(2.26) hold s fo r any x ^ Xo - Summin g ove r character s w e deduce
38 2. E L E M E N T A R Y T H E O R Y O F P R I M E N U M B E R S

THEOREM 2.2 . For any a with (a,q) = 1 we have

A(n) log x
KHJ
n<x ^
n=a(mod q)

where the error term 0(1 ) depends only on q.

REMARK. Th e origina l proo f o f Dirichlet tha t L(l , x) ^ 0 for a real characte r

X is a simpl e consequenc e o f his Class Numbe r Formula : le t K = Q ( A / D ) b e a
quadratic fiel d wit h D = 0, 1 (mo d 4 ) so D is the discriminant o f K. Ther e i s an
associated rea l primitiv e characte r x suc h tha t

(K(s)=Y,(Na)-s = ((s)L(s, X)
a

and ever y primitiv e rea l characte r modul o q ^ 1 arise s i n thi s wa y from K —

l)q). Dirichle t showe d tha t

/ 2nh
"^i f x ( - i) = - i ,
(2-31) L(l,x) = {
^ p ifx(-i ) = i,
where h is the class numbe r o f K, w is the number o f units i n the ring o f integers
of K i f x(—1 ) = —1 , an d e i s the fundamental uni t o f K i f x(~~l ) — 1- Henc e
L(l,x) > -75 - (Se e (22.59) fo r odd characters).
In Chapter s 5 , 22, 23, we develop lowe r bound s fo r 1,(1, x) using technique s of
increasing sophistication. Se e also Section 1 5. 9 for some results about clas s number s
of rea l quadrati c fields .

2.4. Reflection s o n elementary proof s o f the prim e numbe r theorem .

The Prim e Numbe r Theore m wa s prove d i n 1896 independently b y Hadamard

and d e la Valle e Poussi n b y analyti c methods . Th e first elementar y proof s wer e
found abou t fift y year s late r b y Erdos an d Selberg. A t the heart o f their proof s lie s
the formul a (du e t o Selberg)

2
(2.32) V^(logp) + ]T £(logp)(log<7 ) = 2xlog: r + O(x).
p^x PQ^x

To prove thi s w e begin by

2
^ log n = x(log2 x - 2 log x + 2) + 0(log 2 x)
rvCx

= rc(log x) (J2 fc-1 ) ~ & + ( 7 + 2) log k] + 0(log 2 x).

rC<^.0C rZ^^OC
 2. E L E M E N T A R Y T H E O R Y O F P R I M E N U M B E R S 3 9

where 7 is the Eule r constant . Henc e opening the convolutio n A 2 = /x*L 2 w e derive

J2 A 2 (n) = J2 J2 ^ m ) l o ^ 2 n

l^x m\l

J2 J2 ^m ) [^ + (7 + 2) log ^] + O(z)
l^x m\l

J2 - {S(n) logx + A(n) ) + ^ ( 7<J(n) + ( 7 + 2)A(n) ) + 0(x)

n^x n ^

= xlog x H-xlog x + 0{x) = 2xlog x + 0(x)

by Mertens ' formul a (2.1 4 ) an d Tchebyshev' s uppe r boun d (2.1 2 ) (recal l (1 .1 8)) .
EXERCISE. Deriv e from (2.32 ) by induction on k ^ 2 (use the recurrence (1 .47) )
that

(2.33) Y J^k{n) = kx{\ogx)k-l{l + o(^-)}.

n^x

Using Selberg' s formul a (2.32 ) Postniko v an d Romano v [PR ] derive d th e ele -

gant inequalit y

(2.34) | M ( x ) | l o g x < ^|M(-)|+0(xloglog3x) .

n^.x

Compare thi s wit h th e secon d par t o f (2.1 7) . No w th e Prim e Numbe r Theore m

can b e derive d i n th e for m ip(x) ~ x fro m (2.32) , o r i n th e for m M(x) = o(x) fro m
(2.34) b y elementary , howeve r lengthy , tauberia n typ e arguments .
E. Bombier i [Bol ] an d E . Wirsin g [Wi2 ] refine d th e Erdos-Selber g metho d t o
show
A
(2.35) i/;(x) = x + 0(x(\ogx)- )
for an y A > 0 , th e implie d constan t dependin g o n A. Alon g th e sam e line s H .
Diamond an d J . Steini g [DS ] succeede d i n showin g tha t
2
(2.36) \rp(x)-x\ <rrexp|-(logx)^(loglogx)" j

for al l x ^ e e . Th e argument s o f Erdos-Selberg ar e truly elementar y (n o comple x

numbers ar e harme d i n th e cours e o f th e proof ) nevertheles s w e express ou r mixe d
feelings, becaus e th e earlie r analyti c method s base d o n contou r integratio n insid e
the zero-fre e regio n o f th e Rieman n zet a functio n ar e considerabl y simple r an d th e
results ar e superio r (se e Section s 5. 4 an d 5.6) . I n 1 89 9 d e l a Valle e Poussi n prove d
quite easil y tha t

(2.37) I/J(X) = x + 0(x exp(-c^/\ogx))

where c > 0 and th e implie d constan t ar e absolute . Th e subsequen t improvement s
of th e PN T cam e fro m sharpe r uppe r bound s fo r £(s ) nea r th e lin e R e (s) — 1 ,
 40 2. E L E M E N T A R Y T H E O R Y O F P R I M E N U M B E R S

particularly b y Vinogradov' s estimat e fo r exponentia l sum s (se e Chapte r 8) . How -

ever, w e woul d lik e t o sa y tha t Vinogradov' s metho d o f exponentia l sum s shoul d
not b e regarde d a s a ne w techniqu e fo r non-vanishin g o f the zet a function . I n fact ,
all the know n method s o f converting th e uppe r bound s int o th e lowe r bounds , fro m
which t o dra w a zero-fre e region , remai n i n principle th e sam e on e a s originate d b y
Hadamard an d d e l a Vallee Poussin . Yet , som e variations ar e mor e sensitiv e t o th e
input tha n th e others , givin g close r connection s betwee n estimate s an d zeros . Fo r
example th e Borel-Caratheodor y theore m i n comple x functio n theor y doe s th e jo b
very well . Th e bes t erro r ter m i n (2.37 ) know n toda y i s due t o Korobo v [Kor ] an d
Vinogradov [VI ] (se e Corollar y 8.31 ) .
The asymptoti c formul a I/J(X) ~ x i s equivalen t t o th e non-vanishin g o f ((s)
on th e lin e R e (s) = 1 , an d th e latte r ca n b e establishe d relativel y easil y withou t
appealing t o analyti c propertie s o f £(s ) .
Here is how one can arrang e th e analyti c idea s for a proof o f the Prim e Numbe r
Theorem i n the for m (2.35 ) whic h a broad-minded reade r ma y stil l find elementary .
We avoi d comple x variabl e analysis , althoug h fo r clarit y w e d o no t hesitat e t o
employ £(s ) a t comple x number s s = a + it, bu t onl y wit h a > 1 wher e th e
Dirichlet serie s and th e Euler produc t converg e absolutely. W e begin by considerin g
the functio n G(s) = (—l) k(l/((s))(k\ Thi s i s give n b y th e serie s

(2.38) G (s)= £^M (iogm)*.

m

Prom her e w e go to th e finite su m

(2.39) F(x) = ] T /i(m)(logm) fc log —

m^.x

where th e facto r log(rr/m ) play s a tas k o f smoothin g a t th e cut-of f point . Thi s

particular cuttin g i s produced b y mean s o f th e integra l formul a

l o g + y = ^~ / y s
s~2ds

where a > 1 , givin g

s
(2.40) F(x) = ^- f x G{s)s~2ds.
2m J

For th e purpos e o f estimatin g G(s) w e consider (*( s) — (s ~ 1 )C( 5)- Not e tha t
(fe) (fc) ik l)
1
(2.41) ( /C(*)) = (s - l)(l/C*(*)) + k(l/C(s)) -.
Next w e appea l t o th e formul a fro m differentia l calculu s

JJ
a 1 +2a2+---=k L Z J J

where a i , a 2 , . . . ru n ove r non-negativ e integers . Thi s formul a fo r /($ ) = C*( 5)

reduces the problem t o uppe r bound s fo r th e derivative s o f C*( s) a n d a lower boun d
 2. E L E M E N T A R Y T H E O R Y O F P R I M E N U M B E R S 41

for |C*( 5)|- Firs t w e derive b y partia l summatio n tha t fo r R e (s) > 1 ,
/'OO I I

(-l)eCW(s) = ]Tn- s(logn/ + y rr- s

(logx)^ + o(^(logX)' +1
)

= y°°x- s(loga;)^ + o ( (l + ^)(logX/ +1

).
Taking X = 2|s | w e ge t

(2.43) (-l) e
CW(s) = ^ • A+ 0((log2|s|)^ 1 ).
(s-iy
Hence
W
(2.44) (C*(«)) = ( * - 1 )C (0(«) + iC ( ' _1 ) («) « M(log2| a |)'
+1
.
Next w e infe r fro m th e Eule r produc t b y positivit y tha t

i< ^ ( i - K i + ^ + p - ^ V7 )
= JJ ( l + (3 + 2plt + 2p~u + p2it + p~ 2it
)p-^
p

xC 3 (a)|C(a + ii)| 4|C(cr + 2zO| 2

where th e implie d constan t i s absolut e ( a simila r formul a i s als o presen t i n th e
arguments o f Hadamar d an d d e l a Valle e Poussin , se e Theore m 5.26) . Henc e an d
by (2.43 ) w e get th e lowe r boun d

(2.45) | C » | » ( < r - l ) ? H ( l o g 2 | S |)-l

From (2.44 ) an d (2.45 ) w e derive usin g (2.42 ) tha t

(1/C*(s))(fe) < ( a - l)-t( fe 1

+ )| s |-1 (log2| S |) K
where K and th e implie d constan t depen d onl y o n k (i n fact , K = (5A ; + l)/ 2 come s
out b y th e abov e arguments , bu t i t doe s no t matte r fo r application) . Henc e b y
(2.41)
G(s)<(cr-l)-t(fe+1)(log2|s|)K.
Inserting thi s boun d t o (2.40 ) an d integratin g triviall y w e ge t
F(x)<^xa(a-l)-^k+1\
Finally takin g a = 1 + ( l o g x ) - 1 w e conclud e th e followin g estimat e fo r th e su m
(2.39)
1fc+
(2.46) F(x)<rr(logx)4( ).
Note tha t thi s i s non-trivia l i f k > 3 .
Now it i s a completely elementar y exercis e to derive the PN T fro m (2.46) . Firs t
we d o th e su m
fc
(2.47) ff(x) = £>(m)(logm) .
42 2. E L E M E N T A R Y T H E O R Y O F P R I M E N U M B E R S

which i s obtained b y differencin g F(x) a s follow s

fc
F(x + y)- F{x) = # ( x ) lo g -—^- + ] T /i(m)(logm) lo g
m
xKm^x+y
= [ff( a;) + 0(y(logx) fc )llog^^.
LJ x X
Hence
H(x)<^y{\ogx)k + y- 1 x2(\ogx)i{k+l) =2x(\ogx) k A
~
by choosing y = x(lo g x)~A wit h ^4 = (k— 3)/8. Finall y we get by partial summatio n
A
(2.48) ] ^ /x(m ) < x(logx)~

where th e implie d constan t depend s onl y o n k. Sinc e k i s arbitrar y s o i s A. Thi s

proves th e PN T fo r th e Mobiu s function . Henc e th e PN T fo r th e vo n Mangold t
function follow s b y th e formul a (2.8) .
 http://dx.doi.org/10.1090/coll/053/04

CHAPTER 3

CHARACTERS

3.1. Introduction .

The character s o n residu e classes , bot h additiv e an d multiplicative , pla y in -

strumental part s i n analyti c numbe r theory . Othe r character s suc h a s th e Heck e
characters o f a number field, o r the character s o f a finite field ar e als o indispensabl e
in th e moder n theory . W e shal l introduc e thes e i n du e course . However , t o avoi d
redundancy w e start her e b y givin g basic definition s i n a general contex t o f a finite
abelian group . (Character s fo r non-abelia n group s occu r also , bu t mor e naturall y
as Galoi s group s an d ar e bes t see n fro m th e automorphi c perspectiv e i n analyti c
number theory) .
:
Let G b e a finite abelia n group . A homomorphis m x G —) > C* i s calle d a
na s
character o f G. Therefor e (i n th e multiplicativ e notation ) \ th e propertie s

X(xy) = x(x)x(y) for 1a 1

x,y eG,

x(l) = 1 and x( x)m — 1? where m = \G\ is the orde r o f G , therefor e x( x) i s a ro

°t
of unity .

The character s o f G for m a grou p G wit h multiplicatio n give n b y

(XiX2){x) = Xi(x)X2{x) fo r al l x G G.

G i s called th e dua l group , it s identit y elemen t i s the trivia l characte r

Xo{%) — 1 fo r al l x G G.

Throughout x denote s th e comple x conjugate , henc e als o th e inverse . I f G i s cyclic

of orde r m an d g i s a generato r o f G , the n ever y characte r o f G i s o f typ e

Xa{x) = e 2™y'm, iix = gy

for som e fixed residu e clas s a(mo d m) . Thes e ar e distinc t characters , therefor e G
is als o cycli c o f orde r m , s o G i s isomorphic t o G . Th e isomorphis m G ~ G ca n b e
established fo r an y finite abelia n grou p b y writing G a s the direc t produc t o f cycli c
groups. Ther e i s a canonica l isomorphis m betwee n G an d G give n b y x F- > X where

Hx) = X(x) fo r al l x € G.
43
44 3. CHARACTER S

The raison d'etre o f these characters is found i n the following orthogonality relation s

5>(*) = {| d | if * =1'
which allow s u s t o detec t th e grou p identit y element .
Suppose d divide s th e orde r o f G . A n elemen t g G G i s sai d t o hav e exponen t
d i f g d i s th e identity . Th e orde r o f g i s it s smalles t exponent . Th e character s o f
exponent d ar e exactl y thes e whic h ar e trivia l o n th e subgrou p
Gd = {x d : x G G} ,
therefore the y ma y b e viewe d a s characters o f the facto r grou p G/G d. Th e orthog -
onality relation s becom e
[G : Gd] i f y G Gd ,
2^ XW I n0 i f y\t».dnd
0 G d.

This wil l allo w u s t o detec t th e d-th. powers o f G .

3.2. Dirichle t characters .

First w e consider th e character s o n th e additiv e grou p Z/ra Z o f residue classe s
modulo m. The y ar e give n b y

where e(z) = e 27Tlz. Thi s formul a make s a n additiv e characte r a functio n o n Z ,

periodic o f perio d m. Th e orthogonalit y propert y become s

CM
mi f n = 0(mo d ra),
E
a(mod m)
0 otherwise .

We shal l cal l ^ a (n) = e(an/m) primitiv e i f (a , ra) = 1 . Addin g al l th e additiv e

primitive character s w e obtain th e Ramanuja n su m

(3.1) S(n,0;m ) = c m(n)= ^ * e ( ^fan\ y

a(mod m )

(see (1 .56)) . Recal l tha t throughou t thi s boo k Y^ restrict s th e summatio n t o

"primitive" elements , i n th e abov e cas e amon g additiv e character s modul o ra. On e
shows b y Mobiu s inversio n tha t

(3.2) c m{n)= ] T d/x(^) .

d\(m,n)

Hence
 3. C H A R A C T E R S 45

In particular ,
(3.4) c m(n) = /i(ra ) i f (ra , n) = 1 .
In general ,
(3.5) \cm{n)\ < (ra,n) .

Next w e conside r character s o n th e multiplicativ e grou p o f residu e classe s

a(mod ra) wit h (a , ra) = 1,
X : (Z/raZ)* -> C.
As wit h th e additive characters , w e wish t o consider x a s a functio n o n Z; we d o
so b y setting x( n) = 0 whenever n i s not prime t o ra. Thi s make s x a Dirichle t
character, a function o n Z, periodic modul o ra and completely multiplicative . The
corresponding extensio n o f the trivial characte r x o (mod ra) i s called th e principal
character t o modulu s ra. Th e group (Z/raZ) * an d its dual hav e (/?(ra ) elements .
The orthogonalit y relation s becom e
ip(m) if x = Xo,
E X(a ) = {0 otherwise ,
a ( mnod
o d ra)

(/?(ra) i f a = l(modra) ,
E *(« ) = {0 otherwise
%(mod ra)
,

If ra = vfi\Vfi2 wit h (mi,771 2 ) = 1 , then (Z/mZ) * ~ (Z/raiZ) * x (Z/TTT^Z)* ,

so ever y multiplicativ e characte r x ( m ° d m ) i s a produc t X1 X 2 o f multiplicativ e
characters xi(nio d m\) an d X2(mod 777,2).
The Dirichle t serie s associate d t o Dirichlet character s ar e of paramount impor -
tance in analytic numbe r theory . The y ar e called Dirichle t //-function s an d denoted
L(s,x), o r L(x, s) depending o n emphasis:

(3.6) L(s, x) = E *( n ) n ~ S = II (! - X{P)P- S)-1 ;

the series and the Euler produc t ar e absolutely convergen t fo r Re (s) > 1. As is well
known, thes e function s ca n be analytically continue d t o C. Se e Section 4. 6 for a
proof. Man y othe r analyti c properties , application s an d generalizations o f Dirichlet
//-functions will be considered throughou t th e book.
Note tha t b y total multiplicativity , (1 .1 5) , (1 .38 ) giv e
1 L'
— — - - ^ / x ( n ) x ( n ) n - s , ~j;(s,x) = ^ A(n)x(n)n~ 5

for R e (s) > 1.

3.3. Primitiv e characters .
Associated t o eac h characte r x > m additio n t o it s modulu s 777 , is a natura l
number ra*, its conductor. Th e conductor i s the smallest diviso r o f ra such tha t x
may be written a s x = XoX * where x o is the principal characte r t o modulus ra and
X* is a characte r t o modulu s 777* . Fo r some character s th e conductor i s equal t o
the modulus . Suc h character s ar e called primitive . I n the above factorizatio n x * i s
 46 3. CHARACTER S

a primitiv e characte r uniquel y determine d b y \, W e shal l sa y tha t \ i s induced b y

X* or tha t x * induce s x - W e hav e

X(a) = x*(a ) i f (a,m ) = 1 .

The numbe r o f primitiv e character s t o modulu s m i s give n b y

i
(3.7) ^ H = " » n ( -|)n(i-J)a-2
p\\m p \m

The proo f i s a simpl e exercis e i n Mobiu s inversion . Indeed , w e hav e (p = 1 * <£*,

whence ip* = \i * ip, i.e.,

giving (3.7 ) b y multiplicativity . I n th e sam e wa y usin g th e orthogonalit y o f char -

acters w e infe r a mor e genera l resul t
if
(3.8) J2* *(«) = £ ^Hl) (a<m) = L
x(mod m) d\(a—l,m)
Incidentally, th e formul a (3.7 ) show s tha t th e primitiv e character s %(mo d m) exis t
precisely whe n m ^ 2 (mod 4) .
The primitiv e character s ar e pleasan t t o dea l with . Fo r example , w e hav e a
convenient formul a (whic h i s not vali d fo r % not primitive )

1(3-9) £ X (ac+b) = \X{b) 'I™ 1 ;*

K} m
, ,, I 0i f rafa.
Indeed, lettin g S b e th e abov e su m w e obtai n x( l + mix)S — S fo r an y x , wher e
mi = m(a , ra)/(a 2, m) . I f 5 7 ^ 0, this yields X(1 +TOI:E ) — 1 for an y x, whic h implie s
that x i s periodi c o f perio d mi . Sinc e x i s primitive , w e mus t hav e m\ = m , i.e. ,
m I a. Thu s S = 0 if m \ a, an d (3.9 ) i s obviou s otherwise .
The Dirichle t character s o f exponen t tw o ar e jus t th e rea l character s (real -
valued). I n a numbe r o f way s the y pla y a specia l role , particularl y the y ar e fun -
damental i n th e theor y o f quadrati c forms . Belo w w e giv e a complet e lis t o f rea l
primitive characters . I f m — p i s a n od d prime , the n ther e exist s exactl y on e
real primitiv e characte r o f conducto r p , namel y th e quadrati c residu e symbo l (th e
Legendre symbol )

s
This ca n b e define d b y sayin g tha t 1 + Xp( n) * t n e numbe r o f solution s x(mo d p)
to x 2 = n(mo d p).
For m — 4 we hav e exactl y on e primitiv e characte r define d b y

X 4 (n) = ( - l ) ^ i if2fn .
If m — 8, w e hav e tw o primitiv e character s
Xs(n) = ( - l ) * ( n - 1 ) ( n + 1 ) i f 2 f n ,
X4Xs(n) = ( - l ) * ( " - 1 ) ( r e + 5 ) H2\n.
 3. CHARACTER S 47

If m i s prim e power , ther e ar e n o rea l primitiv e character s o f conducto r m othe r

than X4?X8 5X4X8 an d \ v . Ever y rea l primitiv e characte r x ( m ° d m) i s obtaine d a s
the product o f the above ones. Therefor e th e conductor o f a real primitive characte r
is a numbe r o f typ e 1 , /c, 4/c, Sk wher e k i s a positiv e od d an d squarefre e integer .
3.4. Gaus s sums .
Let u s com e bac k fo r a momen t t o a genera l finite abelia n grou p G. Th e
characters o f G for m a complete orthogona l syste m s o that an y functio n / : G—» C
has th e Fourie r expansio n

ipeG
with coefficient s
</,</>> = £ / ( ^ ( s ) -
geG

EXERCISE 1 Prov e tha t

(DU E T O D E D E K I N D ) .
TJ </,</>> = det(f(gh- 1 )).

In th e cas e o f function s o n residu e classes , an d o n a finite field, bot h additiv e

and multiplicativ e character s ar e presen t an d i t i s ofte n necessar y t o transfor m
the Fourie r expansio n i n on e o f thes e system s t o th e other . I n doin g s o w e en -
counter Gaus s sum s (x^) f° r an y pai r o f multiplicativ e an d additiv e character s
X, ip respectively. W e shal l presen t th e Gaus s sum s fo r finite fields i n du e time .
Now we consider Gaus s sum s associate d wit h character s o n residu e classes , sa y
modulo ra. Fo r an y multiplicativ e characte r x ( m ° d ra) w e pu t

(3-10) r ( X)= £ X(b)e(~).

b(mod ra)

Multiplying (3.1 0 ) b y x( a) an d summin g ove r x w e derive b y orthogonalit y

(3.11) e ( £ )= ^ £ x(a)r( X) if(a,m )= l .

x(mod m)

This i s th e Fourie r expansio n o f additiv e character s i n term s o f th e multiplicativ e

ones. Similarl y

6
(3.12) X(a)r(x)= £ *( ) e ( ^ ) '^(a,m) = l
b(mod m)

which give s th e Fourie r expansio n o f x i n term s o f additiv e character s provide d

r(x) 7 ^ 0 . Not e tha t th e conditio n (a , ra) = 1 i n (3.1 2 ) ca n b e droppe d i f x i s
primitive becaus e bot h side s vanis h i f (a , ra) ^ 1 (apply (3.9)) .
48 3. CHARACTER S

LEMMA 3.1 . Suppose the character x modulo m is induced by the primitive

character x * modulo m*. Then

(3.i3) ^>=<0**(^K) -
If x (mo d ra) is primitive, then
(3.14) |r( X )| = yM-

PROOF. W e hav e
e
r(x)= E * X » e ( ^ ) = E ^ ) x * WE * » ( ^ ) -
o(mod m ) d\m a(mo d *jj-)

The innermos t su m vanishe s unles s d = m/m* giving (3.1 3) . T o prov e (3.1 4 ) w e

write
2 'a — 6>
"(X)| = E E xWxWe(^ ) 771
a,b(mod m )

a(mod 7n ) 6(mo d ra)

The inner su m is the Ramanuja n su m (whic h is also the Gaus s sum for th e principa l
character). Insertin g (3.2 ) w e ge t

|r(X)|2 = E ^ ( 7) £ X(0)
-
d\m a(mo d ra)
aEl(mod d)
If x i s primitive o f conducto r ra, the n th e las t su m vanishe s unles s d — m (appl y
(3.9)), givin g (3.1 4) . D

One ca n se e b y Lemm a 3. 1 tha t r(x ) doe s no t vanis h exactl y whe n m/m* i s

squarefree an d prim e t o ra*. In thi s cas e (3.1 2 ) become s a true formul a fo r x( a )-
We now conside r th e mor e genera l su m

r(x,^a)= E x(b)Mb)
6(mod ra)

where ip a{b) = e(ab/m) i s an additiv e character .

LEMMA 3.2 . Let x modulo m be a non-principal character induced by the prim-
itive character x * modulo m* . Let a ^ 1 . We have

T(X, ^O) = r(x) Yl dx*{a/d)fi(m/dm*),

d\(a,m/m*)

in particular,

if (a , ra) =1 , i.e. if ipa i> s a primitive additive character.

Note tha t r(x ) = x ( - l ) r ( x ) > henc e i f x (m ° d ra) is primitive, the n (3.1 4 ) ca n

be writte n as
(3.15) T(X)T(X ) = x(-l)m .
 3. CHARACTER S 4 9

For an y character s xi(mo d mi ) an d X2(mo d 7722 ) with (mi , 7722) = 1 we hav e

(3.16) r(xiX2) = Xi(^2)X2(mi)r(xi)r(x2) .
This multiplicatio n rul e fail s i f the modul i ar e no t relativel y prime . Fo r character s
to th e sam e modulu s th e correctio n facto r i s th e Jacob i su m
(3-17) J{Xi,X2)= Y, Xi(a)X2(l-a) .
a(mod m)

Precisely i f xi > X2 are tw o character s modul o m suc h tha t X1 X 2 is primitive, the n

(3.18) r(xi)r( X 2) = J(xu X2MX1 X2) .
Hence i f al l Xi > X25 X1X2 are primitiv e o f modulu s m , the n
(3.19) |«/(xi,X2) | = Vm" .
If x(mo d m ) i s primitive, the n
(3.20) J ( x , x ) = x("l)Mm) .

3.5. Rea l characters .

For a primitiv e characte r x ( m ° d m ) w e kno w tha t |r(x) | = y/in, however ,
the determinatio n o f th e argumen t o f r(x ) i s a difficul t problem . Usin g Deligne' s
estimate fo r multipl e Kloosterma n sum s on e ca n sho w tha t r(x)/ 'y/rfi ar e asymp -
totically equidistribute d o n th e uni t circl e a s x ( m ° d m ) range s ove r al l primitiv e
characters an d m tend s t o infinit y ove r prime s (se e Chapte r 21 ) .
In th e cas e o f rea l character s r(x ) wer e evaluate d completel y b y Gauss . I t i s
easy t o se e tha t T(XA) = 2i , T(XS) — 2 A/2 an d r(x4Xs ) — 2y/2i.
T H E O R E M 3. 3 ( G A U S S ) . Form odd squarefree

(3.21) r(x ) = emVm~

where
mod4
(3.22) e rn = {1 ^ m
= l
( )>
171
\ i if m = — 1 (mo d 4) .
We shal l giv e a n analyti c proo f o f (3.21 ) . Bu t firs t w e evaluat e othe r Gaus s
sums o f typ e

(3.23) G(m)= £ e(£ )

n(mod ra)

for an y intege r m ^ 1 . I f m i s even w e derive b y shiftin g n t o n + | tha t G(m) =

imG(m), becaus e ^ ( n + f ) 2 = * £ + f (mo d 1 ) . Henc e
(3.24) G(m )= 0,i f m = 2(mo d 4) .
Next w e show tha t
3
(3.25) G(m ) = mG(m), i f m ^ 2(mo d 4) .
This follow s b y splittin g th e summatio n i n G(m 3) int o n = a + 6m 2 (mod m 3 ) i f
2 { m o r n = a- f 6^-(mo d m 3 ) i f 4 | m. No w we ar e read y t o prov e
50 3. C H A R A C T E R S

T H E O R E M 3. 4 (DIRICHLET) . For any m e N ,

1— 4- ? m
(3.26) G(m) = - ^ — ^ .

PROOF. W e hav e
o2>
c
G(m) = 2 £ Gr) + 01( ) -
0<n<-^

In th e segmen t ^ < n < y w e chang e n int o [y ] — n = y — { y } — n . Sinc e

^ ( [ f ] - ^ ) 2 = ^ ( ^ + { f } ) 2 - f + [ f ] - " = ^ ^ + { f } ) 2 - f ( m o d l ) , w e obtai n

<*»>-* E <£)+*- - E .(fit±^l£) +o(.,.

Here, an d a s before , th e erro r ter m 0(1 ) account s fo r th e term s whic h ar e no t
covered i n th e displaye d summatio n (a t mos t tw o o f them) . B y Lemm a 8. 8 eac h
of th e tw o shor t sum s abov e i s approximate d (u p t o a bounde d erro r term ) b y th e
integral
00
rm/4 2 .f 1 / T2 \ + 7
/ e( — )dx= / el— )<f a + 0 ( l ) = ^ >M+ 0(l) .
Jo Vr ™/ J o V m '4
Hence addin g u p thes e result s w e ge t

G(m) = \±^-V^ + 0(1 ).

1+2 i
It remain s t o sho w tha t th e las t erro r ter m 0(1 ) vanishes . Thi s follow s b y iterate d
application o f (3.25 ) i f m ^ 2(mo d 4) ; otherwis e (3.26 ) i s obviou s becaus e bot h
sides vanish . •

We generaliz e G(m) b y settin g

<-> o - E •(£ )
n(mod m )

for an y a wit h (a,m) = 1 . I n particular , G ( —) = G(m). I f ra = 771 1 777 2 wit h

(mi,777,2) = 1 , we obtai n

(3.28) G ( _ £ _ ) = G ( ^ W ^ )
by writin g 7 7 = 77-1 777, 2 + 772777 1 with 77 1 an d 71 2 rangin g modul o m\ an d 777 2 re-
spectively. Thi s formul a reduce s th e proble m o f evaluatin g generalize d Gaus s su m
G(a/777) to tha t o f a prim e powe r modulus . I f 77 7 = p i s a n od d prim e w e can writ e

o(f)- E K)Mf )
6(mod p)

since 1 + ( | ) i s th e numbe r o f solution s t o n 2 = 6(mo d p). Therefor e

<»•»> °(? ) - E $«(£ ) - ( X )" (? M

£>(mod p )
 3. CHARACTER S 51

by changin g th e variabl e an d usin g (3.26) . Nex t fo r distinc t od d prime s p , q w e

derive b y (3.26),(3.28),(3.29 ) tha t

^VH- 0 (i)- C (|M|)-(f)(J)^VH.

Note als o tha t
E l£
(3.30) e pq=epeq(-l) ^ ^.

Combining thes e formula s w e deduc e

THEOREM 3. 5 (QUADRATI C RECIPROCIT Y L A W ) . For any odd primes p ^q
we have

1
1 5 5
<>•»> (D^-t- )" " -
Since G(-l/p) = G(l/p) i t follow s fro m (3.29 ) tha t fo r od d p

(3-32) ( ^ ) = e 2= (_i)^ =X4 (p).

EXERCISE 2 . Prov e tha t fo r od d p,

(3-33) ( ? ) = (-i)t( P2-i) = X8 (p).

Now (3.21 ) follow s b y r(x) = G(p) = e py/p i f m = p i s od d prime , an d i n

general b y th e multiplicativit y o n usin g (3.1 6),(3.30 ) an d (3.31 ) .
Note w e have prove d tha t

^ \m) \m) ~ ^ \m)

x(mod m) x(mo d m)

if m i s od d squarefree . I f m i s od d bu t no t squarefree , th e lef t sid e vanishe s whil e

the righ t sid e doe s not .
For convenienc e w e extend th e Legendr e symbo l (^ ) t o p — 2 by settin g

(§H; :;*!:
Then fo r an y b > 0 we defin e

<»*> (i)-n(D' -
Pv\\b F

If b is odd, thi s symbo l ( | ) wa s introduce d b y Jacobi .

52 3. C H A R A C T E R S

EXERCISE 3 . Deriv e fro m (3.31 ) tha t fo r an y od d integer s a, b relatively prim e

(3.36 ) (i)(s)-H)-"W W
Here (x^y)^ i s the Hilber t symbo l define d fo r xy ^ 0 by
f -1i f x < 0 and y < 0,
(3.37 ) (*,y)oo = i ^ •
(^ 1 otherwise .
Notice tha t 2{x, y)^ — 1 4- sign x - f sig n ? / — sign #?/.

EXERCISE 4 . Prov e tha t fo r an y a,r a wit h (2a , m) = 1 ,

(3.38) G(£ ) = f-Wv^.

;c
m
Finally w e exten d th e symbo l (f ) t o al l integer s a , b except fo r a — b = 0 . I f
ab ^ 0, w e se t

< 339 > © -(H)'"-"-

Then w e se t

and
(3.41) ( | ) = 0 if (a, 6 ) ^ 1.

For 6 ^ 0 thi s symbo l wa s introduce d b y Shimur a [Shi ] i n th e contex t o f

modular form s o f half-integra l weight . H e showe d tha t (1 .53 ) hold s wit h

(3.42) i / ( c , d ) = e d Q.

EXERCISE 5 . Chec k th e consistenc y o f th e abov e settings . Prov e tha t fo r an y

6 ^ 1 th e ma p a h->- ( |) i s a Dirichle t characte r modul o b. Prov e tha t fo r an y a ^ O
the ma p b i-> ( | ) i s a Dirichle t characte r o f conducto r a * | 4a.

The abov e extensio n o f ( | ) wil l b e calle d th e Jacob i symbol .

Any intege r A / 0 wit h A = 0, l(mod 4 ) i s calle d a discriminant . I f A = 1
or A i s th e discriminan t o f a quadrati c field, the n A i s sai d t o b e fundamental .
Any discriminan t ca n b e writte n uniquel y a s A — e 2D wit h D fundamenta l an d
e ^ 1 . A prim e discriminan t i s a fundamental discriminan t havin g exactl y on e
prime factor , thu s i t i s a numbe r o f typ e —4 , —8, 8 an d X<±(P)V with P > 2. Ever y
fundamental discriminan t factor s int o prim e discriminants .

To an y discriminan t A on e associate s th e Kronecke r symbo l ( — J whic h i s

defined fo r an y c ^ 0 by mean s o f th e Jacob i symbo l a s follow s

(3.43 ) ^ ^ ^
cJK \ A j \ b
 3. CHARACTER S 53

where c = 2 vb wit h b odd. Not e tha t ( f ) K = ( f ) = Xs(A) , explicitl y

'1 i f A = l(mod8) ,
(3-44) | ' | j . { - 1 if A = 5(mod8) ,
0 if A = 0(mod4) .
We als o exten d th e definitio n o f th e Kronecke r symbo l fo r c = 0 by settin g
A\f 1 i fA = l ,
V^ 7 xl 0 otherwise .
Therefore (^) K i s defined fo r al l integer s c and A ^ 0 , A E 0 , l(mod 4) .
We shal l dro p th e subscrip t K i n th e Kronecke r symbo l notatio n an d explai n
in an y relevan t cas e tha t w e ar e dealin g wit h th e Kronecke r symbol . Remembe r
the Kronecke r symbo l i s no t define d fo r A' s othe r tha n discriminants . Whe n th e
symbol ( —) appear s withou t comment s i t stand s fo r th e Jacob i symbol .

EXERCISE 6 . Prov e tha t fo r a fundamental discriminan t A th e Kronecke r sym -

bol (^ ) i s a primitiv e characte r o f conducto r |A| .

3.6. Th e quarti c residu e symbol .

Next w e construc t certai n character s o f orde r four . Thes e ar e associate d wit h
Q(i), th e imaginar y quadrati c fiel d o f discriminan t D — —4. Th e rin g o f integer s
of Q(i) i s
Z[i] = {z = a + bi:a,beZ}
and th e grou p o f unit s o f Z[i] is { l , i , i 2 , i 3 } . Th e irreducibl e element s o f Z[i] ar e
(up t o th e units) ; 1 + i wit h N(l - f i) = 2 , th e rationa l prime s q = — l(mod 4 ) wit h
Nq = q 2 an d th e comple x number s TT = a + bi wit h
(3.46) NTT = TTT T = a 2 + b 2 = p = l(mo d 4) .
Note tha t TT an d 7 f are coprime , thes e ar e calle d Gaussia n primes . Ever y elemen t
of Z[i] different fro m zer o factor s uniquel y (u p t o th e unit s an d permutations ) int o
powers o f irreducibl e elements .
For ever y Gaussia n prim e TT w e define a ma p (th e quarti c residu e symbol )
2 3
(3.47) ( - ) :Z[i]^{0,l,M i }

which satisfie s
(3.48) (-] = a 2^1 (mo d TT) .

Note tha t ( ^ ) = 0 i f 7 r | a . I f 7 r f a , the n a?~ l = l(mo d TT) becaus e th e residu e

class ring Z[i]/7rZ[i] i s a finite fiel d wit h p = NTT elements. Thi s propert y (th e littl e
theorem o f Fermat ) implie s th e existenc e an d th e uniquenes s o f solution s t o (3.48 )
with (^ ) = i m fo r som e 0 ^ m < 4. I n particular , w e hav e

(3.49) [ - ] = i ^ ,i f P = TTTT = l(mod4) .

54 3. C H A R A C T E R S

EXERCISE 7 . Prov e th e followin g propertie s o f th e quarti c residu e symbol :

<""> (¥)-(;)(;) •
(3.51) a = /?(mo d n) => ( - ) = ( - ) ,

(3,2) £ ) = (!! ) i f V=J» ,

<«»> a (I
(3.54) ( 4
) = (")•
— 1 = 1 if and onl y if z = a (mod IT) ha s a solution i n Z[i]* .

If ^ = 7T i • • -7rr i s the produc t o f Gaussian prime s (no t necessaril y distinct) ,

then w e se t th e symbo l

<»» (?) = £)-£)•

Clearly th e propertie s (3.50)-(3.53 ) hol d tru e wit h TT replaced b y 7 . I n particular ,
(3.53) give s
(f)=l i f
(«.P ) = 1 -
Note tha t ( 1 + i)2 = 2i . W e sa y tha t a G Z[i] i s odd i f ( 1 + i) \ a an d
primary i f a = l(mo d 2( 1 + z)). Ever y od d numbe r i n Z[i] i s associated wit h
exactly on e primary element , i.e. , ea = l(mo d 2( 1 +z)) fo r exactl y on e unit e = i m.
A primar y elemen t ca n b e writte n a s th e produc t o f primar y irreducible s uniquel y
up t o permutations . W e hav e th e followin g
T H E O R E M 3. 6 ( T H E LA W O F QUARTI C R E C I P R O C I T Y ) . If TTI^2 are distinct
primary Gaussian primes, then

where p\ = NTTI and p 2 = N7T2 . If TT = a + bi is primary, then

2 2 >
' l + ^)\ .(„.-•[-h-h
Q) =i (l-)/ a> (i±l }/d /
= j (-l-6-^)/4 f ( ^ r6'2.

The quarti c residu e symbo l (— ) is a multiplicative characte r o f th e finite field

¥p ~ Z[i]/irZ[i\. Th e Gaus s su m o f thi s characte r is

,3,7) ,< „ = £ (=).(| )

a ( m o d p)

where a runs ove r th e rationa l residu e classe s modul o p and a is the representativ e
of a i n Z[i] modulo 7r . If 7r is primary, the n
2
(3.58) S M= -(-l)^V^ .

We shal l b e intereste d i n th e quarti c residu e symbo l a t rational integer s

(3.59) X7r (n)=(-), forneZ .

 3. C H A R A C T E R S 55

This is a Dirichlet characte r o f conductor p = TTTT and o f order four. Sinc e (71 - , n) = 1 ,
we have xi( n) = n^~ (mo d p). Therefor e

(3.60 ) ^») = G )
is th e quadrati c residu e symbol . Clearl y Xir{ n) an d X7f( n) a r e distinc t Dirichle t
characters bu t thei r square s yiel d th e sam e quadrati c character . Mor e generally ,
if q = p\ • • - pr i s th e produc t o f distinc t prime s pj = 1 (mod 4) , the n ther e ar e 2 r
distinct Dirichle t character s x ( m ° d q) satisfying \ 2 — Xqi namely x = Xm ''' Xn r =
X7, say , fo r an y 7 = 7r i • • • irr wit h 7 7 = q.
3.7. Th e Jacobi-Dirichle t an d th e Jacobi-Kubot a symbols .
Let q b e th e produc t o f prime s = l(mo d 4 ) (no t necessaril y distinc t primes) .
The Jacob i symbo l Xq( n) = (~ ) c a n b e extended t o Gaussia n domai n Z[i ] in severa l
ways correspondin g t o representation s o f q as th e su m o f tw o square s
(3.61) q = u 2+v2 with(u,v )= l .
By requirin g w — u + iv t o b e primar y w e distinguis h u fro m v an d w e fix the sig n
of u. Not e that u = l(mo d 2 ) and i ; = u — l(mod 4) . A Gaussian intege r w — u + iv
is sai d t o b e primitiv e i f (u,v) = 1 , thu s a n od d it ; i s primitiv e i f (w,w) — 1. Fo r
any primar y primitiv e w w e defin e th e symbo l (^ ) : Z[i]—> > {0,1 , —1} by
z\ /Re wz^
<"»> (=) - ( I = F )
where o n th e righ t sid e i s the Jacob i symbol . Mor e explicitl y w e hav e
/ ~ Nf z\ fur — vs\
(3.63) ( _ )= ( _ _ ) , li z = r + rs

where q = ww. I f q i s prim e = l(mo d 4) , w e ge t tw o differen t symbol s (^ ) an d

(=|) whic h wer e considere d b y Dirichle t [Dir] . Throughou t w e call (~) th e Jacobi -
Dirichlet symbo l wheneve r w i s primary an d primitive .
We coul d introduc e th e Jacobi-Dirichle t symbol s usin g root s o f quadrati c con -
gruences: ther e i s one-to-on e correspondenc e betwee n th e root s o f
2
(3.64) UJ + 1 = 0(mo d q)
and th e factorization s q = ww wit h w — u + iv primar y whic h i s given b y
(3.65) u = —uv(mod q).
Here u denote s th e multiplicativ e invers e o f u modul o q. W e obtai n

i i z = r + l s
(3.66) ( « )= ( —)

by (f ) = ( ? ) = ( f t ) = ( g ) = 1 - Henc e i t i s clea r tha t

1(3.67) ( ) = Q if reZ.

For z = i we hav e

(«8> (s) l2 .
56 3. C H A R A C T E R S

Clearly (^ ) i s periodi c i n z o f perio d q, an d i t i s multiplicativ e a s wel l sinc e

(r*i + ujsi)(r 2 + ws 2) = r i r 2 - sis 2 + o;(ris 2 + r 2 si) (mo d g) .

EXERCISE 8 . Deriv e fro m th e quadrati c reciprocit y la w (3.31 ) th e followin g

reciprocity la w fo r th e Jacobi-Dirichle t symbo l

<-> (j ) - (f)
for an y z an d w which ar e primar y an d primitive .
For Gaussia n integer s z = r -\- is = 1 (mo d 2) , w e defin e

(3.70) W=^(H )
where ( A ) stand s for the Jacob i symbol. Not e that [z] vanishe s if z i s not primitive .
We are intereste d i n the multiplicativ e structur e o f [z] withi n th e Gaussia n rin g Z[i]
rather tha n wit h respec t t o th e co-ordinate s (r , s) G Z x Z. Fo r thi s reaso n w e refe r
to [z] a s th e Jacobi-Kubot a symbol . Indeed , thi s symbo l i s relevant t o th e Kubot a
homomorphism SX(2,Z )—> • {±1 } whic h ha s muc h t o d o wit h metaplecti c modula r
forms. O f course , [z] is not multiplicativ e i n th e stric t sense , ye t i t i s nearl y so , u p
to a facto r whic h i s the Jacobi-Dirichle t symbol .
EXERCISE 9 . Prov e tha t i f w i s primary primitiv e an d z = \ (mo d 2) , the n

(3.71) [wz]=e[w\{z}(^)

with e = ± 1 dependin g onl y o n th e quadrant s t o whic h z , w an d wz belon g (se e

the detail s i n [FI1 ]) .
Both symbol s o f Jacobi-Dirichle t an d Jacobi-Kubot a ar e utilize d i n th e proo f
of th e asymptoti c formul a fo r prime s p = X 2 + Y" 4. The y pla y a rol e throug h th e
estimation fo r genera l bilinea r form s i n [wz] (see Propositio n 21 . 4 o f [FI1 ]) .

3.8. Heck e c h a r a c t e r s .
E. Heck e [Heel ] generalize d th e Dirichle t character s t o an y numbe r field. Hi s
characters ar e multiplicativ e function s o n ideals . W e restric t ou r consideration s t o
imaginary quadrati c fields. Ever y suc h field is obtained b y a n extension o f rational s
by y/d, wher e d i s a negative , squarefre e integer . Denot e K — Q(\/d). Th e rin g o f
integers o f K i s a fre e Z-modul e O = Z + o;Z , wher e

d = 2(mod4) ,
d= l ( m o d 4 ) .

The discriminan t o f K
if d = 2,3(mod4) ,
D=
if d= l ( m o d 4 ) .
Note that i n an y cas e K = Q(y/D) an d O = Z + \{D + y/D)Z. Th e extensio n
has tw o automorphisms , th e identit y an d th e comple x conjugation . Th e grou p o f
 3. C H A R A C T E R S 5 7

units U C O i s finite, cyclic , generate d b y th e roo t o f unit y ( w — e27Tl^w o f orde r

w = 4,6 , 2, precisel y

I7 = { ± l , ± i } i f d = - l ,
(3.72) U = {±1 ,±UJ,±UJ 2} i f d = - 3 ,
U = {±1 } i f d < —3 .

The Kronecke r symbo l Xjr>(n ) = ( ^ ) i s a real primitiv e characte r o f conductor —D,

we cal l i t th e field character . Notic e tha t XD{— 1 ) = — 1 (se e (3.39)) . Ever y idea l
a ^ 0 i n O i s a fre e Z-module , an d th e rin g o f residu e classe s O/a i s finite, th e
number o f it s element s bein g th e nor m o f a

(3.73) Na = |{a(mo d a ) : a e 0}\.

The nor m i s multiplicative . I f a — (a) i s a principa l ideal , the n iV a = Na = \a\ 2.

Every idea l a ^ 0 factor s uniquel y int o power s o f prim e ideals . Ever y prim e idea l
of K appear s a s a facto r o f a rationa l prime . Th e la w o f factorizatio n o f rationa l
primes int o prim e ideal s i n K assert s th e following :
- i f XD(P) — 0 > then p ramifies ; (p) — p 2 wit h Np — p,
- i f XD(P) — 1? then p splits ; (p) = pp wit h p ^ p an d Np = p ,
- i f XD(P) — —1 > then p i s inert; (p) — p with Np = p 2.
An integra l idea l a ^ 0 which ha s n o rational integra l divisor s othe r tha n ± 1 i s
said to be primitive. Ever y primitiv e idea l can b e written uniquel y a s the Z-modul e

(3.74) a = a Z + \{b + y/D)Z = [a , \{b + V3)]

with a = iV a an d 6 the intege r determine d b y — a < b < a , b 2 = D(mo d 4a) . Not e

that w e hav e 6 2 — 4a c = D wit h (a , 6, c) = 1 . Thi s i s clea r fro m th e followin g
computation:

(a) = (aa ) = [a, \{b + y/D)][a, \{b - VB)]

2
= [a ,^(b+VD),^(b-VD),ac}
C [a 2, afr, ac] = a [a, 6, c] C (a) .

For a primitiv e idea l a we se t th e poin t i n th e uppe r half-plan e

b
(3.75) z a = A^~ e H.

The invers e idea l a - 1 i s th e fre e Z-modul e generate d b y 1 and z a ,

1
(3.76) a" =Z+6~Z .
-1
Let D denot e th e differen t o f K, s o by definitio n c> i s th e fractiona l idea l

d'1 = {a £ K : Tr a £ Z } .

In ou r cas e th e differen t o f th e quadrati c field K = Q(yD) i s th e principa l idea l

X> = (y/D). Not e tha t d = d an d Nd = |£>| .
58 3. C H A R A C T E R S

Let / b e th e grou p o f al l non-zer o fractiona l ideal s

— : ai,a2 C C 1 a1 a2 ^ 0
<*2

and P e l th e subgrou p o f principa l ideal s (a ) = aO wit h a G K*. Th e facto r

group H = I/P i s calle d th e clas s grou p o f if . I t i s a finit e abelia n grou p o f orde r
h = \H\ = [I : P] which is called the class number o f K (als o of the discriminant D).
Every clas s A i s represente d b y a uniqu e primitiv e idea l a wit h z a i n th e standar d
fundamental domai n (suc h idea l i s called reduced ) F = F~ U F+ , wher e
(3.77) F~ = {z = x + iy G H I : -\ < x < 0 an d \z\ > 1 }
+
F = {z = x + iy e M : 0 ^ x ^ | an d \z\ = 1 } .
Therefore th e clas s numbe r ft = h(D) i s equa l t o th e numbe r o f primitiv e reduce d
ideals. Fo r a primitiv e idea l writte n a s a = [a , ±(b + VD)} wit h b 2 -4ac = D w e
have \z a\2 = ^ , whenc e fo r a t o b e reduce d i t say s tha t eithe r — a < b ^ a < c or
0 ^ b < a = c . Conversel y an y integra l solutio n t o b 2 — Aac = D satisfyin g th e
above inequalitie s yield s a primitiv e reduce d idea l define d b y (3.74) .
Now we go to characters. Firs t w e consider Dirichle t character s o n O. Le t m be
a non-zer o integra l idea l i n O. Th e residu e classe s a (mo d m) , a G O , (a,m ) = 1
form a multiplicativ e grou p (0/m)* . An y homomorphis m

X : (O/m)* - > C*
is calle d a Dirichle t characte r modul o m . Naturall y w e exten d x t o O b y settin g
x(oi) — 0 i f (a,m ) ^ 1 . I f x i s a characte r t o modulu s m then (b y natura l restric -
tions) i t ca n b e als o viewe d a s a characte r t o an y modulu s n C m (recal l tha t thi s
inclusion o f ideal s mean s m divides n ) . Fo r an y x(mo d m ) an d an y f3 G (Dm)-1 w e
define th e Gaus s su m
(3.78) r x(/3)= J2 x(<*)e(Tr(a0))
a. (mo d m )

This i s well defined, becaus e i f a\ = a 2 (mo d m ) then (5{ai — a2) G ( M )_ 1m = D _1

and T r (/3(ai — a 2)) G Z . Th e sam e argumen t show s tha t r x (/3) depend s onl y o n f3
(mod D" 1 ), i.e . r x{pi) = r x (/3) i f 7 - / 3 G D"1. Clearl y
rx(a/3) = x(a)r x (/3) i f (a,m ) = 1 .
The conducto r o f x (mo d m ) i s the larges t idea l f Dm (tha t i s the smalles t diviso r
f o f m ) suc h tha t x factor s throug h (0/f)* . I f f = m the n th e characte r i s sai d t o
be primitive .
PROPOSITION 3.7 . If x (mo d m ) is primitive, then

2 J
(3.79) | r x(/3)| -n \ ^ \
I 0 otherwise.

After havin g introduce d character s o f (0/m)* , w e look fo r (unitary ) character s

of C*. B y definition thes e ar e continuou s homomorphism s o f C* into the uni t circl e
S1 = {a G C I \a\ = 1}.
 3. CHARACTER S 59

Any suc h characte r Xo o : ^ * —> S 1 i s give n b y

for som e £ G Z . W e cal l £ the frequenc y o f Xoo-

Next w e turn t o character s o n idea l classes . Recal l w e have alread y introduce d
the followin g groups :
- / th e multiplicativ e grou p o f fractiona l ideal s ^ 0 .
- P th e subgrou p o f principa l ideals .
- H = I/P th e idea l clas s group .
Now fo r an y integra l idea l m C 0 , m ^ 0 , w e se t th e subgroup s
Im = {ael | (a,m ) = l } c J ,
Pm = {(a)eP I (a,m ) = l } c P .
Here (a , m) = 1 means a = be - 1 fo r som e b , c C (9 , with (bc,m ) = 1 .
EXERCISE 1 0 . Sho w tha t (a,m ) = 1 mean s a = ^~ x fo r som e /3, 7 G O,
(/37,m) = l .
The facto r group s I m/Pm ar e isomorphi c t o I/P — H. An y homomorphis m o f
H int o S 1 i s calle d a clas s grou p character . Ther e ar e exactl y h = \H\ = \H\ clas s
group characters .
Finally w e introduc e th e Heck e character s ("Grossencharakteren") . A Heck e
character modul o m i s a continuou s homomorphis m ip : I m—> • S 1 fo r whic h ther e
exist tw o character s
X : {O/mf - > S 1
l
Xoo : C * - > S
such tha t
^((«)) = x(a)Xoo(c O
for ever y a G 0 , ( a , t n) = l . Not e that \ (mo d m ) an d Xo o are determined uniquel y
by ijj. Indeed w e hav e
XOO{OL) = %l)([a)) i f a = l (mo d m) .

Moreover, th e grou p { a G K * | a = 1 (mo d m) } i s dens e i n C * s o th e abov e

formula define s Xo o uniquely o n C * by continuity . Afte r havin g determine d Xo o we
find x b y
X(a) = ^((a))/xoo(a) .
Not ever y pai r o f character s x (mo d m ) an d Xo o come fro m a Heck e characte r
ip (mo d m) . Indeed , i f e G J7, the n w e hav e x(^)Xoo(^ ) = ^((^) ) = 1 - O n e c a n
show tha t th e conditio n
(3.80) X (e)Xoo{e) = 1 for all e G U
is sufficien t fo r th e pai r o f x (mo d m ) an d Xo o to resul t fro m a Heck e characte r ip
(mod m) . However , a pai r o f character s x (mo d m ) an d Xo o satisfyin g th e abov e
"units consistenc y condition " determin e ip (mo d m ) onl y u p t o a clas s grou p char -
acter, i.e . i t induce s exactl y h Heck e characters .
60 3. CHARACTER S

In ou r cas e th e grou p U i s cycli c o f orde r w = 4 , 6 or 2 generated b y th e roo t

of unity £ = e 2nl/w. Therefor e th e conditio n (3.80 ) reduce s t o tha t fo r e = £ , whic h
becomes

(3.81) X(C ) = C £-
In othe r words , thi s i s a conditio n fo r £ (mo d w). I f D < —4 , we hav e ( = — 1, s o
the unit s consistenc y conditio n become s
(3.82) < ? = A ( 1
1X ( - ) - ) (mo d 2) .

The conducto r o f a Heck e characte r ip (mo d m ) i s th e smalles t diviso r f o f m

such tha t tp is th e restrictio n o f a Heck e characte r t o modulu s f . Thi s i s th e sam e
as th e conducto r o f x (mo d m) . I f f = m , the n ip is said t o b e primitive . Therefor e
tp (mo d m) i s primitiv e i f an d onl y i f th e correspondin g \ (mo d m ) i s primitive .
The Heck e character s o f modulu s m = (1 ) an d o f th e frequenc y £ = 0 correspon d
exactly to the class group characters. Henc e there can be many primitive character s
of conducto r one .
Let ^ b e a primitiv e Heck e characte r o f conducto r m an d o f frequenc y £. W e
consider ip as a multiplicativ e functio n o n al l non-zer o integra l ideal s b y settin g
i/j(a) = 0 if (a , m) ^ 1 . T o ip we associat e th e Heck e L- function

L(s,V0= Yl V>(a)(iVa)- s.
0/aCO

This serie s converges absolutel y fo r R e (s) > 1 and ha s a n Eule r produc t ove r prim e
ideals

P
We als o introduc e th e loca l facto r a t th e infinit e plac e

(3.83) Loo(*,V 0 = (\D\Nm) s/2(27r)-sr(s+ ±\£\).

THEOREM 3. 8 ( H E C K E ) . Let \j) (mod m ) be primitive. Then the complete
product A(s,'0 ) = L OG(s,%p)L(s,ip) has analytic continuation to the whole complex
s-plane except for a simple pole at s = 1 when ip = ipo is the trivial character. In
this case m = (1 ) , £ = 0 and L(s, ipo) = CRT(S ) = ((S)L(S,XD) is just the Dedekind
zeta function of K, so
res A(s, t/>o) = hw 1_ .
s=l

Moreover, A(s , ip) satisfies the functional equation

(3.84) A(s , V) = WWOA( l - s ,V O
with the root number W(ip) given by the corresponding Gauss sum
(3.85) W{iP) = rV(^)(iVm)- 1 /2.
The Gaus s su m T(I/J) associate d wit h ip (mo d m ) i s define d b y

(3.86) ^ = W£ ^K^f )
aGc/cm
 3. CHARACTER S 61

where 7 G O and c C 0 ar e such that (c , m) = 1 , cDm = (7) . Not e that th e definitio n

of r(ip) doe s no t depen d o n th e choic e o f 7 an d c du e t o th e unit s consistenc y
condition (3.80) .
EXERCISE 1 1 . Sho w tha t fo r a primitiv e Heck e characte r ij; (mod m ) whos e
Dirichlet componen t i s \ (mo d m ) w e hav e
(3.87) r(V 0 = V>(a)r x(/3)
-1
for som e (o,m ) = 1 and / 3 G (Dm) .
EXERCISE 1 2 . Prov e tha t fo r a primitiv e Heck e characte r ip (mo d m) ,
(3.88) \T($)\ = (TVm) 1 /2.

EXERCISE 1 3 . I f ip modulo m an d £ modul o n ar e primitiv e Heck e character s

with (m , n) — 1 , the n
T ( ^ ) = V(n)£(m)TtyOr(£) .
REMARK. I f ip is a primitiv e Heck e characte r o f conducto r m = (1 ) an d fre -
quency £, then r(ip) = i l. I n thi s cas e th e functiona l equatio n (3.84 ) ha s th e roo t
number W(ip) = 1 . Sinc e ip(a) = ^(a ) w e find tha t L(s,ip) = L{s^) b y changin g
a int o a i n th e Dirichle t series .

If ^(nio d m ) ha s frequenc y £, the n it s comple x conjugat e ip (mod m ) ha s fre -

quency —£. Therefore , withou t los s o f generality , w e ca n assum e tha t ^(nioc l m )
has frequenc y £ ^ 0 . The n th e functio n / : HI -> C give n b y

(3.89) f( z) = J2 ^ ( *) (Na) * e{zNa)

a

is a n automorphi c for m o f weigh t k — £ + 1 o n TQ(N) o f leve l A ^ = |Z)|AT m an d

character x(mo d N) define d b y
(3.90) X{n) = XD(n)i;(n) i f n G Z.
If £ > 0 , the n / i s a cus p form , whic h i s primitiv e i f th e characte r ^(mo d m ) i s
primitive. I n particular , i f m = (1 ) , then N = \D\ an d x = XD-
REMARKS. W e say a few words abou t th e proof s o f the abov e results. Firs t th e
Fourier serie s (3.89 ) splits into a finite sum over ideal classes. I n each class we get a n
infinite serie s ove r equivalen t ideal s whic h i s a thet a functio n fo r a positiv e definit e
binary quadrati c for m twiste d b y a suitable harmonic polynomial (se e Section 1 4.3) .
Applying th e Poisso n summatio n formul a on e show s tha t thes e individua l thet a
functions ar e automorphi c form s o f th e sam e typ e (cf . Chapte r 1 0 of [1 4]) . Henc e
the / , bein g a linea r combinatio n o f thet a functions , i s als o a n automorphi c for m
of th e sam e type . Anothe r applicatio n o f Poisso n summatio n show s tha t th e thet a
functions satisf y a Jacob i typ e involutio n formula . Summin g ove r th e classe s thes e
formulas yiel d
„ k f ( z \ _ iT W^f - 1
i
' ( ^ )
(at thi s poin t i t i s crucia l tha t -0(mo d m ) i s primitive) , wher e r ( ^ ) i s th e Gaus s
sum an d N = \D\Nm . Pro m this , on e derive s th e functiona l equatio n (3.84 ) b y
taking th e Melli n transfor m wit h respec t t o z = iy. Se e als o Chapte r 1 4 .
62 3. C H A R A C T E R S

We en d thi s sectio n b y givin g specifi c example s o f Heck e character s o n imagi -

nary quadrati c fields.
EXAMPLE 1 . Fo r K = Q(i) th e discriminan t i s D — —4 , the rin g o f integer s i s
O — Z[i], the grou p o f unit s i s U — {±1, ±i} an d th e clas s numbe r i s h = 1 , thu s
every idea l i s principal. Fi x a n intege r £. Pu t m = (1 ) o r m = 2(1 -h i) accordin g t o
4 | £ or 4 \ L Defin e £ : Im -» C * b y

(3.91) « a )= (^) '

where a is the unique primary elemen t whic h generates a , with a = l(mo d 2( 1 -\-i)).
Then £ is a primitive characte r o f frequenc y £ and conducto r m .

EXAMPLE 2 . Le t K — Q(i ) b e a s above . Tak e a positiv e intege r q = 0(mod 4 )

and x : C^fy]/Q^IA)* ~* ^ * a characte r o n th e multiplicativ e grou p o f primitive
residue classe s modul o q in Z[i]. Defin e £ : Iq —> C * b y

««0 = X(«)Q'
where a is the uniqu e primar y elemen t whic h generate s a . The n £ is a Hecke
character (no t necessaril y primitive ) o f frequenc y £ and modulu s q.

EXAMPLE 3 . Le t K = Q(v^D ) hav e discriminan t D < - 3, D = l(mod 4) .

Then th e rin g o f integer s i s

(3.92) 0 = <U(r a + n\/l>) : ra,nG Z, ra = n ( m o d 2 ) j ,

the grou p o f units i s U = {±1} while th e clas s number h = /i(D) ca n b e arbitraril y

large. Sinc e the principa l idea l (y/~D) ha s norm \D\ the rin g inclusion Z c O define s
an isomorphis m

0/(VD)^Z/\D\Z
( 3 - 93 ) V { m + ny/D m , , , ^ IN
2^ y ( m o d I^D -
Composing \i with th e Jacob i symbo l we get a quadratic characte r e : O —> {0 , ±1},
explicitly

,3,4) , ( = ± ^) =@ .

Note tha t £ i s odd , i.e. , e(—a) = —£(a ) becaus e |D | = —l(mo d 4) . Give n £ odd,
we hav e h Hecke character s £ : J—>• C suc h tha t o n principa l ideal s a = (a),

e
(3.95) £(a) = e( a)(f:) .

These ar e primitiv e character s o f frequenc y £ and conducto r m = (VD).

 3. C H A R A C T E R S 63

EXAMPLE 4 . Le t K = Q(VD) b e an y imaginar y quadrati c field wit h w units .

Given £ = O(mo d w), w e hav e h = h(D) Heck e character s £ : / - > C * suc h tha t
£(a) = a/|a|) ^ o n principa l ideal s a = (a) . Thes e ar e primitiv e character s o f
frequency £ and conducto r m = (1 ) . Amon g thes e ther e i s a specia l characte r o f
frequency £ = hw define d b y £(a ) = (a/\a\) w fo r an y o G / , wher e (a) = a h.

EXAMPLE 5 . Le t K — Q(\/D) b e any imaginary quadrati c field of discriminan t

D, an d x(nio d q) a primitiv e Dirichle t characte r o f conducto r q wit h (q,D) = 1 .
Define th e homomorphis m \ ° N : Iq— > C * b y
(3.96) ( X oA0(a) = X (7Va).

Then £ = %oT V is a Hecke character o f frequency zer o which is primitive of conductor

m = (q). I n thi s cas e
2
(3.97) r(0 = XD(q)x(\D\)r (X )
where r(x ) i s th e Gaus s su m fo r th e Dirichle t characte r x(mo d q).
This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/05

CHAPTER 4

SUMMATION F O R M U L A S

4.1. Introduction .

Series o f arithmeti c function s ofte n underg o som e kin d o f involutar y transfor -

mations. Thes e ar e derived no t b y re-grouping an d changin g th e orde r o f terms bu t
deeper change s ar e mad e b y applyin g Fourie r analysis . Fo r example , th e equatio n

(4.1) Yl e-™2/2/ = VvYl z~™ 2y

raGZ n E Z

cannot b e verifie d simpl y b y a combinatorial argument . A ric h sourc e o f relation s

of this kin d lie s i n automorphi c theory . Fo r instance , fo r a classical cus p for m f(z)
of weight k on th e modula r grou p T = SL2 (Z) whos e Fourie r serie s is

00

(4.2) f(z) = ^2\(n)n^ 1 e(nz)

1

the automorph y equatio n f(z) = z~kf(—l/z) yield s (b y integratio n alon g th e ver -

tical lin e z — iy) th e formul a

00 0 0

(4.3) ^ A ( m ) ? H = ^A(n)Mn )
11

where g(x) i s an y smooth , compactl y supporte d functio n o n R + an d

/»oo

(4.4) h(y) = 2m k / J 1k- (47ry/^)g(x)dx.

Jo

See mor e genera l result s i n (4.71 ) an d (4.72) .

Modular transformation s ar e typica l bu t ther e ar e plent y o f other importan t

cases. Fo r exampl e th e formula s (1 5.30 ) (th e Selber g trac e formula ) o r (1 6.34 ) (th e
Kuznetsov formul a fo r sum s o f Kloosterma n sums ) d o no t com e fro m automorph y
equations. Naturall y a n equatio n connectin g on e serie s o f an arithmeti c functio n
(weighted b y a test functio n o f certai n class ) wit h anothe r i s called a summatio n
formula. I n thi s chapte r w e develo p a few suc h formula s an d giv e standar d appli -
cations.

65
66 4. SUMMATIO N F O R M U L A S

4.2. Th e Euler-Maclauri n formula .

We begi n b y considerin g a functio n define d o n a segmen t o f rea l numbers .
Suppose a , b £ Z an d / i s continuous i n [a , b\. The n th e su m o f f(n) ca n b e writte n
as th e Stieltje s integra l

(4.5) £ /(n) = f f(x)d[x]

where [x] denotes th e integra l par t o f x (se e (1 .65)) . Puttin g

(4.6) ^(x ) = x-[x]-\

we derive b y partia l integratio n th e followin g Euler-Maclauri n formula .

LEMMA 4.1 . For f of class C 1 on [a, b] with a < b,a,b £ Z we have

(4.7) Yl / W = / V ( s ) + Hx)f'(x))dx + |(/(6 ) - /(a)) .

This formul a produce s goo d result s whe n f f(x) i s relatively small . I f the highe r
order derivative s o f / exis t an d ar e ver y small , i t ca n b e profitabl e t o continu e
integrating b y parts . T o this en d w e nee d th e Bernoull i polynomial s Bk{X).
We defin e Bk(X) G Q[X] b y th e followin g recurrenc e conditions :

(4.8) B 0(X) = 1,
(4.9) B' k(X) = kB h-
1 (X),

(4.10) f B k(x)dx = 0,
Jo

ifA ; ^ 1 . Th e conditio n (4.9 ) define s B' k(X) i n term s o f Bk-i(X) u p t o a constan t

while the last conditio n (4.1 0 ) (th e orthogonality o f Bk(X) t o constants) determine s
the constant . Th e monomia l X k satisfie s th e firs t tw o condition s bu t i t fail s th e
third one . W e fin d tha t B k(X) = X k - | X f c _ 1 - f . . . . I n particular , w e hav e
Bi(X) = X — | an d B 2(X) = X 2 — X + | . Th e generatin g powe r serie s fo r th e
Bernoulli polynomial s Bk(X) i s

(4.ii) ^ > * )= I>Wfc!=^rr

The las t equalit y follow s b y noticin g tha t

— F(t,X) = Y 1iBk1
. (X)jj— yl=tF(t,X).
 4. SUMMATIO N F O R M U L A S 67

EXERCISE 1 . Prov e tha t

k
Bk(l-X) = (-l) Bk(X)
Bk{X + l)-B k{X) = kX k~1
x k
£B k(x+-)=q ~ Bk{qX).
0^a<q

Next w e define th e Bernoull i number s a s constan t term s o f the Bernoull i poly -

nomials

(4.12) B k = B k(0).

The generatin g powe r serie s fo r Bernoull i number s i s

tk t
(4.13) F(t) = Y,Bk
o fc! e * -1 '

Since F(-t) = t + F(t), i t follows tha t B k = 0ifk>l,k odd .

EXERCISE 2 . Prov e tha t

k k i
(4.14) B k(X) = ^2( e)BlX - .

REMARK. Th e Bernoull i polynomial s an d number s hav e bee n generalize d wit h

respect t o a character x ( m ° d q) a s follows :

0' l^^ g

The constan t ter m o f B ka(X) i s the generalize d Bernoull i number , namel y

Bka = B k,x(0) = q k~l Y. x(o)Bfe(-) -

0^a<q V

The generalize d Bernoull i number s occu r i n values o f L-function s a t specia l points .

They trul y belon g t o th e theor y o f p-adi c L-function s (du e t o Leopoldt , se e e.g .
[W]).

For an y fc^Owe pu t

(4.15) M*) = B k({x})

where {x} = x — [x] is the fractiona l par t o f x. Thes e function s ar e periodi c o f

period one , s o i s th e generatin g functio n
k 1
^t te ^
k=0
68 4. S U M M A T I O N F O R M U L A S

Therefore the y ca n b e represente d b y Fourier series . Th e n-t h Fourie r coefficien t o f

the generatin g functio n i s

/ ( E ^ y e ( - ^ ) ^ = IF3T/ el*-™****
JO 1 r, ' ^0

tW t \ *
t — 27rin V(—Y-
fe=i
Hence th e n-t h Fourie r coefficien t o f ^ ( x ) i s — k\(2irin)~k givin g th e expansio n
27Tin
(4.16) i/; k(x) = -k\ J2( )~ke(nx)

(there i s n o ter m n = 0 by (4.1 0)) . Thi s serie s converge s absolutel y i f k ^ 2 . Fo r

k = 1 we arrang e th e term s symmetricall y gettin g th e serie s
oo
-1
(4.17) ^(x) = - ^(Trn) sin(27rn^ )
l

which converge s pointwis e an d boundedl y excep t fo r x G Z.

EXERCISE 3 . Prov e tha t fo r an y N ^ 1 and x eR
N

(4.18) VO ) = - X^ 7™)"1 sin 27rnx

( ) + °((! + INI N)-1)
1

where ||x| | i s the distanc e o f x t o th e neares t intege r

(4.19) \\x\\ = min{| z - m\ : m G Z} .

REMARK. Evaluatin g (4.1 6 ) a t x = 0 we ge t

«2™> - ^ ^
for m ^ 1 . Thi s turn s ou t t o b e als o tru e fo r m = 0 b y th e functiona l equatio n
(4.75). Moreover , b y th e functiona l equatio n (4.75 ) i t follow s tha t fo r an y m ^ 1 ,
C(l-2m) = --?-B 2 m .
Furthermore, fo r a primitiv e characte r \ o f conducto r g > 1 one ha s
- 1 / - 2 7 r n m , N _J
rn,x

L(l-m,x)= — 5 m ,x

for an y m ^ 1 with ra = ±( 1 - x ( - l ) ) (mo d 2 ) (se e e.g . [IR] , [W]) .

Integrating b y part s fc — 1 times i n (4.7 ) an d usin g (4.9 ) on e derive s th e Euler -

Maclaurin formul a o f orde r k.
 4. SUMMATIO N F O R M U L A S 6 9

THEOREM 4.2 . Suppose f e C k([a, b]) with a < b, a, b e Z . Then

b / (_l\k
{k
(4.20) J2 fW = Ja (f(x)-t-j^M*)f \x))d*

Very ofte n th e firs t orde r Euler-Maclauri n formul a (4.7 ) i s a s goo d a s (4.20) .

Inserting (4.1 8 ) int o (4.7 ) an d integratin g b y part s w e ge t
COROLLARY 4.3 . For f e ^([a, b}) with a <b,a,b e Z we have

where N is any positive integer and the implied constant is absolute. Here and
hereafter the dash indicates that the terms at the end-points of summation are taken
with half values.

4.3. Th e Poisso n summatio n formula .

This sectio n i s based o n classical Fourier analysi s a s described i n the Appendix .
Recall tha t fo r an y functio n / G LX(IR) it s Fourie r transfor m i s denne d b y

(4.22) f(v) = I f(x)e(-xy)dx.

THEOREM 4.4 . Suppose that both /, / are in L l(M) and have bounded variation.
Then
m
(4-23) £ / ( ) = E/>0
where both series converge absolutely.
PROOF. Conside r th e functio n

which i s periodi c o f perio d one . Thi s ha s th e absolutel y convergen t Fourie r serie s

expansion
F(x) = y^CF(n)e(nx)
nez
with coefficient s give n b y
pl /»C O

cF(n)= F(t)e(-nt)dt= f(t)e(-nt)dt = f(n).

JO J-oo

Taking F(0 ) w e ge t th e Poisso n summatio n formul a (4.23) . •

70 4 . S U M M A T I ON F O R M U L A S

EXERCISE 4 . Deriv e (4.23 ) fro m (4.21 ) .

Changing f(x) into f(vx + u) wit h v G M + and u G R the formul a (4.23 )
generalizes to

(4.24) E / ( m , + .)_IE/ (=) e (!2).

rnGZ n £ Z

By th e Fourie r inversio n f[x) = f(—x) thi s ca n als o b e writte n a s

EXERCISE 5 . Usin g (4.24 ) prov e tha t fo r a primitive characte r \ (mo d g) ,

(4-26) J ; /MxM = ^ £ /©*(")

where r(x ) i s the Gaus s sum .
EXAMPLES. Pro m th e Fourie r pair s (4.83) , (4.84) , (4.85 ) w e obtain th e follow -
ing equations :
2
(4,7) E(l -M ) e ( „ l H „ E (Ei^>)
27r| l?
(4.28) ^ e(na;)e- " ' = (TO/)- 1 5 3 I ™ + x +i>\-2.
Tl 77 1

(4.29) Yl e(nx)e-Tn2/y = ^/y^e-*^*?*.

nm

for an y x G R and y G M + . Th e thir d equatio n generalize s (4.1 ) . Th e secon d

equation ha s a geometric serie s o n it s lef t side , s o b y a direct summatio n

sin h 2ny
Yje(nx)e-2^n ly
cosh 2ny — cos 27TX

The firs t su m ca n b e als o execute d b y usin g geometri c series ; it yields

1
(s[n7rxiy}\2 , {y} sm7rx(2[y] + 1)
E (l\ y- M\J(y\) e nx =
si n TTX ) y si n TTX

If y is a positive intege r thi s reduce s t o y _1 (sin7rx?//sin7rx) 2 .

The sam e metho d (averagin g o f integral translations ) work s in several variable s

giving
 4. S U M M A T I O N F O R M U L A S 7 1

THEOREM 4.5 . Suppose f is in the Schwartz class S(R £). Then

(4.30) £ /(m ) = £ /(„) .

The formula s (4.24 ) an d (4.25 ) exten d t o severa l variable s i n a simila r fashion .

The assumptio n tha t / i s a Schwart z functio n ca n b e weakene d considerably . I n
two dimensions th e Poisso n formul a (4.30 ) i s just th e trac e formul a fo r th e Laplac e
operator an d a n invarian t integra l operato r o n th e toru s IR 2 /Z 2 , se e [14].

4.4. Summatio n formula s fo r th e ball .

If / i s radial , the n s o i s / (se e Lemm a 4.1 7) . I n thi s cas e th e Poisso n formul a
(4.30) become s a summatio n formul a fo r th e arithmeti c functio n
fi{m) = | { m i , . . . , rri£ G Z : m\ + • • • + mj = m}\.
Precisely, i f g(x) i s smooth an d compactl y supporte d o n R + , the n
oo oo
(4.31) ^2r 2k(m)g(m)mi-i = ^r 2fc (n)/i(n)n^
o o
where h(y) i s the Hankel transform o f g(x) give n by (4.1 03) . Throughou t w e assume
that £ = 2f c is a n intege r > 1 , so fc is a hal f integer .
REMARKS. Th e conditio n tha t g i s compactl y supporte d ca n b e considerabl y
weakened. I t suffice s tha t bot h g an d h ar e o f type (a , /?) with a < | + \ < (3 (see
(4.104)). T o this en d notic e tha t
1 +e
r2k(m) < r a * -
for m ^ 1 , th e implie d constan t dependin g o n e an d k (i f k ^ 2 , the n thi s hold s
without e).

Although ther e i s n o contributio n fro m m = 0 o n th e lef t sid e o f (4.31 ) ther e

is stil l a contributio n fro m n = 0 o n th e righ t sid e o f (4.31 ) becaus e h(y) i s no t
compactly supporte d i n R + eve n i f g{x) is . I n practic e thi s contributio n yield s a
main term , s o it i s desire d t o singl e ou t th e zero-t h term . W e hav e r 2fc(0) = 1 an d
rk
ft y 2 dx
( )~ Tuj J ^){^y)
as y —> 0 by the asymptoti c J u(x) ~ x v jTY{y + 1). Henc e (4.31 ) ca n b e writte n a s
THEOREM 4. 6 (SUMMATIO N FORMUL A FO R A BALL) . Suppose g is smooth and
compactly supported on R + . Then
OO £. OO

(4.32) ^r 2k{rn)g{m)m^-i = ^ y M ( ^ ) + ^r 2/c(n)/i(n)n^-*,

where M(g) is the Mellin transform of g at s — ^^, i.e.,

POO
h
(4.33) M(g) = / g{x)x ^dx,
Jo
72 4 . SUMMATIO N F O R M U L A S

and h(y) is the Hankel type transform of g

/>o o

(4.34) h(y) = 7 T / g(x)J 1k- (27Ty/xy)dx.

Jo

COROLLARY 4. 7 (SUMMATIO N FORMUL A FO R A CIRCLE) . Suppose g is smooth

and compactly supported on R+ . Then
OO pQQ OO
m
(4.35) 2 _ V ( ) # ( m ) = 7 r / g(x)dx + y^r(m)h(m)
l^ °l
where
•>o o

(4.36) % ) = IT / g(x)J 0(27ry/x^)dx,

Jo
and both series converge absolutely.
Summation formula s fo r the circle wer e establishe d i n one form o r another b y
Hardy-Landau [HaLa ] an d Voronoi [Vor] . Ther e ar e plenty o f expressions fo r the
Bessel functio n i n th e kerne l o f (4.36) . Her e w e selec t a fe w particularl y usefu l
integral representations :

TTJO(Z) = / cos(zsmO)dO
Jo
OO

{t2 -l)-^sm(zt)dt
/
« r °. , , N , ~ f°° • f zw\ f z\dw
= 2 / smize n t)dt = 2 / si n — co s — — .
Jo Jo V2 / \2wJ w
Moreover, w e have th e following asymptoti c expansio n (se e (23.451.1 ) o f [GR])

(4.37) TT J0(z) = ( ^ ) * {cos(z - f ) + 1 sin( s - f ) + O (1 ) }

valid fo r z > 0. Insertin g thi s int o (4.36 ) w e get

/•O O
/
/i(y) = / (xt/)"^(x)cos(27r v Sy- l)dx
Jo
+ x
1 67 T / ( y)~*9(x)sin(2n\^-j)dx + °{ (xy)-i\g(x)\dxy

Integrating b y parts i n the second integra l w e obtain

/>o o

(4.38) h(y) = / (x2/)- i (/(x) C O S ( 2 T T V ^ - f )dx + 0(R(y))

Jo
where
/>0 0

(4.39) R(y)= (xy)-i(\g(x)\+x\g'(x)\)dx.

Jo
Next integratin g b y parts i n the first integra l w e obtain
1 f°°
(4.40) h{y) =/ (a;y)ty(x)sin(27rV^ - f )ds + 0(R(y)).
ty Jo
 4. SUMMATIO N F O R M U L A S 73

In the case of sphere (£ = 3 , k = § ) w e encounter th e Bessel function o f order | ,

which i s an elementar y function , namel y Ji{z) = (2/nz)^ sinz , s o (4.34 ) simplifie s
to
/»oo
h(y)= /
(xy)~ig(x)sin(27ry/xy)dx.
Jo
Setting G(x) = g(x)x~* an d H(y) = h(y)y* th e formula (4.32 ) withA; = | become s
COROLLARY 4. 8 (SUMMATIO N FORMUL A FO R A SPHERE) . Suppose G is smooth
and compactly supported on R + . Then

(4.41) y V 3(m)G(m) = 2T T / x*G(a:)da : + yV3(n)n-*JJ(n )

i^ oi
w/iere
/•OO

(4.42) # ( y ) = / G(x) sm{2<Ky/xy)dx,

Jo
and both series converge absolutely.
To illustrate th e result s w e apply th e summatio n formul a (4.35 ) t o improv e th e
error ter m i n th e Gaus s circl e proble m (se e (1 .70)) .
COROLLARY 4.9 . We have
r m = nX
(4.43) Yl ( ) + °(x*)-

P R O O F . W e shall establis h a n uppe r boun d fo r th e su m i n question , th e lowe r

bound ca n b e derive d b y applyin g simila r arguments . I n wha t follow s i t i s crucia l
that r(n) i s nonnegative, s o we can smoot h ou t th e su m b y enlargin g it s range . T o
this end w e choose the tes t functio n g{x) ^ 0 supported o n O ^ : r < X - H y i n whic h
segment g{x) — min{x , 1 , (X+ Y — x)Y~1}. Thi s i s not smooth , nevertheles s (4.35 )
holds true . Her e Y i s a paramete r a t ou r disposa l t o b e chose n late r t o minimiz e
the resultin g uppe r bound . W e assum e tha t H 7 ^ l 2 . B y (4.35 ) w e obtai n

r
r(m) < ^r(m)g(m) = 7 r(x+ - y - J + ^ (n)^n)'

where h(n) i s th e integra l transfor m o f g give n b y (4.36) . Usin g (4.40 ) an d (4.39 )

one show s tha t
(4.44) % ) < j r W ( l + 3//Z)- i
where Z = XY~ 2. Henc e
OO

1a b

Therefore
] T r(m) < TTX + 0(Y + X*Y~*).
m^X
Choosing Y = X^ w e balanc e th e erro r term s t o O(X^). B y a simila r argumen t
we establis h a lowe r boun d wit h th e sam e mai n an d erro r terms . Thi s complete s
the proo f o f (4.43) . •
 74 4. S U M M A T I O N F O R M U L A S

EXERCISE 6 . Prov e tha t

r(n)
(4.45) T r(m) (l - - ) = ^x + V ^ J 2(27rv^) = \x + 0(x~i).
1m^.x

Note th e erro r ter m i s very smal l du e t o smoothing .

Similarly t o (4.43 ) on e ca n deriv e fro m (4.41 ) th e followin g approximat e for -
mula:

(4.46) £r 3 (m) = yX§+0(X*).

The strongest erro r term her e was obtained b y Chamizo - Iwaniec [Chi ] and Heath -
Brown [HB5] . Interestingl y enoug h i n eithe r pape r characte r sum s wer e brough t
into pla y independently ; the y obtaine d th e exponent s | | an d | | i n plac e o f |
respectively.

4.5. Summatio n formula s fo r th e hyperbola .

First w e establis h summatio n formula s fo r th e diviso r functio n r(n). The y
are essentiall y du e t o Vorono i [Vor ] excep t fo r ou r arrangement s o f th e integra l
transforms.
THEOREM 4.1 0 . Suppose g(x) is smooth and compactly supported on M + . Let
ad=l (mo d c) . Then

(4.47 )

Y^r(m)9[—)cos(——) = /\ log
~ + 2yg(x)dx
, x (n\ /27rdn\
Ex r(n) Hc) C0S l^-)'
(4.48)
m

Ei r (tm\ ) nf 7\)•s i n/27rara \ ^- v , fn\ . /27rdn\

x
( — ) = X > <n H c M — ) •
where p(y) and q(y) are the integral transforms
/>oo

p(y) = / C(27Ty/xy)g(x)dx,
Jo
^(y) = / S(27r^xy)g(x)dx,
Jo
with kernels

C(z) = 4 / cos^w ) cos ( —y ) —

Jo v ^ w
S(z) — 4 / sin(2:u> ) sin( — ) —.
wJ w
Jo ^
 4. SUMMATIO N F O R M U L A S 7 5

These kernel s ar e also give n by (see (3.864.1) an d (3.864.2) o f [GRj)

C(z) = AK0(2z) - 2TTY 0(2Z),

S(z) = AK0(2z) + 27rY0{2z);

therefore bot h result s ca n also be written i n the following singl e formula :

^r(m)e(^ jg(m) = - / (logx+ 2j - 2logc)g(x)dx

(4.49) _ ^ g r(n)e(-^)jf~y0(^Vni)^)dx

4 ^. N ( dn\ f°° T ^ /4T T ^ - \ , N J

+ - /. r\n)e[—) / -ft-ol— \Jnx\g(x)dx.

We shal l deriv e (4.49 ) fro m a stil l mor e genera l summatio n formul a fo r the
tempered diviso r functio n
(4.50) T g(m)= ] P g(m um2).
mim2=m

Here g(x,y) i s a functio n o f clas s C 1 , compactl y supporte d o n R 2 . B y the way ,

the arithmeti c functio n r g(m) i s a prototype o f Fourier coefficient s o f automorphi c
forms i n the space o f continuous spectrum .
PROPOSITION 4.1 1 . Suppose g is a compactly supported function of class C l
2
onR . Let ad =1(mode) . Then

= ( n ) e
(4.5i) £ ^ M < v ) ^ (-T)
where h is given by the Fourier transform of g, namely

(4.52) h(x,y) = -g(-J).

c V c cJ
For n — 0 we have
(453) (
^<°>=/^OlH!}|;> ^>'"<'

P R O O F . Firs t w e open r g(m) o n the left sid e of (4.51) a s in (4.50) an d w e spli t

the summation int o residue classes (mi , 777-2) = (^1 , ^2) (mo d c) . The n in each clas s
we appl y th e Poisson summatio n formul a (4.24 ) gettin g

Y^rg(m)e(^ J = ^ T ^ e(^-u 1 u2J ^ ^ # ( ^ 1 + cv u u 2 + cv 2)

771
ui,u 2 (mo d c) « iV 2

= c~ 2 ^ ^ ] P ^2 e
c(auiu2 + niixi + n 2u2)gy—, — J
^ 1 ™ 2 iii,tt 2 ( m o d e )

^"TE e (44(7'7) = S^(-7)-

76 4. SUMMATIO N FORMULA S

The ter m n = 0 is special sinc e ther e ar e infinitel y man y divisor s o f zero . W e com -
pute Th(0) by reversing th e abov e analysis (o f course, thi s analysi s coul d b e avoide d
altogether, nevertheless , fo r th e econom y o f presentation sometime s a repeate d ap -
plication o f a n involutio n i s justified). Firs t w e arrang e th e summatio n ove r ni , r&2
as follows :

..I«o, 0 ) + il:j („,=) + i E ,(=,o).

nn

Then w e apply Poisson' s summatio n an d th e Euler-Maclauri n formul a (4.7 ) gettin g

/ / ( -c + { l 9{x y)dxdy
c}^) ' -
Similarly w e execut e th e summatio n o f #(^,0) . Gatherin g th e result s w e obtai n
(4.53). •

P R O O F O F (4.49) . On e woul d lik e to appl y (4.51 ) directl y fo r g(x,y) = g(xy),

but thi s i s no t possibl e becaus e suc h a functio n doe s no t hav e compac t suppor t
in R 2 eve n i f g(t) doe s hav e compac t suppor t i n R . Fo r thi s reaso n w e attac h
to #(771 1 777,2 ) redundan t factor s 77(7771 ) , 77(7712) t o localiz e th e rea l variable s x,y i n
compact segment s o f R + . Precisel y w e se t g(x,y) = 7](x)rj(y)g(xy) wher e rj(t) =
min{£/£, 1 } wit h 0 < e < 1 s o tha t r g(m) = r(m)g(m). Not e tha t rf(t) ha s
discontinuity a t t = e, bu t thi s i s allowed fo r applicatio n o f Proposition 4.1 1 as ca n
be see n fro m th e proo f b y inspectio n (w e coul d choos e rj(t) smooth , however , th e
involved computation s woul d have been less explicit). Fo r th e abov e choice we have
by (4.53 )

MO) = JJ{\ + {-C}1 + {\ }^Hy)9( Xy)dxdy

= - 9{y)L{y)dy,

where

WC^'G)^}! 3 ^))*-
Since y i s bounded fro m belo w an d above , an d e is small, w e find by examinin g th e
support o f th e involve d function s th e followin g representation :

^-/('-^{fD-x^Of-^AfWDi-
 4. S U M M A T I O N F O R M U L A S 7 7

The las t integra l range s ove r x > s~ 1 y, s o it is bounded triviall y b y 0(ec/y). The n
we hav e
f°° , N /y\dx n [^ , ,dx n ,
/ ^ H x J "/ v(x)— = 2 + logy-2loge,

The par t o f th e las t integra l wit h ^ < x < 1 equals lo g | an d th e remainin g par t
equals 1 — 7, wher e 7 is the Eule r constan t (se e (1 .69)) . Fro m thes e evaluation s w e
get
L(y) = logy + 27 - 21 og c + O ( - ) .

Hence

(4.55) 77,(0 ) = - c Jg(y)(logy + 27 - 2 log c)dy + 0(e).

As £ tends t o zer o thi s give s th e leadin g ter m o n th e righ t sid e o f (4.49) .

Now w e comput e Th(n) fo r n ^ 0 . Sinc e th e factorizatio n nin 2 = n permit s
the sig n chang e o f both n\,U2 w e ca n replac e g in (4.52) b y th e cosine-Fourie r
transform, i.e. , w e ma y tak e
9 f f 9-7T

h(u,v) = - / / rj(x)rj(y)g(xy) cos— (ux + vy)dxdy.

Changing variable s w e write thi s a s

M«,«) = f/«(»)(/^(*h(f)oo8(^(« a :+^))|)dy.

If w e omit r)(x)rj(y/x), the n th e pur e cosin e integra l equal s
0
f°° /2TT / ^ ^ ^ - i " ^ ^ ^ ifw>0 ,

(see (3.871 .2 ) an d (3.871 .4 ) o f [GR], respectively , o r ou r Appendix) . Thi s yield s

exactly th e remainin g term s o n th e righ t sid e o f (4.49) .
Now w e onl y nee d t o show tha t th e distortio n introduce d b y rj(x)rj(y/x) i s
negligible. W e denot e th e differenc e b y ho(u,v) = h\(u,v) + h\(v,u), wher e

hi(u,v) = -/ ( 1 ) / g{xy) cos— (ux + vy)dydx.

c v £J c
Jo Jo
Integrating b y parts twic e with respec t t o y and the n applyin g Fubini' s theore m w e
get
e r°° f£ f x\ 2TT
2
hi(u,v)= 2 2 / / [l )x g"(xy)cos—(ux + xy)dxdy.
78 4. SUMMATIO N F O R M U L A S

Now integrating b y part s twic e wit h respec t t o x w e estimate th e inne r integra l b y

0(eu~2). Th e inne r integra l i s als o estimate d triviall y b y
2 ,f
/x \g (xy)\dx<^y-3.
Jo
Combining bot h estimate s w e arriv e a t
2 2
hi(u,v) <^v~ / mm(su~ ,y~3)dy
Jo
Next, addin g th e correspondin g boun d fo r h\{v,u) w e obtain ho(u,v) < C £ 3 | i ^ | ~ 3.
Therefore th e tota l distortio n o f the righ t sid e of (4.51 ) , which wa s caused b y intro -
duction of the localizing factors r](x)r](y) t o the left side , is bounded b y 0(e 2/3) sinc e
the serie s ^2r(n)n~ 4^3 converges . Lettin g e ten d t o zer o th e distortio n vanishe s
and w e complete th e proo f o f (4.49) , henc e als o o f Theore m 4.1 0 . •

The summatio n formul a (4.51 ) fo r th e tempere d diviso r functio n (4.50 ) wa s

established i n [DI] . Th e specia l cas e (4.49 ) i s les s flexible , bu t i t i s quit e effectiv e
when applicabl e ( a different, rathe r indirect , proo f o f (4.49 ) wa s give n b y M . Jutil a
[Jul]). W e procee d t o compos e som e variation s o n (4.49) .
First, w e ge t a summatio n formul a fo r certai n serie s o f Kloosterma n sums : i f
(c,q) = 1 , the n

2^ T(m)S(h,m;c)g(m) = -5(/i,0;c ) / (lo g \-^)g(x)dx

m^l C J 0 \ C /

(4.56) - — Vr(n)S(/i-n,0;c ) / Y 0( — y/r^)g(x)dx

C
n>l J 0 \ C J

-\—Y^ T~(n)S(h -f- n,0;c) / K$( —y/nx)g(x)dx.

for an y intege r h. T o prov e this , w e open th e Kloosterma n su m (1 .56 )

e
S(fc,ra;c) = ^ {J >
a (mo d c )

then fo r eac h a , w e appl y (4.49 ) an d the n exchang e th e orde r o f summatio n again .

Note tha t th e Kloosterma n sum s o n th e lef t sid e collapse d t o th e Ramanuja n sum s
on th e lef t side . Thi s featur e i s significant i n applications , se e e.g. [DFI2] , [KM1 ] ,
[KMV1].
Next, playin g wit h additiv e character s w e derive fro m (4.49 ) a summatio n for -
mula fo r th e diviso r functio n i n a n arithmeti c progression .
COROLLARY 4.1 2 . Let (g,r ) = 1 , and g be smooth, compactly supported on
R+. Then
(4.57 )
V r(m)g(m) = ^ / (log x + 27 - 2 V(q))g(x)dx
=±1,,^
ra=r(mod q) ? Jo
+ \ E \c E r(n) {S(r, n ; c)T+ ( J) + S{r, - „; c)T~ ( £) }
QV
c\q
 4. SUMMATIO N F O R M U L A S 7 9

where rj(q) is the additive function

(4-58 ) -K.)-E^ .
P\Q

S(r,n\c) is the Kloosterman sum, and T +(y), T~(y) are the integral transforms

(4.59) T+(y) = -2T T / Y 0(4iryfty)g(x)dx,

Jo
POO

(4.60) T~(y) = 4 / K 0(4Tryfty)g(x)dx.

Jo

REMARKS. Sinc e YQ(Z) ha s th e asymptoti c expansio n (4.37 ) bu t wit h co s an d

sin interchange d (se e (23.451 .2 ) o f [GR]) , w e ge t th e approximat e formul a

T+(y) = — / (xy)ig'(x) c o s ( 4 7 r ^ - -)dx + 0(R(y))

ny Jo 4
by th e sam e argumen t whic h le d u s t o (4.40) , wher e R(y) i s given b y (4.39) . Simi -
larly on e derive s
4
T~(y) = — / (xy)ig'(x)e- ^dx + 0(R(y)).
^y Jo

EXERCISE 7 . Usin g the Wei l bound fo r Kloosterma n sum s (1 1 .1 6 ) an d th e tes t

function g(x) a s i n th e proo f o f (4.42) , deriv e fro m (4.57 ) tha t fo r (q,r) — 1 an d
X^2,
(4.61)
2
J2 r ( m ) - ^ X { l o g X + 27-l-277(g)}+ 0(r (g)(^+X*)logX)

m=r (mo d q)

where th e implie d constan t i s absolute . I n particular ,

r
(4.62) Y, (m) = x l
( °gx + 2 7 - 1 ) + O p d logX) .

This improve s th e erro r ter m i n th e Dirichle t diviso r proble m (1 .75) .

EXERCISE 8 . Prov e tha t fo r an y primitiv e characte r x(mo d q) with q > 1 , an d
any smoot h functio n g(x) compactl y supporte d o n 3R + w e hav e
oo oo
r
(4.63) ^ r(m)x(m)g(m) = 2
r (x)Q~ J2 2
(n)x(n)Mn2~2)
ii
where r(% ) i s th e Gaus s sum , an d h{y) i s th e integra l transfor m
/•O O

(4.64) h(y) = / K(2nyfty)g(x)dx

Jo
with kerne l
(4.65) K(z) = A X(-l)K0(2z) - 2TTY 0(2Z).

[Hint: Expan d \ int o additiv e character s (se e (3.1 2)). ]

80 4. SUMMATIO N FORMULA S

Next w e deriv e a summation formul a fo r

n
(4-66) T„{n, X)= £^ ^{~Y
niri2=n

where x(niod q) is a primitive character o f conductor q > 1 and v is a fixed complex

number. Thes e ar e eigenvalues o f the Heck e operators Tfi i n the spac e o f Eisenstei n
series fo r To(q) an d characte r x- I* 1 principle th e metho d tha t w e use d fo r r(n)
applies fo r 7v(n,x) , thoug h som e arithmetica l argument s nee d t o b e revise d an d
a fe w modification s i n integra l transform s ar e required . However , th e analysi s o f
convergence is the same, so here we do not repea t it . W e treat onl y the two extrema l
cases q\c and (g , c) = 1. Se e als o [KMVI ] fo r othe r simila r formulas .

THEOREM 4.1 3 . Let ad = 1 (mo d c ) and x be a primitive character of con-

ductor q > 1 with q\c. Then for any smooth, compactly supported function g on R +
we have

f^ r„(m , X )e(^)g(m) = ^ {^\{ X)L{

1 + 2i/, x) J°° g(x)x»dx

(4.67) + mf2T„(n,X)e(-^) J™g(x)J±^V^c)dx

+ ^ E ^ x ) e ( f) [ 9{x)Kt(^)dX

where r(x) is the Gauss sum, L( l + 2z/,x ) is the Dirichlet L-function, and

Kl(z) =4cos(7ru)K 2l/(z), ifx(-l) = 1,

K;v(z) = 4ism(irv)K 2v(z), » / x ( - l ) = - 1-

REMARK. Fo r i / = 0we hav e

(4-68) r(n )X ) = £x(d)-

If x is °d d (th e cas e o f modula r form s o f weigh t one) , ou r formul a reduce s t o

f>(m,x)e(^Wm) = ^-T(X)L(l,x) f°°g{x)dx

1 \ c /c J o

+ 27n^jrT(n,x)e(-^) J g(x)J 0(^-V^)dx.

 4. SUMMATIO N FORMULA S 8 1

P R O O F . O n th e lef t sid e o f (4.67 ) w e spli t th e summatio n int o residu e classe s

(mi, 777,2 ) = (ui, U2) (mo d c) and fo r eac h clas s we apply th e Poisso n formul a (4.24 )
getting

— ^2^2
1 Y1Y x(ui)e c(auiu2 - rixui - n 2U2)I(n
1 ,n2)
Tii n 2 ui,u 2 (mo d c )

where I(1 21 ,71 2) i s th e Fourie r integra l

I(nlln2)= / / (- J g(xy)e c(xni+yn2)dxdy.

JO JO y

The su m ove r w 2 (mo d c ) vanishes unles s au\ = 1 1 2 (mo d c) in which cas e it equal s
c. Therefor e w e showed tha t th e lef t sid e o f (4.67 ) i s equal t o

]P^x(n2)e(—ni7i2J/(ni,n2).
n\ n 2

Note tha t x{°) — x{d)- Fro m nin 2 = 0w e ge t

X(d) "
£X(n)[/(0,n) + x(-lK(0,-n) ]
c
1

2l/
= ^(^) L ( - 2 ^ , X ) ( 1 ° ° g ( x ) x v d x ) j°°{jv + xC-lJc^Jy-^^dy .

The las t integra l i s equa l t o 2T(—2v) COSTTV or — 2ir(—2v) smirv accordin g t o

x(—1) = 1 o r x(—1 ) = — 1 b y (4.1 08 ) an d (4.1 09 ) respectively . Her e w e trans -
form L(—2i/,x) m t ° L(l + 2i/,x ) b y th e functiona l equatio n

n -1
«-*-*-^Tw^ \*Tw ,VTT VTT / t -iv(i + ^ ) / r ( f - 1 / ) , i
It!! '.
f x ( - i ) = - 1,
see (4.73) . I n th e cas e o f eve n characte r w e encounte r

2 r
(^+^)w o ^ _ , _ o =
r(-2i/)cos7n/ - 2 2 -„ 2 ^
r(-i/)
and i n th e cas e o f od d characte r w e encounte r
-2r(i + i/)w 0 ; ^ _._ 0 _ ,2
-T(-2v)smTTV = 2 -2 v
^.
r a + i/)
Hence w e conclude tha t i n bot h case s the term s wit h n i n 2 = 0 contribut e th e sam e
amount whic h i s exactl y th e firs t ter m o n th e righ t sid e o f (4.67) . Pro m n i n 2 ^ 0
we ge t

^E E X(n 2)e(--cnin2) [J(m , n2 ) + x( - l ) / ( - n i , -n 2)]

+c ^ - ^ X(»2)e(-nin 2) [ J ( - n i , n2) + X ( - l ) J(ni, - n 2 ) ] -

82 4 . SUMMATIO N F O R M U L A S

Changing th e variables (x,y) - * {u^Jvn2/n\,u~ l\Jvn\jn2) w e ge t

J ( ± n i , ±772 ) = ( — ) / g(y) I /e c(>/?jn 177,2(±u ± u~ 1 ))u2v~ldu J dt> .

Hence th e remaining term s o n the right sid e of (4.67) emerg e b y applying (4.1 1 2) -
(4.115). •

REMARK. I n the proof o f Theorem 4.1 3 we assumed tha t — v is a small positiv e

number, bu t the result extend s t o all complex v by analytic continuation .
THEOREM 4.1 4 . Let ad = 1 (mod c) and x (mo d q) be a primitive character
of conductor q > 1 with (c , g) = 1 . Tfte n /o r any smooth, compactly supported
function g on M+ we have

J2 TA™, X)e[°f)g{rn) = x ( c ) c 2 ^ 1 L ( l - 2i/, X) J 9(x)x~ v

dx

1
(4.69) +^(x)^ £-^)e(-^)/o 5 (^(^>
c

X(c).
» T W -i2 T _v(n>x) C (^)jr f l (a:)^(^)dx
c

where qq = 1 (mo d c), and J ^ , K^ ar e £fte sam e a s m Theorem 4.1 3.

REMARK. Fo r v — 0 and x °d d our formula reduce s t o
(4.70)
^c / \ /»o o
£ r ( m , X ) e ( ^ ) c , ( m ) = ^ L ( l > x ) / sO*) ^

-2^iM£T(„,Be(^)^9(l)Jo(^)fc
P R O O F . O n the left sid e o f (4.69 ) w e split th e summation int o classe s rai=
Til (mod cq)i ^ 2 = U2 (mod c) an d fo r eac h clas s w e apply th e Poisso n formul a
(4.24) gettin g

^ E E ILYI x(wi)ec(awit£ 2-yMi-n2U2)/I/(y,n2)

n\ ™ 2 n i ( m o d cq)
u-2 (mo d c)

where Iv(z\,Z2) i s the same integra l a s in the proof o f Theorem 4.1 3 . Not e tha t

The su m over u 2 (mo d c) vanishes unles s au\ = 772 (mo d c) in which cas e it equal s
c. Puttin g u\ = dn2qq + ct , where £ ranges modul o g we obtain

=
E ^ S ce(-—nin 2Jx(-cni)r(x).
 4. SUMMATIO N FORMULA S 83

Hence th e lef t sid e o f (4.69 ) i s equa l t o

v v
n i ri2

This look s ver y simila r t o th e su m w e hav e alread y deal t wit h i n th e proo f o f

Theorem 4.1 3 . Therefor e (4.69 ) i s derive d fro m (4.67 ) b y makin g th e substitutio n
(x,v,d)— » (x , — v, dq) i n appropriat e places . •

Theorems 4.6 , 4.1 0 , 4.1 3 an d 4.1 4 ar e specia l case s o f summatio n formula s

for th e Fourie r coefficient s o f modula r forms . Actually , thes e result s confir m i n
themselves th e modularit y o f relevan t forms . A s a matte r o f fac t w e wer e abl e t o
establish suc h modula r relation s directl y b y applyin g th e two-dimensiona l Poisso n
summation onl y because in each case our arithmeti c functio n appear s i n the Fourie r
coefficients o f a GL2 for m whic h i s lifte d fro m GL\ forms , i.e. , th e correspondin g
L-function factor s int o two Dirichlet L- functions. Fo r genuine cusp form coefficient s
one canno t appl y th e ordinar y Poisson' s summatio n formula . I f on e know s tha t a
Fourier serie s

n
is a modular form , sa y / G Sk(q,x) ( s e e Chapte r 1 4 fo r notatio n an d surve y o f
holomorphic modula r forms) , the n th e correspondin g summatio n formul a fo r th e
coefficients A/(n ) woul d b e just anothe r expressio n fo r th e equatio n

EXERCISE 9 . Suppos e / i s a cusp for m o f weigh t k ^ 1 , leve l q ^ 1 an d

character x niodulo q. Le t c > 1 , c = 0(mo d q) an d ad = 1 (mo d c) . Prov e tha t
the Fourie r coefficient s o f / satisf y

(4.71) ±XM)e(^)9(rn) = ^±X f(n)e(^)h(n)

for an y smooth , compactl y supporte d functio n g o n R + , wher e

h(y) = 2m h / g(x)J k-i( — y/xy)dx.

Jo c
[Hint.] Appl y th e modula r equatio n fo r z = — - + — and integrat e th e resultin g
Fourier serie s o n bot h side s wit h respec t t o y agains t a suitabl e tes t functio n G(y).

If / i s primitiv e (se e Sectio n 1 4.7) , s o tha t i t satisfies th e Frick e involutio n

equation

with 1 77/ 1 = 1 (Proposition 1 4.1 4) , the n w e als o hav e

00 0 0
m
(4.72) 2 > / ( Mm) = W-
1 /2
^A/(n)%)
11
84 4. S U M M A T I O N F O R M U L A S

where h i s th e sam e Hankel-typ e integra l transfor m o f g a s abov e wit h c = ^Jq

(apply (4.71 ) fo r
(a b\ = ( 0 -l/y/q\
\c d) \y/q o ;
and chang e \(d) int o rjf t o ge t (4.70)) .
In moder n analyti c numbe r theor y ther e i s a grea t deman d fo r summatio n for -
mulas fo r arithmeti c function s o f typ e y(m) = a(m)f3(m - f h) wher e a(ra),/3(n )
are essentiall y modula r form s coefficient s an d h i s a fixe d intege r (bu t a unifor -
mity i n h i s crucial) . Thes e kin d o f problem s ar e ofte n encountere d whe n eval -
uating power-moment s o f familie s o f L-function s o n th e critica l line . See , fo r in -
stance, [DFI1 ] , [Sal] , [Ml] .
4.6. Functiona l equation s o f Dirichle t L-functions .
A summatio n formul a fo r arithmeti c function s ofte n ha s it s realizatio n a s a
functional equatio n fo r th e correspondin g generatin g Dirichle t series . Conversely ,
a functiona l equatio n connectin g tw o Dirichle t serie s lead s (b y wa y o f takin g th e
inverse o f Melli n transform ) t o a summatio n formul a fo r th e coefficien t sequences .
In thi s sectio n w e establis h thi s correspondenc e i n ful l detai l fo r th e Dirichle t L -
functions (3.6 ) usin g th e origina l metho d o f Riemann . Thes e idea s wer e develope d
further b y E . Heck e [Heel ] fo r L-function s o f numbe r field s an d fo r automorphi c
L-functions, an d late r pu t i n th e genera l adeli c settin g i n Tate' s Thesi s [Tal] .
Furthermore th e connection s betwee n functiona l equation s fo r twiste d L-function s
and modularit y vi a summatio n formula s i s th e essenc e o f th e so-calle d convers e
theorems i n th e theor y o f automorphi c form s i n general . Ther e ar e delicat e issue s
of compatibilit y whic h w e d o no t discus s here ; se e fo r exampl e th e Wei l convers e
theorem fo r GLi form s i n Theore m 1 4.21 , o r [1 4] .
Let q ^ 1 an d le t x b e a primitiv e characte r modul o q. W e le t S(x) = 0 i f
q jtz 1 and S(x) = 1 if < / = 1 , i n whic h cas e x = 1 a n d L(x , s) = ((s). Als o w e pu t
K = | ( 1 — x(—1 )) , s o K = 0 if x i s eve n an d n — 1 if x i s odd .
T H E O R E M 4.1 5 . With the above notation, L(s,x ) extends to a meromorphic
function on C which is entire if x 7 ^ 1 an d otherwise admits a unique simple pole
with residue 1 at s = 1 . The completed L-function

A(„)-(J)^r(i±i)L(.,x)
is entire ifqj^l and has simple poles with residue 1 at s = 0 and s = 1 otherwise.
Moreover, it satisfies the functional equation

(4.73) A ( s , X)=£(x)A(l-5,x)

where

(4.74) £ (x) = i -«I&.

Recall tha t th e Gaus s su m r(x ) i s o f modulu s yfq (se e (3.1 0) , (3.1 4) ) fo r x

primitive, s o tha t th e "roo t number " e(x) i s o f modulu s 1 . Theore m 3. 3 o f Gaus s
shows tha t e(x) = 1 for an y rea l primitiv e characte r x -
 4. SUMMATIO N F O R M U L A S 85

For th e Rieman n zet a function , th e functiona l equatio n i s therefor e

(4.75) A(s ) - 7r- s / 2 r(s/2)C(s) - A( l - s) .

PROOF. W e sho w th e argument s fo r q =/= 1 , th e cas e o f ((s) i s onl y slightl y

different becaus e o f th e pol e a t s — 1 . Le t 9(y,x) t> e the thet a functio n

nGZ

for y > 0. B y splittin g int o progression s modul o g , w e hav e

a (mo d q)

where
K
0(y,q,a)= J2 ri e-™2y'«.
n=a (mo d q)

We appl y th e Poisso n formul a (se e (4.24) ) t o thi s sum . Th e functio n e~ nx i s it s

own Fourier transfor m (4.29) . Differentiatin g th e correspondin g Fourie r integral , i t
follows tha t th e Fourie r transfor m o f f y(x) = x Ke~nx y i s give n b y

fy(t)=i-Ky-K-hy->{t).

By (4.24 ) w e obtai n

e(y,q,a.) = i-"y-"-hq-* £ e (-)e(-;q,b)

o (mo d q)

after splittin g agai n i n progressions . Usin g th e formul a (3.1 2 ) fo r a primitive char -

acter i t follow s tha t

(4.76) HV,X) = e(x)v- K~* 0 (~>x)-

Now w e appea l t o th e Melli n transfor m formul a (se e (4.1 07) )

+ 00 7
Kp-nn2y/q (S + K)/2 ^V_
ne

to deriv e

v
^'2 J0 y
for R e (s) > 1 . Spli t th e integra l a t y = 1 and appl y (4.76 ) fo r 0 < y < 1 getting

K/2
(4.77) (~) Hs,x) = \ l +°° {e(x)0(y,x)y {1 -s+K)/2 + 0(y,x)y (s+K)/2
}j
from whic h bot h th e analyti c continuatio n o f A(s,x ) (henc e tha t o f L(s,x) ) an d
the functiona l equatio n (4.73 ) follo w sinc e 9(y,\) decay s exponentiall y fas t a t oo. D
86 4. SUMMATIO N F O R M U L A S

REMARK. Thi s proof shoul d b e compared wit h that o f Theorem 1 4.7 , the func -
tional equatio n o f Heck e L-functions .

Quite ofte n i n application s a formul a fo r th e su m o f a n arithmeti c functio n

A(n) wit h a shar p cu t n ^ x i s mor e convenien t tha n tha t wit h smoot h ending . I n
this situatio n i t i s better t o wor k wit h a n approximat e formul a havin g a good erro r
term rather tha n wit h an exact infinit e serie s which is only conditionally convergent .
Here w e stat e a classica l resul t fo r a twiste d diviso r functio n whic h ca n b e derive d
from th e functiona l equatio n fo r th e Dirichle t L-serie s (se e [FI2]) .

THEOREM 4.1 6 . Let Xi,...,X d be primitive characters to moduli q\,...,qd re-

spectively. Put
Kn) = Yl Xi(ni)>-Xd(n d).
n\ •••nd=n

For any 1 ^ y ^ x we have

Y X(n) = R(x) + w ^- ] T -jz(-) cos(2^(- j+ - )

Here q = q\ • • • q d, u = d — 3 — 2(« i + • • • + no) and w is the root number of the

L-function L(s) = L(s , xi) • • • L(s, Xd) ( w is a product of the normalized Gauss
sums). The main term R(x) is equal to the residue of s~ 1 xsL(s) at s = 1 , so
R(x) = xP(\ogx), where P(x) is a polynomial of degree ^ d. The implied constant
depends only on d and e.

1
1 d-
Estimating th e dua l su m triviall y an d choosin g y = qd+i xd+i w e ge t

h£
] T X(n) = R{x)^0(qdTixdTT- )

where the implie d constan t depend s onl y o n d and e. This , of course, i s not th e bes t
error term . Usin g exponentia l sum s method s on e ca n improv e th e resul t slightly .
The bes t erro r ter m wit h respec t t o x whic h on e ca n hop e fo r i s 0(x2 -2d - — ++ £
).
Indeed, th e first ter m o f th e dua l su m i s abou t a s large . Not e tha t th e Exponen t
Pair Hypothesi s (se e Chapte r 8 ) doe s yiel d th e bes t estimate , whil e th e Rieman n
Hypothesis doe s not !

4.A. Appendix : Fourie r integral s an d series .

Let L 1 (M) denot e th e spac e o f Lebesgue integrabl e function s o n R. Th e Fourie r

transform o f / G L1 (R) i s define d b y

(4.78) Tf(y) = f(y) = f f(x)e(-xy)dx.

Jm
 4. SUMMATIO N F O R M U L A S 87

Though f(x) i s no t necessaril y continuou s it s Fourie r transfor m f(y) i s uniforml y

continuous an d f(y) —) • 0 as \y\ —> o o (th e Riemann-Lebesgu e lemma) . Th e convo -
lution o f tw o function s / , g G L1 (R) i s defined b y

(4.79) (f*g)(x)= If{*-y)g{y)dy

for almos t al l x, an d i t belong s t o L 1 (R), indee d th e L 1 -norm satisfie s | | / * p | | i ^

H/llill^lli- Th e convolutio n i s a smoothin g operator , precisely , i f g ha s th e firs t j
derivatives i n L 1 (R), the n s o doe s f*g fo r an y / i n L X (R).
The Fourier transform f(y) determine s its original f(x). Precisely , i f / G L1 (R),
then
/ (l-y)Kv)e(xy)dy^f{x)

as Y —> o o i n th e L 1 -norm. I f bot h / , / ar e i n L 1 (R), the n th e abov e integra l

converges uniforml y i n x givin g

(4.80) f(x) = j f{y)e(xy)dy.

JR

In othe r words , f(x) = f(—x). Thi s i s the inversio n formul a fo r th e Fourie r trans -
form. Th e Fourie r transfor m o f p(Y) = max{l — M^ , 0 } is calle d th e Feje r kerne l
Y

We have f*g = f-g, i n particular, <p*f = p-f, whic h show s tha t th e linea r spac e
of function s i n L 1 (R) havin g compactl y supporte d Fourie r transfor m i s dense .
Here ar e basi c Fourie r pairs :
{ 1 if \x\ < 1
sin 2iry
(4.82) f{x) = \ i f M = i f(y) iry
v 0 if |x | > 1 ,
2
(4.83) / ( * ) = m a x { l - H , 0 } , f(y) = ( ! ^ ) ,

(4.84) /(x ) = e- 2^l, /(y)=7r- 1

(l + y 2)-1 ,
v
(4.85) f(x) = e-™\ Hv) = e-* \
1
(4-86) /(* ) = -J—, f(y)
ch TTX c h Try

There i s a correspondin g Fourie r analysi s o f function s o n th e circl e T =

These ar e regarded a s functions o n R which ar e periodic o f period one . Th e Fourie r
coefficient o f a functio n / G Ll(T) i s define d b y

(4.87) c n(f) = [ f(x)e{-nx)dx.

Jo
88 4 . SUMMATIO N FORMULA S

They for m th e Fourie r serie s fo r /

(4.88) f(x)^Y,Cn(f)e(nx)
n£Z

which i s nothin g bu t a forma l expressio n a s lon g a s th e convergenc e i s no t estab -

lished. Th e Riemann-Lebesgu e lemm a assert s tha t c n{f)— > 0 a s \n\ —> oc , and , o f
course, thi s i s not sufficien t fo r th e convergenc e o f (4.88) . Nevertheless , th e Fourie r
coefficients c n(f) determin e / , precisely , i f c n(f) = 0 fo r al l n , the n / = 0 almos t
everywhere.
The convolutio n

(4.89) (f*9)(*)= I f(x-y)g(y)dy

Jo
is defined almos t everywhere , an d ha s it s Fourie r coefficient s equa l t o th e produc t

Cn(f*g) =Cn(f)Cn(9)-
In particular , convolvin g / G Ll{T) wit h a trigonometri c polynomia l
P c
N(X) = Yl n(P)e(nx)
\n\^N

gives a trigonometri c polynomia l

(f*PN)(x)= ^2 c n{f)cn(PN)e(nx).
\n\^N

This trick solve s many summabilit y problems . Th e mos t popula r i s the Feje r kerne l

N
(4.90) F N(x) = £ ( l " y )e(nx) = {-^r) •
\n\^N

This produce s th e Feje r partia l sum s

(4.91) J2 (l-^)c„(/)e(nx )
\n\^N

which d o converg e t o f(x) i n th e L 1 -norm o n T . Th e exac t partia l sum s o f th e

Fourier serie s fo r / ca n b e tailore d similarl y b y convolvin g / agains t th e Dirichle t
kernel
/.^s^ / x v ^ / x sin7r(27 V + l)x
4.92 D N(x) = > e(nx) = ^ —
*-^ sin7n r

for an y positiv e intege r N.

Concerning a pointwis e convergenc e fo r symmetri c partia l sum s w e hav e
c
(4.93) J2 nU)<nx) -+ § ( / (* + 0) + f{x - 0) )
\n\^N

for al l x G T provide d / G L l{T) ha s bounde d variation . Th e convergenc e i s

uniform o n close d interval s o f continuit y o f / .
 4. SUMMATIO N F O R M U L A S 89

In man y way s th e proble m o f representin g a periodic functio n b y it s Fourie r

series simplifies i f / i s square integrabl e becaus e L 2(T) i s a Hilber t space , th e inne r
product bein g
(f,9) = / f(x)g(x)da
and th e additiv e character s e n(x) = e(nx) for m a complete orthonorma l system .
Note tha t ||/|| i ^ ||/|| 2 b y Cauchy-Schwar z inequality , s o L 2(T) C L1^). Fo r an y
/ , g i n L 2 (T) w e have th e Parseva l formul a

(4.94) 5> n (/)c n ( 5 ) = {f,g).

n

Hence th e Fourie r coefficient s ar e squar e summable . Th e symmetri c partia l sum s

of th e Fourie r serie s converg e t o th e functio n i n th e L 2 -norm.
Next w e conside r function s i n severa l variable s x = ( x i , . .. ,#& ) G R fc. Say
/ : Rk -» C i s a Schwartz functio n i f
A
(4.95) f^(x)^\x\-
for an y a = ( a i , . . . , o^) an d A ^ 0, wher e \x\ 2 = x\ -f • • • + x\. Le t 5(R fc ) denot e
the clas s o f Schwart z functions . Fo r an y / G S(W k) w e se t th e Fourie r transfor m

(4.96) f(y) = f /(.x)e(—x • y)dx

where x -y — x\y\ -\ h XkVk is the scala r produc t i n R k. Th e Fourie r transfor m

A;
maps th e spac e *S(IR ) into itsel f (chec k (4.95 ) fo r f(y) b y partia l integration) , an d

it ha s th e followin g properties : f(x) — f(—x\ (f,g) = (f,g), explicitly ,

(4.97) f{x) = [ f(y)e(x • y)dy,

(4.98) / f{x)g{x)dx = f f(y)g(y)dy.

JRk JR k

In othe r words , th e Fourie r transfor m i s a n isometr y o n <S(M fc). Th e convolutio n

(4.99) (f*9)(*)= [ f(x-y)g(y)dy

is turne d int o th e produc t T(f • g) = f • g.

One ma y thin k o f R^ a s a homogeneous spac e acte d o n b y th e grou p o f trans -
lations G = R k. I n additio n t o th e translation s th e orthogona l grou p SO(R k) act s
on R k by rotation s makin g R k the euclidea n space . Th e riemannia n metri c o n R k
(of curvatur e zero ) i s given b y th e differentia l ds 2 = (dxi) 2 H h (dxk)2:, an d th e
corresponding Laplac e operato r i s
d2 d 2

^• 10 °) D =
M + '" + M-
The exponentia l function s <p(x) = e(x • y) fo r y G R fc ar e eigenfunction s o f D,
precisely (D + X((p))(f = 0, wit h eigenvalu e X(ip) — 47r 2|y|2. A n importan t poin t i s
that th e Laplac e operato r i s rotation invarian t (i.e. , th e rotation s ar e isometrie s o f
90 4 . SUMMATIO N F O R M U L A S

Rfc), therefor e th e Fourie r transfor m o f a radia l functio n i s als o radial . Precisely ,

employing th e Fourier inversio n (4.97 ) on e shows
LEMMA 4.1 7 . Let k^2. Put v = § - 1 . Suppose

(4.101) f(x)=g(\x\*)\xr
where g is a smooth compactly supported function on R+ . Then
2
(4.102) f(y) = h(\y\ )\y\-"

with

(4.103) h(y) =7T / J v(2v^y)g(x)dx

Jo

where J v(x) is the Bess el function of order v.

If P i s a homogeneous polynomia l i n k variable s an d complex coefficient s suc h
that DP = 0 (it i s called a spherical harmonic) , the n w e have th e Fourier pai r
d
F(x) = P{x)e~^\ F(y) = i P(y)e-*M2

where d = de g P. Thi s self-dualit y (du e to Hecke ) actuall y characterize s spherica l

harmonics. I n principl e i t say s tha t th e multiplication b y spherical harmonic s ha s
little effec t o n Fourie r transform . Fo r example , Lemm a 4.1 7 generalize s t o th e
following equatio n (du e to Bochner )

T(P(x)g(\x\2)\xD = P(y)h(\y\ 2
)\yr
where h i s give n b y (4.1 03 ) wit h th e Besse l functio n o f orde r increase d fro m v t o
v + d (recall tha t v = § — 1). Thes e fact s appea r naturall y i n the context o f unitary
representations.
We en d thi s surve y o f classica l Fourie r analysi s b y comment s o n th e Melli n
transform o n R+. Sa y / : R+ - > C is of type (a, /?) if

(4.104) f(y)y 1s
~ e L X (R+) fo r a < Re s < f3.

For suc h / t h e Melli n transfor m

/>oo
1s
(4.105) M(f)(s) = / f(y)y - dy
Jo

is defined fo r complex variable s in the vertical strip a < Re s < /?. Clearly M(f)(s)
is holomorphi c i n this strip . Sinc e th e Melli n transfor m o n R + i s merely a versio n
of th e Fourie r transfor m o n R derive d b y th e chang e o f variable s y — ex, s = it,
the result s fo r th e latte r translat e appropriatel y fo r th e former . Fo r example i f /
is continuou s o f typ e (a , ft) and ha s bounded variation , the n th e Melli n inversio n
formula holds :

(4.106) /(!/ ) = 5 ^ / M(f)(8)y-ds

where (a) indicate s tha t th e integration i s on the vertical lin e R e (s) = a.

 4. S U M M A T I O N F O R M U L A S 9 1

Here i s a selectio n o f Melli n transform s fro m Chapte r 3 o f [GR ] (formula s

(381.4), (761 .9) , (761 .4) , (41 1 .3) , (523.3) , (222.2) , (293.3) , (871 .4) , (871 .3) , (871 .2) ,
(871.1) respectively )
•>o o
y1
8
(4.107) / e- y - dy = T(s), a > 0,
Jo
1
s
(4.108) / (cosy)y - dy = T(s)cos~- , 0 < cr < 1 ,
Jo 2
/>00
s l
(4.109) / (smy)y - dy = 1
r(s)sm—, - <a < 1,

(4.110) / (l + y)- 1 ys-1 dy=-. , 0 < a < 1,

(4.111) / ^( l + y ) ^ - 1 dy
. = —:, - 1 < cr < 0.

For x > 0 we hav e

1 7TS
(4-112) J cos(^(y-^)y- dy = 2K(x)cos—,
s(:

7T 5
(4.113) ^ sin( | (y - fyv^'dy = 2K s(x) si n

7TS
(4.114) / cos(-(y + -jjy s~1 dy = -7rJ s(x)sm-— -7ry s (x)cos

7T 5
(4.115) / sin( | [y + -^y s~xdy = TTJ S{X) CO S y - 7ry s(s) sin
T"
In th e las t fou r formula s — 1 < cr < 1 an d J s , / ^ , ! ^ ar e th e standar d Besse l
functions.
In th e las t tw o formula s w e als o hav e

(4.116) J s (x)Siny+ys(a;)cosT= 2 s i n ^•

and

(4.117) J J x j c o s — — Y s(x) sm — = — .
v) s V) y J
2 2 2cos ^
This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/06

CHAPTER 5

CLASSICAL ANALYTI C
THEORY O F L-FUNCTION S
We wil l no w discus s a ver y classica l par t o f analyti c numbe r theory . Indeed ,
most o f th e result s discusse d i n thi s chapte r ca n b e trace d bac k i n som e for m
to Riemann' s Memoi r o n th e zet a function . The y wer e subsequentl y extende d
to Dirichle t characters , wit h tw o essentiall y ne w phenomeno n appearing . On e i s
that uniformit y o f estimate s wit h respec t t o th e conducto r i s vital . Th e other ,
not unrelated , i s th e stubbor n appearanc e o f th e so-calle d exceptiona l zero s (o r
Landau-Siegel zeros ) fo r quadrati c characters . Rulin g ou t th e existenc e o f suc h
zeros remain s on e o f th e mai n problem s o f numbe r theory .
Nowadays mor e refine d L-function s occu r frequentl y (se e Chapter s 1 4 an d 1 5
for a survey o f automorphic forms , on e of the sources for L-functions) , bu t i t i s well-
known tha t th e classica l analyti c result s remai n vali d i n grea t generality . Findin g
explicit statement s i n th e literature , unifor m i n al l parameters , i s howeve r some -
times difficult . Fo r thi s reaso n w e prov e al l mai n result s i n a genera l abstrac t
context. Then , startin g fro m Sectio n 5. 9 , we survey th e mos t importan t classe s o f
L-functions tha t aris e i n practice :
(1) Dirichle t L-functions ; se e Sectio n 5.9 .
(2) Dedekin d zet a function s an d L-function s o f Hecke Grossencharaktere n o f
number fields ; se e Sectio n 5.1 0 .
(3) Th e automorphi c L-function s o f classica l cus p form s (tha t i s t o sa y o f
holomorphic an d Maas s form s o n GL(2)) ; se e Sectio n 5.1 1 .
(4) Genera l automorphi c L-function s o n GL(ra) , m ^ 1 ; see Sectio n 5.1 2 .
(5) Th e L-function s o f algebrai c varietie s an d Galoi s representation s (al -
though mos t o f thos e ar e onl y conjecture d t o fi t th e framewor k discusse d i n thi s
chapter); se e Section s 5.1 3 an d 5.1 4 .
In eac h o f these case s we spell out th e concret e meanin g o f the abstrac t results ,
and giv e a fe w additiona l results .

5.1. Definition s an d preliminaries .

In orde r t o trea t al l th e L-function s tha t interes t u s i n on e stroke , w e in -

troduce som e abstrac t definitions , althoug h w e ar e no t tryin g t o creat e axioms .
The alternativ e i s to introduc e automorphi c L-function s a t th e outse t (se e Sectio n
5.12), bu t som e translation woul d stil l be require d fo r th e mos t classica l case s (an d
Rankin-Selberg convolution s woul d requir e a separat e treatment) . W e tr y t o bal -
ance familiarit y an d eas e o f us e wit h sufficien t generality . Th e reade r unfamilia r
with L-functions , o r acquainte d onl y wit h Dirichle t L-functions , shoul d rea d th e
introductory paragraph s o f Sectio n 5. 9 first , the n rea d wha t follow s wit h thi s im -
portant cas e i n mind . Basi c idea s ar e alread y presen t i n th e contex t o f Dirichle t
L-functions. W e the n encourag e hi m o r he r t o ge t acquainte d wit h automorphi c
93
94 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

L-functions (se e [BG ] for instance) . Th e analog y with Dirichlet o r Hecke character s
on ideal s i n a number field d o no t reall y revea l th e beaut y an d th e dept h o f the
genuine modula r forms . Thi s i s especially tru e whe n Rankin-Selber g L- functions
occur, whic h ar e instrumenta l fo r establishin g zero-fre e regions .
We denot e L-function s b y L(/ , s), L(g, s) , althoug h th e symbol s /, g carry n o
specific meanin g unti l Sectio n 5.9 . The y ar e use d merel y fo r th e suggestio n tha t
L-functions usuall y aris e a s attache d t o som e interestin g arithmeti c object .
We sa y tha t L(f,s) i s an L- function i f we hav e th e followin g dat a an d condi -
tions:
(1) A Dirichle t serie s wit h Eule r produc t o f degre e d ^ 1 ,

(5.1) L(f, a) = £ X f(n)n-S = JJ (1 - ^(p)?-)" 1 • • • (1 - ^(pjp-)" 1

with A/(l ) = 1 , Xf(n) G C, cti(p) G C. Th e serie s an d Eule r product s mus t be

absolutely convergen t fo r Re (s) > 1. Th e cti(p), 1 < i < d, are calle d th e loca l
roots o r loca l parameter s o f L(f, s) a t p , an d the y satisf y
for a1 1
(5.2) |<*i(p) l <P P-
(2) A gamm a facto r

ds/2
(5-3) 7 ( / , S)=7r- nr(£^)
J =l

where th e number s Kj G C are calle d th e loca l parameter s o f L(/, s) a t infinity.

We assum e thes e number s ar e eithe r rea l o r come i n conjugate pairs . Moreover ,
Re (K,J) > — 1. Thi s las t conditio n tell s u s tha t 7 ( / , s) ha s n o zer o i n C an d n o pol e
for R e (s) ^ 1 .
(3) A n intege r q(f) ^ 1 , called th e conducto r o f L(/ , s), suc h tha t ai{p) ^ 0
for p \ q(f) an d 1 ^ i ^ d. A prime p \ q(f) i s sai d t o b e unramified .
Prom these , th e so-calle d complet e L- function

(5.4) A(f,s) = q(f)2 1 (f,s)L(f,s)

is defined . Clearly , i t is holomorphic i n the half-plan e R e (s) > 1, ye t it mus t
admit analyti c continuatio n t o a meromorphic functio n fo r s G C of order 1 (see
the Appendix) , wit h a t most pole s a t s = 0 and s = 1. Moreover , i t must satisf y
the functiona l equatio n
(5.5) A(f, 3)=e(f)A(f,l-s).

where / i s an objec t associate d wit h / (th e dua l o f /) fo r whic h Xf(n) = A/(n),

7 ( / , 5 ) = y(f.s). q(f) = q(f) and s(f) is a complex numbe r o f absolute valu e 1,
called th e "roo t number " o f L(/ , s).
We le t r(f) denote th e orde r o f the pol e (i f positive ) o r zero (i f negative) of
A(/, s) at s = 0 and 5 = 1 , equal because o f (5.5) . Sinc e 7(7,1 ) 7 ^ 0, 00, r(f) is also
the orde r o f the pol e o r zer o of L(/ , s) at s =1. Late r w e shall sho w i n al l concret e
cases tha t L(/ , s) doe s no t vanis h a t s = 1, so r ( / ) ^ 0.
We see k unifor m estimate s fo r various analyti c quantitie s relate d t o L(f,s).
For a n individua l L- function, th e onl y paramete r i s s G C, bu t whe n L(f,s) varies ,
we also have to deal with th e degree, conductor , loca l parameters, an d eac h of thes e
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 95

or a combinatio n ma y be unde r investigation . I t turn s ou t tha t mos t result s fo r

L(f, s) are expressed convenientl y i n terms o f the analytic conductor . Firs t pu t
d

(5.6) qoo(5 ) = n^ + ^' l + 3 )-

i=i

Multiplying thi s b y q(f) w e get the analytic conducto r

d
(5.7) q(/ , s) = g(/)qo o W - q(f) U (\s + K,\ + 3).
j=i

We als o denot e
d
q(/) = q(/,0) = <?(/) n(K-1 + 3)-
Note tha t q(/ ) ^ 3 d g(/), s o d < logq(/), an d tha t
d
(5.8) q ( / , S)<q(/)(|S|+3) ,
so estimates could be performed i n terms of this last quantit y without compromisin g
much o f strength .
From th e definition , w e see that i f L(/ , s) an d L(g,s) ar e L-functions , the n
L(f,s)L(g,s) i s one with conducto r q(f)q(g) an d analytic conducto r q(/ , s)q(#, s),
gamma facto r 7(^,5)7(^,5) , roo t numbe r s(f)e(g). Moreover , L ( / , 5) i s als o a n
L-function b y construction, wit h sam e degree , conductor , gamm a facto r an d wit h
s(f) — e(f). Als o i f L(/ , s) i s entire , the n fo r an y fixe d t G R, th e shifte d L -
function L(g, s) = L(/, s + zt)L(/, 5 — it) is another L-function , wit h gamm a facto r
7(17, 5) = 7'(/ , s + it)j(f, s — ii), conducto r q(g) = q(f) 2, roo t numbe r e{g) — 1. The
analytic conducto r i s q(# , 5) = q(/ , 5 + it)q(f, s — it). Th e condition tha t L(/ , 5)
be entir e i s required becaus e w e are not allowe d t o shif t th e poles. W e take th e
product o f two shifte d L-function s b y it an d —it to maintain th e local parameter s
coming i n complex pairs . O f course, thes e ar e only cosmeti c arrangements .
In the sequel we often simplif y th e above notatio n b y not displaying th e depen-
dence o n /, 5 ; we write q = q(f), q^ = q ^ s) , q = q(/ , 5), r = r(f).
If / = / , the n L(/ , 5) is said to be self-dual. Thi s means that th e Dirichlet serie s
of the L-function ha s real coefficients . Fo r a self-dual L-function , th e root numbe r
is real , henc e e(f) = ± 1 . I t i s the n calle d th e sig n o f th e functiona l equation .
The followin g observatio n i s originall y du e to Shimura ; despit e it s simplicity , i t
is significan t i n a numbe r o f application s (se e Sectio n 23. 6 and Chapte r 2 6 for
instance):
PROPOSITION 5.1 . Let L(/, 5) be self-dual with s(f) = - 1. Then L(/ , | ) = 0.
PROOF. B y definition w e have 7 ( /, |) 7 ^ 0, and the functional equatio n applie d
to th e central poin t 5 = \ yield s L ( / , |) = — L(/, | ) , hence L ( / , \) = 0 . •

The loca l root s ai(p) ar e well define d u p to permutatio n o f the indices. I f for
any i w e have |a^(p) | = 1 for al l p \ q(f) an d |a^(p) | ^ 1 otherwise, the n L(/ , 5)
is sai d t o satisfy th e Ramanujan-Petersson conjecture . Thi s implies , i n particular ,
96 5 . CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

that |A/(n) | < Td(n). Similarly , i f fo r an y j w e hav e Re(Kj) ^ 0 , the n L(f,s) i s

said t o satisf y th e (generalized ) Selber g Conjecture , o r th e Ramanujan-Petersso n
conjecture a t th e infinit e place . I n thi s cas e j(f, s) ha s n o pole s fo r R e (s) > 0.
Because A(/ , s) = 7 ( / , s)L(/, s) i s assumed t o b e holomorphic , excep t possibl y
at s = 0 and s = 1 , it follows that pole s of 7(/, s) a t s ^ 0 are zeros of L ( /, s). Thes e
are called th e "trivia l zeros" . The y ar e located exactl y a t th e point s —2 m — K,J ^ 0 ,
1 ^ j ^ d , wher e ra ^ 0 i s a n integer . Fo r example , th e trivia l zero s o f £(s ) ar e
at s = - 2 , - 4 , - 6 , . . . whil e £(0 ) = —\. Th e othe r zero s o f L ( / , s) ar e calle d th e
non-trivial zeros . The y ar e locate d i n the critica l stri p 0 ^ R e (s) ^ 1 . Presumabl y
some trivia l zero s ma y la y i n th e critica l stri p a s well , however , on e conjecture s
they ar e neve r i n th e interior .
As mentioned i n th e introduction , w e will discus s example s startin g i n Sectio n
5.9. I t i s important t o note that man y interestin g L-serie s exis t beyon d ou r context ,
either becaus e o f lac k o f strengt h i n ou r fingers t o prov e th e require d assumption s
(for instanc e Arti n L-function s fo r non-trivia l irreducibl e character s ar e no t know n
to b e entir e i n general , an d genera l L-function s o f varietie s ar e no t eve n know n t o
be meromorphic , se e Section s 5.1 3 an d 5.1 4) , o r becaus e som e o f th e condition s
fail. Fo r instance , i n th e secon d clas s com e Dirichle t L-function s o f non-primitiv e
characters (the y hav e Eule r produc t bu t th e functiona l equatio n ha s extr a factors) ,
L-functions o f genera l modula r form s (se e Theore m 1 4. 7 fo r instance ) whic h hav e
no Euler product s an d a functional equatio n relatin g L(/ , s) wit h som e function no t
directly relate d t o L ( / , s). Stil l interesting ar e partial zet a function s o f ideal classe s
of numbe r fields (se e (22.55 ) fo r imaginar y quadrati c fields) whic h hav e functiona l
equations (relatin g a n idea l clas s a wit h D — a , wher e c ) is th e different) , bu t n o
Euler produc t i f the clas s numbe r i s > 1 . However , thes e partia l zet a function s ca n
be reconstructe d a s linea r combination s o f th e ful l L-function s fo r th e clas s grou p
characters.
The moder n analyti c theor y o f L-functions i s much influence d b y two concepts :
that o f a famil y o f L-functions , an d tha t o f (Langlands ) functoriality . Th e forme r
lacks a formal definitio n bu t i s a strong guiding principle (se e [M2]) . I n this book we
will indicat e whe n example s o f familie s appear . Not e tha t i n thi s context , varyin g
the imaginar y par t t o f s should coun t a s varying i n a family. Th e mai n parameter s
in a famil y o f L-function s ar e encode d i n th e analyti c conductor .
Functoriality, o n th e othe r hand , reflect s th e principl e tha t sometime s ne w L -
functions ar e buil t ou t o f ol d one s b y makin g som e operatio n (simpl e looking , ye t
quite subtle ) o n thei r coefficient s A/(n) , suc h a s raisin g t o som e powe r (Xf(n k)
or Xf(n) k). A prope r formalizatio n require s quit e majo r notatio n an d setup , an d
the naiv e idea s mus t usuall y b e changed ; moreover , muc h o f th e globa l settin g i s
still conjectural. W e will do here with a simple-minded definitio n o f Rankin-Selber g
type L-functions, an d mentio n elsewhere the symmetric square an d other symmetri c
power L-function s (se e Sectio n 5.1 2 an d th e discussio n o f the Sato-Tat e conjectur e
in Chapte r 21 ) .
Let L ( / , s) an d L(g,s) b e L-functions o f degrees d and e , with local component s
at infinit y KI an d v^ an d loca l root s (a^(p) ) an d {/3j(p)) respectively . Fo r p \
q(f)q(9), le t
1
(5.9) L p(f ® g, s) = J] ( 1 - OiipWjWp-')- .
 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S 97

We say that / an d g have a Rankin-Selberg convolutio n i f there exists an L-functio n

L(f ® g,s) o f degre e de suc h tha t

8
L(f®g,s)= n L p(f®g,s) J ]H p(p~ )
p\q(f)q(g) p\q(f)q(g)

where
de
s
(5.10) fl p(p- ) = n ( l - 7 j ( p ) p - a r 1 wit h| 7 ,-(P)|<P.
3=1

The gamm a facto r mus t b e writte n a s

7(/®,lS) = , -f a / 2 n r ( ^ )
with Re(fiij) < Re(« i + fj) an d |/x^- | ^ |K; | + | ^ j | , where ^ an d Vj ar e th e loca l
components a t infinit y o f / an d g. I n addition , th e conducto r o f / ® g must divid e
q(f)eq(g)d an d L(f<8>g, s) mus t hav e a pole at s — 1 if # = / . I n fact, unles s L(/ , 5)
or L(# , 5) factor , w e will se e tha t i n concret e case s L(f (g ) g, s) i s entire i f g ^ / .
We cal l L(f ® g,s) th e Rankin-Selber g L-functio n o r convolutio n o f / an d g ,
or th e Rankin-Selber g squar e i f g = f.
Note tha t i f L{f <S> / , 5) o r L(f 0 / , s) exists , the n th e estimate s o n th e loca l
roots an d o n Kj ca n b e improve d t o |c^(p) | < -y/p and R e (KJ) > — \.
The condition s als o impl y th e inequalit y fo r th e analyti c conducto r
e
(5-11) *(f®g:S)^q(f) q(g)d(\s\ + 3) de

(see (5.8)) . Althoug h it i s by no means obvious that L(f®g,s) shoul d exist, i t is the
case whe n L(/ , s) an d L(g, s) ar e automorphi c //-functions , a s discusse d i n Sectio n
5.12. Thi s cover s mos t case s o f importanc e i n analyti c numbe r theor y (includin g
the L-function s o f numbe r fields). Notic e tha t th e dua l L-functio n i s L ( / 0 g, s).
5.2. Approximation s t o L-functions .
We begi n wit h a crud e resul t whic h will b e use d extensively , an d improve d
later.
LEMMA 5.2 . Any L-function L(f, s) is polynomially bounded in vertical strips
s ~ a + it with a ^ a ^b, \t\ > 1 .
PROOF. Th e L-functio n i s bounded i n the half-plan e a ^ 1 + e by the absolut e
convergence of the Dirichle t series . Henc e we deduce a polynomial boun d fo r L(/ , s)
in the half-plane a < — e by the functional equatio n an d the Stirling formula (5.1 1 4) .
Then the polynomial bound in the remaining strip follows by the Phragmen-Lindelo f
principle (se e Theore m 5.53) . •

We no w stat e a n exac t formula , know n a s th e "approximat e functiona l equa -

tion", whic h gives analytically convenien t expression s fo r L(/ , s) i n the critica l stri p
where th e serie s doe s no t converg e absolutely .
98 5 . CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

THEOREM 5.3 . Let L(f,s) be an L-function. Let G(u) be any function which
is holomorphic and bounded in the strip — 4 < Re (u) < 4, even, and normalized by
G(0) = 1 . Let X > 0. Then for s in the strip 0 ^ a ^ 1 we have

( , 12 , «,,., =? M0V.(^)+^..,E^-.(^)+I»
where V s (y) is a smooth function defined by

j(f,s + u)du
(5.i3) v s(y) = ^i J v uG(u)
7C/» u
(3)

and

1
(5-14) e(f,s)=e(f)q(f) 2-^^s)S).

The last term R = 0 if A(fJs) is entire, otherwise

« = ( re s + „ )^L^m x..
S
\u=l — s u= — s/ 0 /Z7 f,S ) U

PROOF. Conside r th e integra l

I(XJ,s) = -±i [x»A(f,s + u)G(u)^.

(3)

This integra l exist s becaus e A(/ , 5) decay s rapidl y a s t —> +0 0 fo r fixed a b y

Stirling's formul a (se e (5.1 1 3)) . Fo r th e sam e reaso n on e ca n mov e th e integratio n
to R e (u) = — 3 by Lemm a 5.2 . Applyin g th e functiona l equatio n ther e yield s

s/
(5.15) A(/ , s) = I(X, f, s) + e(f)I(X-\f, 1 - s) + Rq Hf,«)

where A(/ , s) come s fro m th e simpl e pol e o f u~ 1 G(u) a t u = 0 and th e las t ter m R
comes fro m possibl e residue s at u = 1 — 5 and u = — s if A(/, s) i s not entire .
By expandin g int o absolutel y convergen t Dirichle t serie s w e hav e

n>l (3 )

We d o th e sam e fo r I(X : , /, 1 — 5) an d combin e the m wit h (5.1 5) . Dividin g bot h

sides b y g s / 2 7(/, s) yield s (5.1 2) . •
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 99

EXERCISE 1 . Let L(f,s) b e an L-function whic h i s entire. Fo r a smooth ,

compactly supporte d functio n F on M+ with Melli n transfor m F, prove tha t

(5.16) Y. WW = e_U±
^ £ M»W»)
where

(3)

(More explici t formulas , i n the case of classical automorphi c forms , ar e described

in Sectio n 4.5 ; in particular, se e the formula (4.71 ) i n which th e test functio n goe s
through th e Hankel transform) .
Show tha t i f L(/, s) = £(s) is the Riemann zet a function , (5.1 6 ) hold s wit h an
additional ter m
/>+o o
R = re s C(s)F(s) = / F(y)dy
s==1
Jo
on th e right side . Prov e tha t i n this cas e
r+oc
H(y) = 2 / F(y) cos(2irxy)dy,
Jo
/o
and (5.1 6 ) i s the Poisson summatio n formula .
E X E R C I S E 2 . Le t L(/, s) be the L-function associate d wit h a holomorphic prim -
itive cus p for m / o f weight k = 2 and level g (think o f the Hasse-Wei l zet a functio n
of a n elliptic curve) . Deriv e fro m (5.1 2 ) th e following formul a fo r the central valu e

where X is any positive number [Hint : tak e G(u) = 1.] Differentiatin g wit h respec t
to X this yield s the summation formul a

A/(n) / 2nn \2 xV^A/(n) / 2irnX\

E /jP

Taking X = 1 we get for self-dual L-functio n wit h roo t numbe r e(f) — 1 ,

I( /.i)= S E^-p(-^)-

For suitabl e tes t function s G(u), both sum s in (5.12) ar e effectively limite d to
the term s wit h n < C \Al(/> s )* W e shall se e this clearl y fo r a particular choic e of
G(u),
G(u) = (co s —J
where A is a positive integer .
100 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

PROPOSITION 5.4 . Suppose R e (s - f K,J) ^ 3 a > 0 for 1 < j ^ d. Then the

derivatives ofV s(y) satisfy
a
(5-17) y V}a\y)^(l + J=y A
,
a
(5-18) y V}a\y) = 5 a + 0((^=y) V
^ \Alo o 7 7

where 5Q = 1 , S a = 0 if a > 0 an d £/i e implied constants depend only on a, a, A

and d. Recall that q^, = qoo(s ) is given by (5.6).
P R O O F . Fo r s an d u wit h R e (s) = a > 0 an d R e (u) = (3 > —a we deriv e b y
Stirling's formul a (5.1 1 2 )

~ W« ' i exp (2 (|s| -| s + u | ) )

«(H+3)' 3 exp(|H;
where th e implie d constan t depend s onl y o n a an d /3 . Henc e
7(/,s + w ) 5/ /7r d
<q^2exp(y|U|).
7(/,s)
We hav e
s
yavs{a\y) = ^ Jy- u
G{u){-ufsalUi + u )du
7(/,s) ^ '
(3)
Moving th e integratio n t o th e lin e R e (u) = fi = —a we derive (5.1 7 ) whil e movin g
to th e lin e R e (u) = A. we deriv e th e boun d 0((y/q^/y) A). Thi s combine d wit h
(5.17) yield s (5.1 8) . •

The formul a (5.1 2 ) expresse s L(/ , s) i n the critica l strip , essentially , a s two par -
tial sum s o f th e Dirichle t serie s L(f,s) an d th e dua l one , eac h o f lengt h \ / q ( / , s).
Hence on e ca n easil y deriv e estimate s i n th e critica l strip . W e giv e her e th e indi -
vidual estimate , an d th e unifor m versio n fo r L-function s satisfyin g th e Ramanujan -
Petersson conjecture . Se e Sectio n 5.1 2 fo r stronge r result s i n th e cas e o f automor -
phic L-functions .
Because L(f,s) ca n hav e a pol e o r zer o a t s = 1 of orde r r , i t i s appropriat e t o
kill this singularity befor e estimating. Therefor e w e shall often estimat e p r(s)L(f) s)
rather tha n L(/ , s), wher e
r
(5.19) Pr(s)= ( i _ il V
\

EXERCISE 3 . Prov e tha t fo r s wit h 0 < a < 1 we hav e

1
(5.20) p r(5)L(/,,)«q(/,5)( -)/2^
for an y £ > 0 , wit h th e implie d constan t dependin g o n e an d / . I f L ( / , s) satisfie s
the Ramanujan-Petersso n conjecture , the n th e implie d constan t depend s onl y o n e
and th e degre e o f th e L-function . [Hint : Us e th e absolut e convergenc e o f L(/ , s)
in a > 1 (o r th e boun d |Aj(n) | ^ T^(n) , respectively) , th e functiona l equation ,
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 101

Stirling's estimat e fo r th e gamm a functio n an d th e Phragmen-Lindelo f convexit y

principle.]

In particular , (5.20 ) yield s th e followin g boun d o n th e critica l lin e

+£
(5.21) L ( / , S)«q(/,S)I

which i s known a s the convexit y bound. Usin g the approximat e functiona l equatio n
rather tha n th e convexit y principle , an d th e Ramanujan-Petersso n conjecture , on e
can deriv e a slightl y bette r estimat e

(5.22) L(f, s) « q(/ , s) i (lo g q(/, s ) ) 4 " 1

where the implied constan t depend s only on the degree d of L(/, s). I t i s conjecture d
that
L(/,S)«q(/,s)£
for R e (s) = \ wit h an y e > 0 , th e implie d constan t dependin g onl y o n e. Thi s i s
the so-calle d Lindelo f Hypothesis , whic h i s on e o f th e consequence s o f th e Gran d
Riemann Hypothesi s (se e Corollary 5.20) . However , i n many application s the majo r
step i s t o improv e o n th e convexit y boun d fo r relevan t L- functions, i n th e sens e o f
reducing th e exponen t \ b y a positive number , howeve r smal l that numbe r ma y be .
Such result s g o back t o H . Wey l fo r th e Rieman n zet a function , namel y
+£
C(s) « \s\-*
(see (8.22)) . Fo r Dirichle t characters , on e derive s fro m Burgess' s estimat e fo r shor t
character sum s tha t
L(X,s)^\s\qi6+£
(see Theore m 1 2.9) , whic h i s wors e tha n (5.21 ) i n ^-aspect , bu t muc h bette r i n
^-aspect: th e latte r i s ofte n th e mos t importan t (e.g . fo r application s t o algebrai c
number theory) .
In vie w o f th e proo f o f (5.22 ) b y th e formul a (5.1 2 ) i t i s clea r tha t an y sub -
convexity estimat e amount s t o provin g tha t ther e i s considerable cancellatio n whe n
_i_.
summing th e oscillatin g value s Xf(n)n 2 % ? and thi s prove s t o b e a n arithmetica l
problem.
Currently, i t is known that a subconvexity boun d hold s in ^-aspect an d ^-aspec t
separately (bu t no t ye t fo r bot h jointly ) fo r an y L-functio n attache d t o classi -
cal modula r form s (se e th e survey s [M2] , [IS1 ] , an d th e paper s [DFI2] , [DFI3] ,
[KMV2], [PSa] , [Ml] , amon g others) .
5.3. Countin g zero s o f L- functions.
One o f th e deepes t subject s o f th e theor y o f L-function s i s th e distributio n o f
the zero s o f L(f,s). W e discus s firs t wha t ca n b e don e usin g basi c propertie s o f
holomorphic function s throug h th e metho d o f Hadamard an d d e l a Valle e Poussin .
LEMMA 5.5 . Let L(f,s) be an L-function. All zeros p of A(f Js) are in the
critical strip 0 ^ a ^ 1 . For any £ > 0, we have
102 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

PROOF. Sinc e th e Eule r produc t expansio n o f L(/ , s) i s absolutel y convergen t

and 7 ( / , s) doe s no t vanis h fo r R e (s) > 1 , ther e ar e n o zero s o f A(/ , s) i n thi s
region. B y th e functiona l equation , th e sam e hold s fo r R e (s) < 0 . Th e secon d
statement i s a genera l resul t o f comple x analysi s fo r entir e function s o f orde r 1 . •

Notice tha t i f p = ft + r y i s a zer o o f A(/ , s), the n p i s a zer o o f th e dua l

A(/, s). Hence , b y th e functiona l equation , th e poin t reflecte d i n th e critica l lin e
p* = 1 — p = 1 — / 3 -H7 i s als o a zer o o f A(/ , s) (wit h correspondin g multiplicity) .
Of course, ther e i s no symmetry amon g th e trivia l zero s of L(/, s) whic h com e fro m
the pole s o f T ( ^ ) .
THEOREM 5.6 . Let L(f,s) be an L-function. There exist constants a = a(f)
and b = b(f) such that
ss
(5.23) (,( 1 - s)) rA(f, s) = e^ bs J J ( l - -)e P /p

where p ranges over all zeros of A(/, s) different from 0,1 . Hence

(5.24) --(^ = 1^ ^ ) - ^ ^ ^- E (^- + l).

both expressions being uniformly absolutely convergent in compact subsets which
have no zeros or poles (recall that r is the order of the pole or zero of L ( / , s) at
8=1).
PROOF. Th e expansio n (5.23 ) i s simpl y th e applicatio n o f th e Hadamar d fac -
torization theore m o f entir e function s o f finite order , an d (5.24 ) follow s b y takin g
the logarithmi c derivative . •

We denot e
s
(5.25) _ ^ ( /)S) = £ A/(n)n-

the expansio n o f th e logarithmi c derivativ e o f a n L-functio n i n Dirichle t serie s

supported o n prim e powers . I n term s o f th e loca l root s ai{p) o f th e Eule r produc t
(5.7) w e hav e
d
k k
(5.26) A f(p ) = Y^^(p) logp.
3 =1

For Dirichle t characters , A x (n) = x(n)A(n) , an d i n genera l A/(p ) = Xf(p) logp fo r

p prime. Not e tha t Aj(n ) = A/(n) .
PROPOSITION 5.7 . Let L(/, s) be an L-function of degree d ^ 1 and let p denote
the zeros of A(/, s) different from 0,1 .
(1) The number of zeros p = (3 + ij such that \j — T\ < 1 , say m(T, f), satisfies

(5.27) m ( T , / ) < lo g q(/,zT)

with an absolute implied constant.
 5. CLASSICA L ANALYTI C T H E O R Y O F L
1 -FUNCTION S 0 3

(2) For any s in the strip — ^^a^2we have

\S + Kj\<l J | S —p|<l

with an absolute implied constant.

(3) The constant b(f) in (5.23) satisfies
1
(5.29) Re(6(/) )= -^Re(p- ).

PROOF. Firs t w e prove (5.29) . Takin g th e logarithmic derivativ e o f the func -

tional equatio n (5( 1 - s)) rA(f, s) = e(f)(s(l - s)) rA(f, 1 - 5) , we get from (5.23 )
11
(5.30) 2R e (6(/)) = b(f) + b(f) = - £ (-^ -+ * + ^- 4+- -
z 5 — p l — s — p p p
0 , l \s — p
p #--' 1 —s— p p

We have (s — p) x — (1 — 5 — p) l < C |p| 2 for 5 ^ p, the implied constan t dependin g

on s , an d similarly p~ 1 + p _ 1<C | p |- 2 so the series

are absolutel y convergen t b y Lemm a 5.5 . S o we can separat e the m i n (5.30 ) an d

the firs t on e vanishes (th e terms cance l ou t sinc e p and 1 — p are zeros o f A(/, s)),
giving (5.29) .
Let T ^ 2 and 5 = 3 + iT. B y (5.26 ) w e have |A/(n) | < dnlogn . Henc e

(5.31) | ^ ( / , * ) | < < ( 2 ) < l o g q ( / ) .

By Stirling' s formul a (5.1 1 6 ) w e have ^logq + — (/, s) < C logq(/, s). Observ e als o
that fo r any zero p = / 3 + 2 7 we have

< Re <
9 + ( T -7 ) 2 Vs-p 74 + (T- 7)2'
Hence w e can tak e th e rea l par t i n (5.24 ) an d rearrang e th e resultin g absolutel y
convergent serie s t o derive usin g (5.29 ) tha t

(5.32) £ _ _ ! _ _ <<logq(/ ,iT).

This implie s (5.27) . T o derive (5.28) , we write s = a + it an d

~(f,s) = ~(f,s) + j(f,3 + it) + +0{\ogqV,s)).

by (5.31 ) , and we get by (5.24 ) an d (5.1 1 6 ) agai n

-£</,.) -£</,.) + r- + ^ - E (7=7 - ST^;)+OP°«'(/-'» -

104 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

In the series, keep the zeros with \s — p\ < 1 and estimate the remainder b y log q(/, s)
using
i 1 1 i 3

\s — p 3 -h it — p l+ (T-7)2
and (5.32) . Moreove r w e have (se e (5.116))
7' v d ^1 v^T' /S + K

JJ 3

-E -^ - + O(iogqoo(s)).
5 + ft-?
|S+«J|<1 ^

Thus (5.28 ) follows . •

T H E O R E M 5.8 . Le t L(f,s) be an L-function of degree d. Let N(T,f) be the

number of zeros p = f3 + ry o / L(/, 5) ST/C/ I t/m t 0 ^ / 3 < 1 and |7| < T. W^ e ftave

(5.33) N(T, f) = l lo g ^^ + 0(log q(/, iT))

for T ^ 1 wit/i a n absolute implied constant.

P R O O F . Le t N'(T, f) b e the number o f zeros p of A(/, s) wit h 0 ^ / 3 ^ 1 an d
0 < 7 < T. W e hav e
(5.34) JV(T , /) = N'(T, f) + N'(T, f) + 0(log q(/) )
where the error term account s for possible real and trivia l zeros of L(f, s) in 0 ^ a <
1. B y (5.27) w e can assume tha t A(/ , s) does no t vanish o n Im (s) = T b y adding
to T a n arbitraril y smal l numbe r i f necessary , modifyin g N(T, f) b y a quantit y
<C log q(/, iT). Choosin g smal l 5 > 0 such tha t A(/ , s) ^ 0 in — 5 < I m (s) < 0, w e
have N(T, f) = I(T) + 0(logq(/,tT)) wit h

'^db/£<'••>*
where C is the rectangle wit h vertice s 3 — iS, 3 + iT, — 2 + iT and —2 — i<5 . W e cut
this rectangl e symmetricall y a t the points \ — iS and \ + iT . B y the functiona l
equation, th e integral o n the left par t o f the contou r equal s tha t o n the right part .
Therefore I(T) equal s twice the integra l ove r th e righ t par t o f C whic h we are goin g
to estimat e b y observing th e variation o f arguments o f each facto r in

Mf,*) = qs/Hf,s)L(f,s) = Tr-^y2 rj r(i±^)L(/,*). ' S + Kj

3=1
ds 2 s 2
The variatio n o f the argumen t o f ^~ l q l give s a contribution t o I(T) equa l to

(5.35) £ lo g £ + 0 ( 1 ) .

By th e Stirling formul a (5.1 1 3) , th e argument o f T(a + it) i s tlogt — t + O(l) i f

t ^ 1 , so the contribution o f gamma factor s t o I(T) i s
lo
(5.36) ^(f ^-f)+°( l o ^(/))-
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 105

For L(/ , s) o n th e vertica l segmen t fro m 3 — id t o 3 + iT , w e estimat e

s
log L(f, s ) = - £ ^ - n ~ « lo g q(/)
t-^ log n

as i n (5.31 ) , s o th e integra l o n thi s segmen t i s < C logq(/). The n o n th e remainin g

horizontal part s fro m \ — id t o 3 — id and 3 + iT t o \ + iT, w e appeal t o (5.28 ) an d
(5.27) t o ge t th e estimat e 0(logq(/,£T) ) .
By (5.34) , (5.35 ) an d (5.36 ) applie d t o L(f,s) an d L ( / , s ) , w e ge t (5.33) . •

REMARKS. Th e mai n ter m o f (5.33 ) come s ou t fro m ou r computatio n o f th e

variation i n the argumen t of q s/2ry(f.s) alon g certain (right ) segments , the contribu -
tion fro m th e L-functio n itsel f bein g smal l (well , onl y fro m th e relevan t segments) .
This illustrate s ho w th e L-functio n i s intrinsicall y connecte d wit h th e companio n
at th e infinit e place .

5.4. T h e zero-fre e region .

Analytic applications of L-functions mak e use of complex integration i n a region
free o f zero s of L(f,s) fo r variou s / . Th e result s ar e stronge r i f the zero-fre e regio n
spreads deepe r int o th e critica l strip . Th e basi c metho d t o cu t on e i s stil l tha t
of Hadamar d an d d e l a Valle e Poussin . I t i s remarkabl e tha t thi s metho d als o
reappears i n Deligne' s secon d proo f o f th e Rieman n Hypothesi s fo r varietie s ove r
finite fields [De2] (th e basic trigonometric inequality (5.69 ) is misprinted there) . W e
present th e traditiona l arguments , bu t i n somewha t mor e elegan t fashion , startin g
with a usefu l genera l lemm a o f Goldfeld , Hoffstei n an d Liehma n [GHL] .

LEMMA 5.9 . Let L(f, s) be an L-function of degree d with R e (A/(n)) > 0 for
(n,q(f)) = 1 . Suppose that at ramified primes \aj(p)\ < p/2. Then L(/ , 1 ) ^ 0 .
Let r ^ 0 be the order of the pole of L(f,s) at s = 1 . There exists an absolute
constant c > 0 such that L ( / , s) has at most r real zeros in the interval
r
S > 1
d ( r + 1 ) log q ( / )'

P R O O F . Le t fy b e zero s o f L(f,s) i n th e segmen t \ ^ f3j < 1 . B y (5.28 ) fo r

s = a > 1 we ge t

(5.37) Yl -V< -^ + ^ y (/,a ) + 0(logq(/))

J
j

on ignorin g zero s othe r tha n /3j's , by positivit y sinc e

Re ^- = ^ ^ 0
a-p \cr-p\2

for an y zer o p — j3 + ij i n (5.28) . Sinc e R e (A/(n)) ^ 0 for (n , q(f)) = 1 we hav e

(5.38) R e ^ ( / , c r ) < 0
106 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

where L nr(f, s) i s the Eule r produc t o f L(/ , s) restricte d t o unramifie d primes . W e

estimate th e contributio n o f ramifie d prime s b y

IE E^SS«E>°^<<>^</> .
p\q(f)KJ^d JKFJF p\q(f)

Hence b y (5.37 )

E— <
a — Bj a
jJ
7 +0(dlogq(/))
— 1

where r i s th e orde r o f th e pol e o r zer o o f L(/ , s) a t s = 1 . Th e abov e inequalit y

shows that fij = 1 is not possible (i n this case r < 0) so we conclude that L(/ , 1 ) ^ 0 ,
in othe r word s r ^ 0 . Suppos e fy > 1 — c(d(r + 1 ) l o g q ( / ) )_ 1 fo r 1 ^ j < n . W e
choose a — 1 + 2c(dlogq(/)) _ 1 an d ge t

ndlogq(/) <
£+0(l))dlogq(/)
2c + c/( r + l )

which implie s n < r + r/2( r + 1 ) + 0(c ) an d n ^ r i f c is smal l enough . •

T H E O R E M 5.1 0 . Let L(/ , s) 6 e an L-function of degree d such that the Rankin-

Selberg convolutions L ( / ® /, s ) and L(f 0 / , s) exist, and the latter has a simple
pole at s = 1 while the former is entire if f ^ f . Suppose that at the ramified
primes \aj(p)\ 2 ^ p/2. There exists an absolute constant c > 0 such that L(f,s)
has no zeros in the region

(5.39) c r ^ l
d 4 log(q(/)(|*| + 3) )
except possibly for one simple real zero f3f < 1 , in which case f is self-dual.
PROOF. Fo r t e R , le t L(g, s) b e th e L-functio n

L(g, s) = C(s)L(f, s + it) 2L(f, s - it) 2L(f<g > /, s + 2it)L(f®f, s - 2it)L(f ® /, s) 2

of degre e ( 1 + 2d) 2. It s analyti c conducto r satisfie s

(5.40) q(g) < q(/ , it)\(f 0 / , 2it) 4 q(/ 0 f) 2 < q(/) 4 + 1 2 d (|t| + 3) 6d2

by (5.1 1 ) .
The produc t i s arrange d s o tha t it s Dirichle t serie s coefficient s ar e real , non -
negative. I t i s sufficient fo r u s to chec k this a t unramifie d places . Fo r a n unramifie d
prime p , th e loca l Eule r facto r a t p fo r L(g , 5) i s o f th e for m (5.1 ) wit h "roots " 1
with multiplicity one , c\jp lt an d ajp~ lt wit h multiplicity two , ctjOck wit h multiplicit y
two, aj(XkP 2lt an d OLjC\^p~ 2lt wit h multiplicit y one , where a^ , a^ ru n ove r th e root s
(5.1) fo r L(/ , s). Therefor e fo r an y k ^ 1 , the su m o f the fc-th powers o f these root s
is
2
I + X ^ + E ^ V -kit >0
JJ
so that A g(n) ^ 0 for an y n coprim e wit h q(f)
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 107

Let p — /3 + 27 be a zero of L(/, s) with / 3 ^ | an d 77^ 0. The n w e take £ = 7

so L(g , s) has a pol e a t s = 1 of order ^ 3 whereas f3 is a zer o o f L(g, 5) o f order
^ 4 . Henc e b y Lemma 5. 9 we have

cd
(5 41) < l <
' ^ ~ dHog^g) * " dMog(q(/)(|t|+3) )

for som e absolut e constant s c > 0 and c' > 0.

Now consider rea l zero s of L(/, 5). Fo r this purpose w e take the product L(g , s)
with ^ — 0 . An y real zer o o f L(/, 5) i s a zer o o f !/(#, 5) o f multiplicity a t leas t 4.
On th e other hand , th e point s = 1 is a pole o f L(g, s) of order 3 if / ^ f an d o f
order 5 if / = / . Henc e by Lemma 5. 9 we conclud e the proof, excep t tha t w e must
prove tha t th e possible exceptiona l zer o of a self-dual L(/ , 5) satisfie s /3f < 1.
Consider instea d

L(h,s)=as)L(f,s)2L(f®f,s).

Assuming L(/ , 1 ) = 0 , it follow s tha t L(h,s) i s entire sinc e th e double zer o com -
pensates th e poles o f ((s) an d L(f<g > / , s). I t als o ha s non-negative coefficients ,
precisely fo r an unramified prim e p and fc^lwe hav e

fc
(5.42) A h(p )=(l + ^ ^ ) % 0 ,
3

so the series

A
(5.43) log L(h,s) = J2Yl \ h(Pk)P~ks
p k>0 *

has non-negativ e coefficients . Le t <r0 ^ 1 be the largest rea l zer o o f L(h, s), s o th e
first singularit y o f log L(h, s) when s decreases. Not e ao exists sinc e £(—2 ) = 0 . By
Landau's Lemm a (se e Lemma 5.56 ) th e series (5.43 ) converge s for any rea l a > cr0,
so b y (5.42 ) w e have log L(h, a) ^ 0 and |L(/i , <r)| ^ 1 . Lettin g a —> cr 0, we get a
contradiction. •

REMARK 1 . I f L(/, s) i s self-dual on e may refin e th e rea l zer o issue of Theorem

5.10 slightly, arguin g with L(h, s) more efficiently. Indee d (5.42 ) shows that A/ l(n) ^
0 fo r all n wit h (n , q(/)) = 1 . Le t I denote th e order o f the pole o f L(f ® /, s) at
5 = 1 . The n L(/i , s) has a pole of order / + 1 at s = 1 while an y real zer o of L(/, s)
is a zero o f L(h, s) of order ^ 2 . Henc e ther e ar e at mos t ^ rea l zero s o f L(/, s)
in th e region (5.39) .

REMARK 2 . Not e tha t i f the Rankin-Selber g L-function s exist , bu t L ( /0 / , 5)

has a pol e o f order > 1 at s — 1, it i s often th e case tha t L(f, s) i s a produc t of
other L-functions . I t is then mor e efficien t t o apply Theore m 5.1 0 t o those tha n to
L(f,s) itself . Fo r instance, conside r th e Dedekind zet a functio n o f a quadratic fiel d
(K(S) = C(s)L(s,x) . The n th e Rankin-Selberg L-functio n i s £(s) 2 L(s, x)2 whic h
has a double pol e a t s = 1.
108 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

EXERCISE 4 . Le t L(f,s) an d L(g,s) b e tw o L-function s o f degree s d an d e

respectively suc h tha t g ^ f an d g ^ f. Assum e tha t Rankin-Selber g L-function s
L{f ® g,s), L(f ® #,s), an d henc e thei r dual , exist , an d tha t th e loca l root s a t
ramified prime s satisf y |a^(p)| 2 < p/2. Sho w tha t ther e exist s a n absolut e constan t
c > 0 such tha t L(f ® g,s) ha s n o zer o i n th e regio n
r.
(d + e)*log(q(/)q(s)(|i|+3)) '
What ca n yo u prov e i f / = g o r / = g? [Hint : Us e th e auxiliar y functio n
L(gis) = L(f^ g)s+^)L(f0g,s)L(f^g,s)L(f^g,s-^)

and appl y Lemm a 5.9. ]

In genera l nothin g stronge r i s know n tha n Theore m 5.1 0 . Fo r certai n impor -
tant L-functions , on e ca n d o better , individually , an d als o b y considerin g families .
Examples, includin g th e mos t importan t on e o f Dirichle t characters , ar e discusse d
in Sectio n 5. 7 an d 5.1 2 .
5.5. Explici t formula .
A summatio n formul a somewha t analogou s t o (5.1 2 ) ca n b e derive d b y inte -
grating th e Dirichle t serie s fo r th e logarithmi c derivativ e o f a n L-function . Fo r
historical reasons , thi s typ e o f formul a i s usuall y calle d a n "explici t formula" , em -
phasizing th e lin k i t expresse s betwee n sum s ove r th e zero s o f a n L-functio n wit h
sums o f coefficient s a t prim e powers .
THEOREM 5.1 1 . Let tp :]0,+oo[— > C be a C°° function with compact support,
and
/*+oo
(p(s) — / (p(x)x 1 s
~ dx,
Jo
be its Mellin transform. Put ip(x) = x~ 1 <p(x~1 ), which means ifj(s) = <p(l — s).
Then we have

(5.44) V (A f(n)<p(n) + Aj(n)^(n)) = ^ ( l ) l o g g ( / ) + r / ^{x)dx

(1/2) '

where p runs over the zeros of L(f, s) in the strip 0 < a < 1 (the non-trivial and
the trivial ones) with relevant multiplicity.

PROOF. Star t wit h the first ter m on the left sid e of (5.44) using Mellin inversio n

] T A/(n)<p(n ) = — / - — (f,s)<p(s)ds.
(2)

Let c be a smal l positiv e numbe r s o tha t L ( / , s) ha s n o zero s i n — c ^ R e (s) < 0 .

Move the lin e of integration t o R e (s) = —c , which i s legitimate b y (5.27) . A simpl e
pole i s picke d u p a t s — 1 with residu e r J ip(x)dx, an d a t th e zero s s = p wit h
residues m(p{p), wher e m i s the multiplicity . S o those residue s giv e the secon d an d
the fourt h term s o n th e righ t sid e o f (5.44) .
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 109

The functiona l equatio n implie s

-^(f,s) = logq(f) + ^ ( /, s) + ^ ( / , 1 - s) + ^(f,1 - s)

L7 7 L

for a = — c. Integratin g ove r a = —c , the facto r logg(/ ) give s th e firs t ter m o n th e

right sid e o f (5.41 ) b y Melli n inversion . Moreover , b y absolut e convergenc e

n).
(-c)

And finally on e ca n mov e th e lin e o f integratio n fo r th e gamm a factor s t o a = |

getting the integral of gamma factors i n (5.44 ) without residue s (th e poles of — (/, 5)
cancel ou t wit h thos e o f — (/, 1 — s)). •

R E M A R K . Becaus e o f th e appearanc e o f th e conducto r q(f) o n th e righ t sid e

of (5.41 ) , thi s formul a turn s ou t t o provid e a goo d analyti c too l fo r it s study . Se e
Theorem 5.3 2 below fo r a n example . However , i t i s quit e usefu l fo r othe r things .
First o f al l th e explici t formul a i s seen a s expressin g a su m ove r prime s b y th e su m
over the zero s and vic e versa. Certainl y ther e ar e other explici t formula s connectin g
primes wit h zeros .

EXERCISE 5 . Le t <p be smoot h an d compactl y supporte d o n [l,+oo[ . Sho w

that

g A(nMn ) = ^ ( 1 - (g _ 1 )g(g +!)>(*)** " I>(P) .

where p run s ove r th e non-trivia l zero s o f C(s) . [Hint : Star t a s above , bu t mov e
the lin e o f integratio n fa r t o th e lef t instea d o f usin g th e functiona l equation ; th e
first ter m come s fro m th e pol e a t s = 1 and th e trivia l zero s o f £(s ) a t s = — 2/c,
k ^ 1 an integer. ]

Sometimes i t i s more convenien t t o wor k wit h th e explici t formul a i n a Fourie r

form rathe r tha n th e Melli n for m (5.44) . Thi s ca n b e derived fro m (5.44 ) b y simpl e
substitutions (p(x) = x _1 //2 #(log:r) an d x — e y gettin g

THEOREM 5.1 2 . Let g(y) be a function of Schwartz class on R which is even.

Let h(t) be the Fourier transform of g{y). We have

(5.45) ^ ( A/(n)+A7(n))^|^=5(0)logg(/) + r/i(^ )

+
5FSZ (^ + *>+ T"'* " a)M5F)* " ?*(£)
where p = | + ^ 7 runs over the zeros of L(f, s) in the strip 0 ^ a ^ 1 (the non-trivial
and the trivial ones) with relevant multiplicity.
110 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

5.6. Th e prim e numbe r theorem .

By Prime Numbe r Theore m fo r a n L- function L(/ , s), we mean first th e asymp -
totic behavio r o f th e su m
(5.46) <&(/,* ) = £ A,(n )
n^.x

which i s essentiall y th e su m o f Xf(p)\ogp ove r primes . Whe n L(f,s) — C(s), thi s

amounts indeed to counting primes p < x. Muc h of the arithmetic interest lie s in th e
dependency o f th e erro r ter m i n asymptoti c approximation s o n extr a parameters ,
notably o n the conductor whe n / varie s in a family. Therefor e w e are seeking result s
which ar e explici t i n ever y possibl e parameter , althoug h no t th e stronges t ones .
The strengt h o f results depend s o n the dept h o f the zero-fre e regio n fo r L(/ , s).
The bes t w e can hope for today i s that L(/ , s) ha s no zeros in the region (5.39 ) wit h
at mos t on e exceptio n o f a simpl e zer o fif i n cas e o f / self-dual . B y ou r definitio n
(only fo r th e purpos e i n thi s section ) th e exceptiona l zer o i s i n th e segmen t

where c i s th e absolut e positiv e constan t fro m (5.39) . Thi s definitio n ma y caus e

some discomfor t fo r ne w enthusiast s o f analyti c numbe r theory , becaus e i t i s loos e
with respec t t o th e constan t c. However , th e practic e show s tha t on e benefit s fro m
such a flexibl e concept . Afte r all , c ca n b e improve d b y ne w tool s i n th e future ,
not t o mentio n tha t w e don't believ e i n th e existenc e o f the exceptiona l zer o of an y
L-function.
In ful l generalit y w e d o no t ye t hav e a stron g boun d fo r th e individua l coeffi -
cients o f L(/ , s). Bu t a crud e boun d fo r A/(n ) o n averag e will suffice . Specifically ,
we postulate th e following :
2
(5.48) ^ | A /(n)| «^2log2(a;q(/))
n^x

for x ^ 1 , where th e implie d constan t i s absolute .

EXERCISE 6 . Deriv e fro m (5.48 ) tha t
2
(5.49) W , a ) = £ A/(p ) logp + 0{^d 2 log (xq(/))
p^x

where th e implie d constan t i s absolute .

Recall tha t th e postulate d zero-fre e regio n (5.39 ) wa s establishe d i n Theore m

5.10 unde r th e assumptio n tha t th e Rankin-Selber g convolutio n L(f 0 / , s) exist s
and ha s a simpl e pol e a t s — 1 . W e shal l se e tha t th e sam e assumptio n i s als o
sufficient fo r th e secon d postulat e (5.48) . T o thi s en d w e appl y Propositio n 5. 7 fo r
L{f ® f,s) rathe r tha n fo r L(f,s). Combinin g (5.28 ) an d (5.27 ) (se e als o (5.1 1 ) )
we ge t
l o
-T(/®/»«^T gl(/)
LG — 1
for 1 < < J ^ 2 , where th e implie d constan t i s absolute. Thi s wit h a = 1 + (log3x) _ 1
yields (5.49 ) b y th e non-negativit y o f coefficients .
 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S 111

From (5.49 ) w e derive usin g Cauchy' s inequalit y tha t

n
(5.50) J2 IM )l « d^ylog(xq(f))
x<n^.x+y

for x ^ y ^ 1 where th e implie d constan t i s absolute .

Now w e ar e read y t o stat e an d t o prov e th e mai n Prim e Numbe r Theorem .
THEOREM 5.1 3 . Let L(f,s) be an L-function for which (5.39) is a zero-free
region with at most one exceptional real zero f3f in the segment (5.47), where c is a
positive absolute constant. Suppose (5.48) holds with an implied constant absolute.
Then we have

(5.51) « / . , )- „ - %+°('^(^',%; ( / ) )(''l°«"l(/>>').

for x ^ 1 , the implied constant being absolute. By convention the term —xP f/(3f is
not in (5.51 ) if the exceptional zero does not exist.
REMARKS. Th e approximat e formul a (5.51 ) ha s meanin g whe n th e erro r ter m
is smaller tha n th e mai n term , an d thi s i s the cas e fo r
1
x ^ ^4c- d 4 log(dlog < ? )

where q = q(/ ) i s th e conductor , d i s th e degre e an d c i s th e relevan t absolut e

positive constant . On e ca n simplif y th e erro r ter m b y compromisin g slightl y th e
strength o f th e result . Fo r example , w e hav e

/
(5.52) V ( / , ^ ) = r - x - ^ + 0 ( v q(7)^exp(-^iV^g^))

where th e implie d constan t i s absolute . Indee d thi s hold s true , becaus e whe n th e
error ter m her e i s smaller tha n tha t i n (5.51 ) th e resul t i s trivial. Th e condition s o f
Theorem 5.1 2 ar e satisfied i f L ( / 0 / , s ) exist s an d ha s a simpl e pol e a t s — 1, bu t
we d o no t restric t (5.51 ) t o thi s cas e only . Indee d w e shal l establis h th e zero-fre e
region (5.39 ) an d w e shall se e the crud e boun d (5.48 ) directl y withou t appealin g t o
the Rankin-Selber g convolutio n L-functio n i n man y classica l cases .
P R O O F O F THEORE M 5.1 2 . Firs t w e smoot h thing s ou t b y writin g

where (j>(z) i s a functio n suppor t o n [0 , x + y] , such tha t </>(z) = l i f l ^ 2 ; < x an d

\(b(z)\ ^ 1 elsewhere. Th e paramete r y wil l b e chose n late r subjec t t o 1 ^ y ^ x.
For example , w e tak e
(p(z) = m i n f - , 1 , 1 H )
\y y J
for 0 ^ z ^ x + y an d <p(z) = 0 if z > x + y. Th e Melli n transfor m o f <f)(z) satisfie s
rx+y x <y / x \
1
s
(t)(s)= / (P(z)z ~ dz < —mi n 1 ,—)
Jo \s \ v
\s \y J
for s = a + it, \ ^ a < 2 .
112 5 . CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

Now w e ca n star t wit h

£A f(n)4>(n) = ±-[-*L(f, 3)fa)ds,

(2)

the integra l convergin g absolutely . W e shall b e operatin g i n th e regio n

{s = a + it | a > 1 - Cl /d4 log(q(/)(|t | + 3)) }
where c\ = 2c/3 , or c\ — c/3, dependin g o n whethe r th e exceptiona l zer o i s i n th e
right hal f o f th e segmen t (5.47) , o r elsewher e (wit h possibilit y o f non-existence) .
Let Z denot e th e boundar y o f this region . Th e poin t i s that al l the zero s o f L(/ , s)
are distance d fro m Z b y a t leas t c/6d 4 log(q(/)(|t| + 3)) . Henc e i t follow s b y (5.27 )
and (5.28 ) tha t fo r se Z,

- ^ ( / , 5 ) « d 4 l o g 2 ( q ( / ) ( M + 3)) .

Moving th e integratio n fro m th e lin e R e (s) = 2 to Z w e ge t b y Cauchy' s theore m

£ A/(n)0(n ) = r0(l) - 4>(M + 4- / ~T if' S ^(s)ds-

z
Here th e presenc e o f th e exceptiona l zer o ter m depend s o n whethe r w e passe d th e
point s = (3f or not . Fo r s = 1 , or s = f3f we hav e

0 ( 5 ) = I* sz s-1 dz + 0(y) = -+0(y)

Jo
while th e contou r integra l ove r Z i s estimate d b y

d4 J ^ m vi n ( l , r ^ ) l o g 2 ( q ( / ) ( | t | + 3 ) ) | ^ | < < ^ 4 x ^ T ) l o g 3 ( q ( / ) T )
7z M \s\y'
where T = x/y an d cr{T) = 1 — ci/d 4 log(q(/)T) . Gatherin g th e abov e result s w e
arrive a t
^ /1 x
^(/, a) = r x - ^ + 0 ( V (xT" 2 + x'(T) ) log 3 (xq(/))J
where T i s a t ou r disposa l subjec t t o 1 ^ T < x . W e choos e T = exp(|v / l°g x )
getting (5.51 ) . W e stil l hav e t o addres s th e questio n o f th e exceptiona l zer o term .
An explanatio n abou t it s contributio n t o th e obtaine d formul a i s neede d i f th e
exceptional zer o doe s exis t an d i t lie s i n th e lef t hal f o f th e segmen t (5.47) , bu t i n
this cas e th e ter m —x^f/fif i s consume d b y th e existin g erro r term , s o i t ca n b e
ignored. Thi s complete s th e proo f o f Theore m 5.1 2 . •

If on e i s willing t o emplo y mor e zero s o f L(/ , 5), then on e ca n deriv e alon g th e

above line s stronge r approximation s t o ip(f,x).
EXERCISE 7 . Assum e tha t L(f,s) satisfie s th e Ramanujan-Petersso n conjec -
ture. Deriv e (usin g Perron' s formul a (5.1 1 1 ) ) th e followin g approximat e expansio n

(5.53) tjj(f,x)=rx- £ ^ - ±+ o(|(logx)log(x dq(/)))

 5. CLASSICA L ANALYTI C T H E O R Y O F 1
L1
-FUNCTION S 3

where p = ft + 2 7 runs ove r th e zero s o f L(/ , s) i n th e critica l stri p o f heigh t u p t o

T, wit h an y 1 ^ T ^ x , an d th e implie d constan t i s absolute .
EXERCISE 8 . Assum e tha t L(f,s) satisfie s th e Ramanujan-Petersso n conjec -
ture an d Theore m 5.1 3 hold s fo r / . Prov e tha t r ^ d.
5.7. Th e Gran d Rieman n Hypothesis .
The Gran d Rieman n Hypothesi s (GR H fo r short ) refer s t o th e followin g con -
jectural statemen t abou t zero s o f L-functions :
GRAND RIEMAN N HYPOTHESIS . Let L(f,s) be an L-function. Then all zeros
of L(/, s) in the critical strip 0 < R e (s) < 1 are on the critical line R e (s) = \.
REMARKS. Needles s t o sa y w e believ e tha t ever y L-functio n (subjec t t o ou r
definition i n Sectio n 5.1 ) satisfie s th e Gran d Rieman n Hypothesis . Yet , provin g
this eve n fo r on e L-functio n woul d b e a n achievemen t o n a historica l scal e fo r
human beings . Not e tha t a n L-functio n ma y hav e zero s o n th e lin e R e (s) = 0
(certainly som e trivia l zeros , probabl y no t th e genuin e ones) , bu t a s w e alread y
proved no t o n R e (s) — 1 . Th e GR H als o ensure s tha t th e trivia l zero s i n th e
open critica l stri p ar e al l o n th e critica l lin e a s well ; d o the y exis t i s no t clear ,
most likel y the y d o not . I t i s ver y importan t t o conside r onl y L-function s wit h
Euler products . Indee d i t i s know n tha t Dirichle t serie s withou t a n Eule r produc t
(still civilized ones ) hav e many zero s even in the half-plan e o f absolute convergence .
Arithmetic geometer s shoul d no t b e concerned , becaus e thei r L-function s originat e
from a n Eule r produc t o f som e kind . O f course , thi s eas y star t make s i t harde r
to investigat e L-function s furthe r i n analyti c aspects . Fo r example , i t wa s onl y
recently tha t analyti c continuatio n o f th e Hasse-Wei l L-function s o f ellipti c curve s
was establishe d b y wa y o f modularity . On e shoul d no t underestimat e th e analyti c
continuation o f an Euler produc t beyon d th e regio n of absolute convergence, fo r on e
often implement s dee p arithmeti c resource s i n on e wa y o r another . A s envisione d
by Riemann , thi s ventur e int o th e critica l stri p illuminate s prim e number s an d
relevant arithmeti c object s buil t o n the m b y revealin g thei r dua l companion s - th e
zeros of L-functions. Mysteriou s a s they ar e today, thes e zero s (mos t likel y comple x
transcendental numbers ) wil l eventuall y b e cracke d fo r furthe r analysis . However ,
for th e tim e being , a s fa r a s th e GR H goes , w e ar e onl y sur e the y ar e fine tune d
to yiel d surprisingl y stron g estimates . Yes , th e uniformit y o f estimate s i n term s o f
involved parameter s whic h ca n b e derive d fro m th e GR H ar e astonishing . O n a
negative not e thi s super b uniformit y coul d b e a reminde r ho w sli m prospect s ar e
to prov e th e GR H i n ou r lifetime . N o tool s o f curren t analyti c numbe r theor y ar e
capable o f treatin g th e question s fo r whic h th e GR H provide s eas y answers . I n
this sectio n w e present a few traditiona l analyti c consequence s o f the GRH , payin g
attention t o th e uniformit y i n term s o f th e conductor .

First w e state a fe w equivalen t fact s whic h hel p t o gras p th e analyti c meanin g

of th e Rieman n hypothesis .
PROPOSITION 5.1 4 . Let \ ^ a < 1 . The following statements are equivalent
for an L-function:
(1) There are neither zeros nor poles of (s — l ) r L ( / , s) in a > a, where r is a
non-negative integer.
(2) The inverse, the logarithmic derivative and the logarithm of (s — l ) r L ( / , s),
the latter normalized so that\ogL(f, s) - » 0 as a - » +oo ? are holomorphic in a > a.
 114 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

(3) Let fJ>f(n) denote the coefficients in the Dirichlet series expansion for the
inverse L(f,s)~ x. Then
+£
(5.54) M(f,x) = £ > / ( r c ) « *" ,

the implied constant depending on f and e > 0.

(4) Let r ^ 0 be the order of the pole of L(/, s) at s — 1. Then
a
(5.55) xl>(f,x) = rx + 0{x +e),

£fte implied constant depending on f and e > 0.

The Riemann hypothesi s assert s that th e abov e statements ar e true with a = \.
What i s interestin g (perhap s somewha t surprisin g fo r a beginner ) i s tha t th e esti -
mate (5.55 ) i s self-improving b y passag e throug h zeros . Indeed , first (5.55 ) implie s
that th e zero s ar e o n th e lin e R e (s) = | , whenc e on e derive s b y th e expansio n
(5.53) wit h T — x usin g th e crud e boun d (5.26 ) th e followin g
THEOREM 5.1 5 . Assuming the Riemann Hypothesis for L ( / , 5) and the Rama-
nujan-Petersson Conjecture for L(f, s) we have

d
1(5.56) >(f,x) =rx + 0(xh\ogx)\og(x q(f)),
for x ^ 1 , the implied constant being absolute.
The Rieman n hypothesi s show s tha t L(/ , s), th e invers e L ( / , $ ) - 1 , th e loga -
rithmic derivativ e ^ r ( / , s) an d th e logarith m lo g L ( / , s) ca n b e wel l approximate d
by extremel y shor t partia l sum s o f th e correspondin g Dirichle t serie s uniforml y i n
Re 5 = a ^ a > \. Clearl y thi s i s no t possibl e fo r s o n th e critica l line , wher e th e
zeros are . W e begi n b y considerin g

Let <j)(y) b e a continuou s functio n o n [0 , 00[ whose Melli n transfor m (j>(w) satisfie s

(5.57) w(w + l)(f>(w) < 1 , fo r w suc h tha t - \ < R e (w) ^ \.

Suppose <p{w) ha s a simpl e pol e a t w — 0 wit h residu e 1 (normalization) . Fo r

example, </>(y) = e~ y wit h 4>(w) = T(w), o r <j)(y) = max( l — T/,0 ) wit h <f>(w) =
w~1(w+ 1 ) _1 .
PROPOSITION 5.1 6 . For \ < Re (5) < § we have

(,58) ^ ( / , s ) . ? ^ ( | )

/or an y X ^ 1 , where the implied constant depends only on that in (5.57).

 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION
11 S 5

PROOF. B y no w it shoul d b e a standar d argumen t fo r a reade r t o deriv e th e

formula (5.58 ) b y contour integratio n wit h th e error ter m bein g equa l exactl y

Indeed thi s i s obtained b y starting wit h the above integral o n the line Re (w) =
\ an d movin g i t t o th e lin e R e (w) = — \. Th e simple pole s a t w = 0,1 — s , p — 5
produce th e corresponding term s i n (5.58) . No w on the line R e (w) = — \ w e have
Re (s -f w) — a — | , an d we deduce b y (5.26 ) an d (5.27 ) tha t

- \ ( / , s + w) < (2 a - 1 ) - 1 log q(/, s + w).

Integrating thi s agains t th e bound fro m (5.57) , we get the error ter m i n (5.58) . D

The su m ove r th e zero s i n (5.58 ) i s estimate d directl y (withou t cancellation )

by mean s o f (5.26) , (5.27 ) an d (5.57 ) givin g

(5.59) - £ ( / , , )= £ ^ Q )- , 0 ( 1 - ,)X*- + o(^±x\-°)

n

for an y X ^ 1 , where th e implied constan t depend s onl y o n that i n (5.57) .

Next w e are going to estimate th e sum of Aj(n). Essentiall y an y crude bound ,
but independen t o f the conductor , lik e |A/(n) | ^ n , woul d suffic e fo r ou r purpose .
Nevertheless, sinc e w e already assume d th e GRH, there i s no point i n ignoring the
Ramanujan-Petersson conjectur e fo r the local roots , whic h yield s |A/(n) | ^ dA(n).
Prom thi s w e get

E *£>*(£)«'» ,-'+-"<«*•
n
Moreover, th e polar ter m i n (5.59 ) i s

- n £ ( l - s)X 1 ~s = r{s - l ) " 1 + OirX 1 -* + rlogX).

Hence (5.59 ) simplifie s t o

Finally choosin g X = log 2 q(/, s) w e obtain

T H E O R E M 5.1 7 . Assume the Riemann Hypothesis and the Ramanujan-Peters-
son conjecture for L ( / , s). We have

"T ( / ' s ) = 7 ^ 1 + 0 (2^T ( l o g q ( / ' s ) ) 2 " 2 < 7 + dloglogq(/ s)

' )
for any s with \ < a ^ | 7 the implied constant being absolute.
116 5 . CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

COROLLARY 5.1 8 . For a ^ \ + (lo g log log q)/(log log q) we have

where q = q(/ , s) is the analytic conductor and the implied constant is absolute.

To ge t an estimate fo r the logarithm o f the L-function w e integrate th e loga-

rithmic derivativ e alon g th e horizontal lin e
z.5/4 rj
log L(/, s) = log L(/, I + zt) - y - (/ , a + it)da .

Applying (5.59 ) w e deduce th e following estimate s

T H E O R E M 5.1 9 . Assume the Riemann Hypothesis and the Ramanujan-Peters-
son conjecture for L(/ , s). We have

]og(pr(8)L(f,8)) « f ° g q ( /' ^ " /T ,x + d log log q (/, s)

(2cr- l)loglogq(/,s )
/or am / 5 wi£/ i | < a ^ | , £/i e implied constant being absolute (recall p r(s) =
(s-l)r(s + l)- r.

As immediat e consequence s o f Theorem 5.1 9 w e have lowe r an d upper bound s

for L(/ , s) in the half-plane a ^ ^ + £:
q(/> *)"" < Pr(s)L(f, s) < q(/ , s) £
with implie d constant s dependin g onl y on £. Th e upper boun d ca n be extended up
to th e critical lin e b y the convexity principle . I n particular, w e obtain
COROLLARY 5.20 . For s with R e (s) = | and any e > 0 we have
(5.60) L ( / , S )«q(/,Sr
where the implied constant depends only on e.

The las t estimat e receive d considerabl e attentio n (i t is known a s the Lindelof

Hypothesis), becaus e it is sufficient fo r solving many fundamental problem s of arith-
metical nature du e to its uniformity i n the conductor. W e just prove d tha t i t follows
from th e Riemann hypothesis . On e ma y thin k tha t th e Lindelof hypothesi s shoul d
be easie r t o establish, bu t there i s a popula r opinio n tha t th e Riemann hypothe -
sis mus t b e solved first , th e point bein g tha t th e Riemann hypothesi s i s linked t o
more natura l mathematica l structures . Clearl y th e Lindelof conjectur e implie s the
following estimat e fo r sums o f coefficients o f the L-functio n

(5.61) ^ A /(n)=xP/(logx) + 0(xi(xq(/)) e)

where Pf(X) i s the polynomial o f degree r — 1 given by

s
Pf(logx) = res L(f,s)x -\
S =l

Some unconditional results , weaker yet useful, ar e given in Chapter 4 (see Corollar y
4.9 fo r ((s)L(s, x±) a n d Exercis e 7 of Chapter 4 for ((s)2 fo r instance).
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 117

5.8. Simpl e consequence s o f GRH .

We illustrat e som e o f the simpl e arithmeti c consequence s o f GRH, beside s the
Lindelof Hypothesis , whic h ha s important application s o f its own.
The firs t proble m i s to estimate th e order o f vanishing o f an L-function a t a
given poin t o n the critical line , i n term s o f the conductor o f /. W e consider th e
case of the central poin t s = 1 /2 . Not e tha t (5.27 ) implie s

ord L ( / , 5)^m(l,/)«logq(/)
s=l/2

unconditionally, whic h w e can improve slightly .

PROPOSITION 5.21 . Let L(f,s) be an entire L-function satisfying GRH. We
have
logq(/)
(5-62) or d L(/,s)« C , „ ,
' -i/ 2 log(j|logq(/) )
with an absolute implied constant (recall that logq(/ ) > d).
PROOF. W e apply the explicit formul a (5.45 ) t o a test functio n g(y) supporte d
on [-2Y,2Y] wit h 0 ^ g(y) < 2 and g(y) ^ 1 if \y\ ^ Y suc h tha t it s Fourie r
transform h(t) i s non-negative o n R. B y positivity w e can drop al l the zeros p =
\ + 27 ^ \ i n (5.45 ) gettin g

mh(0) ^ - 2 ^ R e ( A / ( n ) ) n - ^ ( l o g n ) + 0(logq(/)) ,
n

where m i s the multiplicity o f the zer o p — \. O n the left sid e we have h(0) x Y.
Since we have |A/(n) | ^ dA(n)n b y (5.2), the sum ove r prim e power s n on the right
side i s bounded by
2d ^2 A(n)n 2 < de Y',
logn^2Y

where th e implied constan t i s absolute. Choosin g 2Y = l o g ( | logq(/)) w e ge t

logq(/)
log(|logq(/))
with a n absolute implie d constant . D

Proposition 5.2 1 is optima l althoug h th e estimate o f the sum over prime s i s

quite crude , whic h migh t sugges t som e improvemen t i s possible. Th e issue i s tha t
this su m is extremely shor t wit h th e choice o f y, i n fact n ^ e 2Y ^ | logq(/ ) and
it i s conceivable tha t Af(n) doe s not change sig n i n this range . I f we take L(f, s) k
for k —> +o o instead o f L(/, s), w e see that (5.62 ) canno t hol d withou t th e facto r
| . However , i n a number o f cases, i t is possible t o do better; se e Proposition 5.34
and Exercis e 1 3 below.
The proble m o f detectin g sig n change s o f Af(n) fo r ver y smal l n i s relate d
to anothe r classica l applicatio n o f the Grand Rieman n Hypothesis : ho w many co-
efficients (i n term s o f th e conducto r q(f)) o f L(f,s) a t prime s ar e require d t o
distinguish L(/ , s) amon g L-function s o f a certain class ?
118 5 . CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

PROPOSITION 5.22 . Let L(f,s) and L(g,s) be two L-functions with the same
degree d and gamma factor. Assume that L(f <S> / , s) and L(f 0 g, s) exist and
the latter is entire, that GRH holds for both and that the local roots at ramified
primes for L(f (g ) / , s) and L(f 0 g, s) are of modulus ^ 1 . There exists a prime
p ^ C(dlogq(f)q(g)) 2 unramified for f and g such that the local roots of L(f,s)
and L(g, s) at p are different. Here C is some absolute positive constant.
P R O O F . Th e poin t i s that b y assumptio n th e L-functio n L ( /<g> /, s) ha s a pol e
at s = 1 whereas w e assum e tha t L(f ® g, s) i s entire . Assum e th e loca l root s o f
/ an d g coincid e fo r al l prime s p ^ 2X whic h ar e unramifie d fo r / an d g. The n
Af®g(n) — Af^f(n) fo r n ^ 2X suc h that (n , q(f)q(g)) = 1 . B y the explicit formul a
(5.44) applie d t o a functio n o f th e typ e 0(n/X ) wher e 0 ^ 0 is smooth, compactl y
supported o n [1 , 2], an d 0 / 0 , w e hav e

^Am(n)<l>(n/X)<z:Vxiogq{f®g),
n
J2 K^f{n)4>{n/X) = 4>(0)X + 0(VX\og q( / ® / ) )
n

where 0(0 ) = J <p(t)dt > 0 . Thes e estimate s ar e smoot h version s o f th e prim e

number theore m fo r th e Rankin-Selber g convolutio n L-functions . W e use d GR H
to estimat e th e su m ove r zeros . O n th e othe r hand , th e tw o lef t side s coincid e
by assumptio n u p t o th e contributio n fro m ramifie d primes . Thi s contributio n i s
small, precisel y i t i s
A 2
«E l / 0 / » l + |A/ M(n)| « d (\ogq{f)q{g))\ogX.
n^2X,{n,q{f)q{g))^l

Therefore 0(O) X « VX\ogq(f 0 f)q(f ® g) + d 2 (iog q{f)q(g)) logX , o r X <

2
(d\ogq{f)q(g)) wher e th e implie d constan t i s absolut e (recal l tha t th e definitio n
of th e Rankin-Selber g convolutio n implie s th e boun d q( / 0 g) ^ (q{f)q(g)) d)- •

REMARK. Th e bes t genera l unconditiona l resul t i s muc h weaker . Suppos e

L(/, s) ha s degre e d with root s |QJ^(p ) | < p 1 / 4 an d tha t i t admit s L(f (g ) /, s). The n
for ever y e > 0 ther e exist s C > 0 , dependin g o n e an d o n / wit h th e followin g
property: I f L(g, s) ha s th e sam e gamm a facto r a s L(/ , s) (s o th e sam e degree) , i f
L{f<8 > g, s) i s entire an d i f X g(p) = Xf(p) fo r al l p ^ Cq(g) d/2+£, the n g = / .
To prove this, notice that i f X ^ 1 is the largest intege r fo r whic h Xf(p) = X g(p)
for p < X , the n b y multiplicativit y w e hav e Xf(n) = X g(n) fo r n \ N(X) wher e
N(X) i s th e produc t o f prime s ^ X. Fo r h = / , o r ft = g, on e ca n facto r

L(f<g > ft,s) = L\f <g > ft,s)H(f,ft, s)

where L b(f ® ft, s) i s th e Dirichle t serie s restricte d t o squarefre e integer s an d

H(f,h,s) converge s absolutel y fo r a > \. I n particular , L b(f ® ft, 5) i s entir e i f
ft = g an d ha s a pol e a t 5 = 1 for ft = / , an d satisfie s th e sam e convexit y boun d a s
the Rankin-Selberg L-function . W e get by complex integration o n the line a = | + 5 ,
with S > 0 small enoug h

J2 A / ^ ( n / X ) = 6(f,g)XP(\ogX) + 0(xh^{f ® ff )i+=)

 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 119

where r i s the orde r o f the pol e o f L ( / ® / , s) a t s = 1 and P ^ 0 is the polynomia l

of degre e r — 1 given b y

P(logX) = re s L b(f<g > h, s)X s

~\
s=l

Comparing th e formul a fo r h = f an d tha t foih = g yield s th e result .

5.9. Th e Rieman n zet a functio n an d Dirichle t L-functions .
The Rieman n zet a functio n ((s) i s a self-dua l L-function , wit h conducto r 1 ,
gamma facto r 7(5 ) = 7r - s / 2 r(s/2) an d roo t numbe r 1 . It s Rankin-Selber g squar e
is als o C{s).
More generally, le t x b e a primitive Dirichlet characte r modul o q. Th e Dirichle t
//-function L(s , x) i s a n L-functio n o f degre e 1 with conducto r g , gamm a facto r

7(,) = ^ / 2 r ( ^ ) ,
where 6 = 0 i f x(~~l ) = 1 an
d< 5 = 1 i f x ( — 1 ) = ~1 > a n d it s roo t numbe r i s
e(x) = r ( x ) / ^ w h e r e

a: (mo d q)

is the Gaus s su m associate d t o x (se e Sectio n 3.4) . I f x i s non-trivial , the n L(s , x)

is entire , otherwis e i t ha s a simpl e pol e a t 5 = 1 with residu e 1 .
Any L(s , x) satisfie s triviall y th e Ramanujan-Petersso n conjecture . Th e ana -
lytic conductor i s given by q(x?s) = <?(| £ + 5| + 3) x g(|t | + 3) and q(x ) = (2 + 5)q x g ;
see Theore m 4.1 5 .
A Dirichle t L-functio n L(«s,x ) i s self-dua l i f an d onl y i f x i s quadratic . I n thi s
case e(x) = 1 by Theore m 3.3 .
Any tw o Dirichle t character s x i a n d X 2 admi t a Rankin-Selber g convolutio n
given simpl y b y L(\i <8> X2, s) — L(X3i 5 )> where X 3 is th e primitiv e characte r tha t
induces th e produc t XiX2 - W e hav e X 3 — X1 X 2 if ? for instance , th e conductor s qi
of X i ar e coprime , bu t no t i n general . I n particular , th e Rankin-Selber g squar e o f
L(s,x) i s C( 5) a n d th e finite produc t (5.1 0 ) i s

Fixing a modulu s q and considerin g al l primitiv e character s modul o q gives a n

example o f a "family " o f L-functions . Fixin g X ^ 2 an d considerin g al l primitiv e
quadratic character s o f conducto r q < X give s anothe r family .
By (5.22 ) w e deduc e th e convexit y boun d fo r Dirichle t L-functions .
THEOREM 5.23 . Let x modulo q be a primitive Dirichlet character. We have

(5.63) L( 5 ,x)«(<7|*|)i

for Re s — \, the implied constant being absolute.

The followin g exercis e provide s ver y simpl e estimate s whic h ar e quit e hand y
anyway.
 120 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

EXERCISE 9 . Le t x modul o q be a non-trivia l Dirichle t character . Sho w tha t

for 0 ^ ( J < 1 we have
1
1 k+
(5.64) LW(<T, X)<q -''(logq)

for / c ^ 1 , th e implie d constan t dependin g onl y o n k. [Hint : Us e the boun d

l E n ^ x X W I ^min(x,g). ]
Theorem 5. 8 becomes:
T H E O R E M 5.24 . Let x modulo q be a primitive Dirichlet character, N(T,x)
the number of zeros p = (3 + i^ of L(s,x) ^ n the critical strip 0 ^ / 3 ^ 1 with
|7| ^ T. We have
7V(T,X) = - l o g | ^ + 0 ( g ( T + 3)),
7T lire
with an absolute implied constant.

In additio n t o th e general explici t formul a o f Theore m 5.1 1 , i t i s sometime s

more practica l t o work wit h th e following approximat e formul a fo r

A
ip(x,x) = ^ (n^(n)-

PROPOSITION 5.25 . For any character we have

(5.65) >(x,
1 X) = 5xx- ] T ?—± + o(|(log^) 2)

|Im(p)|<T

where S x — 1 if X — Xo and Sx = 0 otherwise, 1 ^ T ^ x and the implied constant

is absolute.
P R O O F . Thi s i s a specia l cas e o f (5.53 ) fo r primitiv e characters . However ,
(5.65) als o hold s a s it i s fo r x n ° t primitive . Indeed , suppos e %*(mo d q*) wit h
q* | q is the primitive characte r whic h induce s x(mo d q). The n

pa^.x,p\q

Moreover, th e zero s o f L(s,x) agre e wit h thes e o f L(s,x*) excep t fo r th e one s

coming fro m th e product
n v-p")
which ar e p = 27ril/\ogp. Henc e thes e zero s contribut e a t mos t

E E | £^L|«aog^)2.
p\q pill^T
logp ^

The abov e correction s ar e absorbe d b y th e erro r ter m i n (5.65 ) completin g th e

proof. •
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 121

By (5.65 ) w e derive a strong approximatio n t o

A
ip(x,q,a)= ^2 (n)
n=a (mo d q)

in term s o f zeros o f L-functions. Firs t b y the orthogonality o f characters w e have

ip(x, q, a) = —- Y^ X{a)ip{x, x)
(P[Q)
X (mo d q)

Applying (5.65 ) w e obtain

(5.66) ^ ( W )= ^ -^ ^ m ^ £^ l+ 0 (£(Iogarg)a)
WJ r\HJ x ( m o d q) L{p ,x)=0 F
|Im(p)|^T

for q ^ 1 , (a, q) — 1 and 1 ^ T ^ x , with a n absolute implie d constant .

Theorem 5.1 0 gives th e zero-fre e regio n fo r Dirichle t L- functions, excep t fo r
the possibl e exceptiona l zer o /3 X for a rea l primitiv e characte r x - Becaus e o f its
importance w e include her e th e slightly differen t traditiona l proof .
T H E O R E M 5.26 . There exists an absolute constant c > 0 such that for any
primitive Dirichlet character x modulo q, L(s, x) has at most one zero in the region

(5.67) a ^1 -
logq(\t\ + 3)'
rea
The exceptional zero may occur only if x is h an d it ^ s then a simple real zero,
say (3 X, with

(5.68) ! _ _ £_ </3x < !.

P R O O F . Assum e x i s non-trivia l (th e case o f ((s) i s similar , bu t i t need s a

change of vocabulary becaus e our definition o f L-functions doe s not include ((s+it)
with t ^ 0 for the reason tha t i t has a pole a t s = 1 — it ^ 1 ) . The n conside r the
L-function o f degree 8
L(/, s) = C(s) 3L(s + it, X)4L(s + 2% X2)
for som e t G R where X 2 is the primitive characte r modul o # 2 | q which induce s x 2 •
By the trigonometric inequalit y o f de la Vallee Poussi n
(5.69) 3 H + 4 R e ( ^ ) + Re(£ 2 ) = |z|( 3 + 4cos# + cos2(9 ) = 2|z|( l + cos(<9)) 2 ^ 0
for z = \z\ exp(i#), i t follows tha t R e (A/(n)) ^ 0 for all n > 1 such tha t (n , q) = 1 .
In fact , a simpl e computatio n usin g th e definition o f X2 show s tha t thi s remain s
valid fo r all n ^ 1 .
Let p = / ? + ry be a zer o o f L(/, 5) wit h / 3 ^ \. I f x i s not real, o r if 7 ^ 0 ,
then takin g t = 7 it follow s tha t L(/ , s) has a pole o f order 3 at s = 1 while / 3 is a
real zer o of order 4 of L ( /, s). Henc e b y Lemma 5. 9 we have

/3<1-T -4VT<1~
logq(/,p) log( 9 (|7| + 3))
122 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

with a n absolut e positiv e constan t c. I f \ 2 — 1 > P — P i s r e a l w e tak e t — 0, the n

L(f,s) ha s a pol e o f orde r 4 a t s — 1 . B y Lemm a 5. 9 again , L(f,s) ca n hav e a t
most fou r rea l zero s wit h multiplicit y satisfyin g th e abov e inequality , whic h mean s
that ft mus t b e a simpl e zer o o f L(s , x). •

REMARK. On e ca n als o recove r tha t / ? < 1 by Theore m 2.1 .

As a n applicatio n o f Theore m 5.1 3 , on e derive s th e followin g prim e numbe r

theorem .
THEOREM 5.27 . Let x modulo q be a primitive Dirichlet character. We have
r@x / / —c\0gX
(5.70) Y. x(n)Mn) = S xx - ^ + + O
0 I( xx exp (-7=i^f-)(log^4)
e x p( —
V
^P x VVlog x + logg / /
for any x ^ 1 , wt/ i som e absolute effective constant c > 0, the implied constant
being absolute.
For any q ^ 1 an d (a , g) = 1 we /iai/ e

S 4
(5.71) V A(» )= 4,- ^ ^ + 0 ( x e x p ( -!!° f )(log g) )
n = a (mo d g )

for x ^ 1 , where x (mo d q) is the possible exceptional real character modulo q and
fix the corresponding exceptional zero.
Have i n min d tha t ther e coul d b e on e o r tw o rea l primitiv e character s modul o
q, fo r example , i f q = Sr wit h r od d squarefre e w e hav e tw o suc h character s XsXr
and X4:XsXr- Nevertheless, a t mos t on e o f thes e ma y giv e a n exceptiona l zero . I n
fact, th e exceptiona l character s ar e very rare , an d th e exceptiona l zer o itsel f ca n b e
somewhat controlled .
T H E O R E M 5.28 . (1 ) (Landau) Let xi and X2 be distinct real primitive charac-
ters modulo q\ and q^. Assume Xi an d X2 have real zeros fii and fa respectively.
There exists an absolute constant c > 0 such that
(5.72) m i n O S u & K l -°
loggig 2
(2) (Siegel) For any primitive real Dirichlet character \ modulo q, we have

(5.73) P x < 1 -C M

for any e > 0 , with a constant c(e) > 0 depending only on e. If e < \, this constant
is ineffective, meaning it is not numerically computable.
PROOF. (1 ) Conside r th e L-funetio n o f degre e 4
(5.74) L(/ , s) = ((s)L(s, X i)L(s, X 2 )L(s, X1X2)
which ha s conducto r a t mos t (gi^) 2 - Sinc e th e X1 5X 2 are non-trivial , an d distinct ,
L(/, s) ha s a pol e o f orde r 1 at s = 1 . Moreover ,

~U, s) = E ( ! + X i ^ K 1 + X2(n))A(n)n-°
 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S 123

has non-negativ e coefficients . B y Lemm a 5.9 , ther e exist s c > 0 suc h tha t L(/ , s)
has a t mos t on e rea l zer o with (3 > 1 — c/logq(/). Sinc e q(/ ) ^ 9(^i^2) 27 thi s give s
(5.72).
(2) Fo r th e proo f o f (5.73) , w e follo w th e nic e argumen t o f Goldfel d [Gol] .
Let xi-> X2 be (fo r th e moment ) arbitrar y primitiv e rea l character s modul o q\ an d
#2 an d conside r agai n th e L-functio n (5.74) . W e hav e A/(n ) ^ 0 . Le t r/(x ) b e a
function o n [0 , -foc[ suc h tha t 0 < rj ^ 1 , rj(x) = 1 for 0 ^ x ^ 1 and r)(x) = 0 fo r
x ^ 2 . Conside r th e sum s

Z x
^)= T,^ rl(^)>1 (sinc e A/(l) = 1 )
n

for | ^ / S ^ 1 and x ^ 1 . B y contou r integratio n w e hav e

Z(f3, x) = -^ij Hf, * + P)x sf)(s)ds = L(f, /3 )

(2)

+ L(l,xi)L(l,X2)i(l,XiX2)r?(l-/3)^ 1 - / 3 + ^ J L(f,s + p)f)(s)x s

ds.

By (5.63) , th e las t integra l i s < C Q i ^ ^2 • Thus w e hav e

L(f,p) + L(l, X i)L(l, X 2)L(1 , XiX2)fi(l - f3)x l-P > 1 + Ofatoxh-?).

Now assum e tha t /3 i > | i s a zer o o f L(s,xi) - The n i t follow s

L ( l , x i ) £ ( l , Xa)i(l , XiX2)r/(l - A)* 1 "' 3 1 > 1 + 0( qiq2x^).

4
We take x = c(qiq2) wit h c sufficiently large , s o the ter m o n th e lef t i s > | . Pro m
(5.64) w e have L ( l , x i ) < logg i an d L(l,xiX2 ) < ^ogq 1 q2. Moreover , 77( 1 —/3i) <
(1 — /3i) _ 1 sinc e 7/(5 ) ha s a simpl e pol e a t s = 0 wit h residu e 1 . Gatherin g thes e
estimates w e ge t ou r basi c lowe r boun d
1
(5.75) L( , X2) » ( 1 -/3i)(9i9 2 )- 4 ( 1 - / 3 l ) (log9i92)- 2 .
We will use (5.75 ) t o prov e (5.73) , usin g th e hypothetica l zer o Pi o f L(s , xi) a s
a too l t o contro l othe r characters . Le t 0 < e < ~. I f al l L(s,x) f° r X r e al modul o
# d o no t vanis h o n th e rea l segmen t s ^ 1 — £ , the n (5.73 ) i s obviou s (jus t tak e
c(s) = e) . Suppos e otherwis e tha t fo r som e x i modul o qi , th e L-functio n L(s,xi )
has a zer o s = f3\ ^ 1 — £. The n fo r an y othe r rea l primitiv e characte r X 2 modul o
#2, we can writ e (5.75 ) a s
L(l,x2)»g2-4£(logg2)-2
where th e implie d constan t depend s onl y o n £ , sinc e #i , x i a n d /^ i ar e fixe d onc e
e i s chosen . O n th e othe r hand , i f P2 i s a rea l zer o o f L(s,X2 ) w e hav e b y th e
mean-value theore m
_/32
L(1,X2) = ( 1 - 02)L'(lT, X2) « ( 1 - /?2)<?2 0og<Z2)2
for som e a wit h /3 2 < c r ^ 1 by (5.64) . Combinin g thos e tw o bounds give s 1 — fa ^>
g^"4e, henc e (5.73 ) afte r re-definin g e. D
124 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

REMARK. Essentiall y th e boun d (5.73 ) wa s first prove d wit h e = | b y Landa u

[La2] (als o no t effective) , an d wit h an y e > 0 shortly afterward s b y Siege l [Siel] .
The ineffectivenes s o f th e Siege l boun d (5.73 ) i s apparen t i n applications . Fo r
instance, i t i s no t possibl e t o us e i t t o solv e th e clas s numbe r on e proble m fo r
i_
imaginary quadrati c fields. Yes , b y Siegel' s lowe r boun d L(l,x ) > A 2 £ an d b y
the Dirichle t Clas s Numbe r Formul a (2.31 ) , o r (22.59) , w e ge t

(5.76) h(Q(V-A)) = ^ i ( l , X) » A * " '

for — A a negativ e fundamenta l discriminant , \ bein g th e correspondin g Kronecke r

symbol. Henc e h tend s t o infinit y a s doe s |A| . Bu t whe n computin g al l A wit h
h(Q(\/—A)) = 1 it i s no t possibl e t o kno w whe n t o stop . I n Chapter s 2 2 an d 23 ,
we us e muc h mor e sophisticate d idea s o f Goldfel d involvin g L-function s o f ellipti c
curves t o prov e a n effectiv e bound , althoug h muc h weake r (se e Theore m 23.2) .

Siegel's estimat e implie s th e Siegel-Walfis z theore m abou t prime s i n arithmeti c

progressions.
COROLLARY 5.29 . Let q ^ 1 and A > 0 . We have

(5.77) „ (w) = !g + 0(_i_),

< 5 - 78 ' ^ ;
«'«» = ^ )+ 0 ( ( i ^ F ) '
for any x ^ 2 . Moreover, for any primitive Dirichlet character % modulo q > 2, we
have
(5.79) ^ x ( p ) « v ^x(logx)-'4)
p^x
A
(5.80) Yl x(nMn) < y/qx(\ogx)-
n^x

for any A > 0 . The implied constants depend only on A, but are ineffective.
PROOF. I f q ^ (lo g x)A+1 on e gets (5.78 ) from th e prime number formul a (5.71 )
by employin g th e boun d (5.73) . Fo r large r q there i s nothing t o prove , becaus e th e
formula (5.78 ) i s trivial . The n (5.77 ) follow s fro m (5.78 ) b y partia l summation .
Similarly on e derive s (5.79 ) fro m (5.52 ) b y employin g th e boun d (5.73 )
The cas e o f (5.80 ) i s a bi t harder . On e coul d follo w th e proo f o f (5.51 ) usin g
I / - 1 instea d o f V jL, bu t ther e ar e seriou s complication s whic h occu r nea r multipl e
zeros or cluster s o f zeros (s o far w e cannot rul e ou t thi s scenario) . Therefor e w e ar e
going t o deriv e (5.80 ) directl y fro m (5.79) . Firs t w e writ e

Yl XWM W = 1 - Yl X{™>)v(m) ^ xip)

n^.x m^x Pm<P^^

where p m stand s fo r th e larges t prim e diviso r o f m . Le t r ^ 0 b e th e numbe r o f

prime divisor s o f m . The n th e summatio n condition s impl y m 1 //r ^ p m < x/m, s o
m < x r^r+l\ Hence , usin g (5.79 ) th e inne r su m ove r p satisfie s
A
^fqx / x\~ ^/qx /r + l\ A r(m) s/qx
y
*—* m m \ mJ m \V log x / 7 m(logx)
log2
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 125

because (r + l)A < C 2r = r{m). Finally , summin g ove r m trivially w e obtain (5.80 )
with extr a facto r (lo g a;)2. Thi s complete s th e proo f b y resettin g A. •

For Dirichle t characters , th e problem discusse d i n Proposition 5.2 2 becomes

that o f finding th e smallest prim e p wit h x(p) ¥" 1 f°r a non-trivia l characte r x
modulo q > 2, say p m in(x)- On e derive s directl y fro m variou s estimate s fo r short
character sum s th e followin g bounds :

+£
Pmmix) < Vqiogq, 4* , <f , (logg) 2
using respectively th e estimate s o f Poly a-Vinogradov, Burgess , Lindelof hypothesis ,
Riemann hypothesis . Actuall y w e only ge t (logg) 4 by (5.56), but wit h a little extr a
effort on e ca n squeez e fro m th e GRH th e bound (logg) 2 . On e ca n also d o better
unconditionally usin g Burgess's estimate enhanced b y combinatorial argument s (se e
e.g. [HI]) . Linni k proved (b y the large sieve inequality) tha t th e smalles t quadrati c
non-residue fo r a primitive rea l characte r modul o q ^ N, q prime, i s < N £ wit h at
most finitely man y exceptions , th e numbe r o f exceptions dependin g onl y o n e > 0.
See Sectio n 7. 4 for the proo f an d a n analogue fo r elliptic curves .

The Rieman n zet a functio n i s further discusse d i n Chapter s 1 0 , 24, 25, an d

Dirichlet character s i n Chapters 1 7 , 1 8 .
5.10. L- functions o f number fields .
Let K/Q b e a number field of degree d. Th e Dedekin d zet a functio n

at*) = £ (;va)- s = n a - (^prr1

defines a self-dual L- function o f degree d — [K : Q] wit h conducto r D = |dx| , the
absolute valu e o f the discriminan t o f K, and roo t numbe r + 1 . Th e gamm a facto r
is

dS/2r
(5.81) ^ ) - - (I)n+r2r(^)r2
where r\ i s the number o f real embedding s o f K an d r2 the number o f pairs of
complex embedding s s o that d = T\ + 2r2- Th e zet a functio n ha s simple pol e at
s — 1 with residu e
„ , N 2 ri (2yr) r2 hR
res (K(s) = - =,
5=1 wy/D
where h is the class numbe r o f K, R the regulator an d w the number o f roots of
unity. Fo r a proof, se e e.g . [La] .
The Ramanujan-Petersso n conjectur e i s obviously satisfie d fo r C K ( 5 ) - Th e an -
alytic conducto r is

q(CK, s) = D(\t\ + 3 P +r 2 ( | i + 1 | + 3f2 < D(\t\ + 4) d

and (\(CK) = 2 r i + r 2 3 r 2 D ^ 3 dD. Ther e exist s a n absolut e constan t c > 1 such tha t
D > cd~1 (se e Exercis e 1 0 below) s o we hav e logq(Cx ) x \ogD.
126 5 . CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

REMARK. Dedekin d zet a functions facto r a s products o f Artin L-function s (se e

Section 5.1 3) . Le t E/Q b e th e Galoi s closur e o f K, G = Gal(E/Q) an d H =
Gal(E/K). Le t TQ/H D e th e permutatio n representatio n o f G o n th e cose t spac e
G/H. Thi s is the representation induced by the trivial representation of H, henc e by
the invarianc e o f Arti n L-function s unde r inductio n w e hav e L{TQJ H,S) — (K(S)>
One ca n decompos e TQ/H m term s o f irreducibl e representation s o f G, TQJ H —
(Bpnpp, gettin g th e factorizatio n

n
(5.82) ( K(s) = l[L(s,p) ».
P

In cas e o f a failur e o f th e Arti n conjectur e ( a possibilit y w e canno t ye t rul e out )

this i s not a factorizatio n i n term s o f L-function s i n th e sens e o f Sectio n 5.1 .
If K/Q i s abelian , the n E = K, H — 1 and r G/H = 0 x wher e \ r u n s o v e r a
group o f Dirichle t character s s o tha t

CK(s) = l[L(s,x)
X

is a produc t o f L-function s fo r distinc t primitiv e Dirichle t character s o f conductor s

whose produc t i s D = \CIK\ and exactl y on e characte r i s trivial .

We deduc e fro m th e abov e an d th e result s o f th e previou s sections :

THEOREM 5.30 . Let K/Q be a number field of degree d. We have

(s-l)CK(s)^\Dsd\^-^2+e

for 0 ^ a ^ 1 , the implied constant depending only on d and e.

THEOREM 5.31 . Let K/Q be a number field of degree d. Let N K{T) be the
number of zeros p = (3 -f- ij of Or(s) in the critical strip 0 ^ / 3 ^ 1 with \*y\ ^ T.
We have
d
T DT
NK(T) = - lo g —~ d + 0(log DT d)
for T ^ 2, the implied constant being absolute.

One can use the analyti c properties of L-functions t o deduce information o n th e

discriminant o f numbe r fields. Fo r instance , th e explici t formul a yield s quit e goo d
lower bound s fo r D = |d^| . Thi s powerfu l metho d wa s implemente d b y Odlyzk o
[Od], Poito u [Poi] , Serr e [Se7 ] an d others . Her e i s a simpl e result .
T H E O R E M 5.32 . Assume GRH for the Dedekind zeta functions of number
fields. Then we have

(5.83) liminfi{log D+ (d-r 2 )^(i)+r2^(f)}^log7r,

as d = [K : Q ] = n + 2r 2 -5 - +00 .
 5. CLASSICA L ANALYTI C T H E O R Y O F 1
L-FUNCTION S 2 7

PROOF. Appl y th e explicit formul a (5.44 ) t o (K(S) wit h

a x
(P(x) = x~h- ^ ^\
for som e a > 0. Althoug h thi s tes t functio n i s not of compact support , i t decay s
very fas t a t infinity , s o (5.44) holds . Th e Mellin transfor m i s

<P(± +lt) = \ e 4a

for t G 1. Not e tha t thi s i s positiv e (compar e wit h Propositio n 5.55 ) and tha t
4>(x) = x~ 1 cj){x~l). Sinc e (K(S) is self-dual w e ge t
11 1 p+°° i p v
X)A/f(n)0(n) + - ^ 0 ( p ) = -log£ > + - / <l,{x)dx + —J j-(s)4>(s)ds.
np - .
(5)
By positivity , w e derive
11 1f +°° f V
(5.84) -\ogD + -J 4>(x)dx+— I ^(#)d S )0.

By (5.81 ) an d the Fourier inversio n formul a w e hav e

(5 85) S
' h l ^ M*)ds = ~\ l0g7T + ^_lAj^ {lA + «)^(1 + i t ) d t

When a—>> 0, the function (27r) -1 ^)(^ +it) converge s to the Dirac delt a a t 0, hence

Therefore, dividin g (5.84 ) b y d, then lettin g d—>> +00 , an d then lettin g a —> 0 , w e
derive (5.83) . •

Similar argument s ca n be applie d unconditionall y wit h slightl y weake r resul t

using tes t function s a s constructed i n Proposition 5.55 .
EXERCISE 1 0 . Sho w tha t ther e exist s a n absolut e constan t c > 1 such tha t
D = \dx\ > cd for any number field K/Q o f degree d ^ 2 , and, in particular, D ^ 3
if K 7 ^ Q so if ha s at least on e ramified prime . [Hint : Procee d a s in Theorem 5.51
below.]
One canno t appl y directl y Theore m 5.1 0 becaus e th e Rankin-Selberg squar e of
CK(S) involve s convolution s o f Artin L- functions o f K whic h ar e not known t o be
entire, althoug h the y ar e conjectured t o be. However , on e can adapt th e traditional
arguments (a s in the proof o f Theore m 5.26 ) using (K(S) 3CK(S + ^ ) 4 C K ( S + 2i£ )
128 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

and deriv e th e zero-fre e region , the n th e prim e numbe r theore m fo r K. W e phras e

it i n term s o f
f lo g A^p i f a = p k fo r som e fc ^ 1 ,
\ 0 otherwise .
Thus thes e ar e coefficient s i n th e Dirichle t serie s

THEOREM 5.33 . Let K/Q be a number field of degree d. There exists an abso-
lute constant c > 0 such that (K(S) has no zero in the region

d2\ogD(\t\+3)d
except possibly a simple real zero (3 < 1 .
We have
XP
2
}] Atf(a ) = x - — + 0(vDx exp(-ccT \/logx) )
Na^x P

for x ^ 2, where c > 0 is an absolute constant, and the term —x^/p should be
removed if there is no exceptional zero.
Be awar e tha t th e zero-fre e regio n o f Theore m 5.3 3 i s no t necessaril y th e bes t
accessible. Indee d b y (5.82) , on e ca n factoriz e (K{S) an d i f on e know s tha t th e
factors ar e L- functions, the n zero-fre e region s fo r the m produc e on e fo r (K whic h
can b e muc h bette r becaus e o f a smalle r conductor . Fo r instance , i f K — Q(/ip ) i s
the fiel d o f p-th root s o f unity , w e hav e

(K(S) = ((S)
1 [L(S,X)

where the produc t i s over non-trivial Dirichle t character s modul o p. Thu s b y Theo -
rem 5.2 6 an d Theore m 5.2 8 ther e exist s a n absolut e constan t c > 0 such that 0<r( s)
has a t mos t on e simpl e rea l zer o (3 in th e regio n
r.

logp(|*| + 3 )
whereas d = p — 1 and D = |d# | = p p~2 s o Theore m 5.3 3 i s muc h weaker .
Confirming tha t th e zero s o f rea l Dirichle t character s ar e th e mos t difficul t t o
deal with, Star k [St3 ] has show n tha t i f K/Q doe s not contai n a quadratic subfield ,
then ther e i s n o exceptiona l zer o o f CK(S) .
For Dedekin d zet a functions , Propositio n 5.2 1 ca n b e improved .
PROPOSITION 5.34 . Let K/Q be a number field. Assume GRH holds for ( K(s).
We have
, . , N lo g 3D
s=i lo g log 3D
where D = \dic\ ^nd the implied constant is absolute.
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 129

P R O O F . Le t m b e the order o f vanishing a t \. W e argue a s in the proof of

(5.62). I n applying th e explicit formul a (5.45 ) a n additional ter m arise s fro m th e
simple pol e o f CK(S) a t s = 1 , hence we ge t
/»+o o
mh(0) + J2 h(^)= J 9(y)e y/2
dy

-2J2^7=-9(logNa) + 0(\og3D).
a vW a
Both sum s ove r th e zeros and the prime ideal s can be dropped b y positivity. Sinc e
h(0) x Y an d the integral i s 0(e Y) w e infer tha t mY < C eY + lo g 3D. Takin g
Y = log log 3D yield s th e result. •

Generalizing Dirichle t character s t o number fields ar e Hecke character s (Gros -

sencharakters). Thos e are defined thoroughl y i n Section 3.8 for imaginary quadrati c
fields, bu t her e w e use the basic terminolog y i n the general cas e (fo r characters of
weight 0) . See e.g . [La ] fo r details.
Let K/Q b e a number field an d £ a primitive Heck e Grossencharakte r modul o
(m, f2) wher e m is a non-zero integra l idea l in K an d ft is a set of real infinite place s
where £ is ramified. Th e Hecke L-functio n i s defined by
-1
L(Z,8) = l[(l-Z(p)Np->)

for R e (s) > 1 . Heck e showe d tha t L(£ , s) i s an L-function o f degree d — [K : Q ]

which i s entire i f £ ^ 1 and which coincide s wit h CK(S) for £ = 1 . The conducto r is
A — \dK\N K/qxn, wher e N K/Q i s the norm an d the gamma facto r i s given by

7«,*) = .r--/>r(ip'-|n|r(i±A)"+|01,
so q(£ ) ^ 4 d\dK\NK/qm = 4 d A. The root numbe r ca n be explicitly writte n a s a
normalized Gaus s su m for £.
One get s th e following generalizatio n o f Theorem 5.26 .
T H E O R E M 5.35 . Let K/Q be a number field, £ a Hecke Grossencharakter mod-
ulo (m , ft). There exists an absolute effective constant c > 0 such that !/(£,« ) has
at most a simple real zero in the region

dlogA(|t| + 3)'
The exceptional zero can occur only for a real character and it is < 1.
R E M A R K . Thi s sectio n highlight s th e fac t tha t L-function s o f variou s type s
are naturall y define d ove r an y number field K, usin g Dirichle t serie s ove r non-zer o
integral ideals , an d Euler product s ove r prime s ideals . Thu s ther e ar e Art in L-
functions ove r K, automorphi c L-function s ove r K, L-functions o f varieties ove r K,
whereas i n a sens e o f analytic numbe r theor y w e like t o see these a s L-function s
over Q . Thi s vie w i s possible becaus e an y L-function o f degree d ove r if , wher e
[K : Q ] = / , i s a n L-functio n ove r Q , a s define d i n Sectio n 5.1 , of degre e df.
For instance , (K(S) i s of degre e 1 over K. Thu s i t i s easy fo r the reader t o see
what result s ther e ar e for L-function s ove r numbe r fields, wit h uniformit y i n the
130 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

number field. Sometimes , however , th e proo f ha s t o b e adapted , a s fo r Theore m

5.33, becaus e th e Rankin-Selber g //-function s hav e a differen t for m dependin g o n
whether th e L-function s ar e see n ove r K o r ove r Q . Th e reade r shoul d hav e n o
difficulty supplyin g ne w proof s i f necessary . I t i s wort h pointin g ou t tha t fo r al l
known L- functions, th e conducto r o f a n L-functio n o f degre e d ove r K take s th e
form q = \dK\ dNK/qf fo r som e non-zer o integra l idea l f of K. Thu s th e conducto r
and th e analyti c conducto r contai n th e dependenc y i n th e discriminan t o f the bas e
field K.

Analytic numbe r theor y flourished beautifull y i n th e las t tw o centurie s i n th e

garden o f L- functions ove r th e rationals . Th e extensio n t o numbe r fields i s fruitfu l
as well, bu t i t need s t o b e mor e explored . T o encourage th e newcomer s w e end thi s
section wit h a fe w cut e application s o f th e theor y ove r th e Gaussia n domai n Z[i] .
There ar e fou r unit s i n Z[i], namely 1 , —1, i, —i. Ever y idea l o f Z[i] i s principal.
Consider th e "angle " character s

6(a) = ( Ar = eifcarga
if a = (a ) 7 ^ 0 fo r an y k = 0 (mo d 4) . Thi s i s a primitiv e Heck e characte r o f
conductor (1 ) . Th e associate d L-functio n

has conducto r D = 4 an d i t satisfie s th e self-dua l functiona l equatio n

A(5,&) = TT- ST(S + J|1 )L( 5 , &) = A( l - s, &)•
All characters are complex except the trivial character fo r k = 0 giving the Dedekin d
zeta functio n L(s,£o ) = C( s)L(s,X4)- Therefor e ther e i s n o exceptiona l zero , an d
the Prim e Numbe r Theore m (5.52 ) give s
tk c
J2 (^) = 26kU(x) + 0(\k\xe- ^)

where c > 0 an d th e implie d constan t ar e absolute .

As a n exercis e w e propos e t o deriv e fro m th e abov e formul a th e followin g
equidistribution propert y o f Gaussia n prime s i n sectors .
THEOREM 5.36 . Let 0 < / 3 - a ^ f and x ^ 2 . Then
c
\{TT e Z[i ] | K K x , a < argy r < /3} | = ^ ^ L i ( x ) + o(xe~ ^^^

where c > 0 and the implied constant are absolute.

[Hint: Approximat e the characteristic function o f [a , /?] modul o 2n by a smooth
function, periodi c of period 2-7r , and symmetric with respect t o x— > ±x±n. Expan d
this functio n int o Fourier serie s and appl y fo r x = arg7r . B y the symmetr y th e non -
zero Fourie r term s ar e o n frequencie s k = 0 (mo d 4) . A t eac h frequenc y appl y th e
Prime Numbe r Theorem. ]
See Sectio n 1 1 . 8 fo r a n applicatio n o f thi s resul t t o sho w tha t th e Wei l boun d
for th e numbe r o f points o n a curve over a finite field i s optimal i n horizonta l sense .
 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S 131

EXERCISE 1 1 . Playin g with arg a an d |a | deriv e the equidistributio n o f Gauss -

ian prime s i n regula r domain s o f C .

EXERCISE 1 2 . Prov e tha t ther e ar e infinitel y man y pair s o f prime s p , q suc h

that pq = a 2 + b 2 with 0 < b < ( 3 log a)2 .

We recommen d W . Duk e [Du4 ] fo r a powerfu l treatmen t o f a variet y o f prob -

lems b y mean s o f Heck e characters .
5.11. Classica l automorphi c L- functions.
By classica l automorphi c form s w e mea n th e form s o n GL(2) , eithe r holomor -
phic o r eigenfunction s o f th e Laplac e operato r (th e Maas s forms) .
Let / b e a primitiv e holomorphi c cus p form , o f weigh t k ^ 1 , leve l q, wit h
nebentypus \ ( s e e Chapte r 1 4 for definitions) . Le t

n^l

be it s normalize d Fourie r expansio n a t th e cus p oo . The n

np

is a n L-functio n o f degre e 2 with conducto r q and gamm a facto r give n b y

(5.86, 7(/,„ = » - . r ( l ± f c i ^ ) r ( l ± i ^) _ f t p.)~r(. + ^i)

with Ck = 2^ 3~h^2y/7r b y th e Legendr e duplicatio n formula . Therefor e th e analyti c

conductor i s

q(/, S ) = g ( | S + ^ | + 3 ) ( | s + ^ | + 3 ) < g ( H + |fc|+3) 2 ,

q(/) = g ( ^ i + 3 ) ( ^ i + 3)x 9 fc 2 .

The roo t numbe r i s i kf](f) wher e rj(f) satisfie s Wf = rjf. I t i s known tha t L(f,s)
satisfies th e Ramanujan-Petersso n conjectur e b y wor k o f Delign e [Del ] fo r k ^ 2
and Delign e - Serr e [DeSe ] fo r k = 1 . Th e dua l for m / satisfie s
x n
f( ) = XWA/W, i f (n,q) = 1.
Have i n min d the n / ca n b e self-dua l fo r non-trivia l nebentypus . Fo r example , i f
£ i s a clas s grou p characte r o f a n imaginar y quadrati c field K — Q(>/!}) , the n th e
form / wit h coefficient s give n b y

A/(n) = Y, * W
Na—n

is self-dual o f weight k — 1, of level q = \D\ an d nebentypu s x = XD, the Kronecke r

symbol. Fo r thes e facts , se e Sectio n 1 4. 7 an d Propositio n 1 4.1 3 , an d th e reference s
there.
132 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

Similarly, le t (p b e a primitiv e Maas s for m o f leve l q with nebentypu s x whic h

is an eigenfunctio n o f the Laplac e operato r wit h eigenvalu e A = \ + r 2 , wher e r £ M
or ir e [0 , \{. Writin g it s Fourie r expansio n a t infinit y i n th e for m

ifi(z) = y/y^2 p(n)K ir(27r\n\y)e(nx),

we associat e wit h (p th e L-functio n

n^lp

of conducto r g , wit h gamm a facto r

where 5 = 0 if <p i s even an d S = 1 otherwise. Thu s th e analyti c conducto r i s

q(tp, s) = q(\s + ir\ + 3)(|a - ir\ + 3) < g(|s | + |r | + 3) 2 ,

q(^)=g(|r|+3)2xA(7.

We hav e p(n) — x( n)p(n) ^ ( n> Q) — 1- Se e Chapte r 1 5 for th e definitio n o f Maas s

forms. Althoug h L- functions an d Heck e operator s fo r Maas s form s ar e no t full y
presented i n thi s book , th e theor y i s very simila r t o tha t o f holomorphic forms ; se e
for instanc e [BG] , [Bu ] an d [DFI3] .
In contras t wit h holomorphi c forms , i t i s no t know n tha t th e L- functions o f
primitive Maas s cus p form s satisf y th e Ramanujan-Petersso n conjecture , althoug h
there i s n o doub t becaus e th e conjectur e fit s correctl y i n th e genera l Langland s
functoriality program . Th e curren t bes t know n estimat e i s

(5.87) Kl\P P\ < P 7/64

an
where 1 — p(p)p~ s + X{P)P~ 2S
— ( 1 ~ a pP~s)(l ~ PpP~ s), d correspondingl y
|Reir| ^ ^ , henc e

(5.88) A = i + r»>^=0.238...

This is due to Kim an d Sarna k [KiS] . For earlier results see e.g. [S2] , [Se8] , [DuI2] ,
[BDHI], [LRS] .
The origina l theor y o f Ranki n [Ra3 ] an d Selber g [S5] , extended b y W . L i [L] ,
and it s adeli c versio n [JPS ] sho w tha t an y tw o cus p forms , holomorphi c o r non -
holomorphic, admi t a Rankin-Selberg convolution . Th e convolutio n L ( / ® # , s) ha s
a simpl e pol e a t s — 1 if g = / an d i s entir e otherwise . Assum e / o f weigh t k an d
g of weigh t k ^ £ (by symmetry) . Th e gamm a facto r o f / 0 g i s give n b y
 5. CLASSICA L ANALYTI C T H E O R Y O F 1
L-FUNCTION S 3 3

Let no w / hav e eigenvalu e ^ -f - r 2 , an d parit y 5 = 0 o r 1 , an d g hav e eigenvalu e

\+u2 an d parit y rj = 0 or 1 . The n

/s — i(r - u) + i/\ /s - i(r + u) + v\

where v = 0, 1 accordin g t o whethe r 8 = rj or not . Finall y assum e / i s holomorphi c

of weigh t k an d g i s a Maas s for m wit h eigenvalu e \ + r 2 an d parit y 5 . The n

•r(/8»,.)«-T(l±iiiii)r(l±^i)

These factor s ca n b e derive d b y pleasan t computation s fro m th e fact s convenientl y

gathered i n th e Appendi x t o [RS] .
Although th e loca l factor s o f L(f 0 g, s) a t prime s p | q(f 0 g) canno t b e
expressed i n a simpl e wa y fro m th e Heck e eigenvalue s A/(p) , X g(p) o r p/(p) , Pg{p),
this i s stil l th e cas e i f p divide s exactl y [<?(/) , #(#)]. I n thi s cas e on e ca n sho w tha t
the facto r ove r th e ramifie d place s i n (5.1 0) , o r rathe r 1 /H p(p~s), i s equal t o
(1 - a f(p)ag(p)p-s)(l - a f(p)(3g(p)p-s)(l - f3 f(p)ag(p)p-s)(l - p f(p)(3g(p)p~
s
)-
In othe r words , th e loca l facto r a t al l prime s i s given b y (5.9) . W e emphasize agai n
that thi s i s not tru e i n al l cases. Not e tha t sinc e p \ q(f)q(g), on e or bot h o f a/(p) ,
Pf(p) ( res P- a g(p)i Pg(p)) i s z e r o •
In term s o f Dirichle t series , thi s implie s tha t
s
L{f ®g,s)= L(2s , XfXg) Yl A/(n)A^(n)n"

if [q(f),q(g)] i s squarefree, wher e / an d g can b e either holomorphi c o r Maass form s

and Xfi Xg a r e th e respectiv e nebentypus . Not e tha t L(s,XfXg) ls ^ n e Dirichle t
series o f th e characte r XfXg eve n i f i t i s non-primitive , so , fo r instance , i f Xf =
Xg = 1 the n L(2s,XfXg) = Cq(f)q{g){^ s) 1 S th e Rieman n zet a functio n wit h th e
Euler factor s a t p \ q(f)q(g) removed . I n th e genera l case , th e loca l factor s a t
ramified prime s ca n b e describe d usin g th e loca l Langland s conjectur e fo r GL(2) ,
but thi s i s no t ver y explicit . Se e Sectio n 5.1 2 fo r th e relatio n betwee n L(f 0 / , s)
and th e symmetri c squar e L-function .
REMARK. (1 ) Th e convolutio n o f a Maas s for m an d a n holomorphi c cus p for m
of weight 4 is important i n the Phillips-Sarna k theor y o f spectral deformatio n [PS] .
(2) On e ca n als o conside r convolution s o f GL(1 ) wit h GL(2). Thos e ar e th e
twists define d i n Sectio n 1 4.8 , and th e theor y i s rather elementary . Note , however ,
that th e twiste d modula r for m
1 k )/2
/ 0 i> = J2 ip(n)\ f(n)n( - e(nz)
n

(in case of holomorphic / ) o f level gm 2 , wher e ip is modulo m , ma y not b e primitive ,

although i t i s alway s a n eigenfunctio n o f th e unramifie d Heck e operators . I n thi s
case th e Rankin-Selber g L-functio n L(f 0 -0 , s) i s th e L-functio n o f th e primitiv e
134 5 . CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

form wit h conducto r dividin g qm 2 whic h ha s th e sam e Heck e eigenvalue s a s / 0 ip

at almos t al l prime s (se e Exercis e 5 of Sectio n 1 4.7) .
As an example, le t / b e a theta functio n fo r a class group character (se e Sectio n
14.3) for a n imaginary quadrati c field K — Q(\J—D), an d \D th e Kronecker symbo l
for thi s field, s o that / E Si(D,XD)- The n w e have L(f ® XD-> S) = L(f, s).
More generally , on e ca n expres s th e dua l L-functio n L ( / , s) o f a modula r for m
/ havin g nebentypu s x as a twist
(5.89) L(f,s) = L(f®x,s)
(see (1 4.48)) .
It i s quite instructiv e t o fix a modular for m an d conside r it s famil y o f twist s
/ ® X by character s o f conducto r q, or b y rea l character s o f conducto r < Q.

For ease of reading we restate som e of the general results from previou s section s
now i n th e specifi c contex t o f classica l automorphi c L-function .
THEOREM 5.37 . Let f be a holomorphic primitive cusp form of level q and
weight k ^ 1 as above. We have
1
(5.90) L ( / , s ) « ( v^(|S|+fc)) -+£)
for \ < a < 1.
Let ip be a primitive Maass cusp form of level q with the Laplace eigenvalue
A = \ +r 2. We have

(5.9i) ^ . i X ^ H+ H))1-^
for I ^ a ^ 1 . In both cases, the implied constant depends only on e.
PROOF. Th e inequalit y (5.90 ) i s merel y th e re-formulatio n o f (5.20) wit h th e
claimed uniformity , becaus e L(f,s) satisfie s th e Ramanujan-Petersso n conjecture .
To prove (5.91 ) , on e use s instea d th e estimat e

(5.92) Yl \p( n)\2 < * ( * + ? + \r\) £

proved b y Iwanie c [1 7] , hence agai n th e implie d constan t i n (5.91 ) depend s onl y o n

e, no t o n (p. •

Next (5.33 ) give s

THEOREM 5.38 . Let f be a holomorphic primitive cusp form of level q and
weight k ^ 1 , N(T, f) the number of zeros p = /3 + ry of L(f, s) in the critical strip
0 ^ / 3 ^ 1 with \j\ < T. We have

N(TJ) = ^ l o g ^^ +0(log(i(|T | + fc)),

for T ^ 2 with an absolute implied constant.

Let (f be a primitive Maass form level q with eigenvalue A = \+r2, N(T, <p) the
number of zeros p =/3 + ry of L(f, s) in the critical strip 0 ^ / 3 ^ 1 with \j\ ^ T.
We have
N{T,v) = ~ log j f^ + 0(}ogq(T + \r\))
7r [Z7re)*
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 135

for T ^ 2 with an absolute implied constant.

Now we re-formulate ou r zero-free regio n theorems. Th e first result for modular
forms beyon d th e classical contex t o f Dirichlet character s (an d their extension s t o
number fields) i s due to C. Moreno [Mori] . W e ca n apply Theore m 5.1 0 for both
holomorphic an d non-holomorphic forms , gettin g th e following zero-fre e region :
T H E O R E M 5.39 . Let f be a holomorphic primitive cusp form of level q and
weight k > 1 . There exists an absolute constant c > 0 such that L(/ , s) has no zero
in the region
c
°^ ~log</(|t|+f c + 3)
except possibly a one simple real zero (3 < 1 in which case f is self-dual.
Let (f be a primitive Maass form of level q and eigenvalue A = \ + r 2. There
exists an absolute constant c > 0 such that L(ip, s) has no zero in the region
r
G> 1
logg(|*| + |r | + 3 )'
except possibly a one simple real zero /3 < 1 in which case f is self-dual.
Finally w e state th e prime numbe r theore m fo r classical cus p forms .
T H E O R E M 5.40 . Let f be a holomorphic primitive cusp form of level q and
weight k ^ 1 . We have
x (3
/
] T A / 0 ) logp = - — + 0(y/qx exp(-c v logx))
p^x

for x > 2, where c > 0 is some absolute constant and f3 is the exceptional zero of
L(/, s) if it exists and this term is omitted otherwise.
Let if be a primitive Maass cusp form of level q and eigenvalue A = | + r2. We
have
Y^p{p)^gp= -— + 0(y/qx exp(-c^logx))

for x ^ 2, where c > 0 is some absolute constant and (3 is the exceptional zero of
L(cp, s) if it exists and this term is omitted otherwise.
PROOF. Appl y Theore m 5.1 3 (see th e remark following ) an d (5.92 ) t o verif y
(5.48). •

Classical cusp forms giv e examples of families of L-functions wher e many result s
known otherwis e onl y unde r GR H ca n be proved unconditionally . Fo r instance, the
upper boun d o f Proposition 5.2 1 ca n be improved significantly .
Let q be a prime numbe r an d 5^(q)* th e set of primitive weigh t 2 cusp form s
of leve l q (see Chapte r 1 4) . W e have \S2(q)*\ ~ ^ . Defin e

(5.93) L(J 0(q),s)= n L(f,s).

fes2(q)*

This i s an L-function o f degree d = \S 2(q)*\ wit h q(/ ) = (1 2q)\ S2^*\, henc e d ~ §

and logq(/ ) ~ ^ l o g g . Therefor e i t follow s fro m Propositio n 5.2 1 an d GRH tha t
136 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

the orde r o f vanishin g o f L(Jo(q),s) a t s — \ i s < C (/(loggXloglogg) x . However ,

applying th e Petersso n formul a (1 4.60 ) on e ca n easil y sho w (stil l usin g GRH ) tha t
(5.94) or d L(J 0(q),s) < q.
s=l/2
In Chapte r 26 , w e sho w tha t thi s boun d i s sharp . Actuall y th e boun d (5.94 ) i s
established unconditionall y i n [KM2 ] (withou t recours e t o th e Gran d Rieman n
Hypothesis). Se e Sectio n 5.1 3 fo r link s wit h Hasse-Wei l zet a function s o f abelia n
varieties.
5.12. Genera l automorphi c L-functions .
In this section we talk briefly o f L-functions associate d with cusp forms o f higher
rank, becaus e the y ar e abou t t o mak e a permanen t habita t i n analyti c numbe r
theory i n year s t o come . Le t / b e a cus p for m o n GL(m)/Q wit h m ^ 1 . The n
Godement - Jacque t [GJ ] hav e defined a n L-function L(/ , s) o f degree m associate d
to / . Fo r m = 1 , / correspond s uniquel y t o a primitiv e Dirichle t characte r \
modulo q an d L(f,s) = L(s,x) - Fo r m = 2 , / correspond s eithe r t o a primitiv e
cusp form , o r t o a primitiv e Maas s cus p form . Henc e th e automorphi c L-function s
L(f,s) generaliz e bot h Sectio n 5. 9 an d Sectio n 5.1 1 . Sinc e / i s cuspidal , L(f 1 s)
is entir e excep t fo r L(/ , s) = ((s) wit h m = 1 . I n additio n L-function s o f non -
cuspidal automorphi c representation s ar e als o defined , bu t the y facto r a s product s
of cuspida l ones . W e recommen d Cogdell' s Chapte r 9 o f [BG ] a s a surve y o f L -
functions o n GL(m).
It i s expected that an y automorphic L-functio n satisfie s th e Ramanujan-Peters -
son conjecture . Jacque t an d Shalik a prove d tha t th e loca l component s satisf y
l a i(p)l < yjv a n d Re(ftj ) > — |. Thi s ha s bee n improve d b y Luo , Rudnic k an d
Sarnak [LRS] , namely |c^(p) | < p c an d R e (KJ) > —c with

The absolut e convergenc e o f L(f,s) fo r R e (s) > 1 , du e t o Jacquet-Shalik a an d

Shahidi [JS] , follow s fro m th e existenc e o f Rankin-Selber g L-function s (se e th e
proof o f (5.48)) .
A usefu l featur e o f th e associatio n o f L-function s wit h automorphi c theor y i s
that th e roo t numbe r e(f) i s expressed a s a produc t o f loca l roo t number s

e(f) = T\s P(f), wit h e p(f) = 1 if p i s unramified .

p

In fact , s p(f) depend s o n th e choic e o f a non-trivia l additiv e characte r ip p o f Q p ,

but th e produc t i s independen t o f suc h choices .
For an y tw o cus p form s / an d g o n GL(d) an d GL(e), th e L-functio n L(f 0
g, s) exist s i n ou r sens e b y th e far-reachin g generalizatio n o f th e classica l Rankin -
Selberg metho d du e t o Jacquet , Piatetski-Shapir o an d Shalik a [JPS] . Moegli n an d
Waldspurger [MW ] prove d tha t L(f<8>g, s) i s entire i f g / / , an d ha s a simple pol e
at s = 1 if g = / . Th e boun d q(f <8> g) \ q(f)eq{g)d i s due t o Bushnel l an d Henniar t
[BH].
According t o conjecture s belongin g t o th e Langland s Program , th e "mos t gen -
eral" L-functio n shoul d i n fac t b e a produc t o f function s belongin g t o th e clas s
of L-function s o f cus p form s o n som e GL(m)/Q . Othe r part s o f th e Langland s
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 137

conjectures impl y tha t th e Ramanujan-Petersso n conjectur e shoul d hol d fo r an y

automorphic //-function . Moreover , th e conjecture s impl y th e existenc e (an d de -
termination, i n principle ) o f a n equidistributio n la w fo r th e coefficient s Xf(p) (se e
the discussio n o f th e Sato-Tat e Conjectur e i n Chapte r 21 ) .
The convexit y boun d fo r automorphi c //-function s ca n b e mad e uniform .
T H E O R E M 5.41 . Let L(f,s) be an automorphic L-function of a cusp form of
degree d other than C(s) . Then we have
(5.96) L ( / , s ) « q ( / , S r+£
for a ^ \, where a = m a x ( | ( l — a),0) , the implied constant depending only on e
and d.
PROOF. Thi s follow s fro m (5.1 2 ) wit h X = 1 , Proposition 5. 4 an d th e estimat e
y
£\*f(n)\2«x(xq(f)y

for x ^ 1 with a n implie d constan t dependin g onl y on e and d. Thi s i s the analogu e
of (5.92 ) prove d b y Molten i [Mol] . •

The mos t commonl y use d L-function s o f degre e fou r i n analyti c numbe r the -
ory ar e th e Rankin-Selber g convolution s L(f ® g,s) o f tw o classica l cus p form s
(described i n Sectio n 5.1 1 ) . Clos e relative s t o thi s ar e th e symmetri c squar e an d
adjoint squar e L-function s o f classica l cus p forms . Le t / hav e nebentypu s \- Th e
symmetric squar e o f / ( / i s either holomorphi c o r real-analytic ) i s defined b y
2 1
(5.97) L(Sym /, a) = L(f ® /, s)L(s, x ) "
whereas th e adjoin t squar e i s define d b y
2
(5.98) L(Ad /, s) = L(f ® /, s)C(^)- 1 .
Thus bot h ar e L-function s o f degre e three , wit h conductor s q(Sym 2 f) = q(f ®
fMx)-1 I <7(/)2 an d q(Ad 2f) = q(f ® /) | q(f) 2. I f / ha s rea l coefficient , the n
the symmetri c an d adjoin t square s coincide . On e show s tha t th e roo t numbe r fo r
the adjoin t squar e i s e(Ad 2 / ) = e(f ® /) = 1 . Bot h hav e gamm a facto r equa l t o
< y ( / ® / , s ) / r ( s / 2 ) , henc e

(5.99) 7 (Ad 2 / , s) = 7 (Sym 2 / , s) = - - 3 s / 2

r ( ~ ) r ( i ± | ^ ) r ( ^ )

if / i s holomorphi c o f weight k an d

(5.100) 7 (Ad 2 / , s) = 7 (Sym 2 f,s) = ^3 s / 2

r ( | ) r ( | + t r ) r ( | + ir)

if / i s real-analyti c Maas s for m wit h eigenvalu e A = \ + r 2 (not e th e parit y doe s

not matter) . Th e loca l Eule r factor s fo r L(Sym 2 / , s) an d L(Ad 2 / , s) ar e give n b y
2
(1 - a{p) p-°)-\l - X {P)P-S)-\1 - (3( P)
2
P-Sr\
(i-ffip-rHi-p-r^i-Hftp-r1
respectively i f p \ q(f), wher e a(p) an d /?(p ) are the local roots a t p for L(/ , s). Not e
that fro m th e twis t formul a (5.89 ) o r the relatio n a(p)/3(p) = \{p) for P unramified ,
138 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

one ca n also relat e th e adjoin t an d symmetric square s a s twist s (convolutio n o f

GL(3) b y GL(1 )):
Lp(Ad2f,s) = L 2
p(Sym f®x,s).
If p divide s exactl y g ( / ) , then thi s formul a i s stil l valid . I n particular , fo r q(f)
squarefree a simple computatio n show s tha t
L(Sym 2 /, s) = L(2s, X 2) £ \f(n 2
)n-,

L(Ad 2 /, a) = L(2s, x2) £ X(n)\ f(n

2
)n-S.

At th e point 5 = 1 the adjoint squar e L-functio n take s a special valu e

(6.101) L ( A d V . l ) = resL(/8/,S)= V o l ^ H )

where w — (ATr)k/Y{k) o r w — cosh(7rr)/27 r dependin g o n the weight o r eigenvalue

of / respectively , whil e (/ , /) i s the square o f the Petersso n nor m o f / (se e (14.11)
for holomorphi c cus p form , otherwis e i t i s the standar d L 2 -norm i n th e Hilber t
space of automorphic function s o n ro(g)\H; se e Chapter 1 5 and also (26.47)) . Thi s
formula shows , i n particular, tha t L(A d / , 1 ) > 0 , and it i s well exploite d i n the
spectral analysi s o f cusp form s o n GL(2) (se e for example [Roy]) .
Theorem 5.1 0 gives a zero-fre e regio n fo r automorphic L-function s o f any de -
gree.
T H E O R E M 5.42 . Let L(f,s) be an automorphic L-function of a cusp form of
degree d. There exists an absolute constant c > 0 such that L(/ , s) has no zero in
the region

except possibly a one simple real zero fif < 1. For this exceptional zero to exist, it
is necessary that f be self-dual.
PROOF. B y the properties o f Rankin-Selber g convolution s o f cus p form s re -
called above , al l assumptions o f Theorem 5.1 0 are valid, whic h give s the result. •

R E M A R K . On e ca n show tha t fo r an automorphic L-functio n L(/ , 5), the aux-

iliary L-functio n L(g,s) use d i n the proof o f Theorem 5.1 0 has the property tha t
Ag(n) ^ 0 for all n, no t only unramifie d ones . Th e proof o f this depend s o n the
local Langlands conjectur e fo r GL(d) over p-adic fields, proved by Harris and Taylo r
[HT]. Roughl y speaking , fo r an y prime p , ther e exis t continuou s representation s
pp : W(Q P) - » GL{d,C) ) wher e W i s the Weil-Deligne grou p o f Q p , a varian t of
the Galoi s group , suc h tha t th e p-factor o f L(/, s) is given by
1 s
Lp(f,s) = det(l-p p(Frp)p- y

and tha t o f the Rankin-Selber g convolution s use d i n defining g ar e

s
LP(f ® /, s) = det( l - (p p<g> p p)(Frp)p- )-\
s 1
LpU ® /, s) = det( l - {p p ® / 9 p )(Pr p )p- )- ,

LP(f®f, s) = det(l - (p p ® p p )(Pr p )p- s )- 1 ,

 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 139

where p p is the contragredient representatio n o f pp, and Fvp is the geometric Frobe -
nius a t p (se e Tate [Tat ] for more details) ; mor e precisely , Fr p is meant t o act on
the invariant s unde r inerti a for all the above representations. Thi s show s that, wit h
the sam e convention , th e p-factor o f L(g, s) is
det(l-(p0p)(Frp)p-s)"1
where p is the representation
P = ie(Pp®|-| i*)e(/5p®|-rit).
But w e hav e quit e generall y
2
Tr(p®p)(g) = \Trp(g)\
for an y g and p, and this give s A g(pk) ^ 0 for all k ^ 1 .
This kin d o f very dee p arithmetica l argumen t depend s cruciall y o n the auto-
morphic natur e o f L(/, s) and is completely beyon d reac h whe n usin g onl y the bare
data o f the Dirichle t serie s or Euler produc t L(f, s).

Statements lik e Proposition 5.2 2 for automorphic L-function s ar e usually calle d

"multiplicity on e theorems". Th e most basi c result , calle d th e strong multiplicit y
one principle , i s due to Jacquet an d Shalika [JS] :
PROPOSITION 5.43 . Let L(f, s) and L(g, s) be two automorphic L-functions of
cusp forms ofGL{m). If the local components of f and g coincide at all but finitely
many primes, then f — g.
P R O O F . Suppos e the local components coincide for p <£ S, where S is finite (an d
contains th e ramified prime s fo r / o r g). Sinc e th e local factor s fo r the Rankin -
Selberg convolutio n L(f <S> g,s) hav e n o poles a t s = 1 , the order o f the pole of
L(f <3>g,s) i s the same a s that fo r the product ove r prime s p ^ S. Bu t
l[Lp(f®g,s) = Y[L p(f®f,s)
Pis p$s
by assumption . Therefor e L(f ® g, s) has a pole a t s = 1 , which implie s / = g. •

For classica l modula r forms , thi s i s Theore m 1 4.1 2 . I f th e local component s

satisfy |c^(p) | < p 1 / 4 , then th e remar k afte r Propositio n 5.2 2 gives a n explici t
version. Suc h a boun d i s know n i n ful l generalit y onl y fo r n — 2, but usin g the
logarithmic derivative of L(f®g, s) instead of the squarefree par t o f the convolution ,
Moreno [Mor3 ] gav e a general explici t boun d o n the number o f primes required .
Theorem 5.4 2 can be applied i n particular t o the symmetric an d adjoint squar e
//-functions. On e can also appl y i t t o the Rankin-Selberg L-function s L(f ® g, s)
for an y two classical modula r form s sinc e Ramakrishnan [Ra ] has shown tha t thos e
L-functions o f degre e 4 are automorphic . Exercis e 4 show s tha t on e can in fac t
prove non-vanishin g result s fo r Rankin-Selberg L-function s withou t modularity .
T H E O R E M 5.44 . (1 ) Let f and g be classical primitive modular forms, either
holomorphic or real-analytic. There exists an absolute constant c > 0 such that
L(f 0 g, s) has no zero in the region

logq(/®0)(|*|+3)
140 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

except possibly a one simple real zero /? < 1 . For this exceptional zero to exist, it is
necessary that f (g ) g be self-dual.
(2) Let f be a classical primitive modular forms, either holomorphic or real-
analytic. There exists an absolute constant c > 0 such that L(Sym 2 / , s) and
L(Ad2 / , s) have no zero in the region
r
a> 1
logq(/)(|t|+3)'
except possibly a simple real zero (3 < 1. If f is non-dihedral, then the exceptional
zero does not exist.
PROOF. A S mentioned , th e first point follow s directl y fro m th e fact tha t L(f ®
g,s) i s an automorphic L- function.
For th e second, L(Sy m / , s) is automorphic [GJ] , so we need onl y prov e tha t
there i s no exceptional zer o i f / i s non-dihedral, whic h i s due to Goldfeld , Hoff -
stein an d Liehma n [GHL ] (since / i s self-dua l i f ther e i s a n exceptiona l zero ,
L(Ad 2 /, s) = L(Sym 2 /, s)). T o this en d consider th e L-functio n

L(g, s) = as)L(Sym2 / , s)2 L(Sym 2 / ® Sym2 /, s)

= C(*)£(Sym 2 /, s)3 L(Sym 2 Sym 2 /, s).

The las t L-functio n i s a specia l cas e o f the symmetric squar e o f a cus p for m o n
GL(3) an d has been show n b y Bump an d Ginzburg t o have a simple pol e a t s — 1
if / i s non-dihedral (thi s als o follow s b y the first formul a sinc e Sy m / i s a cus p
form o n GL{3)). Henc e L(g,s) ha s a pol e o f order 2 at s = 1 , whereas an y real
zero o f L(Sym 2 /, s) is a zer o o f L(g, s) of order ^ 3 . B y easy loca l computation s
one check s tha t A g(n) > 0 for (n,q(g)) = 1 , hence th e result follow s fro m Lemm a
5.9. * •

REMARK. Th e formul a (5.1 01 ) als o prove s tha t L(Ad 2 /, 1 ) ^ 0 .

The followin g i s a usefu l corollar y (compar e wit h th e lower bound s fo r clas s

numbers).
COROLLARY 5.45 . Let f be a classical primitive modular form, holomorphic
or real-analytic. We have
2
(5.103) | | / | | »Vol(r0(9)\M)g-£

for any e > 0 . The implied constant depends only on e and the weight k or the
eigenvalue A whatever is relevant.
If f is not dihedral, for instance, if f has trivial nebentypus and squarefree
conductor q, then

2 _ Vol(r 0 ((?)\H)
(5.104) ll/l f»
lo g 2g
where the implied constant is effective and depends only on the weight k or the
eigenvalue A whatever is relevant.
 5. CLASSICA L ANALYTI C T H E O R Y O F L
1 -FUNCTION S 4 1

P R O O F . Th e firs t par t i s prove d b y Hoffstei n an d Lockhar t [HL] , an d fol -

lows fro m (5.1 01 ) th e bound (3 ^ 1 — c(e)q~ £ fo r the possible exceptiona l zer o of
L(Ad 2 /, s), whic h implie s a lowe r boun d fo r L(Ad 2 /, 1). Th e proof i s similar t o
that o f Siegel's boun d (5.73) , an d the implied constan t i s non-effective .
If / i s not dihedral, the n ther e i s no exceptional zer o an d one derives (5.1 04 )
by th e same metho d fro m th e zero-free regio n o f Theorem 5.44 . Finally , on e ca n
easily se e that a dihedral for m wit h trivia l nebentypu s ha s a non-squarefree level. •

Generally speakin g the problem o f exceptional zero s for the L-functions o f cusp
forms o n GL(m) wit h m ^ 2 turns ou t to be accessible, th e hardest an d stubborn
case bein g wit h m = 1 for real Dirichle t characters . Th e above resul t o f Goldfeld ,
Hoffstein an d Liehman ha s been extende d t o the following theore m o f Banks [Ba] :
P R O P O S I T I O N 5.46 . If L(f,s) is the L-function of any cusp form on GL(3),
then the exceptional zero does not exist for L(f, s).
5.13. A r t i n L-functions .
In thi s Sectio n w e survey, mostl y withou t proofs , som e o f the definitions an d
properties o f Arti n L-functions . I n contras t wit h th e previous sections , th e fac t
that thos e are indeed L-function s a s defined i n Section 5. 1 remains open . However ,
there is great confidenc e an d evidence that th e relevant conjecture s ar e correct, and
together wit h th e analytic theor y develope d i n thi s chapter , an d possibly GRH,
this provide s heuristi c evidenc e fo r results tha t sometime s ar e accessible b y othe r
methods. Se e e.g. [CF ] for more abou t Arti n L-functions . T o avoid confusio n i n
terminology, w e will sometime s refe r t o the L-functions define d i n Sectio n 5. 1 as
analytic L-functions .
Artin L-function s ar e associated t o a continuous finit e dimensiona l Galoi s rep -
resentation p : Gal(Q/K )— > GL(d, C ) wher e K/Q i s a numbe r field . Defin e the
Artin L-functio n o f p by the following Eule r produc t ove r prim e ideal s o f K:

L(p, s) = IJdet(l- p(Fr p)Np-)-\

p

where Fr p is the Probenius conjugac y clas s at p, acting o n the invariants unde r the
inertia grou p a t p . Thu s L(p , s) i s an Euler produc t o f degree deg(p ) ove r if , o r
deg(p)[K :Q ] over Q.
Artin L-function s enjo y a formalis m paralle l t o tha t o f representations . Fo r
instance, th e dua l L-functio n i s als o th e L-functio n o f the contragredien t repre -
sentation p , an d th e Rankin-Selber g L-functio n o f p\ an d p2 is th e L-functio n
L(pi ® P2, s) of the tenso r powe r pi ® pi (thu s th e notation i s compatible). More -
over, L(p i 0 p2 , s) = L(pi, s)L(p2, s) , an d if p is induced fro m a representation p'
from a finite inde x subgrou p correspondin g t o a finite extensio n K' JK, the n
(5.105) L(p,s ) = L(p / , 5 ).
In particular, takin g K — Q and K' — K an y number field , i t follows tha t an y Arti n
L-function o f degree d can be seen a s one defined ove r Q of degree deg(p)[K : Q] .
For th e trivia l representatio n o f Gal(Q/K) , w e get L(l,s ) = Or(s) . Mor e
generally i f deg(p ) = 1 , class-fiel d theor y show s tha t t o p i s associate d a uniqu e
Hecke character £ such that L(p , s) = L(£ , s). S o it is an analytic L-functio n i n this
case (se e Section 5.1 0) .
142 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

It i s conjectured tha t fo r an y p , L(p,s) i s an analyti c L- function. Sinc e L(p i ®

p2,s) — L{p\, s)L(p2, s) , one ca n restric t attentio n t o irreducibl e representation s p ,
and on e expect s tha t L{p,s) i s i n fac t automorphic .
It i s know n tha t L(p , s) ha s meromorphi c continuatio n t o C an d al l o f th e
conditions se t i n Sectio n 5. 1 ar e satisfied , excep t tha t som e o f thes e L- functions
may hav e pole s i n th e critica l stri p (becaus e w e stil l d o no t kno w i f L(p,s) i s
holomorphic i n th e stri p 0 < R e (s) < 1 ) . Th e gamm a facto r ca n b e describe d a s a
product ove r infinit e place s v o f K o f loca l gamm a factor s j v(p, s). Le t a v b e th e
Frobenius conjugac y class , whic h i s o f orde r 2 i f v i s a rea l plac e tha t extend s t o
two comple x place s o f L , an d i s = 1 otherwise. The n

7r_ds/2r(-j r f J i f v i s a comple x place ,

lv(p,s) = {
*""da/2r(!) r
(^T^)^ ^ i s a rea l place ,

where d = deg(p) i s th e dimensio n o f p , an d d~ ,d+ ar e th e multiplicities , respec -

tively, o f th e eigenvalu e + 1 , —1 fo r p(cr v).
The conducto r q{p) i s of th e for m
d
q(p) = \d K\ ^NK/Qf(p)
for som e integra l idea l f(p ) o f K, whic h i s called th e Arti n conductor , define d usin g
ramification group s (se e Serr e [Se4]) . Th e roo t numbe r e(p) ca n b e i n a sens e
defined fro m th e functiona l equation , bu t ther e i s als o a n a prior i definitio n du e
to Dwork , Langland s an d Delign e a s a produc t o f loca l factors . Thu s th e analyti c
conductor satisfie s
d e :Q
q(p,s) < q(p)(\s\ + 4 p « ^ = ( | ^ | ( | s | +4)^ l) d e ^)iVf( / 3 ).
Prom th e Braue r inductio n theore m (se e e.g . [Se5]) , w e ca n writ e

Trp = ^T^TrTT ;
where rii G Z an d 71 ^ is a representatio n induce d b y a n abelia n characte r & o f a
finite (i n fact , cyclic ) extensio n Ki/Q. B y th e invarianc e unde r inductio n (5.1 05 )
we hav e
(5.106) L(p, s) = l[ Lfa, s) n> = H L&, s)" *
ii

and th e cas e o f Heck e character s (ove r numbe r fields) implie s tha t L(p , s) i s mero -
morphic wit h pole s onl y i n th e critica l strip , an d satisfie s a functiona l equatio n
relating s an d 1 — s. Tha t thi s functiona l equatio n i s th e righ t on e involvin g th e
gamma facto r abov e an d th e Arti n conducto r follo w fro m th e formalis m o f both ,
especially th e behavio r unde r direc t sum s an d induction .
To sho w tha t L(p , s) doe s no t hav e extr a poles , w e nee d t o prov e th e
ARTIN CONJECTURE . Let p be an irreducible, non-trivial, Galois representa-
tion of a number field K/Q. Then L(p , s) is entire.
Even th e cas e d — 2 and K = Q i s still no t completel y settled . Mor e precisely ,
if a n irreducibl e representatio n p : Gal(Q/Q ) —> GL(2 , C) i s given , the n it s imag e
in PGL(2 , C) ca n b e classified a s either a dihedral group , th e alternatin g grou p A±,
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 143

the symmetri c grou p S4 o r th e alternatin g grou p A$. I n al l cases , excep t A5 , it i s

known that L(p , s) i s holomorphic (eve n automorphic), an d ther e ar e partial result s
in the A§ case. Th e dihedral case is classical, the A4 case is due to Langlands, th e S4
case t o Tunnell . Th e latte r i s crucia l i n Wiles' s proo f o f Fermat' s Grea t Theorem .
In a certain sense , th e Arti n conjectur e i s true o n average : fo r instance , i f K/Q
is a fixed finite Galoi s extension , the n

(5.107) ( K(s) = t(s)l[L(p,s)

where p runs ove r th e non-trivia l irreducibl e representation s o f Gsl(K/Q) (se e

(5.82)). Sinc e bot h ((s) an d CK(S) a re know n t o hav e onl y a simple pol e a t s = 1 ,
the possible poles of L(p, s) must b e compensated b y zeros for other representations .
The followin g i s als o useful .
COROLLARY 5.47 . Let p be a non-trivial irreducible Galois representation of
K/Q. Then L(p , s) has neither poles nor zeros on the line R e (s) = 1 .
PROOF. Thi s follow s b y (5.1 06 ) becaus e th e L- functions o f non-trivia l Heck e
Grossencharakters ar e entir e an d d o no t vanis h o n th e lin e R e (s) = 1 (Theorem
5.35). •

Assuming th e Arti n conjecture , on e ca n deduc e quit e interestin g arithmeti c

corollaries usin g th e propertie s o f analyti c L- functions.
CHEBOTAREV DENSIT Y T H E O R E M . Let L/K be a finite Galois extension of
degree d of number fields. Let C C Gal(L/K) be a union of conjugacy classes and

l
rl>{x,C)= J2 °ZNS>-
Np^x
FrpGC

We have as x—> +00 ,

\C\x
^ ^ ' ^ I G a K W r

/ / GRH holds for Artin L-functions, then for x ^ 2,

desip)[K:Q]
(5.109) i>{x,C) = | G J ^ ) | + o(V£(log*) E M Mx <l(p)))

where p runs over the irreducible representations of Gel(L/K), and

The implied constant is absolute.

PROOF. Althoug h th e proo f belo w depend s o n th e Arti n conjecture , (5.1 08 )
is know n unconditionall y an d (5.1 09 ) doe s no t nee d th e Arti n conjectur e (se e e.g .
[Se6]).
144 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

Let G — Gal(L/K). Function s o f th e for m Trp , fo r p irreducible , for m a n

orthonormal basi s o f the vecto r spac e o f conjugacy-invariant function s o n the finit e
group G wit h th e scala r produc t

</,<?> = ^ £/(*)<?(*) .

Hence th e characteristi c functio n 5c o f C C G i s give n b y

where c p i s a s i n (5.1 1 0) . Fo r th e trivia l representation , w e hav e c\ = |C|/|G| , an d

thus (5.1 08 ) follow s fro m th e Arti n Conjectur e an d th e Prim e Numbe r Theore m
5.13 fo r L(p,s). Similarl y (5.1 09 ) come s fro m (5.56 ) an d (5.1 1 0) . •

The estimat e (5.1 09 ) ca n b e mad e explici t i n variou s ways . Sometime s on e

can comput e c p explicitly . Otherwise , a simpl e genera l boun d o n GR H i s given fo r
instance b y

::
^ , C ) = |J^^+o(v^(logx)(v1 ci(log a;[* < M ) + l 0g|dL|)).

To se e this , notic e tha t b y orthogonalit y w e hav e

and i t i s well know n tha t

£ d e g ( p ) 2 = |Gal(L/tf)| .
P

Hence b y Cauchy' s inequalit y w e deriv e

G L K
] P \c p\ logfr**^***!) < [K : QKlogs) ( ^^/K)\) ' \ ^ I ^

= [K:q]{\ogx)y/\C\.

On th e othe r hand , sinc e |Tr(p(x)) | < deg(p) , w e have \c p\ < deg(p) , henc e

^ \c p\ log q(p) ^ lo g JJ q(p) d^ = lo g | dL|

PP

by compariso n o f th e conductor s i n th e identit y o f L-function s

For othe r estimate s an d arrangement s o f th e erro r ter m o n GRH , an d als o fo r

unconditional estimate s wit h explici t erro r terms ; see [Se6 ] an d [LO] .
Here i s a sampl e applicatio n o f th e Chebotare v Densit y Theorem .
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 145

PROPOSITION 5.48 . Let n ^ 2 and N > 1 . Let D N be the set of monic

polynomials f G 1\X] of degree n whose coefficients have absolute value < N.
Then we have
^ EE log p = e„ + o(l)
f ha s a roo t m o d p

as x— > o o and N—> • oo , w/iere e n = 1 — | H h (—l)n _ 1 ^ fno£ e £/m £ en— > 1 — ^

as n ^ ooj .
PROOF. W e have |L>w | = iV n . Gallaghe r ([Ga4] , se e Exercise 3 of Chapte r
7) ha s shown tha t th e set CN of elements / € D^ suc h tha t eithe r / i s reducibl e
or th e Galois grou p o f the equation f{x) — 0 is not the full symmetri c grou p S n
satisfies
\cN\^Nn~hogN.
By trivia l estimate , element s i n CN do not contribute t o the limit above .
Let / £ CAT . Choosin g a roo t a o f / i n C, le t L — Q(a) an d K th e Galoi s
closure o f L. W e have Gal(tf/Q ) - S n an d # = Gal(tf/L ) ~ S n _i. B y Galoi s
theory, w e see that
1
{p | / ha s a root modul o p] = {p \ Fv p E I ) a' Ha] = { p | Pr p £ J?}

where H i s the set of element s i n S n fixing a t leas t on e element. Precisely , thi s

holds u p t o finitely man y prime s a t whic h th e localization s o f Z[X]/(f) an d the
ring o f integers o f L differ . B y the Chebotarev Densit y Theore m w e have
I fr\
V " logp= — vx + o{x)
p^x
f ha s a roo t m o d p

By exclusion-inclusion , i t follow s tha t \H\/n\ = e n , hence th e result. D

See e.g. [Se6 ] for other application s o f the Chebotarev Densit y Theorem .
5.14. L-function s o f varieties .
The situatio n wit h L-function s o f algebraic varietie s ove r Q is even les s under -
stood tha n tha t fo r Galois representations. Le t us call them "arithmeti c geometry "
L-functions, o r just "geometric " L-functions . W e will not discuss thing s i n general,
but mentio n the important case s of algebraic curve s and abelian varieties . Th e only
satisfactory theor y exist s for the intersection o f these tw o cases, namely th e abelian
varieties o f dimension on e (that i s the elliptic curves) .
Let E/Q b e a n ellipti c curve . Modularit y o f E, prove d b y Wiles , Taylor -
Wiles an d Breuil-Conrad-Diamond-Taylo r [W] , [TW], [BCDT], mean s tha t th e
Hasse-Weil zet a functio n o f E, normalize d t o hav e critica l stri p 0 ^ R e (s) ^ 1 ,
corresponds t o a modula r for m o f weigh t 2 an d leve l equa l t o th e conducto r o f
.E, henc e i s an L-functio n accordin g t o Sectio n 5.1 1 . Se e Section 1 4. 4 for a mor e
detailed statement , a s well a s a description o f the root numbe r i n most cases .
Conjecturally, th e "modularity " extend s t o th e Hasse-Wei l zet a function s o f
arbitrary smoot h projectiv e curve s ove r numbe r fields. Le t C/Q b e such a curve .
Then fo r all primes p \ TV , wher e N ^ 1 is some integer , th e curve ca n be reduce d
146 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

modulo p to a smooth projective curv e Cp/¥p. A s described i n Exercise 2 of Chapte r

11, th e loca l zet a functio n
\Cp(Fp>)\Tk>
Z(CP,T) = exp(^_ , ^ J

is a rationa l functio n o f the for m

Z(C
1 P,T)= { 1_TP){ _pT)

for som e moni c polynomia l P p wit h P p(0) = 1 , o f degre e 2g wher e g > 0 i s th e

genus o f C. Writ e
pp(T)= n (i-«i(p)p5T) .

Then th e Rieman n Hypothesi s fo r curve s ove r finite field s (se e Chapte r 1 1 ) show s
that |<x?(p) | = 1 . Th e Hasse-Wei l zet a functio n o f C outsid e iV , analyticall y nor -
malized, i s defined b y th e Eule r produc t
1
LN(C,S) = JJ ( 1 - a 1 (p)p-sy1 •••(! - ^ ( r i f T

It i s conjecture d tha t afte r multiplyin g b y appropriatel y define d Eule r factor s a t

p | iV, th e complet e L-functio n i s entire , i t i s necessaril y self-dua l an d th e gamm a
factor shoul d b e
S
7(c, s)=n-°n 2+Jm|+lr=^(2^-^(3+\y
for som e constan t c g > 0 (se e (5.86)) . Not e tha t th e Rieman n Hypothesi s fo r th e
curves C p ove r finite fields mean s tha t thi s L-functio n satisfie s th e Ramanujan -
Petersson conjectur e (u p t o th e ramifie d places , bu t i t i s know n tha t thos e stil l
satisfy th e require d condition) . I n genera l fo r an y g > 1 , thi s i s no t ye t know n
except fo r ver y specia l cases .
REMARK. I n arithmeti c geometr y i t i s mor e practica l t o leav e th e loca l root s
otj{p) a s the y occu r naturall y withou t normalization , i.e . t o writ e

PP(T)= J ] ^-PoiP)T)

with \/3j(p)\ = y/p. Thu s

L(c,,)=n^(p-s+i/2)-i=
In thi s notatio n th e critica l lin e i s R e (s) = 1 and th e functiona l equatio n relate s
the value s a t s an d 2 — 5 . Th e sam e remar k refer s t o abelia n varieties .

Abelian varietie s ar e higher-dimensiona l generalization s o f ellipti c curves , see n

as algebrai c groups . A n abelia n variet y A/Q i s a smoot h projectiv e variet y wit h a
given rationa l poin t 0 G A(Q) an d additio n an d invers e morphism s p : A x A— > A,
i : A —> A whic h ar e map s o f algebrai c varietie s ove r Q an d satisf y th e axiom s o f
a commutativ e grou p (wit h i givin g th e opposite) . I f th e dimensio n o f A i s one ,
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 147

this correspond s exactl y to elliptic curve s E/Q wit h the "usual " chord-and-tangen t
description o f the group la w (see [Sil] and Sectio n 1 1 .8) .
There exist s a n integer N ^ 1 such tha t fo r p \ iV, the reduced variet y A p/¥p
is smooth . Wei l showe d tha t th e local zet a functio n

Z(AP,T) = e x p (^ j. J

is a rationa l functio n o f the for m

Z ( A p i J
' "P pfi(T) • • • P P,23(T)

where P pj ar e polynomials. Wei l prove d tha t P p^ i s of degree 2g and of the for m

PP,I(T)= n (i-«i(p)p5r )

with |oj i (p) I = 1 , whic h i s th e loca l Rieman n Hypothesis . I n additio n on e ca n

arrange th e cti(p) suc h tha t ai(p)ai+ g(p) — 1 for 1 ^ i ^ g. Th e first polynomia l
P p? i determine s P pj a s the j - th exterior powe r

PPAT)= n (i-«ii(p)---«ii(p)p2T) .

In particular , P P:0 = 1 - T , P p^g = 1 — P gT. Th e globa l zet a functio n o f A/Q

outside N i s

LN(A, s) = l[ ( 1 - a i ^ p - ) - 1 • • • (1 - ^ ^ ( p j p - ) - 1
.
pfJV

Again it is conjectured tha t afte r multiplyin g b y appropriately define d Eule r factor s

at p | TV the complete product L(A, s) is an L-function, whic h is entire and self-dua l
and ha s gamma facto r

j(A, s) = 7 r - » T (! + \) 9T(S- + l) 9 = c 9 (27r)-*T( S + \Y

for som e constan t c g > 0 (see (5.86)). I t satisfie s th e Ramanujan-Petersson conjec -

ture, sinc e the correctly define d ramifie d factor s als o satisfy th e required condition .
For g = 1 , this conjectur e i s true b y the modularit y o f elliptic curve s ove r Q .
For an y g > 1 , except fo r som e case s involvin g so-calle d CM varieties, th e analyti c
continuation an d functiona l equatio n remai n ope n problems . Th e conductor q(A)
(which divide s N) o f this L-functio n i s predicted a s part o f the conjecture an d is a
close relativ e o f the Artin conducto r fo r Arti n L-functions .
Note th e resemblanc e betwee n th e shap e an d definitio n o f th e L-function s o f
curves an d abelia n varieties . I t i s not a coincidence . Th e theor y o f the jacobia n
variety o f curves associate s t o ever y smoot h projectiv e algebrai c curv e C o f genu s
g ove r a field k a n abelia n variet y J(C)/k o f dimensio n g, calle d it s Jacobian , i n
such a wa y tha t fo r a curv e C/Q w e hav e L(C , s) = L(J(C),s). Fo r instance ,
the L-functio n L(Jo(g),s ) o f (5.93 ) i s the L-functio n o f the Jacobia n Jo(q) o f the
modular curv e Xo(q) = ro(#)\H, b y work o f Eichler an d Shimura .
148 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

REMARK. A n alternativ e definitio n o f the loca l factors o f L(A, s) (o r of L(C, s)

for a curve ) i s give n i n term s o f th e £-adi c Tat e modul e o f A: le t
Te(A) = QimA[r])®Qi,
n

then th e Galoi s grou p o f Q act s o n Tt(A) an d fo r an y prim e p ^ £ where A ha s

good reduction , w e hav e
P p ,i(T) = d e t ( l - T P r p | T £(A))-\

denoting b y Fr p a Frobeniu s conjugac y clas s a t p. Dually , P P:i i s give n i n term s o f

the etal e cohomolog y grou p i f 1 ^ , (Q^ ) (se e Sectio n 1 1 .1 1 ) b y
1
P P t l (T) = d e t ( l - T < 7 p | T^A))-
where o~ v i s th e invers e o f Fr p , becaus e Ti{A) i s th e dua l o f H l(A, Qi).
The Mordell-Wei l theore m state s tha t i f A/Q i s a n abelia n variety , the n th e
group o f rationa l point s A(Q) i s finitely generated . Th e computatio n o f it s ran k i s
one o f th e grea t problem s o f arithmeti c geometry . W e hav e th e famou s
BIRCH AN D SWINNERTON-DYE R CONJECTURE . Let A/Q be an abelian variety.
Then
rank A(Q) = or d L(A, s).

Of cours e thi s conjectur e assume s tha t L(A,s) i s a n L-function , s o L(A, | ) i s

defined. I f the Birch and Swinnerton-Dye r conjectur e holds , one can get informatio n
on th e ran k o f abelia n varietie s usin g analyti c methods . Ofte n thi s suggest s ver y
interesting problems . Fo r instance , b y Propositio n 5.2 1 w e derive :
COROLLARY 5.49 . Let A/Q be an abelian variety of dimension g ^ 1 with
conductor q. Assume that its Hasse-Weil zeta function is an L-function and that
the Birch and Swinnerton-Dyer conjecture holds for A(Q). Then we have
logq
rank A (Q) <C
log(|logg)'
with an absolute implied constant. Note that q > 3 9 by Theorem 5.51 below.
This ca n b e turne d ove r t o giv e a lowe r boun d fo r th e conducto r o f a n abelia n
variety wit h a give n rank , fo r instanc e fo r ellipti c curves .
COROLLARY 5.50 . Assume that the Birch and Swinnerton-Dyer conjecture
holds for elliptic curves over Q. Then if E/Q is an elliptic curve of rank r ^ 1 and
conductor q, we have
q^>rcr
where c > 0 is some absolute constant and the implied constant is also absolute.
It woul d b e ver y interestin g t o prov e an y unconditiona l estimat e comparabl e
to thos e tw o corollaries usin g algebrai c methods . I n Chapte r 2 6 we study th e orde r
of vanishin g o f th e zet a functio n (ofte n calle d th e analyti c rank ) fo r th e abelia n
varieties Jo(q)/Q, obtainin g quit e precis e results .
As fo r th e discriminant s o f numbe r fields i n Sectio n 5.1 0 , on e ca n us e L-
functions t o investigate , conditionally , th e conducto r o f elliptic curves , or of abelia n
varieties ove r © .
 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S 149

T H E O R E M 5.51 . Let A/Q be an abelian variety over Q of dimension g ^ 1

such that L(A, s) is an analytic L-function of conductor q = q(A). Then we have
q ^ e 1 ,2g, and in particular q > 3.

See [Mes] for even stronger bounds . Th e bound q > 1 is known unconditionally ,
by th e celebrate d theore m o f Fontaine [Fon ] accordin g t o whic h ther e i s no abelia n
variety ove r Q everywhere unramified . Fo r a variety o f dimension g — 1 , this mean s
there i s n o ellipti c curv e define d ove r Q whic h i s unramifie d everywhere , i.e . it s
discriminant i s divisibl e b y a t leas t on e prime . I n thi s cas e i t ca n als o b e prove d
without appealin g t o L-functions o r modularity b y showing that th e discriminant o f
the Weierstras s equatio n (1 4.1 8 ) (wit h integra l coefficients ) canno t b e equa l t o ± 1
(see [Sil] , ex . 8.1 5) . I n th e highe r dimensiona l case , n o suc h explici t reformulatio n
is accessible , an d Fontaine' s unconditiona l proo f i s muc h mor e involved .

P R O O F O F THEORE M 5.51 . Appl y (5.24 ) t o L(A,s) a t s = a > 1 , then tak e

the real part. Becaus e L(A, 5) is self-dual, th e constant b disappears by using (5.29) .
Moreover, b y th e positivit y

Re - -, ^ > 0,
P?
we derive th e inequalit y

1V V
- lo g q - g log 2TT + g — (a + \ ) + —

Next b y th e Ramanujan-Petersso n boun d w e hav e

^W)|<<^(a).
Hence w e ge t

^ l o g O log2T T - V -(o + A ) - £ »

for an y a > 1 . Takin g a = 2.3 , one check s tha t th e righ t sid e i s 0.62... > 0.6 , henc e
q ^ e 1 -2^ an d q > 3 . •

REMARK. Sinc e e 2A > 1 1 , and i t i s known (se e (1 4.41 ) ) tha t th e ellipti c curv e
with smalles t conducto r i s

E : y2 + y = x3 - x 2

with conducto r 1 1 , the argumen t implie s tha t 1 1 is i n fac t th e smalles t conducto r

of a n abelia n variet y ove r Q , assumin g thei r zet a function s ar e L-functions .

5.A. Appendix : comple x analysis .

In thi s sectio n w e presen t i n concis e for m th e concept s an d result s o f comple x

analysis tha t ar e use d i n this chapter . W e omit mos t proof s bu t refe r t o [Tl] , [Ru ]
or [Ahl ] (amon g man y references) .
150 5. CLASSICA L ANALYTI C T H E O R Y O F L - F U N C T I O N S

A.l. Functions of finite order.

DEFINITION A . l . Le t / : C -> C be an entire function . The n / i s said t o be
of finit e orde r i f there exist s f3 > 0 such tha t
|/( S )|«exp(| S n
for s e C . I f one ca n take an y /3 > 1 , / i s said t o be of order ^ 1 , and if no /? < 1
is possible, the n / i s said t o be of order 1 .
Let / b e a meromorphic functio n o n C. The n / i s of order < 1 if there exist s
entire function s g and h of order ^ 1 such tha t f = g/h.
THEOREM 5.52 . Let f be an entire function of order 1.
(1) We have

/w=- p n( i -j) e ' / "

uniformly on all compact subsets ofC, where r is the order of the zero of f at s = 0
and p runs over zeros of f different from 0 .
(2) For any e > 0, the series

A. 2. The Phragmen-Lindelof principle for a strip.

The Phragmen-Lindelof "principle " i s a generic term for theorems which exten d
the maximu m modulu s principl e t o various infinit e region s o f the complex plane .
Only th e case o f strips a ^ a < b interests us.
T H E O R E M 5.53 . Let f be a function holomorphic on an open neighborhood of
a strip a ^ a ^ b, for some real numbers a < b, such that \f(s)\ < C exp(|s|^) for
some A ^ 0 and a ^ a ^ b.
(1) Assume that \f(s)\ ^ M for all s on the boundary of the strip, i.e. for
a = a or a = b. Then we have |/(s) | ^ M for all s in the strip.
(2) Assume that
\f(a + it)\^Ma(l + \t\) a,
\f(b + it)\^Mb(l + \t\)P

for t e R. Then

\f(<r + it)\ < M^Ml- t(a\l + W'W+M-'W

for all s in the strip, where £ is the linear function such that £(a) — 1, £{b) = 0.
A.3. The Perron formula.
Let h(x) b e the function give n by
1i fx>1 ,
l
h(x) = < 22 i f x = 1,
0 iix<l.
 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S 151

PROPOSITION 5.54 . We have

i r rc+iT
c+l
1
ds
s /
(5.111) - ^
27r / x — = h(x) + 0[
% J c_iT s V ,T|logx|
for any x > 0 , x ^ 1 , T > 0 and 0 < c ^ 2 , m'£/ i an absolute implied constant. If
x = 1, the factor | logx| z s omitted.
A.4 • The Stirling formula.
The Stirlin g asymptoti c formul a

*•"*> r 1
W-(T) i
(i)'( +0
(R))
is vali d i n th e angl e |args | ^ TT — e wit h th e implie d constan t dependin g o n e.
Hence fo r s = a -f ^, £ ^ 0 , a fixe d

(5.113) I > + it) = V2^(it)a-2 e~2 1*1 ( ^ ^ {l + °(jji) } •

Let

be a gamm a facto r a s in (5.3 ) with Re(/«j ) > —1 . I t ha s no poles i n Re (s) ^ 1 .

Recall th e corresponding conducto r a t infinit y i s
d
qoo(s) = IJ(|t + /^| + 3).
3=1

Let -\ ^ Re(s) ^ 2 . Pro m (5.1 1 2 ) we deduce

(5.114) |7(s)ini * + ^l = M s ) ^ + a ^ ^

ence w e
whereA: = Re Ylnj- H have
S
7(1 ~ s) _ fMh-*)TT + KJ
(5.H5) ^^tM^Tlj^-.
1
7(5) U - 3 S + Kj

Moreover, usin g th e recurrence T(s + 1 ) = sT(s) w e derive

(5.116) ^ ( s ) - £ __L_«i ogqoo (5).

The implie d constant s i n (5.1 1 4 ) an d (5.116) ar e absolute.

A. 5. Existence of test functions.
For application s o f the explicit formula , especiall y i n the absence o f the GRH ,
it i s useful t o know that ther e exist tes t function s wit h certai n positivit y properties .
152 5. CLASSICA L ANALYTI C THEOR Y O F L-FUNCTION S

PROPOSITION 5.55 . There exist positive, non-zero, C°° functions 7 7 on [0,+oo [

with compact support in [e _ 1 , e] such that rj(l) = 1 , r)(0 ) > 0 , n(x) = r](x~ 1 ) for all
x > 0 and moreover either fj(it) ^ 0 for t G R o r R e (r)(s)) ^ 0 /or all s G C wit/ i
ki<i.
SKETCH O F PROOF . Defin e f(y) — rj(e y), whic h i s a n even , compactl y sup -
ported, C°° functio n o n R , an d

V(») = I f(y)e sydy

JR
for al l s G C, henc e
lty
m= / f(y)e dy
JR
Re(fi{s))= / f(y)e aycos(ty)dy= / f(y)cosh(ay)cos(ty)dy.
JR JR
By th e firs t formul a t o hav e r)(it) ^ 0 i t suffice s t o choos e / ^ 0 suc h tha t it s
Fourier transfor m i s positive. Th e Fourie r pai r (4.83) , suitabl y smoothed , doe s th e
job.
From th e secon d formul a an d th e maximu m modulu s principl e fo r harmoni c
functions, th e inequalit y R e (r)(s)) ^ 0 will hold fo r \a\ < 1 if and onl y i f the Fourie r
transform o f th e eve n functio n g{y) = f(y) cosh(y ) i s non-negative . Conversely , i f
a functio n g (even , smoot h an d compactl y supported ) wit h thi s propert y i s given ,
a suitabl e tes t functio n7 7 is easil y obtaine d b y reversin g thi s procedure . No w i f go
is an y smoot h compactl y supporte d positiv e functio n o n R , the n th e convolutio n
square g = go * go wil l work , sinc e g = p§ .
The suppor t o f r\ an d it s valu e a t 1 can b e easil y adjuste d b y homogeneity . •

Functions o f thi s typ e wer e constructe d b y Poito u an d other s fo r th e purpos e

of obtainin g lower-bound s fo r th e discriminan t o f numbe r field s [Poi] . Not e i n a
way the y ar e simpl y bette r behave d version s o f the basi c functio n f(s) = (a — s)~ l
for a > 1 (se e (5.28 ) an d th e proo f o f Theore m 5.51 ) . Se e als o [PP ] fo r anothe r
construction o f a tes t functio n wit h usefu l propertie s i n applications .
A.6. Landau's lemma.
LEMMA 5.56 . Let X n ^ 0 for all n ^ 1 . Suppose the series

D(s) = J2 X nn~s

converges absolutely for all s with R e (s) > o~o, but not for any s with R e (s) < o"o;
in other words, &o is the abscissa of absolute convergence of D(s). Then s — cr 0 is
a singularity of D(s), i.e. D(s) cannot be continued analytically to a neighborhood
ofo-Q.
 http://dx.doi.org/10.1090/coll/053/07

CHAPTER 6

ELEMENTARY SIEV E METHOD S

The siev e method s originate d fro m work s o f Vigg o Bru n o n th e Goldbac h an d

Twin Prim e Conjecture s [Brl] , [Br2] . Moder n siev e method s appl y t o a grea t
variety o f problems, an d whe n combined wit h technique s o f analytic numbe r theor y
they ar e ver y advance d indee d (cf . [FI1 ]) . I n this chapte r w e provid e onl y basi c
results from siev e theory. Firs t w e establish by refining th e combinatorial argument s
of Brun th e so-calle d Fundamenta l Lemma . The n w e develop a n uppe r boun d siev e
of Selber g an d illustrat e it s powe r b y a few popula r applications . Mor e complet e
treatments o f siev e method s ca n b e foun d i n th e book s o f Halbersta m an d Richer t
[HaRi] an d Greave s [Gr] .
6.1. Siev e problems .
Let A = (a n) b e a sequence o f non-negativ e numbers . Th e ultimat e questio n
is ho w ofte n thes e number s ar e supporte d o n primes ? Fo r example , i f A is the
characteristic functio n o f th e shifte d prime s
1 if n = p + 2,
[ 0 otherwise ,
then w e ar e askin g ho w ofte n i s p + 2 prime? I n thi s case , th e heuristi c argument s
described i n Sectio n 1 3. 1 yield th e followin g asymptoti c formul a
2
(6.2) 7T 2(x) = |{ p ^ x ; p + 2 is prime} | ~ Cx(\ogx)~
where C is a positive constan t give n b y

(6.3) C = 2J j ( i _ _ I _ ) =1 -32032.. .

Hence ther e "are " infinitel y man y twi n primes , bu t w e d o no t hav e rigorou s proo f
of thi s ol d conjectur e b y an y method .
The siev e method s alon e (i n thei r classica l format ) ar e incapabl e o f selecting
primes, howeve r the y yiel d satisfactor y result s fo r slightl y modifie d problems . Le t
P b e a se t o f primes , z ^ 2 an d
(6.4) P(z)=Hp.
p£V

Now w e see k estimate s fo r th e sifte d su m

(6.5) S(x,z) -£
(n,P(z))=l

153
154 6. E L E M E N T A R Y SIEV E M E T H O D S

When V i s the se t o f al l prime s an d z — y/x, the n S(x, y/x) agree s wit h

(6.6) S(x) = J2 ap
p^x

up t o a fe w term s a n wit h n < y/x. Th e siev e method s produc e uppe r an d lowe r

bounds fo r S(x,z) o f correc t orde r o f magnitud e a s fa r a s z ^ x a, wher e a > 0 is
a smal l positiv e constant . Not e tha t i f n ^ x ha s n o prim e divisor s les s tha n x a,
then n ha s a t mos t [1 /a ] prim e divisors , s o one ca n sa y n i s almos t prime .
The remarkabl e universality o f sieve is that i t applie s to quite general sequence s
A = (a n). Al l we need ar e asymptotic s fo r th e partia l sum s
a
(6.7) A d(x)= Y, "
n-^x
n=0 (mo d d)

with d | P(z), d < y. W e assum e tha t

(6.8) A d(x)=g(d)X + r d(x)
where g(d)X i s a n expecte d mai n ter m an d rd{x) i s a n erro r ter m whic h w e thin k
of a s being relativel y smal l i f d < y. I n th e mai n ter m X i s approximately equa l t o

(6.9) A(x) = J2a n

n^.x

so g(d) stand s for th e density of the masses a n attache d t o n = 0 (mo d d). Thinkin g
of th e divisibilit y b y distinc t prime s a s independen t events , i t i s not to o surprisin g
that w e assum e tha t g(d) i s a multiplicativ e functio n suc h tha t
(6.10) 0^g(p) < 1 if peV.
Of course , th e approximatio n o f typ e (6.8 ) i s no t unique , howeve r t o ge t goo d
results i n practic e ther e i s no t muc h roo m t o choose g{d) an d X from . Fo r som e
prime modul i q on e ma y hav e a goo d approximatio n (6.8 ) wit h g(q) = 1 , whic h
means tha t almos t al l mas s come s fro m th e a n wit h n = 0 (mo d q). I n thi s case
the chanc e o f finding term s a n supporte d o n primes i s slim; tha t i s why w e exclud e
such prim e modul i fro m th e se t V. Sinc e w e us e (6.8 ) onl y fo r d | P(z), w e ar e
free t o modif y th e multiplicativ e functio n g(d) arbitraril y a t d prim e t o P(z) an d
indeed fo r notationa l simplicit y w e se t
(6.11) g(p)=0ifp£V.
6.2. Exclusion-inclusio n scheme .
Detecting the condition (n , P(z)) = 1 by means of the Mobius inversion formul a
1i fm = 1 ,
(6.12) S(m) = J2^ = { 0 ifra^l ,
d\m ^

the su m (6.5 ) unfold s int o

d
S(x,z) = J2[YJ M ))an-
n<;X d\n
d\P(z)
 6. ELEMENTAR Y SIEV E METHOD S 155

Changing th e orde r o f summatio n w e obtain th e Legendr e formul a

d
(6.13) S(x,z)= ^2 »( )Mx)-
d\P(z)

Next introducin g (6.8 ) w e ge t

(6.14) S(x, z) = XV{z) + i?(x , z)
where
(6.15) V(z)= J ] (l-g(p))
p\P(z)

and
r
(6.16) R(x,z) = J2 ^x)-
d\P(z)

On probabilisti c ground s on e ma y expec t tha t th e produc t V(z) represent s th e

true densit y o f the mas s comin g fro m a n wit h n ^ x havin g n o prime divisor s p < z
in V. However , thi s i s almos t neve r correc t excep t fo r z smal l relativel y t o x i n
the logarithmi c scale . I n othe r words , th e divisibilit y b y distinc t prime s ar e no t
completely independen t event s whe n thes e prime s ar e relativel y large . Technicall y
speaking th e proble m i s that , i f z i s large , th e remainde r su m R(x, z) ha s man y
terms, s o i t exceed s th e expecte d mai n ter m XV(z).
In orde r t o reduc e th e numbe r o f term s i n th e Legendr e formul a (6.1 3) , an d
consequently i n th e remainde r ter m (6.1 6) , Bru n truncate d th e Mobiu s functio n
n{d) t o tw o sets , sa y V + an d V~, an d considere d th e incomplet e convolution s
+
(6.17) S (n) = £ //(d )
d\n
dev+
(6.18) 6~(n)= J2 /*(<* )
d\n
dev~
in plac e o f 8(n). H e los t th e exac t propert y (6.1 2 ) bu t maintaine d th e inequalitie s
(6.19) 5-(n) < 5 ( n ) ^ 5+(n)
for al l n ^ 1 . Henc e on e obtain s th e lowe r an d uppe r bound s
(6.20) XV-{z) + R-(x, z) < 5(x , z) < XV+(z) + R +{x, z)
where
± X
(6.21) V {z) = Y, ^(d),
d\P(z)
± X
(6.22) R (x,z) = Y drd(x)
d\P(z)

and Aj , A^ " denote th e Mobiu s function o n the sets V +, V respectively . Assumin g

that
(6.23 ) V+,V~ c[l,y)
156 6. ELEMENTAR Y SIEV E METHOD S

we can estimat e th e remainde r sum s i? ± (x, z) b y

(6.24) R(x,y)= ] T M*)| .

d<y
d\P(z)

Sometimes w e can exploi t th e intrinsi c structur e o f the siev e weight s A J t o ge t

estimates fo r K^^x^z) whic h ar e bette r tha n tha t fo r R(x,y). Thi s i s du e t o a
possible cancellatio n o f th e erro r term s rd(x) i n R ±(x,z)) bu t w e d o no t ventur e
into thi s sophisticate d are a i n thi s boo k (se e e.g . [1 9] , [1 1 0]) .
For notationa l simplicit y w e ar e writin g squarefre e number s a s th e produc t o f
distinct prime s enumerate d i n decreasin g orde r
(6.25) d = Pi'-'Pr wit h Pi>--->p r -
Put
+
(6.26) D = {d = p i • • -p r ; Pm <Vm fo r al l m odd }
(6.27) V~ — {d = pi • • -p r ; Pm < Vm fo r al l m even }
where y m ar e suitabl e parameters . B y conventio n bot h set s V + an d V~ contai n
d = 1 . W e star t fro m th e recurrenc e formul a
(6.28) V(z ) = 1 - £>(p)V(p) .
p<z

Iterating thi s we derive by the well-know n exclusion-inclusio n procedur e th e follow -

ing identities :

(6.29) V(z) = V +(z)- J2 V

n(z),
n od d

(6.30) V(z) = V-(z)+ Y, V n(z)

n eve n
where
(6.31) V n(z)= J2---Y, 9(pi---Pn)V( Pn).
yn^Pn<---<Pi<z
Pm<ym,'m<n->'nT>=n (mod 2 )
Dropping th e non-negativ e term s V n(z) i n (6.29 ) an d (6.30 ) i t follow s tha t
(6.32) V-(z)^V(z)^V+(z).
Now w e ca n easil y verif y th e lowe r an d th e uppe r boun d siev e condition s (6.1 9) ;
indeed (6.32 ) become s (6.1 9 ) o n takin g P(z) = n an d g(d) = 1 .
Prom no w o n th e combinatoria l sieve s A + = (Aj ) an d A - = (A^~ ) depen d o n
the truncation parameter s y m alone . Bru n trie d variou s parameters, hi s best choic e
was ym = y a(3 fo r suitabl e constant s 0 < a < 1 < /? . W e choos e th e parameter s
ym dependin g o n the prime s p i , . .. pm whic h occurre d prio r t o th e ra-th ste p o f th e
iteration, precisel y

(6.33) ym = (y/pi--Pm)*
where (3 is a fixed number , / 3 > 1 . Clearl y ever y d G V + U T>~ satisfies d < y
except possibl y fo r d = p G V~ . Howeve r assumin g z ^ y i n thi s cas e w e als o hav e
d — p < z ^ y. Ou r choic e (6.33 ) i s inspire d b y heuristi c argument s exploitin g th e
 6. E L E M E N T A R Y S I E V E M E T H O D S 157

concept o f "sievin g limit" , however , a s we are not going for optimal results , w e are
free t o choose an y / 3 > 1.
6.3. Estimation s ofV+(z), V~(z).
We nee d a n upper boun d fo r V^(z) an d a lower boun d fo r V~(z). B y virtue
of (6.29 ) an d (6.30) i n both case s the problems reduc e t o getting uppe r bound s for
Vn(z). Sinc e V(p n) i n (6.31 ) ar e non-negative som e o f the summation condition s
can be relaxed b y positivity. W e get

Vn(z)^ £ / • • ] £ 9(Pl'-Pn)V(p n)
Vn^Pn<---pl<Z
Pi---Pm-iP^n<y

where th e conditions hol d fo r all 1 ^ m < n regardless o f parity. Thes e condition s

imply tha t
1 (P-l\m
p
Pi" 'Pm <y i fm < n,
by inductio n o n m. I n particular, fo r m = n — 1 this yield s th e following lowe r
bound fo r the least prim e divisor s o f d — p\ • • -pn i n V n(z),

Pn > (y/pi-"Pn-i)fi+1 >y? +l{

* }
^y^^ ]
> z n,
/ / 3 - 1 \n
say, wher e z n — z K $ ' wit h
z = ya, 0 / 3 .

Having establishe d th e lower boun d p n ^ z n w e drop th e other condition s and

estimate a s follows

V„(Z)^ J2---1 2 9(Pl---Pn)V(jp n)

Zn^Pn<---<Pl<Z

To proceed furthe r w e make th e followin g assumptio n abou t th e multiplicative

function g(d): fo r any w with w < z, we have

log z \ K
(6 .34) n <!-««>-<*(&)
^-yypj) ^^y '
w^p<z

where K, > 0 and K > 1 do not depend o n w.

REMARKS. Th e conditio n (6.34 ) is relatively eas y to verify i n practice. W e call
the exponen t K the sieve dimension . Notic e tha t hi is not uniquely defined ; an y
larger numbe r i s also qualified . Th e constant K ca n be disturbingly larg e i f the
densities g(p) ar e near on e for many smal l p. Nevertheles s th e result s ca n be refine d
in thi s respect . Indeed , i n practice on e ca n establis h th e following estimat e

<«*> n <>-*)>-.(^r'( 1 +i ^)
158 6. E L E M E N T A R Y SIEV E M E T H O D S

for an y w wit h 2 ^ w < z, where K! and K' ar e positive number s independen t of

w. Thi s estimat e implie s (6.34 ) wit h K = 1 + K' an d

K'
(6.36) K = 1 + logz

Therefore i f ca n be reduced t o a numbe r nea r 1 at th e expense o f increasing the

dimension K by 1.

By (6.34 ) w e derive usin g th e inequality 1 + x ^ e x tha t

V{zn)^KV(z)(-^) <KV(z)e"l
v/3-
where j3 — Kb + 1 . Henc e

K.W < £(= tkgjf)V'VH <£(=««' ) V vW .

Inserting n ! ^ e(n/e) n w e obtain
1
(6.37) F n (z) < e- anKMV(z)

where a = b~1e1+b . We choose b large s o that a < 1 .

The conditio n p n ^ y n i n (6.31 ) implie sp ™ ^ y , an d this show s tha t th e
sum (6.31 ) i s void i s n ^ s — (3. Therefore (6.37 ) yield s

b + 1
(6.38) ] [ > n (2)= £ y n(2) < _^!_K ^)-
v 7
n>0 n>s-/ 3 " ^

We choos e / ? = 9K + 1 , getting 6 = 9 and a < e~l. Henc e we conclude

THEOREM 6.1 . Let A + = (Aj) , A ~ = (A^~ ) be the combinatorial sieves of level
y > 1 and the parameter ft — 9K + 1 . TTie n /or any multiplicative function g(d)
satisfying (6.1 0), (6.1 1 ) and (6.34) an d an V z ^ V 1 ^ we have
+ 1
s 0
(6.39) V (z) < ( 1 + eP- K )V(z),
(6.40) V~(z) > ( 1 - e ^ - K1 0 ) y ( z )
s

where s = lo g y/ log 2.
COROLLARY 6.2 . Le t the conditions be as above. Then

(6.41) S(x, z)<(l + e^sK10)V(z)X + R(x, z s)

1 0
(6.42) S(s , z) > ( 1 - e ^ ' l T ) VO*) * - i?(x , zs )

where f3 — 9K + I, s is any number ^ j3, and R(x,y) denotes the remainder

sum (6.24).
6.4. Fundamenta l lemm a o f sieve theory .
Combining th e inequalities (6.39) , (6.40 ) wit h (6.32 ) w e conclude th e following
 6. ELEMENTAR Y SIEV E METHOD S 159

FUNDAMENTAL LEMM A 6.3 . Let K > 0 and y > 1 . There exist two sets of real
numbers A + = (AJ) and A ~ = (A^~) depending only on K and y with the following
properties:
(6.43) Af = l,
(6.44) | A ± | < 1 if\<d<y,
(6.45) A ± = 0 ifd^y,
and for any integer n > 1,

(6.46) J2x~d^°^J2xt
d\n d\n

Moreover, for any multiplicative function g(d) with 0 ^ g(p) < 1 and satisfying
the dimension condition
logz\K/ K \

/or all 2 ^ w < z ^ y we have

(6.48) J ] A ± 5 ( d ) = ( 1 + O(e-(l + i ^ ; ) 1 0 ))n(l-^))

where P(z) denotes the product of all primes p < z and s = logy/ log z, the implied
constants depending only on n.
REMARK. Th e Fundamenta l Lemm a play s a n assistin g rol e i n siev e theory ; it
usually serve s fo r preliminar y eliminatio n o f the element s o f a sequenc e whic h hav e
small prime divisors, that i s primes p < z — y}l s wit h s —> +oo . I n such application s
the result s ar e asymptoticall y precise , so one loses nothing whil e gaining a practica l
assumption tha t th e sequenc e i n question i s supported o n almos t primes . Quit e
often on e onl y need s a n upper-boun d o f th e righ t orde r o f magnitude , tha t is
w
(6.49) ] T \±g(d)<$:K l[(l-g(p)).
d\P{z) P<z

This holds under th e conditio n (6.34)(whic h i s weaker tha n (6.47) ) wit h th e implie d
constant dependin g onl y o n n.

Although th e sequenc e A = (a n) accompanie d u s whe n constructin g th e sieve s

A + = (Aj) an d A - = (A^"), one shoul d realiz e tha t A + , A - have actuall y nothin g
in commo n wit h A. On e ca n legall y appl y th e particula r sieve s A + , A ~ t o any
sequence o f non-negative numbers , th e onl y risk i s that th e result s ma y no t b e idea l
if th e paramete r K (th e so-calle d dimension ) an d y (th e level ) o f A+ , A " d o no t
match perfectl y th e characteristic s o f A.
Of cours e on e doe s no t nee d t o pus h s to infinity . Th e lowe r boun d (6.42 ) is
already interestin g i f 5 > ft. Assuming that th e remainder ter m R(x, y) i s negligible
for a suitable leve l y, on e infer s fro m (6.42 ) tha t

(6.50) Yl a n »X(logz)-*

(n,P(*)) = l
160 6. E L E M E N T A R Y SIEV E M E T H O D S

where z = y 1 / ^ + £ ) for an y e > 0 provide d y i s sufficientl y larg e i n term s o f e.

Suppose w e ca n contro l R(x,y) fo r y = x a~£ wit h 0 < a ^ 1 . The n (6.50 ) hold s
for z — x ^a _ £ ) / ^ + £ ) , whil e th e conditio n (n,P(z)) = 1 together wit h n ^ x impl y
that n ha s a t mos t a/(3 prim e divisors .
In genera l t o establis h a n uppe r boun d fo r S(x, z) i s muc h easier , becaus e
S(x,z) i s decreasin g i n z s o one ca n reduc e z a t will .
EXERCISE 1 . Deriv e fro m (6.41 ) th e uppe r boun d
(6.51) TT 2(X) <Cx(logx)~ 2 .

EXERCISE 2 . Sho w tha t th e sequenc e (6.1 ) satisfie s th e siev e condition s fo r

dimension K = 1 and leve l a = 1 / 2 (se e Theore m 1 7.1 ) . The n deriv e fro m (6.42 )
that
1 /20 2
(6.52) 7r 2 (x,x ) > x(\ogx)-
where 7^(2 , z) denote s th e numbe r o f prime s p ^ x fo r whic h p + 2 ha s n o prim e
divisors < z. Therefor e ther e ar e infinitel y man y prime s p suc h tha t p + 2 ha s a t
most twent y prim e divisors . Recen t development s o f sieve methods i n this directio n
show (amon g othe r things ) tha t

\{p < x;p + 2 = pi o r P1P2 wit h pi , p 2 > x 3 / / 1 1 }| > x ( l o g x ) -2

for al l sufficientl y larg e x. Henc e p + 2 ha s a t mos t tw o prim e divisor s infinitel y

often. Thi s las t resul t i s du e t o J.R . Che n [Ch] . W e stil l d o no t kno w whic h cas e
p + 2 — p\ o r p + 2 = P1 P2 occur s infinitel y often ? Probabl y both !
6.5. T h e A 2 -sieve.
A powerful an d elegan t metho d o f an upper boun d siev e came from Atl e Selber g
[SI]. Selberg' s metho d yield s result s o f grea t generality , i t i s simple r tha n combi -
natorial sieve s a t th e start , thoug h equall y comple x i n it s advance d forms . W e
modify th e notatio n o f the previou s sectio n slightl y t o displa y clearl y th e econom y
of condition s whic h ar e require d fo r performin g Selberg' s sieve .
Recall tha t a n uppe r boun d siev e o f leve l D i s a sequenc e o f rea l number s A ^
for d < D wit h A i = 1 such tha t
X
(6.53) Yl d > ° fo r al l m > 1 .
d\m

Hereafter th e superscrip t + i s omitted fo r notatio n simplicity . Thi s positivit y con -

dition wa s quit e difficul t t o ge t a hol d o n i n combinatoria l sieve . Selber g mad e i t
very eas y b y choosin g A ^ such tha t
Ad=
(6-54) E (E^)2
d\m d\m

where {pd} i s anothe r sequenc e o f rea l number s wit h

(6.55) p i= 1 .
Since the square s ar e non-negative , suc h a choice guarantee s (6.53 ) n o matter wha t
Pd are!
 6. ELEMENTAR Y SIEV E METHOD S 161

Selberg's choic e amount s t o

(6.56) A d= ] Tp dlpd2.

In orde r t o contro l th e leve l w e assume p d ar e supporte d o n integer s < \fl),

(6.57) p d = 0i f O v^D .

Hence th e resultin g siev e {A d} has leve l o f suppor t D. Followin g Selber g w e cal l i t

the A 2-sieve o f leve l D.
Throughout, A = (a n) i s a sequence of non-negative numbers , P i s a squarefre e
number an d fo r an y d \ P w e se t

(6.58) \A d\= J2 a n = g(d)X + r d(A).

n=0 (mo d d)

Here a s before , g(d) i s a multiplicativ e functio n satisfyin g (6.1 0) .

Applying th e A 2-sieve t o th e sequenc e A = (a n) i n th e siftin g rang e P w e ge t

a
S(A,P) = Yl ^ E a - ( E Pdf
(n,P) = l n d\(n,P)

= E 52pd 1 Pd2\A[dud2]\ = XG + R(A,P)

di,d2|P

where

(6.59) G = ^ S^([di,d2])/9 dlpd2

di,d2|P

and

(6-60) R(A,P) = Y J2 PdiP<hr ldlld2]{A).

diMP

The tas k befor e u s i s to mak e thi s genera l inequalit y optimal . Forgettin g fo r a

moment abou t th e remainde r ter m R(A, P , A2) w e wish to minimize G with respec t
to the unknow n number s p d subjec t t o (6.55 ) an d (6.57) . Th e ensuin g number s wil l
satisfy

(6.61) \p d\ <1 ,

hence th e remainde r ter m i s automaticall y unde r control .

The expressio n (6.59 ) i s a quadrati c for m i n p&. In orde r t o find th e minimu m
of G i t help s t o diagonalize . I n th e presentatio n below i t goe s withou t sayin g tha t
Pd is supporte d on , an d th e relevan t variable s o f summatio n ru n ove r th e divisor s
of P (thu s ove r squarefre e numbers) . Furthermor e w e ca n assum e tha t

K
0 < # ( pJ ) < l i f p|p
1(6.62)
; '
g(p) = o ifpfp .
162 6 . E L E M E N T A R Y SIEV E M E T H O D S

Let h{d) be the multiplicative functio n define d b y

(6.63) k( P) = ^ -y

We obtai n
G = ^2 g(abc)p acpbc
abc\P

(a,6) = l
2
i
^ 5 (c)^/,(%(d) 2 (^ 5 (m)p^cdm
c ;

= E^)_ 1( E 5Mp 2
m) .
d|P mE0(mo d d)

Hence b y the linear chang e o f variables

m
(6.64) £ d = p(d) ^2 g( )pm
m E 0 ( m o d d)

we obtain th e diagonal for m

(6.65) G = 5>(d) -1 &

d|P

We stil l hav e t o reinterpre t th e condition (6.55 ) i n terms o f the new variables

£d- To this en d we use the Mobius inversio n t o convert (6.64 ) int o

M6) « = $ | £ fc.
^ V ; d = 0 ( m o d i)

In particular , fo r £ = 1 this give s the linear equatio n

(6.67) £ & = !•
d|P

Moreover, on e observes b y (6.64 ) an d (6.66 ) tha t th e support condition s (6.57 ) fo r

pd are equivalent t o these fo r £d,
(6.68) f d = 0i f O >/# .

Now ou r targe t i s to minimize (6.65 ) o n th e hyperplan e (6.67) . Applyin g

Cauchy's inequalit y t o (6.67 ) w e derive GH ^ 1 where

(6.69) H= Yl M< 0
d<VD,d\P

so G cannot b e smaller tha n H~ l. The equality

(6.70) GH = 1
holds for
(6.71) & = h^H-1 ifd<y/D.
 6. ELEMENTAR Y SIEV E METHOD S 163

Note tha t

(6.72) H^Yl h
^ = II^1 + %)) = II^1 - 9(p))-\
d\P p\P p\P

(6-73) ^ U ( i _ 0 ( p ) ) .
P\p

Next w e compute pi by inserting (6.71 ) int o (6.66 ) gettin g

li(£)g(£)peH = £ h(m),
m<y/D
ra=0(mod i)

that is
=
(6.74) ^ i^ E W
d<y/D/£

Now we show (6.61 ) . T o this en d w e group th e term s in (6.69) accordin g to the

greatest commo n diviso r o f d and £ getting
H
= T, H M<* ) = £>(*) Y, Mm )
k\t d<VD k\l m<y/~D/k
(d,£)=k (m,i)=l

^(E^))E Hm)=li(t)PiH
m<\fD/£
(m,l) = l

and thi s prove s (6.61 ) (thi s nea t estimat e i s due to J.H. va n Lin t an d H.-E . Richer t
[LR]). Pro m thi s on e get s directl y b y (6.56)
(6.75) \\ d\^r3(d).

Prom th e above result s w e conclud e th e following

T H E O R E M 6.4 . Let A = (an) be a finite sequence of non-negative numbers and
P be a finite product of distinct primes. For every d\P we write

(6.76) |.A d|= Y, a n=g(d)X + rd(A)

n=0(mod d)

where X > 0 and g(d) is a multiplicative function with 0 < g(p) < 1 for p\P. Let
h(d) be the multiplicative function given by h(p) = g(p)(l — #(p)) _ 1 and

(6.77) H= Y, M< 0
d<VD,d\P

for some D > 1. Then we have

(6.78) S{A,P)= Yl CLn^XH^+RiA^P)

(n,P) = l
164 6. ELEMENTAR Y SIEV E METHOD S

where

(6.79) R(A,P) = ^\ drd{A)

d\P

with \ d given by (6.56) and (6.74)-

Using (6.75 ) on e estimate s th e remainde r ter m crudel y b y

(6.80) \R(A,P)\^ Y, r 3(d)\rd(A)\.

d<D,d\P

The A 2-sieve i s ver y genera l indeed . It s uppe r boun d doe s no t requir e an y

regularity i n the distributio n o f the densit y functio n g(p) ove r primes , therefor e th e
dimension o f a siftin g proble m doe s no t pla y explicitl y a role . I f g(p) = u(p)p~ 1 ,
where u>(p) i s th e numbe r o f residu e classe s (mo d p) whic h on e want s t o exclude ,
then h(p) = oo(p)(p — uj(p))~ 1 i s th e rati o o f th e number s o f rejecte d t o admitte d
classes, an d H incorporate s thes e ratios . O f course , th e large r uj(p) is , th e smalle r
the uppe r boun d (6.78) . However , a fe w loca l deviation s o f uo{p) fro m it s averag e
have insignifican t globa l effect . T o th e contrar y ou r estimate s derive d b y th e com -
binatorial siev e ar e quit e sensitiv e i n thi s respec t becaus e th e hypothesi s (6.35 ) i s
assumed t o hol d fo r al l segment s o f prime s w < p < z, n o matte r ho w short .
The A 2-sieve yield s fantasti c result s fo r siftin g problem s whe n th e numbe r o f
residue classe s t o b e exclude d i s large , i t ca n compet e wit h th e larg e siev e metho d
of Linni k (se e Sectio n 7.4) .
REMARKS. Not e tha t th e number s p d depen d o n th e multiplicativ e functio n
g(d) s o (indirectly ) o n th e siftin g sequenc e A. However , assumin g som e regularit y
of g(d) (a s i n th e K dimensional sieve , se e (6.34) ) on e ca n establis h fairl y goo d
approximation (fo r smal l d a t an y rate )

6.6. Estimat e fo r th e mai n ter m o f th e A 2 -sieve.

To brin g th e Selber g uppe r boun d (6.78 ) t o a practica l for m w e nee d a clea r
lower boun d fo r
H = H(D) = Y. M<*) -
d<y/D,d\P
The uppe r boun d (6.72 ) hold s i n genera l bu t a goo d lowe r boun d require s som e
restrictions o n th e densit y functio n g , s o o n /i , an d th e siftin g rang e P. Pro m no w
on P i s the produc t o f al l prime s p < V^D , so we hav e

h
H(D) = Y! (d)
d<Vn

where th e superscrip t b restricts th e summatio n t o squarefre e numbers .

 6. E L E M E N T1
AR Y SIEV E METHOD S 6 5

We ca n estimat e H(D) strongly , elementaril y an d nicel y fo r g(d) = d _ 1 (thi s

occurs whe n on e i s sifting number s i n a n interval) . I n thi s cas e h(d) = (p(d)~ 1 an d

d
H(D)= Y, ~in(l + W + ---)> E m-^logv^ .
d<y/D P\d m<y/D

If g(d) agree s wit h dr x onl y fo r (d , q) — 1 (a s fo r exampl e i n th e cas e o f siftin g a n

arithmetic progression) , the n thi s lowe r boun d hold s wit h th e followin g modifica -
tion:

(6.82) H(D)>l[(l-g(p))(\ogVD).
p\q

The abov e exampl e i s rather special . No w suppos e g(p)p i s n o n average , sa y

(6.83) Yl 9(P) logP = Klog x + 0(1 )

for al l x ^ 2 where K is a positive number . Henc e g(p) logp < C 1. Suppos e als o tha t
2
(6.84) ^ ( p ) logp<oc.
v
Since h(p) = g(p) + 0(g(p) 2), i t follow s tha t (6.83 ) hold s fo r h a s wel l (wit h a
different implie d constant) . Applyin g Theore m 1 . 1 fo r th e multiplicativ e functio n
h i n plac e o f / w e ge t

H{D) = c(lo g VD)" { 1 + 0((logZ?)- 1 )}

where
^rWTylT1-^))-1^-^
P

(the require d condition s (1 .89 ) an d (1 .90 ) ar e satisfie d b y (6.83) , (6.84) ) an d th e

implied constan t depend s o n tha t i n (6.83) . I f D i s larg e i n term s o f thi s constan t
we ca n inver t thi s approximatio n gettin g

(6.85) H(D)- 1
= 2 T ( K + l)H g{logD)-K{l + O^logD)" 1 )}

where

(6-86) H g = H(l-g(p))(l-$)-«.
P

6.7. Estimate s fo r th e remainde r ter m i n th e A 2 -sieve.

Once again w e consider a class of sifting problem s i n which th e individua l erro r
terms satisf y

(6.87) \r d(A)\ ^ g(d)d.

Naturally wit h thi s propert y on e make s th e conditio n

(6.88) g{d)d > 1 i f d\P.

166 6. E L E M E N T A R Y SIEV E M E T H O D S

This implie s g{[d\, c?2])[^i5 ^2] ^ 9(di)g (^2)^1 ^2, therefor e

2 2 2
(6.89) \R(A,P,A )\^ ( ] T \Pd\9iW) < (jj J2 HmMm))
d<\fD m<y/D

by (6.60 ) an d (6.72) , wher e cr(m) denotes th e su m o f divisor s o f m.

Next assum e tha t th e densit y functio n g(p) satisfie s th e followin g crud e boun d
(6.90) J2 9{p)\ogp^\og{2x/y).

Hence w e infer th e sam e propert y fo r h(p)a(p)p~ 1 . The n w e appl y (1 .85 ) gettin g

V^ h(m)a(m) <^i — Y ^ h(m)a(m)m * .

r
m<VD m<\ D

Here w e hav e
1_ 1-
^ /i(ra)cr(m)ra ^ H ^ /i(ra)ra < iJ .

Hence w e conclud e tha t

2
(6.91) R(A,P)^D(logD)- .
Combining wit h (6.78 ) w e get
THEOREM 6.5 . Suppose the conditions of Theorem 6.4 hold. Moreover, assume
(6.87), (6.88) and (6.90). Then we have

(6.92) S { A , P ) ^ + o { ^

where H = H(D) is given by (6.77), D > 1 is arbitrary, and the implied constant
depends only on that in (6.90).
6.8. Selecte d application s o f A 2 -sieve.
Consider th e sequenc e A = (a n) whic h i s th e characteristi c functio n o f a n
arithmetic progressio n i n a shor t interva l
(6.93) n = a(mo d g) , x < n< x +y
where (a , q) = 1 and 1 < q < y. Le t P consis t o f prime s p ^ y/y wit h p \ q. The n
1
(6.94) TT(X + y ; g, a) - TT(X; q, a) ^ S(A, P) + y/yq' .
On th e othe r hand , th e condition s o f Theore m 6. 5 ar e satisfie d wit h X = yq~ x an d
g(d) = d' 1 i f (d,q) = 1 . B y (6.82 ) w e hav e H(D) > ^-logD. Combinin g (6.92 )
with (6.94 ) w e ge t

7T(x + y ; q, a) - TT(X ; q, a) < —— + O - —^ +

z
ip(q) log D \\og Dq
Choosing D = yq~ l w e conclud e
 6. E L E M E N T A R Y SIEV E M E T H O D S 167

THEOREM 6.6. For (a,q) = 1 and 1 ^ q < y we have

(6.95) n(x + y; qja)-7r{x;q,a) < —— —-—- + O . 2

2
<p(q)log(y/q) \q\og (y/q)
where the implied constant is absolute.

This i s calle d th e Brun-Titchmars h inequality . Usin g th e larg e siev e metho d

H.L. Montgomer y an d R.C . Vaugha n [MVl ] hav e show n tha t th e erro r ter m i n
(6.95) ca n b e deleted .
Next w e take A — (an) th e characteristi c functio n o f th e polynomia l value s
n = (r a — a±) — • (m — ctk)
with 1 ^ m ^ x , wher e al l aj ar e distinct . I n thi s cas e g(p) — v(p)p~1 wher e v{p)
is th e numbe r o f root s modul o p. I f p i s sufBcientl y large , i/(p) = k s o w e hav e a
^-dimensional siev e problem . B y (6.85 ) an d (6.92 ) w e deduc e
THEOREM 6.7 . Let a = ( a i , . . . , a^) be distinct integers which do not cover all
residue classes to any prime modulus. Then the number of integers 1 ^ m ^ x for
which ra — c*i,.. . , m — oik are all primes satisfies
fe '1 log
0 glog
l 0x >
(6.96) 7 r ( x ; a ) ^ 2 ^ ! ^ x ( l o g x ) - (l + o( 1 Q ^))
logx
where
k
. n ( _ ^u(p)\ /1 ) \- ~
(6 , 7) B 11 )( _i
p J v p)

REMARKS. Th e uppe r boun d (6.96 ) i s large r b y facto r 2 k k\ tha n th e conjec -

tured asymptoti c
7r(x;a) ~ Bx(logx)~ k.
This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/08

CHAPTER 7

BILINEAR FORM S A N D TH E LARG E SIEV E

7.1. Genera l principle s o f estimatin g doubl e sums .

Certainly th e mos t versatil e instrument s o f analyti c numbe r theor y ar e doubl e
sums
a
(7.1) tf (a , /?) = ^ J2 rnM(m, n) .
mn
Here a = ( a m ) , ft = (/3 n) ar e finit e vector s o f comple x number s an d
(7.2) d > = (0(m,n) )
is a finit e matri x wit h comple x entries . I n matri x notatio n \I/(a , /3) = a$/3* . Give n
<I> the goa l i s t o giv e a shar p estimat e fo r \I/(a,/3 ) whic h i s vali d fo r arbitrar y
vectors a , /3 . Yes , i t i s essentia l t o assum e nothin g abou t th e number s a m , /3 n ,
because i n practic e the y ar e quit e comple x objects . Ver y ofte n a n attemp t t o
utilize specifi c structur e o f these vector s bring s bac k a mirro r imag e o f the origina l
situation. Henc e a question , wha t i s th e benefi t fro m arrangin g th e origina l sum s
into bilinea r forms ? Th e ke y featur e i s tha t th e sequence s a , (3 do no t se e eac h
other! Therefor e a cancellatio n i n \J/(a,/3 ) i s possible , subjec t onl y t o propertie s o f
the matri x coefficient s </>(ra , n). Indee d w e hav e
2
(7.3) | * ( a , / ? ) | ^A||a||2||/3||2
where A = A($ ) i s the nor m o f th e correspondin g linea r operato r an d

(7.4) IMI 2 = £ K , | 2 , |L0|| 2 = $ > „ | 2 .

A more transparent wa y of estimating \I/(a , /3) goes by Cauchy' s inequality . T o thi s
end choos e firs t m t o b e i n th e oute r su m an d n i n th e inne r su m gettin g
2
(7.5) |*(a,/?)| < ||a|| 2 £|£/W( m ' n )f-
mn
Note tha t th e unknow n coefficient s a m i n th e oute r su m ar e gone . A t thi s poin t
one coul d b e mor e flexibl e b y introducin g ne w coefficient s f(m) i n plac e o f a m t o
smooth th e nex t ste p an d t o optimiz e th e fina l bound .
Smoothing is one among many optional devices which are put int o practice wit h
remarkable effect s i n analyti c numbe r theory . Thi s i s not jus t a technical device , i t
can b e viewe d a s a kin d o f spectra l completion .
Next squarin g ou t th e inne r su m ove r n followe d b y interchangin g th e orde r o f
summation w e ge t

169
170 7. BILINEA R FORM S AN D TH E LARG E SIEV E

For ri\ — ri2 , the so-calle d diagona l terms , on e canno t ge t cancellatio n i n th e inne r
sum
2
(7.7) 5>Kn)| ,
m

so the savin g facto r comes solel y fro m th e fac t tha t th e diagona l i s smaller tha n th e
full square ! However , fo r mos t o f ni, ri2 of f the diagonal one may find a considerabl e
cancellation i n th e su m
(7.8) ^ 0(m , ni)<£(ra, n 2 )
771

because o f independen t variation s i n th e sig n o f 0(ra,ni) , 0(rn,n2) . Ho w on e

executes th e summatio n i n (7.8 ) i s a separat e questio n dependin g o n particula r
characteristics o f <£ . A t thi s poin t th e sum s (7.8 ) ca n b e eithe r estimate d directl y
finishing th e job, o r i t ca n b e transforme d t o anothe r su m (s o to spea k a dual sum )
by a kin d o f spectra l formula . I n th e latte r scenari o eve n a trivia l estimatio n o f
the dua l su m ofte n produce s a non-trivia l boun d fo r th e origina l sum . Moreover ,
one ca n als o exploi t th e extr a summatio n ove r ni , 71 2 . Ou r poin t i n thes e remark s
is t o sa y tha t th e genera l estimat e (7.3 ) obstruct s th e visibilit y fo r improvements .
We advis e t o carr y ou t th e argument s i n practic e alway s b y Cauchy' s inequalit y
because i t allow s yo u t o contro l you r position .
Before launchin g Cauchy' s inequalit y t o th e doubl e su m \£(a,/3) , mak e sur e
to choos e th e oute r an d th e inne r sum s variable s convenientl y a s th e result s wil l
be different . Thi s optio n o f interchangin g m , n i n \I/(a,/3 ) i s th e essenc e o f th e
following dualit y principle .
DUALITY PRINCIPLE . Suppose for any complex numbers f3 nj
2
(7.9) ^ | ] T / 3 n0(m,n)j <A||
mn
Then for any complex numbers a m

2
(7.10) ^ | ^ a m</>(m,n) ^A||a|l2,

where A is the same in both inequalities.

PROOF. Th e lef t sid e o f (7.1 0 ) i s equa l t o (7.1 ) wit h

(7.11) f3 n = ^2 a m(f>(m, n).

m

By (7.5 ) an d (7.9 ) w e get #(a,/3) < A||a|| 2 ||/?|| 2 . Bu t ||/3|| 2 = *(<*,/? ) s o (7.1 0 ) i s
2

derived fro m (7.9) . •

Because o f th e applicatio n o f Cauchy' s inequality , i t i s inevitabl e tha t muc h

attention concentrate s o n estimatin g sum s o f type (7.9 ) o r (7.1 0) . Thes e equivalen t
statements fo r genera l coefficient s a m , fi n ar e calle d Larg e Siev e Inequalitie s fo r
historical reason s althoug h the y d o no t resembl e a siev e o f an y kind . I t i s better t o
regard (7.9) , (7.1 0 ) a s a near-orthogonalit y propert y o f <I>.
We end thi s introductio n t o genera l bilinea r form s wit h a fe w comment s abou t
obvious cases for which the ideas described definitel y canno t work . Suppos e 4>{m,n)
 7. BILINEA R FORM S AN D TH E LARG E SIEV E 171

factors, sa y 0(ra,n ) = x(m)'0(^) , o r <p(m,n) is a shor t linea r combinatio n o f suc h

products. I n othe r words , th e variable s m , n o f </)(ra , n) ar e separate d a t no , o r
low, cost . The n w e ma y choos e th e vector s a — ( am ) , /3 = (/3 n) wit h a m = x(ra) ,
/3n = ^(n) . Fo r this biased choice of vectors the bilinear for m i s \I/(a, /3) = ||a|| 2 ||/3|| 2
showing n o cancellatio n o f terms .
The principles of estimating bilinear form s i n analytic number theor y hav e bee n
articulated ver y slowly . Perhap s becaus e i t i s no t th e simplicit y o f th e ideas , bu t
rather th e additiona l device s introduce d b y researcher s i n particula r case s whic h
make thi s technolog y s o effective today . Firs t application s deal t wit h bilinea r form s
in whic h (/>(ra , n) depende d onl y o n th e produc t mn, sa y

(7.12) (p(m,n) = g(mn),

where g i s a n arithmeti c functio n i n on e variable . O f course , i n thi s cas e g canno t

be multiplicative . I.M . Vinogrado v shoul d b e mentione d fo r th e firs t impressiv e
results. Usin g a n exclusion-inclusio n procedur e ( a siev e method ) h e manage d t o
express exponentia l sum s ove r prime s b y doubl e sum s an d estimate d th e latte r
by elementar y mean s (cancellatio n i n geometri c series) . The n utilizin g th e resul t
in th e circl e metho d Vinogrado v gav e a stunnin g solutio n o f th e ternar y Goldbac h
problem. Hi s ideas are still alive today (se e Chapters 1 3 and 1 9) ; however, ther e ar e
more instances which stimulated th e whole business of bilinear forms . Som e of these
will b e explicitl y describe d i n forthcomin g chapters , an d man y ar e embedde d i n
arguments withou t recognition . Linni k [Li2 ] ha s elaborate d th e ide a o f estimatin g
double sum s int o th e dispersio n method , whic h i s missin g i n thi s book .
The notatio n chose n fo r thi s introductio n serve s a s a model . Obviousl y w e d o
not nee d t o conside r 0(ra , n) a s a functio n i n tw o integra l variables . Actuall y i n
some work s i t i s nice r t o inde x th e tw o variable s b y mor e appropriat e parameter s
(rational fractions , characters , eigenvalues , etc.) .
In thi s chapte r w e giv e a selectio n fro m th e grea t variet y o f bilinea r form s
and larg e siev e typ e inequalities . A fe w basi c one s will b e give n detaile d proofs , a
few mor e wil l b e discusse d i n a broa d context , an d man y wil l b e jus t state d wit h
references t o origina l publications .

7.2. Bilinea r form s wit h exponentials .

The bilinea r form s o f typ e (7.1 ) wit h

(7.13) <f>(m,ri) = e{x rnyn)

appear quit e often i n classical analytic number theory . Her e x m , y n ar e real number s
and e(z) = exp(27rz2: ) a s usual .

LEMMA 7.1 . For any complex numbers a m and real numbers x m we have

(7.14) / \^2a me(xmy)\dy^5Y1

Y,Y I
Y
rn 2Y\x rni-xrn2\<l
172 7 . B I L I N E AR F O R M S A N D T HE L A R G E SIEV E

PROOF. Le t g(y) be a non-negative functio n o n R with g(y) ^ 1 if \y\ ^ \ and

supp(#) C [-1,1]. The n th e left sid e of (7.14) i s bounded by

/ #(^)|X^ me
( X m 2 / ) | dy = 2Y^Y^ arni^rn29(<2Y(xrni ~ Xm2))
m mi,?7i 2

2Y\xmi-xrri2\<l

The Fourie r pai r g(y) = ( S1 ^y J , g(y) = ^- max(l — |f|,0 ) satisfie s th e abov e

conditions, givin g th e result. •

Our basi c inequalit y is

THEOREM 7.2 . Let xm ; y m be real numbers with | x m | ^ X, \y m\ ^ Y. Then
for any complex numbers a m, (3 n

\Y^^2a™Pne(Xrnyn)\ < 5 ( 1 7 + 1 ) 2 ^ ^ ^ |^m 1 ^m2|J

™ ™ |a; mi -xm2|y<l

(7.15) ( E E l&^l)
|yni-yn2|X<l

P R O O F . Pu t e = ( 4 X ) "1 an d
7r£mG?m
sin(27r££mJ

so | 7 m | ^ 7rX|am |. Writin g

e(xy) — / eixtjdt
v yj
sm27rsx J y__£ v ;
we find tha t th e left sid e of (7.15) i s bounded by
, rVn+e , rY+e
n
E^ / ^7 me(xmy)d?/ ^ / ^ |/3 n| ^7me(xm7/)

Hence b y Cauchy's inequalit y an d Lemma 7. 1 we find tha t th e square o f the lef t

side o f (7.15) i s bounded by

(/ ( E l^ n0 dy){ \%2'YTn ex
( my)\ d V)

^2e{ E E \PnJn 2\)5(Y + e){*X)2( J2Y, l^a m2 |).

£
\Vn1 -yn 2\< \x mi-xm2\Y<l

Here 1 0e(Y + S)(KX) 2 = *£(XY + \) < 25{XY + 1 ) completin g th e proof. •

If the points x m an d y n ar e well-spaced, s o that onl y the diagonal term s appea r

on the right sid e o f (7.15), we ge t
 7. B I L I N E A R F O R M S A N D T H E L A R G E S I E V E 173

COROLLARY 7.3 . Suppose \x m\ ^ X, \y n\ ^ Y, and for m\ ^ m 2, n\ ^ n 2 we

have \x mi - x m2| ^ A, \y ni - y n2\ ^ B . Then
x
1 A I1
A
Y.Y.arnPne{xmyn)\^h{l + XY)2(Kl + —y( Kl + ^ 2
\\a\\W\\.

PROOF. Appl y (7.1 5 ) wit h max(X , B 1 ) an d max(Y , A1 ) i n plac e o f X an d

Y respectively . •

The spacin g proble m ca n be easily solve d fo r points give n by smooth functions .

For exampl e w e derive b y Corollar y 7. 3 the followin g
COROLLARY 7.4 . Let f(m), g(n) be real functions on [M,2M], [N,2N] respec-
tively, such that f < F, g < G and \f\ ^ FM~ l, \g'\ ^ GN' 1 . Then for any
complex numbers a m, f3 n we have

J2 E <*mMf(m)g(n)) « (FG)-\ (FG + M)\(FG + N)\\\a

PROOF. W e have |/(mi ) - f(m 2)\ > |ra i - m 2\FM~l an d |(j(ni ) - g(n 2)\ ^
|ni - n 2\GN~l henc e th e resul t follow s fro m Corollar y 7. 3 for A = F M _ 1 , J B =
GW" 1 , I = F an d F = G . D

Corollary 7. 3 yields als o interestin g result s fo r exponential sum s wit h monomi -

als, becaus e th e rational point s ar e well-spaced .
LEMMA 7.5 . Let a/3 ^ 0, A > 0, K ^ I, L ^ 1 . The number of integral
quadruples {k,k',t,,l f) with K ^ fc, k' < 2K, L ^ £, f ^ 2 L suc/ i that

'k'\<* /t!\P
^A
™I ( I ) - ( T)
is bounded by 0(K L(AK L + log2K L)) where the implied constant depends only on
a, (3.
EXERCISE 1 . Prov e Lemm a 7.5.
Prom Lemm a 7. 5 we derive
COROLLARY 7.6 . Let a/3^6 ^ 0 , X > 0 , K, L, M, N ^ 1 . Let a k,i and 6 m?n
be complex numbers with \a k^\ ^ 1 an d |6 mjTl | ^ 1 for K ^ k ^ 2K, L ^ £ ^ 2L,
M ^ m ^ 2M ; i V ^ n < 2A \ TAe n

7 17 kaePm^n6
( - ) S E E S a k^b^ne (X-
KaLPM^N6,
k£ rn n

<1 (l ++ — J ( l + — ) (XKLMN)2 \og(2KLMN)

where the implied constant depends only on a, (3, 7, S.

174 7. B I L I N E A R F O R M S A N D T H E L A R G E SIEV E

7.3. Introductio n t o th e larg e sieve .

The larg e siev e typ e inequalitie s ar e ubiquitou s i n moder n analyti c numbe r
theory. Th e nam e i s somewhat misleadin g however , a s sieve s d o no t ente r th e
general picture : i t was derive d fro m Linnik' s origina l ide a (se e [Lil]) , whic h wa s
subsequently transforme d beyon d recognitio n int o genera l L 2 -type estimates , fro m
which siev e result s ca n b e derive d i n som e cases .
First w e explain i n genera l term s th e underlyin g philosophy , whic h i s expecte d
to hav e muc h large r applicability . Le t A ' be a finite set o f "harmonics" , well-suite d
maybe t o solv e analyticall y som e interestin g equation . T o eac h x G X, we assum e
that a sequence (x(n)) i s associated , s o to spea k th e "Fourie r coefficients" .
The larg e siev e proble m fo r X is to find C = C(X, N) ^ 0 such tha t th e
following "larg e siev e inequality " hold s fo r th e L 2 -mean o f a linear for m i n th e
x(n):

(7.18) Yl | Yl a - x ( n )f ^ C(*,JV)||a|| 2
xEX n^N

for an y comple x number s a n , wher e ||a|| 2 = ^2\a n\2. B y Cauchy' s inequalit y on e

gets (7.1 8 ) wit h C(X, N) = A/"^!, bu t thi s i s not usefu l i n applications .
The ide a i s tha t th e "harmonics " ar e orthonormal , o r nearl y so , henc e i f on e
expands th e squar e
i2
] T ] P a nx(n)\ = ] T anian2 ^ x(ni)x(n 2 ),
xeX n^N ni,n 2 xeX

the diagona l term s correspondin g t o ri\ — ri2 should b e dominating , contributin g

roughly |^ f j i|a.||2 becaus e x{n) shoul d b e o f size 1 on average . I n an y cas e i t i s har d
to imagin e tha t C(X, N) woul d b e muc h smalle r tha n th e numbe r o f element s i n
X. Moreover , becaus e o f the dualit y principl e fo r bilinea r forms , whic h mean s tha t
(7.18) i s equivalent wit h

2
(7.19) ^\^2b xx{n)\ ^C(X,N)^2\bx\2
n^N xeX x£X

with diagona l contributio n NJ2\b x\2, i t is no t possibl e t o mak e C(X,N) smalle r

than th e lengt h o f th e vecto r (a n ). Therefor e i n vie w o f thes e diagona l limitation s
the bes t estimat e whic h on e ca n hop e fo r C(X,N) i s roughl y

(7.20) C(X,N)~\X\+N.

This hold s indee d i n a number o f crucia l cases .

A shar p larg e siev e inequalit y (7.20 ) say s tha t th e linea r form s

^a nx(n)

of lengt h i V ^ \X\ wit h \a n\ ^ 1 are, o n averag e ove r x G X, of siz e iV 1 / 2 . Thi s i s

best possible , an d althoug h i n many cases it is expected tha t th e individual sums ar e
of thi s size , ver y fe w case s ar e known . Especiall y interestin g i s th e cas e a n = /x(n)
 7. BILINEA R FORM S AN D TH E LARG E SIEV E 175

and x(n) = A/(n) ar e coefficient s o f som e L- function (se e Chapte r 5) . The n th e

individual boun d
Y^ »{n)\ f{n) < e N%+ e

(for an y e > 0) i s equivalent wit h th e Rieman n Hypothesi s fo r L(/ , s) (se e (5.52)) .

Hence, a shar p larg e siev e type inequalit y fo r a famil y o f L- functions i s as powerfu l
on averag e a s th e Gran d Rieman n Hypothesi s woul d give .
7.4. Additiv e larg e siev e inequalities .
In th e additiv e larg e siev e inequalit y th e harmonic s considere d ar e additiv e
characters o f Z with x(n) = e(an) fo r som e a G R (of cours e onl y a mo d 1
matters). S o the linear forms t o be estimated ar e just th e trigonometric polynomial s

(7.21) S(a) = ^2a ne(an)

n
where th e comple x coefficient s ar e supporte d i n M < n ^ M + N. Not e tha t i f ar
are well-space d modul o 1 , i.e. ,
||a r - a 3\\ ^5 i fr ^ s,
for som e 5 > 0 (wher e \\x\\ is the distanc e t o th e neares t integer) , the n th e numbe r
of distinc t point s a ri i.e. o f harmonics , i s ^ 1 + 6~ l.
T H E O R E M 7.7 . For any set of 8-spaced points a r e M/ Z and any complex
numbers a n with M < n ^ M + N, where 0 < S ^ \ and N ^ 1 is an integer, we
have
2
(7.22) 5 ^ | Yl Oneiarntf^tf-'L+N-yWaW .
r M<n^M + N

This estimat e i s bes t possible ; it was prove d i n th e for m state d independentl y

by Selber g [S3 ] an d Montgomer y an d Vaugha n [MV2] , an d i s o f th e strengt h o f
(7.20). W e giv e th e proo f o f Montgomer y an d Vaugha n whic h i s based o n th e
following generalizatio n o f th e Hilber t inequality .
LEMMA 7.8 . Suppose X r are distinct real numbers with |A r — A s| ^ S ifrj^s.
Then for any complex numbers z r we have

<"»> lEEsfr^jEw' -
PROOF. B y Cauchy' s inequality , it suffices t o sho w tha t
2 ^2

<™> ElEi^d^EiA — A. r s^r

r s S2 *
Squaring ou t w e arrang e th e lef t sid e a s follow s

L = J2Y,ZsZt Yl
1
r^s,t

z2^2z2 {Xr
r^s s^t
-x sy+
^J^xr^s,t
s-xt z2{\ r-\s x r-\J-
176 7. B I L I N E A R F O R M S A N D T H E L A R G E SIEV E

For the last su m we have

y ^ /1
1 \ _ y ^1 y ^1 2
^-—' V Ar — A s \ r — \t / ~* X r — X s *—f X r — Xt X s — Xt

hence
Zszt
+2
- E M * X. v: <., O T, £ ?(Aa"A
t t )2

+
^J- AT^A ; i \ J~^K ~ \t A^A ; / •
s^t ryts r^t
Note no w that th e estimate (7.24 ) amount s to finding the norm of the matri x (^ r?s )
with fjL r,s = (A r — A s ) _ 1 i f r ^ s and \ir^r — 0, thus w e may assume tha t th e vector
(zr) i s extremal . Sinc e th e matri x i s skew-hermitian , th e extrema l vecto r i s an
eigenvector, i.e.
VZR
Ar — A s
for som e purel y imaginar y v G C. Thi s show s tha t th e last tw o sums i n L cance l
out. Therefor e fo r the extremal vecto r w e have
ZsZt
I+ 2TT .
s r^s s^zt

Applying 2\z szt\ ^ \z s\2 + \z t\2, w e obtain

Since w e have |A r — A s| ^ S\r — s\ the innermost su m is bounded b y 25~ 2£(2) =

7T2/352. Thi s give s (7.24 ) completin g th e proof o f (7.23). •

COROLLARY 7.9 . For any set of 5-spaced points a r G R/Z and any complex
numbers z r we have

(v 7J-25) \t-j — T r^r -s;—a Ts) IN^ s~x E Z I .

2
E Esin7r(a
^\
r^s r

W e apply (7.23 ) to the doubly-indexed se t o f numbers z^^ = (— l) m z r

PROOF.
and Xm^r = m + a r wit h 1 < m ^ K , gettin g

17 TTK 2
E E (-i)™" — k^Ew -
(r,m)^(s,n) r

Here w e can replace th e summation conditio n (r , m) ^ (s , n) b y r ^ 5 because for

r = s th e remaining term s cance l ou t pairwis e (ra , n) agains t (n,ra) . I f we put
k = m — n and divide b y if w e derive

r^s k= — Kr
 7. B I L I N E A R F O R M S AN D T H
1 E L A R GE S I E V E 7 7

This yield s (7.25 ) b y letting K —> • +0 0 becaus e fo r a £ Z

fc
y (-i) = *
^ k + a si n na
kez
where th e series is summed symmetrically . •

COROLLARY 7.1 0 . For any real x we have

z z 5 |2
(7.26) \Z^l^ r s—. ; r\^ 1 ^
\*-^ £—j sm7r(a r — a s) I ^-^
r^s \ r sj r
PROOF. Thi s follows by applying Corollar y 7. 9 twice with z r twiste d b y e(xa r)
and e(—xa r). •

Now w e are ready t o prove (7.22) , firs t wit h S~ x + N instea d o f £_ 1 + N — 1 .

By duality , i t is equivalent t o show tha t
z
(7.27) V J \Y, r<nar)\2 ^(N + S-^Wzf
M<n^M+N r

for an y comple x number s z r. Squarin g out , w e see that th e diagonal ter m r = s

contributes iV||2|| 2 , whil e th e off-diagonal term s yiel d
v—v v-^ \~ ^/ / w v ^ v^ /T^/ \\ sin7r7V(ar — a s)
}^^zrzs}^e(n(ar-as)) = }^}^zrzse(K(ar-as)) ^ _

with K — M + \(N + 1). Thi s quantit y i s smaller tha n th e left sid e of (7.26), henc e
is < <5 -1 ||2:||2. Addin g bot h contribution s give s (7.27) .
To save the —1 , w e apply Theore m 7. 7 to the points (a r + k)K~1 wit h i ^ fc ^
K, fo r the trigonometric polynomia l T(a) = S(aK). W e obtai n

Kj2\S(ar)\2 = ^Y,\T(^)\2^(d~lK + NK K +
- VM2
rk r
because th e points (a r + k)K~ l ar e spaced b y 5K_1, an d the sum T(a) range s
over m = nK wit h (MK + K - 1 ) < m < (MK + K - 1 ) + (NK + K - 1 ) . Henc e
dividing b y K an d lettin g K ten d t o infinity w e obtain (7.22 ) a s stated (thi s tric k
is due to Paul Cohen) .
For man y applications , a weaker estimat e wit h th e factor <5 _1 -f N — 1 replaced
by C(5~ 1 + N) wit h a n absolute constan t C is sufficient. Thi s kin d o f large siev e
inequality wa s firs t state d b y H. Davenport an d H. Halberstam [DH1 ] . W e ar e
going to give a quick proo f o f one such estimate . B y duality w e need t o show tha t

J2 | ^ 7 r e ( a r n ) | 2 ^ C ( J -1 + iV)||7|| 2
\n\<N r

for an y comple x number s 7 r . Le t f(x) b e a non-negative functio n wit h f(x) ^ 1

if \x\ ^ 1 . Th e left sid e abov e i s estimated b y the same expressio n wit h extr a
smoothing facto r f(n/N), an d the latter equal s

X^5Z 7r7sCr ( r ' 5 )

178 7. BILINEA R FORM S AN D TH E LARG E SIEV E

where

n
By Poisson' s formul a

(T(r,s) = NY,f((ar-aa + h)N).

h

We choos e f(x) suc h tha t it s Fourie r transfor m satisfie s f(y) < C (1 -f- y 2 ) - 1 givin g
<r(r,s) < N(l + ||a r - a s\\2N2)-\ Finall y applyin g 2 | 7 r 7 s | ^ | 7 r | 2 + | 7 s | 2 an d
oo
1
^2 W(r, s)\ < N ^ i + (SkN) 2)-1 < (T 1 +TV
s k=0
we complete th e proof .
The abov e argument s woul d yiel d th e bes t possibl e constan t C — 1 if th e tes t
function f(x) wa s chose n appropriately . Selber g di d just that , hi s optimal functio n
being quit e complicated .
Now specializ e th e larg e siev e inequalit y (7.4 ) b y takin g fo r th e point s a r th e
rationals a/q wit h 1 ^ q ^ Q an d (a , q) — 1. Thes e point s ar e space d b y 5 = Q - 2 ,
indeed i f a/q •£ a' /qf, the n
\\a/q - a'/q'\\ = ||(ag ' - a W w ' l l > ( w ' ) " 1 £ Q~ 2-
Therefore Theore m 7. 7 yield s
T H E O R E M 7.1 1 . For any complex numbers a n with M < n ^ M -\- N, where
N is a positive integer, we have

( 7 - 28 ) E T jE a ne(^)fW + iV-l)||a||2.
q^Q a (mod q) M<n^M+N ^

Notice tha t i f A = (a n) i s supporte d o n a n arithmeti c progressio n n = £

(mod /c) , and (/c , q) = 1 , we ca n chang e variable s t o deriv e
COROLLARY 7.1 2 . For any complex numbers a n with M<n^M + Nwe
have
2
(7-29) E E *| E a ne(^)\\(Q + k^N) J2 M 2
-
Q^Q a (mo d a ) n=£ (mo d q) n=l (mo d k)
(k,q) = l

EXERCISE 2 . Prov e th e followin g larg e siev e inequalit y i n several variables : le t

d ^ 1 and S > 0 an d le t a r = ( a r , i , . . . , ar ,d) b e 5-space d point s i n R d/Zd, i.e . i f
r ^ s , w e hav e ma x \\a r,i — a s,i\\ ^ S. The n

(7.30) ^ |^> n e ( n • a r ) | 2 < ((T d

+ N d)\\a\\2
rn

where x • y i s th e standar d scala r produc t i n R d , an d a n ar e arbitrar y comple x

numbers fo r n = (rci,.. . , n^) wit h 1 ^ n ^ ^ AT , the implie d constan t dependin g
only o n th e dimensio n d .
 7. BILINEA R FORM S AND TH E LARG E SIEV E 179

7.5. Multiplicativ e larg e siev e inequality .

Instead o f additiv e characters , w e now tak e a s harmonic s th e Dirichle t charac -
ters. I f on e consider s X — {x \ X * s a characte r modul o g} , then b y expandin g th e
square an d applyin g th e orthogonalit y relatio n on e finds immediatel y

2
£ \^2a nX(n)\\(q + N)\\a\\
X (mo d q) n^N

which i s comparabl e t o (7.20) .

Much mor e significant , however , i s th e cas e wher e th e harmonic s ar e al l prim -
itive Dirichle t character s t o modulu s q ^ Q. Her e w e hav e th e basi c resul t o f
Bombieri an d Davenpor t [BD ] (thei r origina l resul t i s slightl y weaker) :
THEOREM 7.1 3 . For any complex numbers a n with M < n ^ M + N, where
N is a positive integer, we have

2
(7.31) E T ]^ I £ a nX(n)\\(Q + N-l)\\af.
q
q^Q ^ ' x (mo d q) M<n^M+N

This i s again a s strong a s (7.20) . Not e tha t withou t th e restrictio n t o primitiv e

X, this inequalit y woul d b e fals e b y a larg e facto r (tak e a n = 1 for al l n , the n th e
trivial character s modul o q ^ Q contribut e alread y abou t N 2Q). A n interestin g
feature o f (7.31 ) i s tha t th e boun d depend s onl y o n th e lengt h o f th e interval , bu t
not o n it s position . I n thi s aspec t th e larg e siev e inequality fo r Dirichle t character s
has som e advantag e ove r th e GRH .
P R O O F . W e ca n i n fac t giv e a slightl y mor e precis e estimate , whic h i s some -
times usefu l i n applications .
The primitiv e multiplicativ e character s x (mo d q) can b e expande d int o addi -
tive character s b y mean s o f Gaus s sum s (se e (3.1 2) )

a (mo d q)

for an y n. I f x i s not primitiv e bu t induce d b y a primitive Xs modulo s with q — rs,

then i f (r , s) = 1 , we hav e

. , /an\ v ^ * v^ */ ,x ( n^ cn \
E ^ a
H j ) =
^ £ X(bs + cr)e[ v + - )
a (mo d q) b (mo d r ) c (mo d s)

= x(n)Xs(r)c r(n)r(xs)

where c r(n) i s a Ramanuja n su m (3.1 ) . W e deriv e

^2aTnx(n)cr(n) = ^—- ^ x{o)S{a/q)
n \Xs) a ( m o d q)

for an y comple x number s a n , wher e S(a) i s th e trigonometri c polynomia l (7.21 ) .

180 7. BILINEA R FORM S AN D TH E LARG E SIEV E

By orthogonalit y o f characters w e then hav e

rs^Q V ;
x (mo d s ) n ^ QX (mo d q ) a
(r,s) = l

= £E * |E-«(f) "
g^Q a mo d g n

Applying Theore m 7.1 1 , we conclude tha t

(7-32) E -77 ^ E * I E «nX(«K(n)| 2

^(Q2 + iV-l)||a||2.
rs^Q X (mod s ) M<n^M+A ^
(r,s) = l
This inequalit y contain s (7.31 ) b y positivit y (ignor e al l term s wit h r ^ l ) . •
REMARK. Bombier i an d Davenpor t di d succee d i n getting interestin g applica -
tions of (7.32) b y exploitin g th e extr a summatio n ove r r.
7.4. Application s o f the large-siev e t o sieving problems .
In thi s sectio n w e explain ho w the larg e siev e inequality fo r additiv e character s
implies a sieve resul t i n the "ordinary " sens e (se e als o Chapte r 6 for background
on sieves) , an d giv e Linnik' s firs t applicatio n o f such a result fo r estimation o f the
least quadrati c non-residue .
Consider a finite se t M of integers an d a finite se t V of prime numbers . For
each p e V, le t Qp C Z/pZ b e a set of residue classe s "t o siev e out" . Th e dat a
(Ai,V,Cl) define s a sieving problem; th e correspondin g sifted set is
(7.33) 5 ( M , P , l ] ) = { m e A 1; m (mod p) 4 0, fo r all p G V\.
The goa l i s to estimate S — \S(M,V,Cl)\, th e cardinality o f S(M,V,fi,). Mor e
generally w e consider , fo r an arbitrary sequenc e o f complex number s a = (a n ), th e
sum
a
(7.34) Z = Yl -'
nes(M.v.n)
and w e see k a bound fo r Z in terms o f the ^2-nor m o f a. Naturall y i n this contex t
the nam e "larg e sieve " i s appropriate , becaus e i n contras t wit h othe r method s (lik e
combinatorial sieves , se e Chapte r 6) , w e ar e looking fo r strong estimate s eve n if
\QP\ is rather larg e whe n p is large.
We deriv e no w fro m Theore m 7.1 1 a result somewha t stronge r tha n Linnik' s
original versio n (se e [Rot] , [Mol] , [Hul] , [Bo2]) .
T H E O R E M 7.1 4 . Suppose M is contained in an interval of length N ^ 1 , and
assume Q p ^ Z/pZ for every p £ V, i.e., uj(p) = \Vtp\ < p. Then we have

(7.35) | Z|2<^±«!|H|2

for any Q ^ \, where

(7.36) H = Y^Kq)
 7. B I L I N E A R F O R M S AN D T H E L A R G E SIEV E 181

and h(q) is the multiplicative function supported on squarefree integers with prime
divisors in V such that
u(p)
(7.37 ) h(p) =
p - uj(p)

In particular, taking for a n the characteristic function ofS(A4, V, Q), we have

N + Q2
(7.38 ) S^
H

Let S(a) b e th e trigonometri c polynomia l (7.21 ) , so S(0) — Z. Firs t w e estab -

lish th e followin g inequalit y fo r individua l moduli .
LEMMA 7.1 5 . For any positive squarefree number q we have
|2
S
(7.39 ) h(q)\S(0)\^ X T \ 0
a (mo d q)

P R O O F . Le t X(q, v), fo r v e Z/qZ, denot e

X(q,v)= ^ a n
n=v (mo d q)

Using additiv e character s w e hav e

av\„(a
*(«,") = ; £> e [ •-T^(i)'
-
a (mo d q)

so b y orthogonalit y (Planchere l formula ) i t follow s tha t

q £ \x( q,u)\*= Y: K^)f .

v (mo d q) a (mo d q)

If q — p i s prime, w e have X(p,v) = 0 for v G Hp , s o by Cauchy' s inequalit y w e ge t

\z\2 = |S(o)l2 = I J2 X
(P>")\
v (mo d p)

va (mo d p)

which yield s (7.39 ) b y subtractin g th e ter m fo r a = 0 (mo d p). No w i f q — q\q2

with (^1 ,^2 ) = I , w e hav e

a (mo d q) a\ (mo d q{) a 2 (mo d q 2)

Hence i f (7.39 ) hold s fo r q\ an d #2 ? we derive b y successiv e application s tha t

2
]T* | s ( ^ ) | % % 2) X) * \s(^)\ >h(qi)h(q2)\S(0)\2,
a (mo d q) a\ (mo d q\)

so the proo f o f (7.39 ) i s finished b y inductio n o n the numbe r o f prime factor s o f q.O
182 7. BILINEA R FORM S AND TH E LARG E SIEV E

P R O O F O F THEORE M 7.1 4 . Su m (7.39 ) ove r q ^ Q an d the n appl y Theore m

7.11. •

Linnik [Lil ] establishe d a n inequalit y somewha t weake r tha n (7.38) , ye t pow -

erful enoug h t o giv e a spectacular applicatio n t o th e proble m o f the leas t quadrati c
non-residue modul o a prim e p , i.e. , th e smalles t positiv e intege r q(p) suc h tha t
(q(p)/p) = — 1. Not e tha t q(p) i s prime . I t i s conjecture d tha t q(p) <C £ p£ wherea s
the bes t know n estimat e i s

q(p) < s / + £ , 0 = l/4V e = 0.1 51 6.. .

From th e Gran d Rieman n Hypothesi s fo r Dirichle t L-function s on e derive s tha t

q{p) <C (logp)2 wit h a n absolut e implie d constant .
THEOREM 7.1 6 (LINNIK) . Let e > 0. The number of primes p ^ N such that
q(p) > N £ is bounded by a constant depending only on e.

PROOF. Le t X £ b e th e numbe r o f prime s ^ yN wit h q(p) > N £, whic h w e

want t o sho w i s bounde d i n term s o f e. T o thi s en d conside r th e sievin g proble m
with

M = {1 ,2,...,7V} ,
£
V = {p ^ y/N | ( - ) = 1 for al l n < N },

ttp = {v (modp) | ( - ) = - 1 } .

Note tha t uj(p) = \{p— 1 ) and h{p) = (p — l)/(p + 1). Thi s i s indeed a "larg e sieve"
because h(p) ^ | .
The sifted se t contains the set, say Z £ , of all n ^ N whic h have no prime divisor s
larger tha n N £. Therefor e b y (7.8 ) wit h Q = y/N w e ge t Z £ = \Z £\ ^ 2NH~ l.
Combining thi s inequalit y wit h

\X£^ Y, h(p)^H
q{p)>Ne

we ge t X £Z£ ^ 67V . I t remain s t o estimat e Z £.

It i s know n tha t Z £ ~ 5(e)N fo r som e 5(e) > 0 a s N — > o c (se e e.g . [Bo2] ,
pp. 8-9) , s o the resul t follows . Alternatively , a sufficient lowe r boun d fo r Z £ follow s
2

by counting in the set Z £ number s o f special type n = mp\ • • • pk ^ N wit h N £~£ <
Pj < N £ fo r 1 ^ j ^ k = e~ x. W e ge t
_N
z£> V »iV.
^ Lpi---PkLz> i
p i , - ,Pk

Hence I £ < 1 . Finall y changin g i V t o N 2 an d e t o e/2 w e conclud e Linnik' s

theorem exactl y a s stated . •

See Proposition 7.3 0 for a n analogu e fo r ellipti c curv e usin g the larg e sieve typ e
inequality fo r symmetri c squar e L-functions .
 7. BILINEA R FORM S AN D TH E LARG E SIEV E 183

EXERCISE 3 . (1 ) Consider th e followin g siev e i n dimension d : fo r any prim e

peP, le t ft(p)c(Z/pZ) d b e a subset wit h uj(p) = \Q(p)\ < p. Le t M C [-N, N}d.
Put
S(M,V,Q) = {m = ( m i , . .. ,m d) e M \ fo r p e V, m (mod p) £ Qp}.
Show tha t
S = \S(M9V,fi)| < (7V d + Q 2d)H-x
where i f i s given b y (7.36) an d the implied constan t depend s o n d [Hint : Us e
(7.30)].
(2) Fi x n > 1 . Let DM be the se t of monic polynomial s / G Z[X] o f degree n
with coefficient s i n absolut e valu e ^ N. Fo r r = ( r i , . . . , r^) wit h ri + 2r2 H = n
let C r C Div b e the subse t o f polynomials / fo r which / (mo d p) doe s no t facto r
as a product o f r\ linea r factors , T2 quadratic factors , ... , for any prim e p. Sho w
that
|cr|<ivn-hogjv,
the implie d constan t dependin g onl y o n n.
(3) Deduc e tha t th e se t C C D^ o f polynomials / fo r which th e Galoi s grou p
of th e splittin g field of / i s not S n satisfie s

\C\ < TV"- 1 logTV

the implie d constan t dependin g onl y o n n . [Hint : Us e the fac t tha t i f / ha s Galoi s
group H whic h i s a proper subgrou p o f 5 n , th e union o f the conjugate s o f H is
distinct fro m S n, henc e ther e i s a "splittin g type " r as in (2) such tha t / G Cr]
This is due to Gallagher [Ga4] .
Another interestin g applicatio n o f two-dimensional larg e siev e i s the proo f b y
Duke [Du5 ] that almos t al l elliptic curves E/Q have their p-torsion field with Galoi s
group GL(2 , Z/pZ) for all prime s p.
7.6. Panoram a o f the larg e siev e inequalities .
There i s a great variet y o f estimates o f the larg e sieve type which ar e used thes e
days i n analytic numbe r theory . I n this sectio n w e selecte d a few importan t one s
(without proofs ) t o make an impression abou t th e scop e of the matter .
Some application s requir e larg e siev e inequalities i n which th e se t o f harmonic s
consists o f characters twiste d b y exponentials o f different kind . Th e first case of
such a hybrid larg e siev e is due to P.X. Gallaghe r [Gal] .
THEOREM 7.1 7 . Let N, Q, T ^ 1 . For any complex numbers a n we have

E E * / T |E a «^ n ) ni f df<< ^ 2T+iV )iHi 2 -

q ^ Q x (mo d g ) T n^N

A discret e versio n o f this result , amon g furthe r generalizations , ca n b e found

in [Mo2] .
Of course , twistin g th e multiplicative character s b y additiv e one s doe s not
produce stronge r results , bu t if the exponentia l componen t i s not linear, the n i t
is likel y t o be orthogonal t o the characters an d one ca n receive extr a savin g (in
Theorem 7.1 7 th e exponentia l componen t i s e(^plogn)). Her e i s another exampl e
of som e importanc e fo r applications (se e [DuI3]) .
184 7. BILINEA R FORM S AN D TH E LARG E SIEV E

THEOREM 7.1 8 . Let v(x) be a real, smooth function on M + such that 0 <
x\v'(x)\ < 1 and \v'(x)\ 2 < \v"(x)\. Then for X ^ Q ^ 1 we have

E E * |Ea^(n)e(x )| 2 < < ( i V + 0 x l o g X )1

' ^ l|a| 2
-
q^Q x (mo d q) n^N
where the implied constant is absolute.
If th e linea r form s Y2 anX(n) a r e lacunar y (man y coefficient s a n vanish) , the n
the powe r o f th e larg e siev e reduce s significantly . Ther e i s a grea t deman d fo r
results whic h woul d b e sensitiv e i n thi s respec t (cf . P . Elliot t [El]) . O . Ramar e
worked ou t man y interestin g estimate s o f suc h kin d (unpublished) .
PROBLEM 7.1 9 . Prove that for any complex numbers a n supported on primes
|2
E E * |E a «*( n )| «(Q 2 + Ar(log^V)-1)IH 2
q^Q X (mo d q) n^N

where the implied constant is absolute.

A mor e genera l proble m i s t o fi x / G Z[X] an d establis h goo d larg e siev e
inequalities fo r linea r form s

J 2 a n e ^ ) , J2a nX(f(n)).
< a
H
2<N n<N

One shoul d realiz e how fortunate i s the cas e o f large sieve for coefficient s o f ful l
density, becaus e th e dualit y principl e work s efficiently. Fo r sequence s o f coefficient s
an whic h ar e ver y sparse , th e dualit y argument s ar e poo r an d ver y littl e ca n b e
done directl y whil e prediction s fo r th e bes t possibl e estimate s ar e risky .
In th e sam e not e i t i s a challengin g proble m t o establis h th e bes t larg e siev e
inequality fo r harmonic s whic h ar e primitiv e character s o f a fixe d order . Her e w e
have a powerfu l resul t o f D . R . Heath-Brow n [HB1 ] fo r quadrati c character s (se e
also th e earlie r resul t o f M . Jutil a [Ju2]) .
THEOREM 7.20 . For any complex numbers a n we have

Eb |E ba
™ Q f «( M7V £ M+iV
)( )Ni
Here the superscript b restricts the summation to squarefree numbers, ( —) stands
for the Jacobi symbol, e is any positive number and the implied constant depends
only on e.
R E M A R K S . Th e fac t tha t w e ca n vie w ( ^ ) a s eithe r a characte r i n n o f mod -
ulus m , o r a characte r i n m o f modulu s 4 n (b y wa y o f th e quadrati c reciprocit y
law) i s th e ke y featur e o f th e proof . Not e tha t th e proble m i s self-dual , i.e . inter -
changing m an d n doe s no t matter . Fo r interestin g application s o f Theore m 7.20 ,
see [HB1 ] , [Sou] , [IM] , [PP] .

In connectio n wit h Kloosterma n sum s on e meet s character s ( ^ p ) wher e / i s

a quadrati c polynomial . Th e correspondin g larg e siev e wa s considere d i n [II ] wit h
the followin g resul t
 7. BILINEA R FORM S AND TH E LARG E SIEV E 185

THEOREM 7.21 . For any complex numbers a n we have

I / n 2_ 4 \| 23 2
an £ |2
X \J2 {—^-)\
m
^(MN) (M2 +M3N)\\a
where e is any positive number and the implied constant depends only on e.
EXERCISE 4 . Le t S(ra , n;c) denot e th e classica l Kloosterma n sum . Fo r an y
complex number s a m , /3 n we hav e

E | X Y^PnSim.^c^iC^M + N^alMl

EXERCISE 5 . Fo r an y comple x number s a m , (3 n we hav e

2
E E * | EE ^ „ e ( = ) | % ( Q + MAr)||a||2||/3||2.
q^Q a (mod q) m^M,n^.N
(mn,q) = l

The estimate s i n th e abov e exercise s ca n b e easil y derive d fro m th e larg e siev e

inequalities (7.1 1 ) , (7.1 3 ) respectively . Muc h harde r i s th e followin g estimat e fo r
bilinear form s wit h Kloosterma n fraction s du e t o W . Duke , J . Friedlande r an d H .
Iwaniec [DFI4] .
THEOREM 7.22 . Let a be a positive integer. For any complex numbers am, (3 n
we have

E E <^0ne(a^)<z(MNy(± + ±) m(a + MN)ka\\

(m,n) = l

where the implied constant depends only on e.

REMARKS. Withou t th e facto r ( M _ 1 + i V - 1 ) 1 / 5 8 , thi s boun d woul d b e jus t
trivial. I t i s ofte n crucia l fo r application s t o hav e a savin g factor , les s importan t
how larg e i t is . Th e proo f o f Theore m 7.2 2 use s th e amplificatio n method ; th e
principle o f thi s i s explaine d a t th e en d o f Chapte r 26 .

Bilinear form s i n comple x domain s ar e als o desire d fo r applications . W e selec t

from [FI1 ] tw o result s fo r th e Jacobi-Dirichle t symbo l (^ ) an d fo r th e Jacobi -
Kubota symbo l [wz] (for definitions , se e Sectio n 3.7) .
T H E O R E M 7.23 . Let a w, (3 Z be any complex numbers with \a w\ ^ 1 , \fi z\ ^ 1
for w, z in the discs \w\ 2 < M, \z\ 2 ^ N. Then

X * X a - ^ ( ^ ) <:(M + N)T2(MNo-
where the * restricts the summation to the primary primitive numbers w, e is any
positive number and the implied constant depends only on e.
Using th e twiste d multiplicativit y (3.71 ) o f th e Jacobi-Kubot a symbo l on e ca n
derive from Theore m 7.2 3 the following estimat e (subjec t t o some minor conditions )

X* X*awf3z [^ < c^+^o ^ (MN) 7o+

^£
186 7. B I L I N E A R F O R M S AN D T H E L A R G E SIEV E

With simila r result s bu t i n a differen t fashio n on e ca n trea t bilinea r form s i n

ijjf(mn) wher e ij)f{() ar e th e normalize d Fourie r coefficient s o f a fixed cus p for m
/ G 5/c(To(AO, ^) o f weight k = hal f a n od d integer , an d v i s the compatibl e thet a
multiplier (se e (1 .53 ) an d (3.42)) . Thes e coefficient s ar e no t multiplicativ e s o a
cancellation i n general bilinear form s i s possible. Th e followin g estimat e wa s proved
in [Dull] ,

b
Y! E <*™Mf{rnn) « (MNY{M\ + MIA0||a||||/3| |

where a m , /3 n are an y comple x number s an d £ is an y positiv e number , th e implie d

constant dependin g onl y o n e an d th e cus p for m / . Probabl y thi s genera l estimat e
holds tru e wit h M 1 / 2 + M 1 /4 7V replace d b y M + N. Hav e i n min d tha t th e boun d
i/jf(l) < C I s i s expected t o be true fo r £ squarefree bu t i t i s not ye t proved ; i t i s essen-
tially equivalent wit h the Lindelof Hypothesi s for a suitable automorphi c L- function
twisted b y th e rea l characte r \i m th e ^-aspec t (vi a Shimur a correspondenc e an d
Waldspurger formula) .
7.7. Large-siev e inequalitie s fo r cus p forms .
The Fourie r coefficient s o f cusp forms, o r better th e eigenvalue s o f Hecke opera -
tors i n the spac e o f cusp forms , ar e analogue s o f Dirichlet characters . A s tools the y
are powerfu l du e t o th e orthogonalit y i n th e sens e o f th e larg e siev e inequalities .
There ar e tw o essentia l aspect s o f th e matter , th e spectra l an d th e leve l aspects .
Moreover, som e hybri d aspect s als o appea r i n practice .
Let (UJ) be a n orthonorma l basi s o f Maass cus p form s fo r T = To(q) wit h q ^ 1
(see Theore m 1 5.5 ) an d le t Xj = Sj(l — Sj) wit h Sj = \ + itj b e th e correspondin g
eigenvalues o f th e Laplac e operator . Le t

Uj(z) = ^/yYl pj(n)K itj(27r\n\y)e(nx)

be th e Fourie r expansio n o f Uj a t th e cus p o o (se e Lemm a 1 5.1 ) . Conside r th e

normalized coefficient s

2
(7.40) Vj(n)= ( k ) Pj(n), fo rn ^ 0 ,t 3 > 0.

These ar e expecte d t o b e essentiall y bounde d o n averag e wit h respec t t o n , q, tj.

Using th e Kuznetso v formul a (Theore m 1 6.3 ) on e ca n sho w
THEOREM 7.24 . Let q ^ 1 , T ^ 1 and N ^ 1 . For any complex numbers a n
we have

(7-41) E | E <W n
) f « (<1 T2
+ NlogN)\\a\\ 2

where the implied constant is absolute.

Many result s o f thi s typ e ar e establishe d i n [DI] , bu t wit h a slightl y weake r
term iV 1 +£ i n plac e o fT V logTV (probabl yT V would suffice) . Not e tha t qT 2 i s
approximately th e numbe r o f cus p form s Uj(z) fo r r 0 (g) wit h 0 < t 3 ^ T b y
Weyl's Law .
 7. BILINEA R F O R M S A N D T H E L A R G E SIEV E 187

Still more orthogonalit y i s expected fro m th e primitiv e cus p forms . Th e follow -

ing estimat e i s a n ope n proble m
PROBLEM 7.25 . Let Q ^ 1 , T ^ 1 and N ^ 1 . Prove that for any complex
numbers a n

2
(7.42) E E * | E «n^-(n)| « (Q 2T2 + N)\\af

where the implied constant is absolute and ^ restricts Uj to the primitive cusp
forms, so T nUj — Xj(n)uj and Vj{n) = \j{n)h>j(l) for all n ^ 1 .
Similar inequalitie s hol d tru e fo r holomorphi c cus p forms . I n thi s cas e mor e
precise result s ca n b e derive d fro m th e Petersso n formul a (Propositio n 1 4.5) . Le t
T b e a n orthonorma l basi s o f Sk(q). Le t

f(z) = J2 af(n)e( nz)

be the Fourie r expansio n o f / a t oo . W e consider th e normalize d Fourie r coefficient s

(7-43) ^ ( n ) = (^zJl)3 a / ( n ) .

Recall tha t i f k ^ 2 we hav e

1
(7.44) \T\ = dim S k{q) x kq J] ( 1 +P' ).
P\Q

THEOREM 7,26 = Let J be any orthonormal basis of Sk{q) with k > 2 . Then
for any complex numbers a n we have

(7.45) ^ | V J a n ^ ( n ) | 2 « ( 9 + iV)||a|| 2

where the implied constant is absolute.

PROOF. B y (7.43 ) an d Propositio n 1 4.5 , th e Petersso n formul a implie s tha t
the lef t sid e o f (7.45) , sa y L(q), equal s

'AiT\/mn
g||a|| 2 + 27rg i k ^1
c .ama^^a rnanS(m,n]c)Jk-i(-
nb\m,n;c)Jk-i[
c = 0 (mo d q) m,n^N

Opening th e Kloosterma n su m w e deriv e b y orthogonalit y o f additiv e character s

and b y Cauchy' s inequalit y tha t

| ^ ^ a m a n S ( r a , n ; c ) | ^ (c + N)\\a\\ 2.

Next (t o separat e th e variables ) w e derive vi a th e powe r serie s expansio n

(-l)exk-1+2e
J f c _ i ( 2 z ) = E OT
OO (fc + £
)
188 7. BILINEA R FORM S AN D TH E LARG E SIEV E

that

\^2^arnanS{m,n]c)Jk.ly—y- J J ^ I k-i{— J (c + N)\\a\\ 2

m,n^.N

where
k — l-\-2£

The conditio n x — 27Ty/mnc~ 1 ^ 2TTNC~ 1

^ 1 holds fo r any c = 0 (mod q) if
g ^ 2TTN, givin g
1
L{q)^q\\a\\2 + 2nq £ c" (^f)\c + N)\\ a l l 2 ,
c=0 (mod q)

hence (7.45 ) follow s i n this case .

To remov e th e condition q ^ 2TTN w e observ e tha t Sk(q) C Sk(pq) an d the
index i s [ro(p# ) : To(g) ] ^ p + 1 . Usin g thi s embeddin g wit h appropriat e re-
normalization w e find tha t L(q) ^ ( 1 + p~1)L{pq). I f q ^ 27r7V , w e ca n choos e p
with 27ri V ^ pq ^ ATTN an d appl y th e resul t fo r L(pq) gettin g (7.45) . •

With a sligh t modificatio n th e above argument s wor k als o fo r S k(q,x) w r t n

any characte r \ (mo d q) an d fo r k = 2 . However , th e case fc = 1 is very differ -
ent becaus e th e spac e Si(q,x) * s small (probabl y dim5i(g,x ) < ^ v ^ ° S t f i f X *s a
real character) . Technicall y speakin g thi s intrinsi c distinctio n manifest s itsel f by
the lac k o f convergence o f the serie s o f Kloosterman sum s i n the Petersson for -
mula. Wha t lie s behin d th e scen e i s the hug e spectru m o f Maass cus p form s nea r
the botto m o f which onl y a small par t (th e holomorphi c form s a t the bottom) is
detected b y the Petersson formul a (th e lac k o f spectral completenes s cause s the
problem!). I n this scenari o on e ca n procee d b y the duality principl e (n o spectra l
completeness i s required then) , bu t a new obstacl e emerge s wit h th e degre e o f the
automorphic harmonic s (GL(2) x GL(2) i s essentially GL(4) ; t o the contrar y for
Dirichlet character s GL(1 ) x GL(1 ) i s still GL(1 )) whic h amplie s th e conductor.
Nevertheless som e non-trivia l orthogonalit y ca n be established fo r sufficiently lon g
vectors relativ e t o th e level . Fo r exampl e W . Duk e succeeded i n proving alon g suc h
lines th e followin g larg e siev e typ e inequalit y (no t perfec t bu t quit e useful )
THEOREM 7.27 . Let x he an odd primitive quadratic character modulo q. Let
T he an orthonormal basis of Si(q,x) an ^ ^ ( n ) ^ e ^e corresponding normalized
coefficients. Then for any complex numbers a n we have
a
(7.46) £ | E " ^ ) | 2 « ( 9 + A0IHI2

where the implied constant is absolute.

Duke [Dul ] gav e an important applicatio n o f (7.46) fo r estimatin g th e dimen -
sion of Si(#, x)- Specificall y h e proved

dimSi(g,x)«411/12+£
for q prime an d x a r e a l primitiv e characte r modul o g , the first improvement ove r
the boun d dimSi(q,x) ^ tf/l°g# whic h come s fro m th e trac e formula .
 7. B I L I N E A R F O R M S A N D T H E L A R G E SIEV E 189

Using th e existenc e o f Rankin-Selber g L- functions fo r arbitrar y degre e auto -

morphic forms , Duk e an d Kowalsk i [DK ] hav e establishe d b y th e dualit y principl e
a ver y genera l typ e o f larg e siev e inequalit y whic h ha s application s especiall y t o
the stud y o f automorphi c //-function s clos e t o th e lin e R e (s) = 1 . W e mention th e
special cas e o f symmetric squar e coefficient s (se e Sectio n 5.1 2 ) an d giv e a sketc h o f
proof t o illustrat e th e wealt h o f ingredients .
THEOREM 7.28 . For q squarefree, let S2(q)* be the basis of primitive forms of
level q and weight 2 . Let Xf(n) be the Hecke eigenvalues for f G S^^)*- For any
complex numbers a n we have

(7-47) E E lE^/^r^^ 1 0
^)1 ^^^^)!!^ 2
q^Qfes2(q)* n<:N

if N ^ Q ^ 1 , with an implied constant depending only on e.

SKETCH O F PROOF . On e use s dualit y an d on e ca n attac h a smoot h tes t func -
tion t o th e su m ove r n , reducin g th e proble m t o estimatin g

where / , g G £ 2 (<?)* > f°r som e fixed smoot h functio n ip with compac t suppor t suc h
that <p(x) = 1 for 0 ^ x ^ 1 . (Th e Heck e eigenvalue s Xf(n) fo r / G Sl(q) ar e real) .
By Melli n inversio n w e hav e

H
(f^) = ^~ [ G(s)0(s)Nsds
2m
(2)

where
G(s) = 5 > / ( n 2 ) A 9 ( n 2 ) n - .
n
2
Using th e fac t tha t A/(n ) are , u p t o smal l perturbation , coefficient s o f th e sym -
metric squar e L- function (se e Sectio n 5.1 2 ) an d th e descriptio n o f Rankin-Selber g
L-functions o f cus p form s o n GL(3) , i t follow s tha t

G(s) = L(Sym 2 / 0 Sym 2 g, 5 )G?i(5)

where Gi(s ) i s given by a n Eule r produc t absolutel y convergen t fo r R e (s) > \. W e

have Deligne' s boun d |A/(n) | ^ r(n) (se e (1 4.54)) , whic h implie s

G1(5)«|C(<T+i)|^«(a-i)^

for R e (5) = a > | , fo r som e absolut e constan t A > 0 , wit h a n absolut e implie d
constant.
It i s know n (b y propertie s o f Rankin-Selber g convolutions ) tha t L(Sym 2 / ®
Sym2 g,s) i s entire excep t i f Sym 2 / = Sym 2 g, i n whic h cas e ther e i s a simpl e pol e
at s — 1 . Moreover , i n th e Appendi x t o [DK] , Ramakrishna n show s tha t thi s
exception i s equivalen t wit h / = g i n th e cas e o f squarefre e level s (i n th e genera l
190 7 . BILINEA R F O R M S AN D T H E L A R G E SIEV E

case, a quadrati c twis t ca n occur) . S o moving th e contou r t o th e lin e a = | + S for

6 > 0, w e ge t
2 2
#(/>$) = 2 ^ / G^LiSym /® Sym g,s)<p(s)N*ds

for / ^ g . Fo r < J = | - f J, w e use th e convexit y boun d (se e Exercise 3 of Chapte r 5 )

L(Sym 2 / ® Sym2 g, a + it ) < q(Sym 2
/ ® Sym2 g, a + 2t) 1 / 4 ~ < 5 / 2 + £
for thi s L-functio n o f degre e 9 , wher e q(-,s ) i s th e analyti c conducto r define d i n
(5.7). W e hav e
q(Sym2 / <g > Sym 2 g, a + tf) < Q 6 (|t| + l ) 9
by the bound (5.1 1 ) o f Bushnell an d Henniar t fo r the conductor o f a Rankin-Selber g
convolution, wit h a n absolut e implie d constant . B y the rapi d deca y o f <£(s) , we get
for f^g,
6
H(f,g) < N^ Q3/2-36+£5-A
for an y S > 0. W e tak e 5 = (logiV) - 1 s o
+£
H(f,g) « iV5 Q3/2
forT V ^ Q an d / ^ # .
If / = <7 , by Deligne' s boun d w e hav e |A/(n 2 )A^(n 2 )| < r(n 2)2 ^ r ( n ) 4 henc e
by direc t estimatio n
# ( / , / ) «iV(logTV) 1 5 .
2
Since ther e ar e < C Q form s o n th e lef t sid e o f (7.47) , i t follow s tha t fo rT V ^ Q we
have (7.47) . •

In thi s connectio n w e suggest th e followin g proble m a s a firs t cas e o f a higher -

degree larg e siev e inequality :
PROBLEM 7.29 . Prove that

J2 | ^ a n A / ( n 2 ) | 2 « ( g i V r ( g + iV)||a|| 2
feHk(q) n<:N

where Hk(q) is a Heche basis of eigenforms on Sk(q), with q squarefree, and the
implied constant depends only on e and k.
We ca n deduc e fro m Theore m 7.2 8 th e analogu e o f Linnik' s resul t (Theore m
7.16) fo r ellipti c curves .
PROPOSITION 7.30 , Let A > 0 and Q > 2 be fixed. For E/Q a semistable
elliptic curve of conductor ^ Q, let M(E) be the number of semistable elliptic
curves F/Q with conductor ^ Q such that ap(p) — Q>E(P) f or a ^ V ^ (l°gQ) A -
Then we have
M{E) < QP' A
where the implied constant depends only on A.
It i s conjectured tha t th e numbe r o f semistable ellipti c curve s of conductor ^ Q
is abou t Q 5 / 6 ; th e lowe r boun d i s eas y b y constructin g explici t families , bu t th e
best know n uppe r boun d i s Q 1 + £ fo r an y e > 0 (se e [DK]) . Therefor e ou r resul t i s
non-trivial fo r an y A > 1 1 .
 7. B I L I N E A R F O R M S AN D T H E L A R G E S I E V E 191

PROOF. Fo r a given weight 2 primitive form / o f conductor q < Q , q squarefree,

let M'(f) b e th e numbe r o f primitive modula r form s g, o f weight 2 with squarefre e
conductor ^ Q suc h tha t A/(p ) = \ g(p) fo r p ^ (logQ) A. B y modularit y o f
(semistable) ellipti c curve s ove r Q , w e hav e M(E) < M^/E) wher e JE i s th e
weight 2 form associate d t o E. Not e tha t A/(n ) = A^(n ) i f al l prim e divisor s p \ n
are < (logQ) A.
As i n Linnik' s proof , th e ide a i s t o fin d coefficient s a n supporte d o n prim e
numbers tha t mak e th e linea r for m

Lf=^T a nXf(n)

large, henc e L Q = Lf i s larg e an d repeate d fo r al l g counte d i n M f(f). O n th e

other hand , th e larg e siev e estimat e (7.47 ) wil l sho w tha t thi s canno t happe n ver y
often producin g a boun d fo r M f(f). Becaus e th e coefficient s \f(p) ca n b e quit e
small (i n contras t wit h Dirichle t characters) , a n = A/(n ) ma y no t work . Henc e
one i s tempte d t o us e Propositio n 1 4.22 , bu tT V ther e woul d b e to o smal l (limite d
to N ^ (\ogQ) A). Instead , notic e tha t b y th e Prim e Numbe r Theore m an d th e
formula
\f(p)2-\f(p2) = lfov(p,q(f)) = l,
one o f th e set s T\ o r T 2 give n b y
Ti = {p < (\ogQ) A I \\ f(p*)\ > i an d (p,q(f)) = 1 }

satisfies \T{\ ^> (log(5) A (^41 oglog(5) _1 wit h a n absolut e implie d constant . Assum e
it i s T2 (th e cas e o f T\ i s similar , bu t simpler) .
We hav e |A/(n) | ^ 2~^^ i f n i s squarefre e wit h al l it s prim e factor s p G T2.
Fix m ^ 1 (t o b e chose n later) . Le t a n = Xf(n) fo r suc h n wit h uj(n) = ra, an d
an — 0 otherwise. I t follow s tha t
2
£/=£w =EiA/(n)i2^2"2mc/
where U i s th e numbe r o f value s o f n ^ T V satisfying th e desire d condition . B y
Theorem 7.28 , w e hav e
1 5
M'{f)L) « iV(logA0 ||a|| 2 = iVaogTV) 1 5
^

for an y N ^ Q s. Henc e fo r N = Q s w e ge t

M'(/)«Q8(logQ)1522mC/-1.

Choosing m = [8(logQ)(AloglogQ)- 1 } w e ge t 2 2 m < Q e

an d

showing tha t
8/A
M'(f) « Q +£ « Q 9/A

for e smal l enough . •

192 7. B I L I N E A R F O R M S A N D T H E L A R G E SIEV E

7.8. Orthogonalit y o f elliptic curves .

A straightforwar d approac h t o the large siev e inequalitie s require s near-ortho -
gonality an d almost completenes s o f the outer harmonics . However , som e space s
contain smalle r subspace s whic h preserv e completenes s wit h respec t t o thei r own
structure. The n provin g a larg e siev e typ e inequalit y fo r the subspace become s a
harder ye t feasible proble m o f independent interest . A clear exampl e i s offered b y
the famil y o f elliptic curve s

(7.48) E : y 2 = x 3 + ax + b
with a, 5 G Z , K a ^ i , 1 ^ K 5 . Th e corresponding modula r form s f ab ar e o f
weight 2 and level q ^Q = 16(4A 3 + 27B 2). W e hav e abou t AB suc h curve s whil e
the tota l numbe r o f all primitive cus p form s o f weight tw o and level q ^ Q is about
Q2. Therefor e w e are dealing wit h a very smal l subse t indeed .
In thi s sectio n w e present a large siev e typ e inequalit y fo r the harmonics asso -
ciated wit h th e family o f elliptic curves . I f m is squarefree th e Hecke eigenvalu e for
the cus p for m associate d t o (7.48 ) i s given b y the character sum

(7.49) X ab{m) = fi(m) ^ (

m
x (mo d m )
(see Sectio n 1 4.4) . However , i t i s mor e interestin g t o conside r th e sum over th e
reduced classe s x (mo d m) , (x,m) = 1 . Put
' x3 + ax + b
(7.50) K b(m)=,(m) £ * ( ~ ^ ) -
x (mo d m )

Note tha t

W = E"M(^(7)-
d\m

THEOREM 7.31 . For any complex numbers a a, /3 & we have

where e is any positive number and the implied constant depends only on e.
PROOF. W e can assume that m is odd squarefree. B y Gauss sums (se e Theorem
3.2)
m 3
Kb(m) = M )"T y ]( — ) Y\ e m (2:(x + ax + b))
z (mo d m) x (mo d m)

e
= /x(m ) ~p= ^2 ™ (ax) Yl ( — )em(z2x3 + zb).
x (mo d m) z (mo d m )

Hence the inner su m in (7.51 ) equal s

a 2 3
//(m)-p= ^2 \^ aern(ax))(^2f3b ^2 e m(z x + zb)j
m
x (mo d m) <*> b z (mo d ra)
 7. B I L I N E A R F O R M S A N D T H E L A R G E S I E V E 193

By Cauchy' s inequalit y th e lef t sid e o f (7.51 ) i s bounde d b y (AB) 1 ^2 wher e

A=
H E * |E a«e™(ax)f < I H | 2 ( A + M 2 )
mx (mo d m) a

and
B
= E^m E * I E A E e m(^ + zb)
x (mo d m) b z (mo d m)

<E~E IE ^ E e ^ v + zft )
ra a : (mo d m ) fc> z (mo d m)

TO z\ = z\ (mo d m ) *> i ,& 2

2
<£73(m)T(m) ^ * |/? 6em(^)| « ||/J|| 2 (B + M 2)M*.
mz (mo d m )

This complete s th e proo f o f (7.51 ) . •

REMARKS. Pro m th e Hass e boun d A* 6(m) < m 1 /2-^?™) (se e Theore m 1 1 .25 )
it follow s tha t th e lef t sid e o f (7.51 ) i s < C \\a\\\\f3\\(AB) 1 /2M3/2log2M, whic hwe
regard a s a trivial bound , becaus e n o cancellation i s used. I n application s on e ofte n
wishes t o hav e a savin g slightl y mor e tha n M 1 / 2 . Ou r theore m guarantee s thi s
saving relative to the trivial bound i f AB i s larger than M 2. Choosin g A = X 1 / 3 an d
B = X 1 !2 w e have the elliptic curves with discriminant A a& = — 16(4a3 +276 2 ) < C X
and w e obtaine d th e desire d savin g i f M < X 5 / 1 2 . Thi s i s quit e a larg e rang e fo r
M, bu t no t ye t completel y satisfactory . Indeed , fo r application s t o L- functions on e
needs t o pas s th e barrie r M — X1/2. Th e followin g conjectur e woul d provid e th e
passage
CONJECTURE 7.32 . For any complex numbers (7 m ) supported on squarefree
numbers we have
I |2

EE 7rnKb{m)\ <(A + M)M J2 1 7

mT(m)

where the implied constant may depend on b slightly.

When b i s a squar e w e ar e havin g ellipti c curve s o f positiv e rank . Simila r
estimates ar e expecte d t o b e tru e fo r specia l familie s o f ellipti c curve s o f a give n
rank.

EXERCISE 6* . Le t m b e a positive , odd , squarefre e number . Prov e tha t fo r

any comple x number s a a , /?& ,

J 2^ 2 ^ a «&A a 5 (m) < r 4(m)(A + m) 2 (B + ra) 2 ||a

B^b^2B
194 7. B I L I N E A R F O R M S AN D T H E L A R G E SIEV E

7.9. P o w e r - m o m e n t s o f L-functions .
We hav e remarke d tha t a shar p larg e siev e typ e inequalit y (7.1 8) , (7.20 ) fo r
coefficients o f a famil y o f L-functions , ha s strengt h comparabl e t o th e Gran d Rie -
mann Hypothesi s o n average . S o it i s not surprisin g tha t th e larg e siev e can als o b e
employed t o deriv e estimate s fo r average s o f L-functions whic h ar e a s stron g a s th e
Lindelof Hypothesi s woul d giv e (se e Corollar y 5.20) . Ther e i s a vas t literatur e o f
relevant results . Fo r classica l L-function s w e recommen d th e boo k b y A . Ivi c [Iv] .
In thi s sectio n w e only gras p th e industr y b y showin g a fe w representativ e results .
THEOREM 7.33 . For T ^ 2 we have

2
(7.52) / |C(|+it)| dt<TlogT.

THEOREM 7.34 . We have

8
(7.53) J2 E * l^+^x)| « Q 2 ( * 2 + l)(logQ(|<|+2))17
g ^ Q x ( m o d q)

for any t G R , where the implied constant is absolute.

THEOREM 7.35 . Let k ^ 2 and J 7 be a Heche orthogonal basis of Sk(q)- We
have

(7.54) Y, l L(/» \ + ft )l4 « <*? + l)(loK^r(|*| + 2))17

fer
for any t E l , where the implied constant depends only on k.
REMARKS. I n al l cases , slightl y mor e precis e estimate s ar e know n (i n som e
cases asymptoti c formulas) . Se e [T2 ] fo r th e Rieman n zet a function , an d [KMV1 ]
for cus p forms . Als o (7.53 ) an d (7.54 ) ar e onl y comparabl e t o th e Lindelo f hy -
pothesis i n q- aspect. I n t-aspec t the y ar e a s goo d a s th e convexit y bound , extr a
averaging ove r t woul d improv e th e situation .
In Chapter 26 , as a by-product o f a deeper study of special values of L-functions ,
we prov e asymptoti c formula s fo r th e first an d secon d moment s o f derivative s
L ; (/, | ) wit h respec t t o th e od d weigh t 2 cusp form s (se e (26.35 ) an d (26.36)) . I n
fact, th e moder n method s use d i n the analyti c stud y o f special values of L-function s
(and thei r zeros ) rel y heavily o n various type s o f averaging, som e of which ar e quit e
sophisticated indee d (the y requir e orthogonalit y o f a capacity beyon d th e diagonal ,
so to speak) . Se e the introductio n t o Chapte r 2 6 for som e references , t o whic h ca n
be adde d [CI1 ] (amon g man y others) . Se e als o Sectio n 9.2 .
SKETCH O F PROO F O F THEOREM S 7.33 , 7.34 , 7.35 . Th e argument s ar e quit e
familiar s o w e onl y indicat e th e essentia l steps . Th e metho d start s wit h suitabl e
short approximation s o f L-function s b y Dirichle t polynomial s (se e Chapte r 5) , an d
then applie s a large sieve inequality t o th e resultin g linea r forms . Withou t compro -
mising estimate s on e ca n tak e a momen t o f orde r 2k a s lon g a s th e lengt h o f th e
approximation fo r th e fc-th power i s bounde d b y th e tota l numbe r o f "harmonics" .
For instance , i n Theorem 7.34 , ther e ar e abou t Q 2 character s involved , an d fo r eac h
one th e approximatio n t o L ( | + it, \) * s °f lengt h abou t (q(\t\ -\- l)) 1 ^ 2 , s o w e ca n
apply Theore m 7.1 3 fo r th e mea n squar e o f L( ^ - f z£,x) 4 , an d ge t a boun d fo r th e
eighth power-momen t whic h i s nearl y bes t possible .
 7. B I L I N E A R F O R M S AN D T H E L A R G E SIEV E 195

We sketc h Theore m 7.3 4 thi s way , leavin g th e othe r tw o theorem s a s exercise s

(use Theore m 7.1 7 wit h Q = 1 and Theore m 9. 1 i n th e firs t case , o r Theore m 7.2 6
in th e las t case) .
Let x b e primitiv e modul o g , 2 ^ q ^ Q. W e hav e
s
L(s,xY = ^x(ji)T A(n)n-

w e se e
for R e (s) > 1 . B y Theore m 5. 3 fo r L(s, \Y tha t L ( | - f z£ , x)4 i s equal t o

^X(n)r4(n) / n \ 4-4^(5-^ ) V - x H ^ w p \

where V(2/ ) — Vs(y) an d W(y ) = W s (y) ar e give n b y (5.1 3 ) an d (5.1 4 ) wit h th e

gamma facto r 7(5 ) = TT~ 2ST(^(S + ft)) 4 fo r n = | ( 1 — x( — l)) an d som e suitabl e
choice o f auxiliar y functio n G(u). Th e function s V(y) an d VF(y ) deca y rapidl y fo r
y ^ > t 2 + 1 (se e Propositio n 5.4 ) s o practicall y w e ar e lef t wit h sum s o f lengt h
TV x q 2(t2 + 1 ) . B y Theore m 7.1 3 w e ge t

£ E * |£^r«« !+»><'<*»> 17

Using thi s estimat e essentiall y fo r N x Q (t + 1 ) on e derive s (7.53) . •

This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/09

CHAPTER 8

EXPONENTIAL SUM S

8.1. Introduction .
As usua l w e denot e th e additiv e characte r o n R b y e(x) = e 27Tlx. Exponentia l
sums whic h w e consider i n thi s chapte r ar e o f typ e

(8.1) S f(N)= £ e(/(n) )

where / i s a smooth, rea l value d functio n o n th e interva l [1 , N] whic h i s calle d th e

amplitude functio n an dT V is a positive intege r calle d th e lengt h o f the sum . We
could begi n th e summatio n a t any poin t an d conside r
e
(8.2) S f(M,N)= Yl (/(n))
M<n^M+N

where M is an y intege r an d / i s a smooth functio n o n [M + 1 , M + N]. O f course ,

the latte r ca n b e transforme d t o th e forme r b y shiftin g th e argument ,

Sf(M,N)= V J e(/( n + M)).

We have th e trivia l boun d |5/(M , N)\ ^ N and ou r goa l wil l b e t o improv e o n thi s
bound a s muc h a s w e can . However , i n man y case s eve n a slight improvemen t o f
the trivia l boun d i s sufficient fo r applications .
Exponential sum s Sf(N) fo r suitable / ar e encountere d i n various area s of
analytic numbe r theory . Fo r example , the y ca n b e use d t o estimat e th e Rieman n
zeta functio n ((s) o n vertica l lines . Indee d b y th e simpl e approximatio n

s
/V 1_s
(8.3) <(*) = J2 n- + j-^+0(N-n

for s = aH- it with a ^ \ an d 1 ^ t ^ N the proble m reduce s t o estimatin g sum s

which ar e o f typ e (8.2 ) wit h f(x) — =£ logx. Anothe r exampl e i s the proble m o f
counting lattic e point s insid e a planar domain . Thi s reduce s t o estimatin g integra l
points unde r a curve y = g(x), wher e g is a smooth, positiv e decreasin g functio n
on [1 , TV]. Th e numbe r o f point s i n questio n equal s

E \9(n)]= E 9(n)~ E ^W ) - f •
l<n<N l<n<N l<n<N

197
198 8. EXPONENTIA L SUM S

On th e right sid e the first su m i s well approximated b y an integral (se e (1 .67) ) whil e
the secon d su m i s equa l t o

by th e truncate d Fourie r serie s (4.1 8 ) fo r ip(x). Henc e th e proble m reduce s t o

estimation o f sum s (8.1 ) wit h f(x) = hg(x). I n particular , i n th e Gaus s circl e
problem (lattic e point s i n th e circl e x 2 + y 2 = R 2) w e encounte r sum s Sf(R) wit h
/(x) = hy/R 2 — x 2, an d i n th e Dirichle t diviso r proble m (lattic e point s unde r th e
hyperbola xy = N) w e encounte r sum s Sf(N) wit h f(x) = hNx~ x.
In this chapte r w e present classica l method s fo r estimatin g genera l exponentia l
sums S f(M,N) du e t o H . Weyl , J . G . Va n de r Corpu t an d I . M . Vinogradov .
Our result s d o no t depen d o n particula r propertie s o f / bu t onl y o n estimate s fo r
derivatives. I t i s alway s understoo d tha t th e function s involve d ar e o f clas s C fe,
where k i s th e larges t orde r o f derivatio n appearin g i n th e statements .
Note tha t i f / i s wel l approximate d b y g, the n 5/(M , N) ca n b e estimate d b y
Sg(MJ N f) fo r som e N' ^ N. Precisely , i f / = g + /i , then b y partia l integratio n

(8.4) Sf(M,N) = S g(M,N)- / S g(M,x - M)de(h(x)).

JM

Hence w e ge t
f
(8.5) \S f(M,N)\ < C h max \S g{M,N )

where
rM+N
Ch = l + 2ir \ti(x)\dx.
JM
If h i s monotoni c an d bounde d (s o Ch i s bounded) , the n th e origina l su m Sf an d
the modifie d su m S g ar e practicall y equal . Sometime s i n application s w e find tha t
a smal l alteratio n o f / i s necessary t o mee t th e condition s o f ou r results , i n whic h
case (8.5 ) ca n b e used .
8.2. Weyl' s method .
The first application s o f exponential sum s i n numbe r theor y wer e given i n 1 91 6
by H. Weyl [Wl ] t o the problem of equidistribution o f sequences modulo one. I n th e
second pape r o f 1 92 1 Wey l [W2 ] develop s a genera l metho d whic h i s particularl y
good fo r Sf(N) wher e / i s a polynomia l

f{x) = ax k + px 1k
- ^'"eR[x]

with a > 0 . I n thi s cas e w e cal l Sf(N) a Wey l su m o f degre e k.

The Wey l su m fo r a linea r polynomia l f{x) = ax i s a su m ove r geometri c
progression
^ /,r x v ^/ x sin7rai V /a / 7 % r _ \
l<n<N
 8. E X P O N E N T I A L SUM S 199

Since | sin7ra| ^ 2||a|| , wher e ||a| | denote s th e distanc e o f a t o th e neares t integer ,

we obtain fo r a £ Z,

(8.6) I V e(an)\ ^ mi n (iV , — -

I ^— ' I \ 2 a

The Wey l su m fo r a quadratic polynomia l f(x) = ax2 + /3x is also calle d a

Gauss sum . I n the specia l cas e wher e a , b are integer s wit h (2a , N) = 1 we hav e
the exac t formul a (3.38 ) fro m whic h it follows b y completin g th e squar e tha t

MI £ 4°-^"' N.
l<n<N V
Though ther e i s no simpl e expressio n fo r a general Gaus s su m
2
Sf(N)= ] T e(an +/3n),

we ca n estimat e thi s quit e well . Firs t w e arrange \Sf(N)\ 2 |2a s follow s

\Sf(N)\2= J2 e(<*t2 + Pt) J2 e

(2ain)-
\£\<N l^n,n+£^N

Then applyin g (8.6 ) w e ge t

2 1
(8.8) \S f(N)\ ^N + Yl min(2JV,||2crf||- ).

Hence on e ca n deduc e tha t

1, 1
(8.9) \S f{N)\ < 2y/aN + —
— loi ogg r -
y/a a
if 0 < a ^ \, whic h restrictio n ca n alway s b e arranged. However , a more pre -
cise estimat e depend s o n th e diophantin e natur e o f the leadin g coefficien t a . By
Dirichlet's approximatio n theore m ther e exist s a rational approximatio n t o 2a o f
type
l
3.10) ' 2" a"- - " <"
2Nq
'ith (<
with (a,g ) = 1 and 1 < q ^ 2N. Henc e ||2a£| | ^ \\\a£/q\\ fo r an y 1 ^ £ < NJ ^
0(mod q) an d

2 ||2a£||-^2( ^ + l) £
1<*<JV V ^ 7
^(mo
^(mod d cq)
^ 0 ( m o d q) ^ 0 ( m o ^ 0 ( m o d <? )

Then estimatin g th e partia l su m of (8.8) ove r 1 < £ < N, £ = 0(mod q) triviall y b y

2N2q~1 w e arriv e at
\Sf{N)\2 ^N + 2N 2q~l + 4(7V + g) logg.
This togethe r wit h th e trivia l boun d |5/(iV) | ^ N implie s
20 0 8. EXPONENTIA L SUM S

THEOREM 8.1 . If f(x) = ax 2 + fix with 2a satisfying (8.1 0), then

(8.11) \S f(N)\ ^ 2Nq~^ + q* logq.

Theorem 8. 1 i s essentiall y bes t possible . Slightl y bette r result s hol d tru e fo r

almost al l coefficient s a,/ 3 (mo d 1 ) wit h respec t t o th e Lebesqu e measure . Indee d
we have the followin g estimat e fo r th e second , th e fourt h an d th e sixt h powe r mea n
values

2
(8.12) / Yl e(an +/3n) da = TV ,
J° l<n<N

2
(8.13) f \ ^2 e(an +pn)\4da<$:(Nlog2N)2,

2 6
(8.14) [ [ \ Y2 e(an + f3n)\ dadfi<^(Nlog2N)3.

PROOF. Th e firs t formul a i s th e Parseva l identity . Th e secon d integra l i s

bounded b y th e numbe r o f solution s t o n 2 - f n\ — n2 + n\ i n positiv e integer s
ni,ri2,ri3,n4 < TV , and th e latte r i s bounded b y

]T r r yt)2{i)<^{N\og2Nf
^ ^ v iug ^J V
^2N2

The thir d integra l i s equa l t o th e numbe r o f solution s t o th e syste m

J m + n 2 + n 3 = n 4 + n 5 + n 6,
\ n\ + n\ + n\ = n\ + n § + n §
in positiv e integer s 71 1 ,71 2,723,724,71 5,72 6 ^ AT . Puttin g k v — nv — n v+s an d ^ =
71^+71^+3 for 1 ^ 1 / ^ 3 we get the system /ci-f/C 2 + fc3 =0 an d kiii+k2i2 + k$is = 0 .
This reduce s t o on e equatio n k\(i\ — ^3 ) + ^2(^ 2 — ^3) = 0 an d k 3 i s determine d
uniquely b y &i,/c2 - I f &i(^ i — is) — 0> then £2(^ 2 — is) — 0, therefor e ther e ar e a t
most 1 6-/V 3 suc h solutions . I t remain s t o coun t th e solution s wit h k\(i\ — is) ^ 0 .
Given m wit h 1 ^ m ^ 4./V 2 an d £ 3 wit h 1 < is ^ 2i V ther e ar e a t mos t r 2(m)
solutions t o m = fci(£i — £3) = —^2(^ 2 — ^3) i n fci, /C2, ^1, ^2- Thu s th e tota l numbe r
of solution s i n questio n doe s no t excee d
4N 2
Yl r (m)<^(N\og2N)3.
2
Km^4iV

This complete s th e proo f o f (8.1 4) . •

Now w e estimate Weyl' s su m Sf(N) fo r a polynomia l o f an y degre e k ^ 2 . W e

begin b y

\Sf(N)\2= E J2e(f(m)-f(n))
0<ra, n^N

=E E e(/( £ + n)-/(n)).
|£|<AT 0<n^i V
0<^+n^n
 8. EXPONENTIA L SUM S 201

If / i s a polynomia l o f degree k, the n g(x) = f(£ -f x) — f(x) i s of degree k — 1

(for an y £ ^ 0) . B y repeated applicatio n o f the abov e "differencin g process " we
ultimately arriv e a t a Weyl su m of degree on e for which w e can us e th e non-trivia l
bound (8.6) . Precisel y w e deriv e by induction
PROPOSITION 8.2 . If f(x) = axk -\ with k ^ 1 , then

m i n
\Sf(N)\ < 2N{N~k X)"' "S ^ H afc!^ '''^-lU"1)} •
-N<eli...1ek.1<N

PROOF. Fo r k = 1 this boun d i s interpreted a s that i n (8.6) . Moreover , we

have alread y prove d thi s resul t fo r k = 2 in (8.8) . Suppos e th e result i s true for
k ^ 2 . Let f(x) = ax k+1 + . . . , the n g(x) = f(£ + x) - f{x) = a(k + \)£x k + . . .,
therefore
|S/(A0| 2 ^ E | E e(a(k + l)nk + --.)\
\£\<N 0<n^N
0<£+n^N

2l~k
k 1
^2N ^2{N~ J^-"Y
1 niin^Hf c + l)!^..-^-!^- }
\£\<N -iV<£i,...A-i<^ V
by th e inductio n hypothesis . Henc e b y Holder' s inequalit y w e ge t

|5 / (iV)| 2 ^4iV 2 {Ar- f e - 1 £ • • • £ ; m i n ^ M f c + lJWx.-.^ll-1)}21 * .

-JV<£i,...,*fc<N

This yield s th e resul t fo r /c + 1 . •

Next w e use Proposition 8. 2 to estimate sum s Sf(M,N) fo r / whic h ca n b e

well approximate d b y polynomials . Le t k ^ 2 . Suppos e

(8.15) ^ | / « 0 ( a ; )|<F

in M ^ x ^ M + N. The n w e have /(a ; + M) = p(x) + r(x), wher e

p(x) = f(M) + xf'(M) + ••• + ^^/(fc-^M )

and th e derivativ e o f r(x) satisfie s

\r'(x)\ = ^^\f( k)(y)\ < kx k l

~ M'kF

for som e M ^ y ^ M + N. Therefor e b y (8.5) w e obtai n

k
\Sf{M,N)\ ^ (l + 27rFN M-k)\Sp(Nf)\

for som e N f ^ N. Thi s estimat e i s reasonably goo d fo r short sum s S/(M , TV) , i.e.,
if A 7" i s much smalle r tha n M , becaus e w e ca n mak e th e facto r 1 + 2irFN kM~k t o
be bounded , b y choosing k whic h i s not very large . Applyin g Propositio n 8. 2 fo r
the Wey l su m S P(N') w e ge t
202 8. E X P O N E N T I A L SUM S

COROLLARY 8.3 . Let f(x) be a smooth function in the interval [M , M -\- N] in

which it satisfies (8.1 5). Then
k
\Sf(M, N)\ ^ 2N(1 + 27rFN M~k)V22~k

where
V = N'~ k E ' ' ' E min(7V , \\f^ k-l\M)h • • • 4-2H"1)-
-iV<^,...,4-2<^

We shall use the above results to estimate Sf(M, M f) fo r an y M' wit h 1 ^ M' <
M. T o thi s en d w e assum e tha t (8.1 5 ) hold s i n th e whol e interva l M ^ x < 2M.
We choos e 1 ^ N < M an d spli t 5/(M , M') int o shorte r sum s o f lengt hT V getting

\Sf(M,M')\^ E \Sf(M + jN,N)\+2N

where J = [M/N]. Henc e b y Corollar y 8. 3 an d Holder' s inequalit y w e obtai n

\Sf(M,M')\^2N(l + 2nFN kM-k) E vf~ k

+ 27V
o^j<J

^ 2M( 1 + 27rFN kM-k)W22~k + 27V

where Vj ar e th e correspondin g sum s associate d wit h th e point s M + jM an d W

is the mean-valu e o f thes e sum s

0<j<J
1
=J -iNi-k J- E - E min(7V,||/^- )(M + j i V K 1 . . . 4 - 2 i r 1 ) .
0 ^ j < J -N<l!...,£ k-2<N

r an
Gathering term s accordin g t o th e produc t 4 • • • 4 -2 — d fo r r = 0 estimatin g
trivially w e arriv e a t

W^ k2 k 1
N~ +2kN
1 k
- E °r J Y
~ E min(AT , ^ / ^ - ^ (M + jiV)!!- 1 )

where R = N k~2 an d c r denote s th e numbe r o f representations r = 4 • • • 4 -1 wit h

1 ^ J 4 , . . . , 4 -1 < -N" - Fo r eac h r w e conside r separatel y th e inne r su m

U 1
=-j E minfAU^-ir )
J
oc?<J
where y 3 = rf^^^M + jN). Not e tha t1 3/^- 1 ^ rk\M 1 ~kF = y , say . W e wan t th e
points t/ j t o b e wel l spaced , an d fo r thi s reaso n w e requir e a lowe r boun d fo r th e
derivative o f / o f orde r k throughou t th e whol e segmen t [M , 2M]. Precisel y fro m
now o n w e assum e tha t

I/W(l)l<f
(8.16) j 4
 8. E X P O N E N T I A L SUM S 20 3

with A ^ 1 for M ^ x ^ 2M . The n b y th e mean-valu e theore m w e deduc e tha t

\Vj' "Vol ^ &\f - J\ w n e r e A = rk\(2M)~ kNFA~1 . No w i t i s clear tha t

\u\^Y0^j<J
2F + l / _ , v - 2
^ — y - ( ^ + 5 Z T ^ - J ^ ( 2F + J- 1 )(7V + 2A- 1 log3M).
0<j<J

This yield s

W<m^(%^+%Z).
Finally
k 1k
/N \/FN A N AM ~ \22~k

where th e implie d constan t i s absolute . Thi s hold s fo r an yT V with 1 ^ N ^ M.

Suppose M k ^ F ^ 1 , so we may tak eT V = [MF _1 / fc ] givin g
T H E O R E M 8.4 . Let k ^ 2 . Suppose f satisfies (8.1 6) in the segment [M , 2M].
T/ien for I ^ M f ^ M we have

(8.17) S f(M,M') <^ A 4/2k(FM~k + F' 1

)4^2"M\og3M
where the implied constant is absolute.
REMARK. A t th e en d o f th e proo f o f (8.1 7 ) w e assume d tha t M k > F ^ 1 ,
however thi s conditio n i s no t necessar y becaus e otherwis e th e resul t i s trivial .
COROLLARY 8.5 . Suppose f(x) satisfies (8.1 6) for fc = 2 and k — 3 in the
segment [M , 2M]. TTie n for I ^ M' ^ M ^ F we have

(8.18) S f(M,M') < AF^Mi log3M .

where the implied constant is absolute.
PROOF. Thi s follows by taking the minimum o f the two bounds (8.1 7 ) fo r fc = 2
and fc = 3 . •

Note that (8.1 8 ) i s trivial if M ^ F s . However , fo r an y M whic h i s comparabl e

with F i n the logarithmi c scal e Theore m 8. 4 yield s a non-trivial boun d b y choosin g
k appropriately . Fo r example , choosin g fc = [ 2 log Fj lo g M] w e deriv e
COROLLARY 8.6 . Suppose f(x) is smooth in [M , 2M] and all its derivatives
satisfy

2 2 k
(8.19) A - V^|/<*'(i)|a F

with some A^\. Then for 1 < M' ^ M < F we have

(8.20) S f(M,M') < A*M l~4~" log3 M

where 7 = lo g Fj lo g M and the implied constant is absolute.
20 4 8. EXPONENTIA L SUM S

As application s o f th e abov e result s w e deriv e estimate s fo r th e Rieman n zet a

iction o n th e line s a — \ an d a — 1 . Fo r t ^ 3 we hav e

21) ((a + it)=J2 n ~a~U + 0{t- a).

2<t

Hence w e deriv e b y (8.1 8 ) th e followin g subconvexit y boun d

2
(8.22) C ( | + ^ ) < ^ ( l o g t ) .

Moreover, w e derive b y (8.20 ) th e followin g approximation :

(8.23) C( l + it) = Y, n ~1 ~U + 0(e~ VT rt

° )

with y = t 3 / l o g log f, whic h yield s b y trivia l estimatio n

8.3. Va n de r Corpu t method .

This metho d emerge d fro m tw o paper s o f 1 92 1 an d 1 92 2 b y J . G . va n de r

Corput [Corl] , [Cor2] . Briefl y speakin g th e novelt y consist s i n replacin g th e su m
by integral s an d estimatin g th e latter . Thi s i s execute d b y a n applicatio n o f th e
Poisson summatio n formul a followe d b y a n asymptoti c evaluatio n o f th e resultin g
Fourier integral s (b y using various techniques such as stationary phase) . On e arrive s
at a n approximat e equatio n fo r tw o sums o f different length ; thi s transformatio n i s
called th e jB-process . Th e amplitud e function s i n bot h sum s ma y chang e shap e bu t
not size , an d the y hav e comparabl e derivatives . Nex t thes e amplitud e function s
are reduce d b y a differencin g proces s whic h i s a refinemen t o f tha t use d i n Weyl' s
method; thi s routin e i s called th e yl-process . Th e final estimat e fo r th e exponentia l
sum 5/(M , N) i s obtained b y repeate d applicatio n o f th e tw o processes .

Now w e procee d t o th e first ste p o f th e va n de r Corpu t method , whic h i s a

truncated versio n o f th e Poisso n formul a

(8.25) ^ F ( n ) = ^ F ( m ) .
nm

PROPOSITION 8.7 . Let f(x) be a real function with f"{x) > 0 on the interval
[a, b}. We then have

(8.26) J2 e

a<n<b a
(/(n))= E /
— E<m<fi+E
r
e(f(x)-rnx)dx + 0(e- 1 +\og(P-a + 2))

an
where a,(3,e are any numbers with a ^ /'(& ) ^ /'(& ) ^ fi d0 < £ ^ 1 , the
implied constant being absolute.
 8. EXPONENTIA L SUM S 205

PROOF. Le t M stan d fo r th e interva l [a — e,f3 -f e]. Firs t w e giv e th e trivia l

bound

fb\ i P i
dx
/V ^ e(—mx)\dx^2(b — a-\-l) / Y ^ e(—mx)
a
^ meM ^ ° rneM
(8.27)
2
< 2( 6 - a + 1 ) / min(/3 - a + 2 , x_1)cfo < ( 6 - a + 1 ) log(/3 - a + 2) .

This proves (8.26 ) if O < 6 —a < 2 . Therefor e fro m no w on we assume that b — a > 2 .
We appl y th e Poisso n formul a t o th e tempere d su m

fb
^2 9{^)e{f(n)) = Y1 9{x)e(f(x) - mx)dx )
Ja
a<n<b m

where g(x) i s a function whic h planes the edges, precisely g{x) = min( x —a, 1, b—x).
This functio n i s introduce d temporaril y fo r convergenc e o f th e su m o f th e Fourie r
integrals an d wil l b e remove d later . Fo r m ^ M w e write b y partia l integratio n

2TT Z
fb f ( g( x b
)V
/ d(x)e(f(x) - mx)dx = - / I _ (x) ——m;J e(f(x) - mx)d
V X m /
Ja Ja J\ )
b
e(f(x) - mx) g{x)f"{x)
•e(f(x) — mx)dx
\Jb-l J a f'(x)-m Ja (/ ' (x) — my
= Tb(m) - T a(m) + Tab(m),

say. Fo r 7^>(ra ) w e hav e

rb
\Tab(m)\ < / f"(x)(f(x) - m)- 2dx = ( a - m ) " 1 - (/ 3 - m ) " 1 .

For T a(m) w e hav e

/•a+l
|Ta(m)| < / \f'(x) - ml^dx ^ \a - m]' 1 + |/ 3 - ml" 1 .

Integrating b y part s onc e mor e w e obtai n anothe r boun d

\Ta(m)\ < (/'( a + 1 ) - m ) - 2 + (/'(a ) - m ) " 2 + I f + d(f(x) - m) \ "- 2

= 2max{(/ / (a + 1 ) - m)~ 2 , (/'(a) - m ) " 2 } ^ 2( a - ra) -2

+ 2(/ 3 - ra)" 2
.

The sam e bound s hol d fo r Tb(m). Combinin g thes e bound s w e deriv e

\Ta(m)\ + |7i(m) | + |T a 6(m)| <

(/3 — m)(a — m)
for an y m £ M.. Henc e w e conclud e tha t

fb
] T 9(n)e(f(n))= ^ / g(x)e(f(x) - mx)dx + 0{e~ l + log(/ 3 - a + 2)) .
^^ ^Is ™a
a<n<b m£A4
KA J &
206 8. EXPONENTIA L SUM S

It remain s t o remov e th e weight s g(ri) an d g(x). Th e correctio n o n th e lef t sid e i s

0(1) an d o n th e righ t sid e i t i s 0(log(/ 3 - a + 2) ) b y (8.27 ) fo r b = a + 1 . Thi s
completes th e proo f o f (8.26) . •

EXERCISE 1 . Deriv e (8.26 ) fro m th e Euler-Maclaurin formul a (4.7 ) using (4.1 8 )

or directl y fro m (4.21 ) .

It i s easy t o generaliz e (8.26 ) fo r a weighte d su m a s follows :

(8.28) ] T g(n)e(f(n))= £ / g(x)e(f(x) - mx)dx

a
a<n<b a-£<m<P+e

+ 0 ( G ( s - 1 + l o g ( / 3 - a + 2)) )

where g i s an y smoot h functio n o n [a, b] and

(8.29) G=\g(b)\+ f \g'(y)\dy.

Ja
For th e proo f pu t

e V
S(y)= E (/W)-E f e(f(x)-mx)dx.
a
a<n^y a-s<m<(3+e

Thus S(y) < C e~ x 4 - log(/3 — a + 2 ) b y virtu e o f (8.26) . No w applyin g partia l

summation w e obtain (8.28 ) wit h th e erro r ter m

g(y)dS(y) = 5(b)g(b) - [ S(y)g'(y)dy

which i s bounded b y G(e~ l 4 - log(/3 — a -\- 2)) a s claimed .

In a specia l cas e (8.28 ) simplifie s t o
LEMMA 8.8 . Let f(x) be a real function with \f f(x)\ ^ I — 0 and f"{x) ^ 0 on
[a, b]. We then have

1
(8.30) E 9(n)e(f(n))= f g(x)e(f(x))dx + O^- )
a
a<n<b ^

where G is given by (8.29) and the implied constant is absolute.

Our nex t tas k i s t o evaluat e th e exponentia l integral s i n (8.26) . W e begi n b y

proving simpl e estimate s fo r

*.31) I f(a,b)= f e(f(x))da

Ja

where f(x) i s a real , smoot h functio n o n [a , b]. Notic e tha t th e trivia l boun d
|i/(a, 6)| ^ b — a cannot b e improve d substantiall y i f f^ < C (6 — a)~ k fo r al l k ^ 1 ,
because suc h a functio n i s more o r les s constant. However , i f one o f the derivative s
is somewha t larger , the n ou r resul t wil l b e non-trivial .
 8. E X P O N E N T I A L SUM S 20 7

LEMMA 8.9 . Suppose f'(x)f"(x) 7 ^ 0 on [a , 6]. Then

(8.32) \I f(a,b)\ ^\f{a)\-^\f\b)\-\

PROOF. W e ca n assum e withou t los s o f generalit y tha t f"(x) > 0 . The n b y

partial integratio n w e ge t

Hence

^ |//(a ' 6)K M + i?k + 7^-?k

which i s bette r tha n wha t i s claimed. •

LEMMA 8.1 0 . Suppose that for some k ^ 1 we have

(8.33) | / « ( x ) |^ A
for any x on [a , b] with A > 0 . Then
1
k /k
(8.34) | J / M ) I ^k2 A~ .

P R O O F . W e appl y inductio n wit h respec t t o k. Fo r k = 1 the resul t follow s

from (8.32) . Suppos e tha t f( k+1 ) > A > 0 on th e whol e interva l [a , b]. Thus f^ i s
increasing, s o it ma y hav e a t mos t on e zer o i n [a , 6], sa y c if it exists . I f f^ ^ 0 on
[a, 6], w e stil l defin e c by puttin g c — a o r c — b according t o whethe r f^ > 0 o r
f(k^ < 0. Le t 5 be a positive numbe r t o b e chose n later . W e put a\ = max(a , c — S)
and b\ = min( c + 5 , b) and w e spli t //(a , b) = / / ( a , ai ) + If(o>i, &i) + Ifipi, b). Fo r
the integra l ove r th e middl e segmen t w e us e th e trivia l boun d

|//(ai,fci)|Oi-ai ^25 .
To estimat e th e integra l ove r th e lef t segmen t [a , ai] w e ma y assum e tha t a i > a ,
that i s a i = c — S > a , o r els e w e ge t nothing . No w w e verify tha t fo r a ^ x ^ ai ,

-/ ( f c ) (*)= f 1
C {k+ )
f (y)dy-f{k)(c)
Jx
{k+1 )
>I f (y)dy >{c- x)A ^6A>o.
JX

Hence b y th e inductio n hypothesi s

\If(a,a1)\^k2k(5A)-1/k.
The sam e boun d hold s tru e fo r th e integra l ove r th e righ t segmen t [bi , b]. Addin g
these thre e bound s w e ge t
k 1 /k
\If(a,b)\ < 25 + 2k2 (5A)~ .
This i s tru e fo r an y 5 > 0 . O n puttin g 5 = A - 1 /( f e + 1 ) w e complet e th e proo f o f
(8.34) fo rif c + 1 . D

Combining Lemma s 8. 8 an d 8. 9 w e obtai n

208 8. EXPONENTIA L SUM S

COROLLARY 8.1 1 . Let f(x) be a real function with 6 ^ \f'(x)\ ^ I — 0 and

f"(x) ^0 on [a, b]. Then

(8.35) Yl 9(n)e(f(n)) « G0" 1

a<n<b

where g is given by (8.29) and the implied constant is absolute.

EXERCISE2 . Prov e tha t fo r an y real number s f n satisfyin g 6 ^ f n — f n-\ ^
/n+i ~ fn < 1 - 0 we have

(8.36) X X - M ^ cotan -

Combining Propositio n 8. 7 with Lemm a 8.1 0 for k = 2 we derive

COROLLARY 8.1 2 . Let b — a ^ 1 . Let f(x) be a real function on [a , 6] with
f"(x) ^ A > 0 on [a , b]. Then
e
(8.37) J2 (/( n )) « (A 6) " A*) + 1) A _ i
a<n<b
where the implied constant is absolute.
PROOF. B y Lemma 8.1 0 the integrals in (8.26) ar e bounded b y 8A"1/ 2 an d the
number o f these integral s i s less tha n f'(b) — f'(a) + 1 (take a = f'(a),(3 = f f(b)
and e = | ) , hence (8.37 ) follows . D

Note tha t f'(b) - f'{a) = (b - a)f"{y) fo r som e y wit h a ^ y ^ b, thus (8.37 )

implies
COROLLARY 8.1 3 . Let b — a ^ 1 . Let f{x) be a real function on [a, b] such that
A < f"{x) ^ rjA with A > 0 , rj ^ 1 . Tfter c

(8.38) ] T e(/(n))<77A*(6-a ) + A-i

a<n<6
ic/iere the implied constant is absolute.

Now w e ar e goin g t o establis h a n approximat e formul a fo r th e exponentia l

integrals (8.31 ) whic h i s mor e precis e tha n th e estimat e (8.34) . W e begi n b y a
special cas e
LEMMA 8.1 4 . Let h(x) be a real function on [0 , X] such that

(8.39) h{0) = 1 , h{x) > 1 , (xh{x)) f > 1 ,

(8.40) ti(x) < X " 1 , h"(x) < X~ 2
.
Then for a > 0 we have

h dx +0
.41) [<^ ^ = ^Wa ^X
where the constant implied in O depends only on the constant implied in ^> of (8.39)
and that in < C of (8.40).
 8. E X P O N E N T I A L SUM S 20 9

PROOF. Fo r notationa l simplicit y w e se t h(x) = g 2{x). Clearl y g{x) satisfie s

all th e condition s o f h(x). W e chang e th e variabl e x 2g2(x) = t t o obtai n

2 2
/ e(ax g (x))dx = e{at)dU - e(at)f(x)dt
Jo Jo Jo
2
where T - (Xg(X)) an d

f(x) = ((xg(x))'-l)/2xg(x)(xg(x))'.

Using th e condition s (8.39) , (8.40 ) fo r g(x) an d th e Taylo r expansion s

g(x) = 1 + ^ ' ( 0 ) + 0(x 2X~2), g'(x) = g'(0) + 0{xX~ 2

)

we deduc e tha t f(x) < C X~ l an d f'(x) < C X~ 2. Henc e w e obtai n b y partia l

integration
,Tf T
2nia / e(at)f(x)dt = / f(x)de(at)
Jo Jo

= e(aT)f(X) - /(0 ) - / e(at)f'(x)dx(t)

Jo
fx
<^X~X+ \f {x)\dx ^ X~ l

Jo
because x f{t) i s positive. Thi s boun d i s absorbed b y th e erro r ter m i n (8.41 ) . Nex t
we exten d th e firs t integra l t o
1 f°° / 1
/ e(at)dt2
/o vo = e - / y/Sa
Jo ^°
and w e estimat e th e exces s b y partia l integratio n a s follows :
f°° I I f°° I
2nia / e{at)dU = -T~ 2 e(aT) - / e(at)dt~2
JT JT
<r"i = (^(X))- 1 <A:- 1 .

This boun d i s als o absorbe d b y th e erro r ter m i n (8.41 ) completin g th e proof . •

COROLLARY 8.1 5 . Let f(x) be a real function on [a,b] such that

(8.42) /"(c )> A,

3 1 2
(8.43) \f^ \x)\ < AX' , \f«\x)\ < AX"

/or som e A > 0 and X > 0 . Suppose /'(c ) = 0 a i som e poin t c m (a , 6). T/ie n w e

(8.44) / /(0)6)=c(/(c) + | ) / " ( c ) - 5 + 0 ( I ( ^ l -+ - l -+ i

where the implied constant is absolute.

210 8. EXPONENTIA L SUM S

PROOF. W e writ e
pb pb — c pc —a
If(a,b)= / e(f(x))dx= / e(/( c + x))dx + / e(f(c-x))dx.
Ja JO JO

By Taylor' s expansio n w e hav e f(c + x) — f(c) + ax 2h{x), say , wher e a = \f"{c)

and h(x) i s a functio n whic h satisfie s

and /i'(z ) < X " 1 , ^ ^ ) < X - 2 . Thes e estimate s follo w fro m (8.42) , (8.43) .
Hence th e condition s o f Lemm a 8.1 4 ar e satisfied i n th e interva l 0 < x ^ Y wit h
Y = min( 6 — c , X) gettin g

1
j\(f(c + x))dx = e(f(c) + l)(4f(c)r 2+0(Kl^).

For th e remainin g par t o f th e integra l w e appl y Lemm a 8. 9 gettin g

' #+ #« I/'C C + K)!"1 + i/'^r1«(Ar)- 1.

/y
From bot h estimate s w e obtai n
rb
jT e(/(z))ct e = e(/(c ) + §)(4/"(c))- § + o Q (^ + 1 )) .

Similarly w e evaluat e th e integra l ove r th e segmen t [a , c] completin g th e proo f o f

(8.44). D

Now w e ar e read y t o sho w th e mai n resul t o f th e va n de r Corpu t metho d

THEOREM 8.1 6 . Let f(x) be a real function on [a , 6] whose derivatives satisfy
the following conditions: A ^ f ' ^ 77A , |/( 3 >| ^ 7]A(b - a ) " 1 , |/< 4>| < r]A(b - a)~ 2

for some A > 0 and rj ^ 1 . Then we have

e
(8.45) Y, (/H)=E e(/(x m)-mxm + |)r(x m )-i+i?/(a,6)
a<n<b a<ra</ 3

where a — /'(a),/ 3 = f'{b) and x m is the unique solution to f'{x) = m. Here

Rf(a,b) is considered as an error term; it satisfies

(8.46) R f(a, b) < A" 5 + r/ 2 log(/3 - a + 1 )

where the implied constant is absolute.
P R O O F . W e appl y (8.25 ) wit h e = | , a = /'(a ) an d / 3 = /'(b) . Fo r m wit h
a + £ < m < / 3 — s (not e thi s rang e i s voi d i f /3 — a ^ 1 ) w e evaluat e th e involve d
exponential integral s b y (8.44 ) gettin g

fb _ I
/ e(f(x) -mx)dx = e(f(x m) - mx m + \)f"(x m) 2
+ 0
( ? ( r J - + ^ - ) )-
V A V o — x m Xm — a/J
 8. EXPONENTIA L SUM S 211

Here w e hav e rjA(b - x m) > /'(& ) - f'(x m) = f3 - m an d r]A(x rn - a) ^ f(x m) -

f'(a) — m — a , therefor e th e erro r term s contribut e a t mos t

^2 [ ^~ ~ m ) _ 1 + ( m ~ a) _ 1] < rf log(/ 3 - a + 1).

a+e<m</3 — e

For th e tw o missin g integral s wit h a — e < m ^ a -\- e an d / 3 — e ^ r a < / 3 + £w e

use th e boun d 8A - 1 / 2 whic h follow s fro m (8.34 ) wit h k = 2 completin g th e proo f
of (8.46) . •

REMARKS. I n genera l th e boun d (8.46 ) canno t b e significantl y improve d be -

cause i t i s almos t a s smal l a s th e individua l term s i n bot h sum s (8.45) .

EXERCISE 3 . Usin g partial summatio n deriv e from (8.45 ) th e following formul a

for a weighte d su m

(8.47)
n
Y^ g(n)e(f(n))= ^ 9(x m)f {xm)~2e{f{xm) - mx m + | )
a<n<b a<ra</ 3

+ 0(GA~l + Grj 1 log(/3 - a + 1 ) )

where g i s an y smoot h functio n o n [a , 6] an d G i s give n b y (8.29) .

EXAMPLE. Le t X > 0 ,TV > 0 and a> l,v > 1 . W e the n hav e
(8.48)

N<n<uN M<m<fiM

+ 0(./V~ 2 l 0g(7V + 2 ) + M " 2 log( M + 2) )

where ^ + i = 1 , //^ = u a an d MA/ " = X , th e implie d constan t dependin g onl y o n

a, v.
The approximat e formul a (8.47 ) (whic h constitute s th e B-proces s o f va n de r
Corput) transform s a n exponentia l su m o f amplitud e f(x) t o a n exponentia l su m
of amplitud e h(y) relate d b y

(8.49) h(y) = f(x)-xy, y = f'(x).

Thus / an d h hav e essentiall y th e sam e size . However , th e lengt h o f th e sum s

changes fro m b — a t o / 3 — a = f'(b) — f f{a). Not e tha t th e operatio n (8.47 ) i s
involutary, s o it woul d b e useless t o appl y i t tw o times i n a row. I t i s recommende d
to appl y th e ^-proces s i n th e directio n whic h reduce s th e lengt h o f the exponentia l
sum. The n o n th e shorte r sid e on e ca n estimat e triviall y (gettin g fo r exampl e
(8.24)), bu t i t i s not necessar y t o terminat e i n this way . A new possibilit y i s offere d
by a differencing proces s which reduces the amplitud e function , bu t doe s not chang e
the lengt h o f summation. Thi s i s called th e ^-proces s o f Weyl-van de r Corput . W e
begin b y th e followin g genera l inequalit y
212 8. EXPONENTIA L SUM S

LEMMA 8.1 7 . For any complex numbers z n we have

(8.50) ^ ^ ( i+ ^ ^ i - M) £ w „
a<n<b \q\<Q a<n,n+q<b

where Q is any positive integer.

P R O O F . Settin g z n = 0 if n £ (a , b) we obtain

^= / v Zn =
2_J
Z
"a = / j ^n+q
a<n<b n n

for an y g G Z. Su m up thi s ove r q with 0 ^ q < Q getting

a-Q+l<n<b 0^q<Q

Hence b y Cauchy's inequalit y

Q2\S\2 < (b -a + Q) £| £ z
-+^
n 0^q<Q
Z
= (b-d + Q) Yl YJ n+qi~Zn+q2
0^<?i,<?2<Q n

= (b-a + Q) ^2 uti) ]C Zn+^

\q\<Q n
where u(q) is the numbe r o f integers 0 ^ qi,q2 < Q with # i — # 2 — q, i.e., v(q) —
Q — \q\. Thi s complete s th e proof o f (8.50). D
Taking z n = e(f(n)) w e obtain th e differencing proces s
PROPOSITION 8.1 8 . Let f(x) be real function in (a, b) and Q a positive integer.
We have

(8.51) I £ e(f(n))\\ (l + ~) £ (l-^ ) £ «(/( * + «)-/(»))

wftere a(g ) = max(a, a — q) and b(q) = min(6 , b — q).
R E M A R K S . Not e tha t q = 0 is always there , s o the resulting boun d fo r th e
original su m X^ e (/( n )) c a n n e v e r b e better tha n ( 6 — a ) Q - 1 / 2 , i.e. , one cannot
save mor e tha n Q 1 / 2 from th e trivial boun d b — a. I n practice Q is chosen t o b e
much smalle r tha n b — a, so the new amplitud e functio n h(x) = f(x + q) — f(x) has
reduced derivatives , whil e a(q) and b{q) do not change much . Her e th e flexibility
in choosin g Q is another innovatio n introduce d b y van der Corput t o the original
differencing metho d o f Weyl.
1
EXERCISE 4 . Prov e tha t fo r 1 ^ Q ^ ( 6 - a) 2 we have
1
1 1 1 i
e
(8.52) Yl e(/(n))<(6-a)Q"2+(6-a) 2 ^ (M™))
a<n<b a+Q<n<b-Q

where ft (a;) = f(x + q) — f(x — q) for some 0 < q < Q and the implied constan t is
absolute [Hint : Us e only eve n integer s fo r Weyl shifts. ]
 8. EXPONENTIA L SUM S 213

By successiv e applicatio n o f Propositio n 8.1 8 an d Corollar y 8.1 3 w e deriv e

COROLLARY 8.1 9 (VA N DE R C O R P U T ) . Let b - a ^ 1 . Let f(x) be a real
function on (a , b) such that A ^ f^ 3\x) ^ 77 A where b — a > 1 , A > 0 and rj > 1 .
Then
(8.53) ] T e(/(n))«r|2A^&-a)+A^(6-a) 2
a<n<6
where the implied constant is absolute.
PROOF. Fo r 0 < q < Q and a<x<b-qwe hav e h"(x) = f"(x + q) — f"(x) =
qf^(y) fo r som e y G (a, b). Henc e A g < h"(a; ) ^ r]Aq an d b y (8.38 )

£ e(f(n + q)- f(n)) « 7 ] ( A ^ ( 6 - a ) + ( A g ) " i

a<n<6—g

A simila r resul t i s true fo r th e sum s wit h negativ e q. Fo r q = 0 we hav e th e trivia l

bound b — a -f 1 . Collectin g thes e result s w e deduc e b y (8.51 ) tha t
2
J£ e(/(n))| «(l+ ^ ) { f e - a+ V J ( „ A M ( 6 - a )+ (AqH) }
a<n<6 ^ 0<g< Q
&-a\/, . 1
< ( l + - £ - ) ( 6 -o + r;A5Qi(6-a)+A-iQ§ )
~Q
for an y positiv e intege r Q. However , th e abov e estimat e remain s vali d fo r an y
positive rea l numbe r Q fo r obviou s reasons . Takin g Q = A _ 1 / / 3 w e obtai n (8.53 )
provided A ^ (b — a )- 3 , an d otherwis e th e resul t i s trivial . •

The Weyl-va n de r Corpu t inequalit y (8.51 ) ca n b e use d repeatedl y t o reduc e

further th e amplitud e functio n unti l (8.38 ) become s useful . Thi s lead s t o th e fol -
lowing generalizatio n o f (8.38 ) an d (8.53) .
THEOREM 8.2 0 (VA N DE R C O R P U T ) . Let b-a^l. Let f(x) be a real function
on (a , b) and k ^ 2 such that A < f^ k\x) < nA, where A > 0 and rj ^ 1 . Then
n
(8.54) Yl <f( ^< ^ ^ ( b~ a) + A~ K(b - a) 1 2
' *^
a<n<b

where K = (2 k — 2)~ 1 and the implied constant is absolute.

For larg e k th e bound s (8.1 7 ) o f Wey l an d (8.54 ) o f va n de r Corpu t ar e com -
parable.
8.4. Discussio n o f e x p o n e n t pairs .
Throughout Sf (N) denote s a n exponentia l su m o f typ e

Sf(N) = V J e (/(n))
N<n^Nf

with N ^ N' ^ 2N an d / i s a smoot h functio n o n [iV , 2N]. W e hav e establishe d

numerous estimate s fo r Sf(N) whic h depen d essentiall y o n th e lengt hT V and o n
the siz e o f th e amplitud e functio n / . I n thi s sectio n w e describ e thes e result s i n
a unifie d for m an d discus s possibl e extensions . Fo r th e sak e o f simplicit y w e dro p
some minor , ye t necessar y conditions , s o ou r assertion s ar e no t strictl y correc t i n
214 8. E X P O N E N T I A L SUM S

this section. Rigorou s assertion s ca n be found i n the boo k b y S . W. Graha m an d G .

Kolesnik [GK ] a s well a s i n th e origina l paper s b y va n de r Corpu t [Corl] , [Cor2] ,
E. Phillip s [Ph ] an d R . A . Ranki n [Ralj .
Suppose / behave s lik e a monomial, precisel y w e assum e tha t th e derivative s
satisfy
j
(8.55) \fW(x)\^FN~
for N ^ x ^ 2N an d an y j ^ 0 , wher e th e implie d constan t depend s o n j . In
particular, \f(x)\ x F and \f'(x)\ x FN~ X. I fT V i s large r tha n F by a sufficientl y
large factor , the n Corollar y 8.1 1 yield s
1
Sf(N) 4ZNF- .
This boun d i s the bes t possible , becaus e i n thi s cas e the su m Sf(N) i s well approx -
imated b y th e correspondin g integra l (se e Lemma 8.8) . I n what follow s w e stay ou t
of thi s specia l cas e b y assumin g tha t
(8.56) A = FN^ > 1.

Inspired b y man y example s w e postulat e th e existenc e o f universa l number s

(8.57) 0 < p , < K §
having th e propert y tha t
q
(8.58) S f (TV) < APN ^F£
for an y exponentia l su m o f length N and th e frequenc y functio n / satisfyin g (8.55 )
and (8.56) . Her e e i s an y positiv e numbe r an d th e implie d constan t depend s onl y
on e an d o n th e sequenc e o f constants implie d i n (8.55) . W e call (p , q) an exponen t
pair (notic e tha t ou r definitio n o f q differs b y | fro m tha t i n th e literatur e [GK]) .
Clearly i f (p, q) i s a n exponen t pair , the n an y pai r o f large r number s for m a n
exponent pair . Th e trivia l estimat e fo r Sf(N) correspond s t o th e exponen t pai r

(8.59) (P,q)=(0,\).

The followin g statemen t i s the mos t optimisti c fo r thi s aspec t o f th e theor y o f

exponential sums , an d i s widel y believe d t o b e correc t i n it s essential s (mino r as -
sumptions o n / migh t b e require d t o avoi d pathologies) :
E X P O N E N T PAI R HYPOTHESIS . The pair {p,q) = (0,0 ) is an exponent pair.
In other words, an exponential sum of type Sf(N) with / , N satisfying (8.55) and
(8.56) should satisfy
£
(8.60) S f(N)<z:N^F .

This bound , i f tru e fo r reasonabl e functions , woul d lea d t o solution s o f man y

problems i n analyti c numbe r theory . Fo r example , it implies th e Lindelo f Hypoth -
esis
 8. E X P O N E N T I A L SUM S 215

Moreover, i t would solve the Gaus s circle problem an d the Dirichlet diviso r problem ;
namely i t implie s

~ ^ r(n) = TTX + 0 ( > i + £ ) ,

^2r(n) = xlog x + ( 2 7 - l) x + 0(x^ + f f ),

n^.x

where th e exponent s ar e bes t possible . Actuall y fo r thes e problem s on e onl y need s

to kno w tha t (p , q) = (0, \) i s a n exponen t pair . However , ou r curren t knowledg e
is very poor .
Corollary 8.1 3 tell s u s tha t

(8.61) (P,q)=(\,0)

is an exponen t pair . Clearl y a linear conve x combinatio n o f exponent pair s i s agai n

an exponen t pair . Therefor e (8.59 ) an d (8.61 ) impl y tha t (p,q) = (\,\) i s an
exponent pair , however , w e have alread y prove d a stronge r resul t (8.1 8 ) whic h ca n
be no w interprete d a s sayin g tha t

(8.62 ) M = {\>\)
is a n exponen t pair . Theore m 8. 4 yield s th e exponen t pair s
4
/x / 1 4 ( / c -1 ) \

and Theore m 8.2 0 yield s

(8-64) fe ,)=^____, _ _ _ _ _ _

for an y intege r k ^ 2. Notic e that a s k tend s t o infinit y i n both sequences , p — p(k)

tends to zero slightly faste r tha n q — q(k) tend s to |. This featur e make s i t possibl e
to clai m non-trivia l bound s fo r th e exponentia l sum s Sf(N) whos e lengt h N i s
comparable wit h th e amplitud e F in th e logarithmi c scale . I n th e nex t sectio n w e
shall produc e b y Vinogradov' s metho d a sequence o f exponent pair s (p , q) in whic h
p tends t o zer o muc h faste r tha n q tends t o \. S o far w e d o no t hav e a n exponen t
pair wit h p — 0 and q<\\ an y suc h exampl e woul d mar k a majo r breakthroug h i n
the theor y o f exponential sums .
A continuu m o f exponen t pair s ca n b e create d b y alternatin g th e A and B
processes. B y th e Weyl-va n de r Corpu t differencin g inequalit y (8.51 ) on e derive s
A-PROCESS. If(p,q) is an exponent pair, so is

*-65» ^«»-(_?2-_^) -

By th e va n de r Corpu t approximat e functiona l equatio n (8.45 ) on e derive s

216 8. E X P O N E N T I A L SUM S

^-PROCESS. If(p,q) is an exponent pair, so is

(8.66) B(p,q) = (q,p).

Here ar e som e example s o f exponen t; pairs generate d fro m th e trivia l pai r

B(O,
D- -(5.' 0),

AB(O,
2
$ • < • J).
2
A B(O,
\)- -(n ' 7.)/ '
A3B(O,
\)- V30 - )
'3<V'
A4B(O, = (1 - )
\)- '31/'
V62
2
ABABA B(O,
* ) •
»(n • i ) -
For th e purpos e o f estimatin g th e Rieman n zet a functio n o n th e critica l lin e
one need s a n exponen t pai r (p , q) with p + g as smal l a s possibl e sinc e
(8.67) ((l + it ) < *2(P+9)+ e, i f t ^ 1 .

In 1 95 5 Ranki n [Ral ] compute d th e minimu m o f 0 = | ( p + g ) ove r al l exponen t

pairs generate d b y th e A an d B processes , namel y # m i n = 0.1 6451 067... . I n othe r
applications o f exponent pair s one seeks the minimu m o f a fractional linea r functio n
#(p> a) = ( aP + PQ + M ) / ( ^ P + 7 ^ + v). Havin g thes e application s i n min d i n 1 98 5
Graham gav e a truly elegan t algorith m fo r th e solutio n t o thi s genera l problem ; se e
[GK].
Today ther e ar e know n exponen t pair s whic h canno t b e obtaine d b y th e va n
der Corpu t method . Fo r example , Huxle y an d Wat t [HW ] showe d tha t

(8.68) M=G|<S )
is an exponen t pai r b y elaboratin g th e wor k o f Bombier i an d Iwanie c [BI] . Furthe r
improvements alon g thes e line s ar e give n i n [Hu4 ] an d a ver y nic e expositio n i s
contained i n [GK] .
The Wey l shif t (fro m n t o n + q) wa s introduce d fo r th e purpos e o f reducin g
the amplitud e functio n (fro m f(ri) t o h(n) = f(n + q) — /(n)) . However , thi s shif t
also offer s a n additiona l variabl e o f summation . On e ca n tak e advantag e o f this b y
an appea l t o th e theor y o f two-dimensiona l exponen t pairs , bu t th e improvement s
are no t impressiv e (i n th e contex t o f the Rieman n zet a functio n th e result s ar e stil l
weaker tha n wha t follow s fro m th e Huxley-Wat t exponen t pai r (8.68)) .
8.5. Vinogradov' s method .
In th e lat e thirtie s I . M . Vinogrado v [Vi2] , [Vi3 ] invente d a ne w metho d fo r
estimating exponentia l sum s o f Weyl' s typ e whic h i s ver y stron g fo r sum s o f larg e
amplitude (relativ e t o th e length) . H e spen t a considerabl e tim e o f hi s lif e t o
 8. EXPONENTIA L SUM S 217

massage th e results , s o w e hav e toda y severa l elegan t version s du e t o hi m an d hi s

followers.
Let f(x) b e a rea l smoot h functio n i n [iV , 27V]. Ou r ai m i s t o estimat e expo -
nential sum s
(8.69) 5/(0,6) = V J e (/(n))
a<n<b

for an y a, b with N ^ a < b ^ 2N. W e shal l approac h th e proble m i n severa l

distinct steps .

Step I (Smal l Shift) . W e begin, a s i n Weyl's method , b y shiftin g th e summatio n

n
5/(0,6)= J2 <f( + l))-
a—q<n<b—q
We choose q to b e relatively smal l s o we can approximat e f(n - f q) by a polynomia l
in q of reasonably smal l degree. Precisel y w e apply th e Taylo r expansio n f(n + q) =
Fn{q) + Rn(q), wher e F n(q) i s the polynomia l

with coefficient s otj{n) = f^\n)/j\ an d R n{q) satisfie s th e boun d

\Rn(q)\^qk+1f{k+1)/(k+l)l

with / ^ + 1 ) — max|/( fe+1 )(a;)|. Insertin g thi s approximatio n w e ge t

1
k+ 1
(k+ )
5/(o,6)= £ e(F n(q))+0(2q + q f N)
a<n<b

where \9\ ^ 1 . Her e th e first erro r ter m 2q represents th e trivia l boun d fo r expo -
nential sum s ove r tw o non-overlappin g interval s an d th e secon d erro r ter m come s
from th e inequalit y \e(R n(q)) - 1 | < 2n\R n(q)\ < q k+if(k+i)m W e s h a l l r e q u i r e a n
the implie d constant s i n th e forthcomin g estimate s t o b e absolute . Thi s i s impor -
tant, eve n with respec t t o /c , because fo r som e application s (firs t o f all for boundin g
£(1 -|- it)) on e use s k dependin g o n th e lengt h o f summatio n an d th e siz e o f th e
amplitude function .
At thi s poin t Vinogradov' s approac h diverge s fro m tha t o f Weyl . Th e first
new ide a i s tha t no w w e ar e goin g t o averag e ove r q in a specia l subset , sa y Q C
{ 1 , 2 , . . . ,Q} , rather tha n i n th e ful l set . W e obtai n
1
(8.70) 5/(o,6) = J2 i 2 r E e ( F " ^ ) ) + 0 ^ + ^ f e +1 / ( f c +1 )7V
)
a<n<b q£Q

where \Q\ denote s th e numbe r o f point s i n Q. Thoug h F n(q) i s a polynomia l i n q

the innermos t su m
Sn = ^e(F n(q))
q£Q
cannot b e considere d a s Weyl' s su m o f degre e k, becaus e o f restriction s fo r Q (t o
be specifie d soon) . Moreover , w e ar e no t goin g t o appl y th e ide a o f differencin g b y
way o f squarin g th e su m Sf(a,b) ) bu t rathe r w e procee d directl y t o estimatio n o f
218 8. EXPONENTIA L SUM S

each inne r su m S n separatel y fo r a < n < b. I n othe r words , w e no longe r nee d th e

variable n unti l th e las t step , an d eve n there , i t i s no t essential .

Since the variabl e n play s a minor rol e in what follow s w e simplify th e notatio n
by considerin g th e exponentia l su m

S=5>(F(<?))
qeQ

for arbitrar y polynomia l wit h rea l coefficient s

F(q)= Y, a ^-

Obviously S depend s o n th e fractiona l par t o f th e coefficient s aj . Fo r thi s reaso n

we shal l exploi t onl y th e coefficient s wit h \a.j\ < 1 , becaus e thes e ca n b e easil y
controlled. O n th e othe r hand , th e ver y smal l coefficients , specificall y \aj\ < Q~ J,
do no t mak e a variatio n i n th e argumen t o f e(F(q)). Therefor e onl y th e middl e
terms o f ou r polynomia l F(q) pla y a role , tha t i s thes e wit h Q~i < \otj\ < 1 , th e
other term s ar e wasted . I n th e final argument s w e shal l appea l t o th e origina l
coefficients aj — ot.j(n) = f^\n)/j\ whos e particula r shap e i s no t important , w e
only nee d t o estimat e thei r size .

STEP I I (creatio n o f bilinea r forms) . W e choos e

Q = {xy:l^x iy^P}

where th e number s q = xy ar e counte d wit h multiplicit y o f suc h representations ,

thus \Q\ = Q = P 2 an d th e exponentia l su m S become s a bilinea r for m

(8.7i) s= J2 E <n*v))-

It i s importan t tha t th e variable s x , y ru n independentl y ove r tw o set s (o f som e

cardinality), ye t i t i s of n o significanc e wha t th e point s are . Fo r example , w e coul d
restrict x, y t o specia l number s o r conside r a genera l bilinea r for m

with an y comple x coefficient s a(x) , 6(y), wher e X, y ar e th e set s o f point s o f typ e

x = [ 1 , . . . , xk], y = [1 , y , . .. , yk] wit h x,y e Z, 1 ^ x ^ X, 1 ^ y ^ Y an d
(x, y) = F(xy). I f X, y ar e sufficientl y larg e an d well-space d set s the n a non-trivia l
bound fo r B(X,Y) follow s easil y b y Fourie r techniqu e (se e Chapte r 1 0) . However ,
this i s no t th e cas e unde r consideratio n here ; ou r set s ar e quit e sparse .

STEP II I (expandin g th e variables) . Ou r nex t goa l i s t o creat e mor e point s

of summatio n an d mak e man y variable s wit h n o larg e gaps . T o thi s en d w e rais e
the bilinea r for m (8.71 ) t o a hig h power , appl y Holder' s inequalit y an d rearrang e
 8. E X P O N E N T I A L SUM S 219

the resultin g multipl e su m a s follows:

|5|'<P'-1^|^e(F(a;i,))|'
x y

e 1
= P ~ J2(* E e(F(x yi) + • • • + F(xy e))
x yi,---,ye

Al
= pi ~X 1 L ^ ' ''' ' X
^ Yl Cxeia^x + • •• +a k\kx
k
).
A i , . . . ,Af c x

Here £ x ar e complex number s wit h \( x\ = 1 and z/(Ai,... , A&) is the number of

solutions t o the system o f equations
( 2/ 1 H + yt = A i
2/i + • • • + 2/1 = A 2
(8.72 )

yf- + 2/ * = A fe
in integer s 1 ^ yj ^ P fo r j = 1 ,2,.. . ,£ . Thu s £ ^ X 3 ^ £Pi, or else ther e are
no solutions . Sinc e £ will be much large r tha n k (approximatel y £ x k 2) w e expec t
that mos t vector s (Ai,.. . , A&) have goo d numbe r o f representations b y the system
(8.72). Thi s numbe r i/(Ai,.. . , Xk) ma y var y onl y slightly . I n order t o smooth the
outer summatio n w e raise \S\ £ again t o power 2£, getting

\S\2i2 < P 2 1
^-^I 2{e )
t,k(P) - JtAP)Ze,k(P)

where

h,k{P)= £ KAi,...,A fc ),
A i , . . . ,Afc

2
JiAP)= Yl ^ (Ai,---,A fc ),
A i , . . . ,A fe
21
k
Ze,k(P)= Yl \52(xe(a 11\ x+ -<- + a k\kx )
A i , . . . ,A & 3 :

For th e average o f v(\i,... , A&) we hav e th e exact valu e

itAP) = P 1 ,
but th e estimatio n o f the averag e of i/2 (Ai,... , A&) is a much harde r problem . Not e
that Ji,k(P) i s the number o f solutions t o the system o f homogeneous equation s
f 2/ i H Yyi- yi+i • yu = 0
2 , , 2 2
- 2/2 , = o
.73) 2/i + * • * + Vi ~ Vi+i

U i + • • • + yt - 2/ m v\t = o
in integer s 1 ^ yj < P for j = 1 , 2 , . .. , 2£. W e hav e
21
(8.74 ) JiAP)= '" \ Yl e (aix + "- +akx )
k
da\ . . . da k.
J J
° ° l<x<P
22 0 8. E X P O N E N T I A L S U M S

More generally , i f the right-hand sid e of the system (8.73 ) i s replaced b y constants
<Ti, . . ., ak then th e number o f solutions J^k{Pj] cri, • • • , crk) to the system o f inho-
mogeneous equation s i s given b y the integral

/" '/ M C e (aix~\ \-a kx

k
)\ e(a 11a -\ h akak)dai . . . dak.

Hence it is clear tha t

(8.75) J ^ ( P , ; e n , . . . ,<r fc) ^ J ^ ( P j ; 0 , . . . ,Q) = J itk(P).

STEP I V (tunin g dow n t o linear polynomials) . Estimatio n o f Jijk(P) lie s

in the heart o f Vinogradov's metho d (se e the next step) , however , th e vital savin g
comes fro m Z^^ k{P) a s it is the only su m left wit h non-trivia l characters . W e have

e
Ze,k(P) ^ ^2 5 Z ( a i ^ i ^ i H ha fcAfccrfc)
a s i , . . . ,x 2£ A i , . . . ,Afe

where a k = a k(xu . . . ,x2i) = S i (x j _ xh

j+t)-> s o
l°"fe l ^ tP k'- Henc e b y (8.75) we
arrive at

Ze,k{P) ^ Ji,k(P) Yl J 5Z e (^i A i^i + Va k\kak)

C T i , . . . ,CTf c A i , . . . ,Af c

Here a i , . .. , 0>, A i , . .. , A^ run independently ove r al l integers with |<J^ | ^ £P h an d

|A^| ^ £P^. Not e tha t th e multiple su m factors s o we write th e result as

2k
1 k{k+ )
Zi,k{P) ^ Jt, k(P)t P ^

where

A= J ] D(a hJPh),

D{a,X)=X~2 Y, I J2e (amn)

\m\^X \n\^X

Now, as in Weyl's method , th e source of cancellation i s in the exponential sum s for

linear polynomials . Thi s is reached ver y differentl y her e by Vinogradov's techniqu e
of bilinear form s tha n befor e b y Weyl's differencin g process . Bu t above all , the new
process is much faster. Indeed , Weyl' s differencing require s applications of Cauchy's
inequality k — 1 times (eac h applicatio n reduce d th e degree by one), whic h amount s
to raisin g th e original su m to power 2 / e _ 1 , wherea s Vinogradov' s metho d get s the
job don e wit h th e power 2£ 2 x k 4. Anothe r ne w feature i s that w e produced linea r
polynomials ou t of every monomia l i n F, not just on e from th e highest degree ;
consequently th e saving factor s appea r abou t k times.
Gathering th e above estimate s w e obtain
2£2 2k
(8.76) \S\ ^ e P"V-V+k(k+Vjlk(P)A.
 8. E X P O N E N T I A L SUM S 221

STEP V (estimation o f Ji^(P)). Le t u s give some thought abou t th e inequalit y

(8.76). Her e S is the origina l su m

S{au . . . , a k) = ^2 e ( Q i < ?1
+ " a k<ik)

while J^f c i s the L 2^-norm o f

e
P ( a i , . . . ,a k) = ^2 ( a i ^ H \~OL kx
k
)

with respec t t o the coefficient s a i , . . . , a^ ove r the toru s (R/Z) fc . A t first glance we
may thin k tha t th e proble m o f estimatin g Jt, k(P) i s no t muc h differen t tha n tha t
of th e origina l su m S. Yes , this woul d b e a n eve n harde r proble m i f w e dealt wit h
individual sums. But , du e to complete integration ove r the coefficients, w e are reall y
having a diophantine proble m (countin g solution s t o th e syste m o f homogeneou s
polynomial equation s (8.73)) .
By a heuristic argumen t on e i s lead t o th e boun d
12€ fc+
(8.77) J ^ ( P ) < P -^ >
(the linea r equatio n i n (8.73 ) fixes on e variable , th e quadrati c equatio n fixes two
additional variables , etc.) , bu t w e canno t prov e this . Vinogrado v succeede d in
proving a slightly weake r bound .
THEOREM 8.21 . Let k,m be positive integers and £ ^ k(k + m). We then have

for any P ^ /^(i-V*)"" 1 .

We begi n b y provin g
LEMMA 8.22 . Let p be a prime number, p > k, and A i , . . . ,A & be integers.
Then the number T = Tp (Ai,... , Xk) of solutions to the system of congruences

x\ + • • • - f xk = Ai (mo d p)
3.79 )
[xil h ^ E A f c (mo d pk)

in # i , . . . ,x k (mo d p k) such that Xi ^ Xj (mo d p) is bounded by

1
1hih
(8.80) T^k\p ^ .

PROOF. W e writ e eac h x r uniquel y i n th e for m

k l
xr = xrl + xr2p H h x rkp ~
with 0 ^ x ri < P for 1^ r $ J k and 1 < i ^ k. Th e first coordinates x r\ satisf y
the syste m (8.79 ) (mo d p). Pro m th e theor y o f symmetri c polynomial s i t follow s
that th e j - t h symmetri c function s i n £ i i , £ 2 i , . .. , £/ci, sa y Sj , ar e determine d b y
A i , . . . , Afc and tha t x r\ ar e roots o f xk — sixk~x -\ [-(—l) k
sk. Ther e ar e a t mos t
k roots , s o there ar e at most fc!solutions ( x n , x 2 i , . . • , xk\).
22 2 8. EXPONENTIA L SUM S

Now suppos e ( x n , x 2 i , . . . ,Xki), .. • , (xih,x2h,.-- ,Xkh) wit h 1 < ft < k ar e

given. W e shal l find condition s fo r th e coordinate s o f orde r ft + 1 . Lettin g

Xr\ + ^r2 P H + X rhph~X = y r/l

for 1 < r < fc, we hav e

h+1
^r = 2/r/ i + aj^+iP^mo d p ).
Consider th e system (8.79 ) (mo d p h+l) reduce d b y omitting th e first ft congruences,
i.e.,
k
1
h+
Y,(Vrh + Xrh+lP^™ = A m(mod p )
r=l
for ft < m ^ k. Henc e
k
(8.81) ^Xrh+iy?h = Mm(mo d p), ft < ra < /c ,
r=l

where // m ar e som e number s determine d b y th e coordinate s x r j o f orde r 1 ^ j ^ ft

and b y th e constant s A m wit h ft < m < k. Th e number s y rhA ^ r ^ k, ar e
different modul o p becaus e s o ar e x r. I n particular , a t mos t on e y rh vanishe s
modulo p , sa y y ^ , i f any . I f w e choos e £1 /1 +1 ,.. . ,Xhh+i arbitraril y th e remainin g
ones Xh+ih+i, • • -^/c/i+i ar e determine d modul o p uniquel y fro m th e linea r syste m
(8.81) becaus e it s determinan t (Vandermonde ) i s non-zero . Therefore , th e numbe r
of coordinate s o f orde r ft + 1 does no t excee d p h. W e conclud e tha t
O/ 1 fc(fc-l)
T < klpp l---pk-1 = klp—^—.
a
COROLLARY 8.23 . Let p be a prime number, p > k, and let Uk P(P) be the
number of solutions to the system of congruences

(8.82) £ {xt - x^ +k) E E 0 (mo d p m)

l^h^k

for 1 < m ^ k, in X < Xh ^ X + P, 1 < ft ^ 2k, such that Xh 1 ^ Xh 2 (mo d p) for

1 ^ fti 7 ^ h2 ^ k. We have

(8.83) U kp(P) < fe!(pfc 4- P ) f c

P V ^ .

Our nex t ai m i s t o establis h a recurrenc e inequalit y fo r J ^ ( P ) . W e assum e

that £ > k ^ 2 and P ^ pq, wher e p i s a prim e numbe r > k t o b e chose n late r an d
q i s a n integer . Clearly , w e hav e Ji,k{P) ^ Je,k(pq)- Fo r estimatin g Ji,k(PQ) we
divide the solution s o f the syste m (8.73 ) int o two classes . W e say tha t ( y i , . . . , y 2i)
is o f th e first clas s i f bot h sequence s ( y i , . . . , yi) an d (y^+i),.. . , y2i) hav e a t leas t
k differen t number s modul o p. Th e remainin g solution s ar e o f th e secon d class .
Let u s estimate th e numbe r Ji , say , of the first clas s solutions. Sinc e k number s
can b e pu t i n £ places i n £(£ — 1 ) . . . (£ — k + 1 ) way s w e obtain J\ < £ 2kJn, wher e
 8. EXPONENTIA L SUM S 223

J n stand s fo r th e numbe r o f solution s ( y i , . . . , y<ii) suc h tha t al l ( y i , . . . , yu) ar e

different modul o p an d al l (y^+i,.. . , yi+k) ar e differen t modul o p. Lettin g

S(a) = J2 <F a(y)),

where F a(y) = a xy -\ h akyk, w e split

s(<*)=u(mod < F«(v)) = Ed p) 5 ^

E p)El^y^P u(mo
u(mod p) l^y^l
y=u(mod
y=u(moc p)
say. B y Holder' s inequalit y w e obtai n
pi pi. , , 2I 2£-2k
Jn= •• • / Yl S Ul(a) - - S Uk(a)\ ^ S u(a) da
iti,... ,Ufc(mo d p ) u(mo dp)

< p *- a *-i £ Jll(tt) i

0<u<p

say, wher e ^ } * means tha t th e summatio n i s restricte d t o th e residu e classe s ui,

. . . , Uk modulo p tha t ar e al l different , an d

2
Jn(u)= [ •• • I I V*S Ul(a)-'SUk(a)\ \Su(a)\2i-2kda.

This i s equa l t o th e numbe r o f solution s ( y i , . . . ,2/2^ ) o f th e syste m (8.73 ) suc h

that al l y i , . . . , i/fc ar e differen t modul o p , al l y^+i,.. . , yg+i,.. • , y^+/c ar e differen t
modulo p an d al l th e remainin g one s ar e congruen t t o u modul o p. Fo r thes e w e
set yh = u + pvun with 0 ^ Vh < q if k < h < £ or k + £ < ft < 2^ . W e obtai n

E ( ^ - ^ ) + E [(«+ pvh)m - ( « + ? w i = 0,
0<h^k k<h^£

for 1 ^ m ^ /c . No w w e appl y a n obviou s fac t tha t i f ( y i , . . . , y2^) satisfie s th e

homogeneous syste m (8.73) , the n (y i —it,.. . , y2^ — u) als o does . Therefor e w e
obtain
x
E( h-*H+t)+pm E « - v
W = o,
0<h^k k<h^£
for g ^ m ^ fc, where 2 ^ = y ^ — w for 0 < ft ^ k an d £ < h ^ £ + k. I n particular ,
we observ e tha t x ^ satisf y th e syste m o f congruence s (8.79) . Give n a solutio n t o
(8.79) w e se e tha t Vh satisfy th e inhomogeneou s syste m o f equation s

E w - v ™+^ =A-' l < m ^ k>

with som e constan t \ m independen t o f Vh- Thus th e numbe r o f the v^s i s bounde d
by Ji-k,k(q)i se e (8.75) . Combinin g wit h (8.83 ) w e conclud e tha t

J n ( u ) < W(p fc + P) k k1
1
=ii ±
P p S Jt-k,k(Q)
and
(8.84) J x < f kk\V2l-2k-m^1 Pk{pk + P) k
Je-ktk(q).
22 4 8. EXPONENTIA L SUM S

Now w e estimat e J2 , th e numbe r o f solution s o f th e secon d clas s o f (8.73) . Amon g

the number s ( y i , . . . , y^) ther e ar e a t mos t k — 1 different residu e classe s modul o
p, o r amon g th e numbe r (y^+i,.. . ,2/2^ ) ther e ar e a t mos t k — 1 different residu e
classes modul o p. Therefor e

J<2=
" yZ S Ul(a) • - S Ui (a)S Ui+1 (a)... S u^ (a)da

where U i s th e sequenc e o f vector s u = ( i ^ , . . . , i ^) (mo d p) wit h th e relevan t

property, s o \U\ < 2k ipiJrk1
~.
By th e inequalit y (x\ • • • x n)^ ^ -(x\ + • • • 4 - xn) w e obtai n
11

(ui,... ,u 2i)£U h=l

say, wher e

2i 2k
J2(u)= I •• • f \S u(a)\ da ^ J^ k(q) ^ q Je^k(q).
Jo Jo
Hence
t i+k l 2k
(8.85) J 2 < 2k p - q Je.kM{q).
By (8.84 ) an d (8.85 ) w e conclude tha t
2fc
JeAP) < f* f c ! p 2 < - 2 f c - ^ P V + P) fc + 2fcV +fe - 1 9 2fc l J t.ktk(q).

Finally assumin g tha t P ^ k k w e may tak e a prime numbe r p wit h 2 + P 1 / ^ ^ p ^

2p1/k a n ( i th e intege r g = \p~ 1 P] + 1 getting
LEMMA 8.24 . Let k ^ 2 , ,£ > \k(k + 3 ) - 1 and P ^ & fc. We then have
4
(8.86) J,, fc (P) ^ 2 ^+ ^ J ^ - M ^ 1
"*).

Now Theore m 8.2 1 follow s easil y fro m Lemm a 8.1 4 b y inductio n i n m.

Combining (8.76 ) wit h (8.78 ) fo r £ = k(k - f m) w e obtai n
2
(8.87) | 5 | ^ 4P (APfc(H1 )(1
^r)^2(H-)2

where m i s an y positiv e intege r an d P > k k^l~k) m .

STEP V I (estimatio n o f A) . I t remain s t o estimat e A . W e hav e

m i n X
X>D(a,X)^ £ 2||am| 1 ( ' ^ )
, ,< v

1
^^ V '2am-r /
|r|<Cc*A~+i | a m - r | < ^

^ 2 ( a X + l)( X + ^ min(2X , —)).

0<u<l/2a
 8. E X P O N E N T I A L SUM S 22 5

The las t su m i s bounde d b y a n integra l whic h equal s a - 1 logeX , therefor e

D(a, X) ^ 2X~ 2(aX + 1 )(X + a' 1 lo g eX) ^ 4( a + a^X' 2) log3 X
for a > 0 an d X ^ 3 . W e als o hav e trivia l boun d D{a,X) ^ 4 . Fro m bot h
estimates w e conclud e tha t fo r an y 7 c { l , 2 , . . . , / c } ,
k
(8.88) A < JJ(|a,-| + la^p-V) (4k\og3£P) .

STEP VI I (estimatio n o f exponentia l sums) .

We no w assum e tha t f(x) i s a rea l smoot h functio n o n [iV , 2iV] suc h tha t fo r
all j ^ 1

j3
(8.89) a~ F< °^\f U)(x)\ ^ a 3"F

on [TV , 2iV], wher e F ^ i V ^ 2 and a ^ 1 . Henc e fo r a ^ n ) = f {j)(n)/j\ w e ge t

a - j 3 ( 2 i V ) - j F < la^n) ! ^ a j3
N~jF
and
A ^ J^(FAT-^ ' + F - 1 ^ ' p - 2 j > f e 4 ( 8 A ; l o g 3 P ) f c .
jei

We assum e tha t P 4 ^ T V ^ F 3 an d fc ^ 21 ogP/logiV . W e choos e

f. log P log P1
V ' logiY <J^log(7V/P)J
getting
A < Y[(FN- J)(2kek)k2(log3N)k < 7V"i jr(J-1 )(2A:e/c)fc2(log3Ar)/c

where
lQ
7 = m > ^ ^ gF 1 > fog*1) fog ^) , 1
11 2+
" log(N/P) lo gTV " 2(log7V)
Finally
(8.90) A ^ exp(-(logF) 2 (logP) 2 (21 og7V)- 3 )a /c4 (^log3AT) fc .

Now settin g P = iVi , k = [ 4 log P/ lo g TV] and m = 8f c w e infe r fro m (8.70) ,

(8.87) an d (8.90 ) th e followin g
T H E O R E M 8.25 . Let fix) be a smooth function on [TV , 27V] which satisfies
(8.89) with F ^ TV 4 and a ^ 1 . We then have
1 8
(8.91) S f(a,b) < aA^exp(-2- (log7V) 3 (logP)- 2 ),
where the constant implied in < C is absolute.
APPLICATIONS: W e conclud e th e presentatio n o f Vinogradov' s metho d b y
drawing fro m (8.91 ) a fe w basi c applications .
226 8. EXPONENTIA L SUM S

COROLLARY 8.26 . For t^N^2we have

3
(8.92) Yl ^<7Vexp(-/3(log7V) (logt)- 2 ),
l^n^N

where (3 and the constant implied in < C are absolute.

THEOREM 8.27 . There exists an absolute constant a > 0 such that for s = a+it
with t ^ 2 and | ^ a ^ 1 ,
a
(8.93) C ( 5 ) < t ^-^f(logt)i.
The implied constant is absolute.
PROOF. B y (8.21 ) usin g partia l summatio n an d (8.92 ) w e hav e

C(s)= Yl ri- s
+ 0(l)<^ ( x~ a
exp(-^(logx) 3(logt)-2)^

pi pi
f
= ( l o g t ) / t^-^-^du^ilogt) t W-Ku-vfdu
Jo Jo
where f(u) = ( 1 - cr) u - / ^ 3 an d v - ^/( l - cr)/3/ 3 s o /(v ) = 2( 1 - a)?/3y/3@-
This give s (8.93 ) wit h a = 2/3>/3/? . •

COROLLARY 8.2 8 (VINOGRADOV-KOROBOV , 1 957) . There exist absolute con-

stants 7 > 0 , 6 > 0 such that

(8.94) |C(< 7 + zt)|^ 7(log*)§

m the region
(8.95) £ ^ 2, a > l-5(\ogt)-i.

There ar e severa l interestin g device s b y mean s o f whic h on e ca n transfor m

an uppe r boun d fo r £(s ) int o a zero-fre e regio n an d deriv e estimate s fo r l/£(s) ,
(f(s)/((s) i n thi s region . Th e origina l argument s o f Hadamar d an d d e l a Valle e
Poussin (presente d i n Chapte r 5 ) produc e onl y a zero-fre e regio n o f th e typ e a ^
1 — c(logt) - 1 , regardles s o f ho w goo d th e estimat e fo r £(s ) tha t on e has . Howeve r
combining thes e argument s wit h a metho d o f Landa u on e ca n establis h deepe r
relationships (se e Theore m 3.1 0 an d Theore m 3.1 1 o f [T2]) .
LEMMA. Let (j)(t), ip(i) be positive increasing functions for t ^ 0 such that
(f)(t)ip(t) = o(exp((/)(t))) as t - + +oo . / /
C ( 5 ) « e x p ( ^ ) ) for a>l-^(t)-\
then
1 1
COO^O for a^l-cxp(2t + 3)- (/){2t^3)- .
Moreover in this region
1/C00, C'(s)/((s) « 0(2 t + 3)^(2 i + 3) .

Hence usin g Corollar y 8.2 8 w e obtai n

 8. E X P O N E N T I A L SUM S 22 7

THEOREM 8.29 . We have ((s ) ^ 0 and

— <(log£)i(loglogt)* ,
Cw
^<(log*)i(loglog*)*
C(«)
for c r ^ l — c(logt) / (loglog^) _ 1 / 3 ; £ ^ 3 7 where c > 0 is an absolute constant.
_2 3

Now the standar d contou r integra l argumen t (se e Chapte r 5 ) yield s th e follow -
ing stronges t know n Prim e Numbe r Theorem .
COROLLARY 8.30 . We have
ip(x) — x + 0(xexp(—c(logx) s (loglo g x ) - 5 ) )
for x ^ 3, where c > 0 is some absolute constant.

Let Tk(n) stan d fo r th e numbe r o f way s o f factorin g n int o k positiv e integers ,

so the Dirichle t generatin g serie s fo r 7>(n ) i s ( k(s).
THEOREM 8.31 . For any e > 0 and x ^ 2 we have
r
(8.96) D k(x) = J2 * W = xPk(logx) + 0(x^ +£),

where Pk is a polynomial of degree k — 1 , 5k = 1 — 7A: -2 / 3 , ^ is a positive absolute

constant and the constant implied in O depends on e and k only.
PROOF. B y Perron' s formul a (5.1 1 1 ) w e obtai n

k
Dk(x) = — / C (s)-ds + Oix^T- 1 ),
2
™ Ji+e-iT S
where T i s an y numbe r wit h 2 ^ T < x. Mov e th e integratio n t o R e (s) = cr ,
\ < a < 1 , getting b y (8.93 )

Dfc(x) = re s C fc(s)— + 0((x ( 7 T a ; c ( 1 - < j ) § + x T " 1 ) ^ ) .

3=1 5

The residue i s equal t o xPk(logx) an d th e erro r ter m become s 0(x 6k+2s) b y settin g
T = x k'2/3 an d a = 1 - ( 2 a / ^ ) ~ 2 . •

Recall tha t fo r smal l k on e get s stronge r result s usin g va n de r Corpu t method .

"i-i
It i s conjecture d tha t th e bes t exponen t i n (8.96 ) i s 5k = ^iT2k
f° r a n y ^ -
This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/10

CHAPTER 9

THE DIRICHLE T POLYNOMIAL S

9.1. Introduction .
A Dirichle t polynomia l i s a finite Dirichle t serie s
anTl
(9.1) D(s)= Yl ~S

with comple x coefficient s a n. Thi s i s a specia l cas e o f sum s o f typ e ^ a n e _ A ^ n ^ s

where A(n ) ar e distinc t rea l number s calle d "frequencies. " Ou r specia l cas e X(n) —
log n i s distinguished b y th e followin g features :
- X(n) i s smoot h an d slowl y increasing ,
- X(n) i s additive .
The derivativ e o f D(s) i s als o a Dirichle t polynomia l o f lengt h TV,
a
D'(S)= Yl 'nn~S>

its coefficient s o! n — —a^logn ar e change d onl y slightly . Th e additiv e propert y o f

our frequencie s implie s tha t th e produc t o f two Dirichlet polynomial s i s a Dirichle t
polynomial, namel y w e hav e

( E a «n_s)( ^ b ™m-s) = X > r s

with L = MN an d th e coefficient s o f th e produc t ar e

(9.2) a= ^2 a nbm.
nm=£

The main objectiv e o f the theor y o f Dirichlet polynomial s i s to estimate D(s) a t

special points. Sinc e the set of special points is not constructible i n practice (thin k of
the zero s p = / 3 + i7 o f ((s) wit h (3 > | ) i t i s necessary t o deal with arbitrar y points .
These ca n b e electe d b y combinatoria l argument s int o well-space d sets . Moreover ,
the coefficient s a n ar e quit e complicate d i n applications , therefor e on e ha s t o trea t
them a s arbitrar y comple x numbers . Thoug h i n suc h generalit y i t i s impossibl e t o
give a non-trivia l boun d fo r D(s) a t an y fixed 5 , i t i s tru e tha t |D(s) | canno t tak e
large value s fo r man y well-space d points . Henc e ou r fundamenta l questio n wil l b e
how ofte n \D(s)\ i s large r tha n th e mea n valu e ove r th e relevan t se t o f points .
If s = a + it ha s fixed cr , the n b y changin g th e coefficient s a n int o a nn~<7 w e
can assum e withou t los s o f generalit y tha t al l th e point s i n questio n ar e o n th e
229
230 9. T H E D I R I C H L E T P O L Y N O M I A L S

imaginary lin e s = it. Afte r establishin g result s fo r points on the imaginary lin e we
shall exten d the m fo r sets o f well-spaced point s i n vertical strips . Actuall y w e can
also modify th e Dirichlet polynomia l D(s) by twisting its terms by various function s
fs(n) dependin g o n s which chang e onl y slightl y i n n (hav e smal l derivative s i n n),
that i s to say, the theory extend s t o sums o f type

(9.3) ^cinfs(ri)n~

Such a twis t ca n be handle d usin g an y standard Fourie r analysi s (separatio n o f

variables i n / s ( n ) , se e Proposition 9.1 1 ) .
9.2. Th e integral mean-valu e estimates .
By Cauchy' s inequalit y w e ge t

(9.4)
l<n<N
E ani <GN

where

(9.5) G= > El a J 2
.

Of course , thi s boun d i s best possible , however , on e can do better o n average.

THEOREM 9.1 . For any complex numbers a n we have
rT, | 2
il
(9.6) /J ]T a nn dt = (T + 0(N))G
^ ° l<n<N

where the implied constant is absolute.

PROOF. Le t f(t) b e the following continuou s piecewis e linea r functio n
'0 if t ^ -TV ,
1 + -*- if - N < t ^ 0 ,
1
^ N
/(*) 1 if 0 < ^ T ,
1
N if T < t ^ T + JV,
<0 if t > T + N.

Then ou r integral i s majorized b y

u 2
f f(t)\ E a nn \ dt= Yl a manF{^\
l<n<N l^m,n^AT

with
F(x)= f f(t)x ltdt.
We hav e F ( l ) = T + N an d F(x) < iV - 1 (logx)~ 2 i f x ^ 1 . Henc e fo r m ^ n we
have
mfm\A T i A rn\-1
2
/ r a + n \2 N
F — «CA^ _1 log — < C— < T r^ .
\nJ V nJ N\m — nJ (m — n) z
 9. TH E DIRICHLE T POLYNOMIAL S 231

Therefore ou r integra l i s bounde d fro m abov e b y

(T + N)G + O^N Yl Wma n\(m - n)~ 2 ) = ( T + 0(N))G

l^m/n^iV

by using th e inequalit y 2\a rnan\ ^ |a m | 2 + |a n | 2 . Similarl y w e derive the sam e lowe r

bound. Combinin g bot h estimate s w e complet e th e proo f o f (9.6) . •

This resul t shoul d b e compare d wit h th e discussio n o f bilinea r form s an d larg e

sieve i n Chapte r 7 .
Theorem 9. 1 implie s tha t G? i s th e mean-valu e o f \D{ii)\ i n th e followin g
asymptotic sens e

(9.7) li m i / \D(t)\ 2dt = G.

T->oo T J 0
Note tha t th e mean-valu e o f th e produc t o f Dirichle t polynomial s i s essentiall y
bounded b y th e produc t o f mean-values . Precisel y w e deriv e fo r th e coefficient s
(9.2) th e followin g inequalities :

X ^ l 2 ^ ^2Y1 \ ant>ma>nibmi\
£ nm=nimi
a
^Y1^2 \ nbm\2r(nm)
nm

<(£M2TW)(£M2TM).
nm

Hence usin g th e boun d r(n) < C n£ w e obtai n

(9-8) ] > > , | « (MJV) £ (5> n | 2 )(£|& m | 2 ).

For th e produc t o f k polynomial s w e hav e

( E <4V* ) • • • (E ^ fc)»-J) = E <•«» -

n^iVi n^N k n^N
with N = Ni ... Nk an d th e coefficient s ar e

««=E--E<--e
ni---nk=n
The mean-valu e o f thes e coefficient s satisfie s
(fe)
E i ° n i 2 < E " - E K ) - " a ^ ) W n i •••"* ) ^ ^ - - - ^A;
n ni n k

where fo r a give n sequenc e A — (a n) w e pu t

(9.9) G k = £
n
This follow s b y th e inequalit y Tfc(ni---rifc ) ^ ^ ( n i ) • • • Tfc(rifc). Sinc e Tjt(n ) i s
multiplicative, i t suffice s t o chec k thi s inequalit y fo r rij — paj. I n thi s cas e w e
have Tk(p a) — ( 1 + a ) ( l + a/2 ) • • • ( ! + a/(f c - 1 )) , s o i t suffice s t o chec k tha t
232 9. TH E DIRICHLE T POLYNOMIAL S

1+ OiH h ak)/£ ^ ( 1 + OL\jf) • • • (! + a k/tj fo r 1 ^ £ < k, whic h i s obvious .

Note tha t
£
Gk < G7V
by Tk(n) < C ne , wher e th e implie d constan t depend s o n s an d fc; however, w e shal l
often i n practic e hav e a bette r estimat e G k < C G(\og2N)K wit h som e K ^ k.
The approximat e formul a (9.6 ) i s th e bes t possibl e o f it s kin d i n general . I t i s
best use d fo r polynomial s o f lengt hT V whic h i s nea r T i n th e logarithmi c scale . I f
the polynomia l D(s) i s quit e shor t w e ca n bette r appl y (9.6 ) t o a suitabl e powe r
D(s)k gettin g

COROLLARY 9.2 . For any integer k ^ 1 we have

(9.10) /
J0
/ V| Yla a -n \ alt
n
lt
< ( T + N k)Gkk
l<n<N
where G k is given by (9.9) and the implied constant is absolute.
Still w e ma y no t b e abl e t o matc h T an d N k fo r k a n integer , s o (9.1 0 ) i s
not perfec t i n som e ranges . I n thi s connectio n H . Montgomer y [Mo2 ] mad e th e
following
CONJECTURE M k. Suppose \a n\ < 1 . Then for any real k with 1 ^ k ^ 2 we
have

u k
(9.11) / ] Ta nn \ dt<^(T + N )Nk+£

where the implied constant depends only on e.

It i s no t difficul t t o prov e (9.1 1 ) fo r som e specia l polynomials , fo r example ,
when a n — 1 for al l n. However , th e boun d (9.1 0 ) fo r genera l comple x number s a n ,
with N k+£ replace d b y G kN£, wa s show n b y J . Bourgai n [Bou ] t o b e fals e i f k i s
not a n integer. Hi s counter-example i s given by a Dirichlet polynomia l supporte d i n
a shor t interval . Bu t suc h case s ar e no t commo n i n applications , s o the Conjectur e
Mfe, i f true, i s still important . I t yields , amon g othe r things , th e densit y conjectur e
for zero s o f th e Rieman n zet a functio n (se e Chapte r 1 0) .

9.3. Th e discret e mean-valu e estimates .

Next w e establis h discret e version s o f Theore m 9. 1 an d som e o f thei r conse -
quences.
DEFINITION. A set T o f real numbers t r i s said to be well-spaced if \t ri —t r2 \ ^ 1
for an y r\ ^ r<i.
LEMMA 9. 3 (GALLAGHER) . Let T be a set of points t r with \ ^ t r < T - \
which is well-spaced. Let F(t) be a smooth function on [0,T] . Then we have

(9.12) J2 \ F^)\2 <

Jo
[ T(\F(t)\2 + \F(t)F'(t)\)dt.
trer
 9. T H E D I R I C H L E T P O L Y N O M I A L S 23 3

P R O O F . Fo r a smoot h functio n f(x) o n [0,1 ] w e get th e identit y

f(x)= J f(t)dt+ C tf{t)dt + [ (t-l)f(t)dt

JO JO Jx
by partia l integration . Henc e

l/(|)K f\\f{i)\ + \\f\t)\)dt.

Jo
Applying thi s fo r f (t) w e infe r

i/(|)i 2 < / (\f(t)\ 2 + \f(t)f(t)\ dt

Jo
This give s
\F(U)\2^ r +2
(\F(t)\2 + \F(t)F'(t)\)dt,
Jtr-i
and summin g ove r t r i n T w e deduc e (9.1 2 ) becaus e th e interval s (t r — \,tT + | )
do no t overlap . D

Notice tha t b y th e Cauchy-Schwar z inequalit y w e hav e

2
(9.13) f \F(t)nt)\dt^{J \F(t)\ dtY(J \F'(t)\ 2dty.

Applying (9.1 2 ) togethe r wit h (9.1 3 ) fo r F(t) = D(it) on e get s

THEOREM 9.4 . Let T be a set of well-spaced points in the segments [0,T ] with
T ^ 1 , and let a n be any complex numbers. Then we have
a
(9.14) ] T | J2 nnltr\ <^{T + N)Glog2N
tr£T l^n^N

where G is given by (9.5) and the implied constant is absolute.

COROLLARY 9.5 . Let the conditions be as in Theorem 9.4- For any integer
k ^ 1 we have
lU
(9.15) ] T j Y^ ann \< C (T + N k)Gkk \og2N k

where Gk is given by (9.9) and the implied constant is absolute.

REMARKS. Gallagher' s lemm a yield s a transitio n fro m a n individua l valu e
D(it) t o a n integral . I t employ s th e derivativ e D'(it) whic h i s no t acceptabl e
in som e applications . However , w e ca n mak e a transitio n withou t differentiatio n
of D(s). T o thi s en d tak e a smoot h functio n f(x) > 0 supporte d o n [|,2iV ] wit h
f(x) = 1 on [1 ,7V ] suc h tha t x af{a\x) < 1 for 0 ^ a ^ 2 . B y Melli n inversio n

f(x) = —J F{t)x«db
where
Fit) = f f(x)x it-1 dx < C (1 + |i|)" 2 log 2N
234 9 . TH E DIRICHLE T POLYNOMIAL S

by partia l integration . W e multipl y a n b y / ( n ) , whic h i s a redundan t facto r i f

1 ^ n ^ TV , and appl y th e abov e integra l representatio n o f f(n) gettin g
1 f°°
D(s) = — D(s + it)F(t)dt.
2?r J_ 00
Hence
oo
2
\D(it)\(l + \t-t r\)- dt
/ -OO

where th e implie d constan t i s absolute . Th e sam e argument s appl y t o D(s) k wit h

k a positiv e intege r givin g
oo
\D(it)\k(l + \t-t r)-
2
dt
/ -oo

where the implied constan t i s absolute. Henc e the estimates (9.1 4 ) an d (9.1 5 ) follo w
from (9.6 ) an d (9.1 0 ) respectively .

The bound (9.1 4 ) has deficiency o f not dependin g on the cardinality of the set T ;
it i s relatively weake r fo r smalle r sets . Montgomer y [Mon ] succeede d i n improvin g
(9.14) i f R = \T\ i s smalle r tha n T* (not e tha t an y well-space d se t T £ [0,T ] ha s
no mor e tha n T + 1 points). Th e ide a i s based o n th e dualit y principl e fo r bilinea r
forms:
Let X — (xmn) b e a comple x matrix . The n th e followin g statement s abou t X
and a numbe r D ar e equivalent :
(A) Fo r an y comple x number s a nj

I, 2
2
^2 \^2 anXrnn\ ^ D^|a n | .
mn n

(B) Fo r an y comple x number s 6 m ,

.| 2
2
^2\^2brnXrnn\ < ^ ^ | 6 m | .
nm
For th e proof , se e Chapte r 7 .
THEOREM 9. 6 (MONTGOMERY) . Let the condition be as in Theorem 94- Then
we have
a
(9.18) ] T | J2 nnitr\2 < G(N + RT*) log TT
treT l^n^N

where the implied constant is absolute.

PROOF. B y th e dualit y principl e th e proble m i s equivalent t o showin g tha t fo r
any comple x number s c r on e ha s

( 9 - 19 ) J2 \J2 c
r ^ f < (Yl \cr\ 2
)(N + RT*)log2T.
l^n^N t reT t r£T
 9. TH E DIRICHLE T POLYNOMIAL S 235

Squaring ou t an d changin g th e orde r o f summation o n the lef t sid e o f (9.1 9 ) w e ge t

C
/j / j r\ c r2 ^ \*ri ~ tr 2 )
r\ r 2
where
nlt
z(t)= Yl -
By (8.20) , (8.26 ) an d (8.34 ) w e deriv e tha t
1
(9.20) Z(t) < Nlt]' + \t\ 2 lo g |£|, i f \t\ ^ 1 .
This yield s (9.1 8) . •

By th e Lindelo f Hypothesi s £( ^ + it) < C \t\ £ fo r \t\ > 1 , on e shoul d expec t a

better boun d fo r Z(t) tha n (9.20) , namel y tha t
1
(9.21) Z(t) < N^- + N* \t\ £, fo r |t | ^ 1 .
Hence w e get
lU £
(9.22) J2 Yl ctnn \ <^G(N + RN2)T

in plac e o f (9.1 8) . Bu t a n eve n stronge r boun d i s predicte d b y Montgomer y i n th e

following
CONJECTURE M. Suppose \a n\ ^ 1 . Let T be a set of R well-spaced points t r
in the segment [0,T ] with T ^ 1 . Then
a
(9.23) V | J2 nniu\ <^N(N + R)T £.

EXERCISE. Sho w tha t Conjectur e M implie s Conjectur e Mk fo r an y rea l 1 ^

k < 2 (wit h extr a facto r T £).
9.4. Larg e value s o f Dirichle t polynomials .
The discret e mean-valu e estimate s offe r som e answers to the questio n ho w ofte n
a Dirichle t polynomia l assume s larg e value s a t a give n se t o f well-space d points .
Thus (9.1 4 ) yield s
THEOREM 9.7 . Let D(s) be a Dirichlet polynomial of length N > 1 and let T
be a set of well-spaced points t r in the segment [0 , T] with T ^ 1 such that
(9.24) \D(it r)\ ^ V
for some V > 0 . Then the cardinality ofT, say R= \T\, satisfies
(9.25) R<z:(T + N)GV~ 2 lo g 2N
where the implied constant is absolute.
Notice tha t th e boun d (9.25 ) i s non-trivia l i f V ^> G? \og2N. Fo r V muc h
larger tha n thi s Montgomer y go t a bette r resul t b y applyin g (9.1 8) .
23 6 9. T H E D I R I C H L E T P O L Y N O M I A L S

THEOREM 9.8 . Let the conditions be as in Theorem 9.7 with

(9.26) V^G?T*log2T.

Then
2
(9.27) # < c G W " log2T

where the implied constant is absolute.

P R O O F . W e ca n assum e tha t T i s large r tha n N b y a larg e constan t factor ,
because otherwis e (9.27 ) follow s b y (9.25) . Insertin g (9.24 ) int o (9.1 8 ) w e ge t

RV2 <z:G(N + RT^) log2T .

This implie s (9.27 ) unde r th e restrictio n (9.26) . •

The restrictio n (9.26 ) ca n b e removed a t a cos t o f introducing a n extr a ter m i n

the boun d (9.27) .
COROLLARY 9. 9 ( H U X L E Y ) . Let the conditions be as in Theorem 9.7. Then

(9.28) R < {GNV~ 2

+ G 3 iV7V- 6 }(log2T) 6

where the implied constant is absolute.

P R O O F . Huxley' s ide a (th e subdivisio n method ) i s a simpl e tric k o f dividin g
the se t T t o mee t th e conditio n (9.26) . Firs t w e can assum e tha t

(9.29) V^G*log2T
1 i
because otherwis e th e assertio n (9.28 ) i s trivial . The n w e hav e V ^ G*T 04 lo g 2T0
with T 0 = m i n { T , G - 2 y 4 ( l o g 2 T ) - 4 } . Th e conditio n (9.29 ) ensure s u s tha t 1 <
T0 ^ T . Le t Te b e th e subse t o f point s o f T i n th e interva l [IT, (£ + 1 )T 0] wit h
0 ^ I ^ TT 0 _1 . Fo r eac h subse t Tt> Theore m 9. 8 i s applicabl e givin g R £ = \Te\ <
GNV~2\og2T. Henc e
l
R = S£RI4Z TT^ GNV~2 log2 T
t

which yield s (9.28) . •

Applying th e abov e result s fo r th e polynomia l D{it) k wit h k a positiv e intege r

one get s
k
(9.30) i J < (T + N )(GkV-2)k\og2Nk,
(9.31) R < {{G kNV~2)k +T(G 3
kNV-
6 k
) }{log2Tf

where th e implie d constan t i s absolut e (se e (9.25 ) an d (9.28 ) respectively) .

In 1 975 , M . Jutil a cam e u p wit h severa l innovation s t o improv e th e abov e
estimates i n som e ranges . Hi s bes t resul t (whic h w e stat e her e withou t proof ,
see [Ju3] ) i s
 9. TH E DIRICHLE T POLYNOMIAL S 23 7

THEOREM 9.1 0 ( J U T I L A ) . For any positive integer k we have

„3 2 , ^{-.(^-^(f)^},^
for any e > 0, the implied constant depending only on e and k.
REMARKS. Notic e tha t i f
e
(9.33) V ^G*N*(NT)
then Jutila' s boun d (9.32 ) become s essentiall y Huxley' s boun d (9.31 ) b y lettin g k
be sufficientl y larg e i n term s o f e.

The bes t (conditional ) estimate s fo r th e numbe r o f larg e value s o f a Dirichle t

polynomial wit h bounde d coefficient s come s fro m Montgomery' s conjectures . B y
Conjecture M/ ~ one derive s
k
(9.34) R^(T + N )NkV-2k(NT)£
for an y rea l k ^ 1 an d e > 0 , th e implie d constan t dependin g o n e an d k. In
applications (9.34 ) i s best use d fo r k wit h N k = T . B y Conjectur e M on e derive s

(9.35) R < N 2V~2T£, i fV > N^T £

.
Actually th e abov e estimate s fo r R ar e equivalen t t o th e Montgomer y conjectures .
All th e discret e mean-valu e estimate s wit h respec t t o a well-space d se t T —
{ti,... , IR} C [0 , T] ca n b e generalize d slightl y t o sum s o f typ e

(9.36) ] £a nfr(n)n^

where f r(n) i s a nic e functio n whic h doe s no t var y i n n to o much .

PROPOSITION 9.1 1 . Suppose that each f r{x) satisfies
a
(9.37) x \tia\x)\^2
for 1 ^ x ^ N and 0 ^ a ^ 2 . Suppose that

(9.38) I Y,a nfr(n)n^ ^v

l^n<N

for 1 ^ r< R. Then the bounds (9.30) and (9.31 ) hold true with Gk replaced by
Gk(log2N)2. In particular, / r ( n ) = n~ Gr with 0 ^ a r ^ 1 qualifies.
P R O O F . W e ma y assum e tha t f r(x) i s supporte d i n [|,2iV ] wher e i t satisfie s
(9.37). The n w e writ e
lt
fAn) = — / f r(it)n- dt

where f r{s) i s th e Melli n transfor m o f f r(x),

S l
fr{s) = / fr{x)x ~ dx.
JO
238 9. T H E D I R I C H L E T P O L Y N O M I A L S

This satisfie s f r(it) < C (|t | + 1 ) 2 log2A r b y partia l integratio n tw o times . Henc e
we ge t
oo
\D(itr
l<n<N /

« (lo g2AT) ]T \D(it r(m))\(\m\ + 1)"

where t r{rn) = t r — t i s th e poin t a t whic h \D(it r — it)\ i s maxima l wit h m < t ^

m + 1 . Notic e tha t |£ r (ra)| < T + \m\ + 1 . B y th e hypothesi s (9.38 ) i t follow s tha t

1
\D(itr(m))\ > (|ra | + l ^ ^ k ^ A O "

for som e m G Z. Le t 7^ i b e th e se t o f suc h point s t r(m) an d Rm it s cardinality ,

so we hav e R ^ ]T m i? m . Th e se t Tm ca n b e divide d int o tw o well-space d subsets .
Applying (9.30 ) an d (9.31 ) fo r each of these subsets we obtain correspondin g bound s
for R rni an d summin g ove r m w e get th e asserte d bound s fo r R. •

9.5. Dirichle t polynomial s wit h characters .

A larg e par t o f th e theor y o f Dirichle t polynomial s (9.1 ) extend s t o sum s o f
type

a
(9.39) D(s, X)= J2 nX(n)n-s

where % is som e sor t o f arithmeti c harmonic , fo r example , v(n ) ca n b e a Dirichle t

character, o r th e Fourie r coefficien t o f a modula r for m (se e Chapte r 1 4) . However ,
in th e latte r exampl e on e i s face d wit h technica l difficultie s cause d b y th e lac k o f
complete multiplicativit y (th e generatin g L-functio n ha s Eule r produc t o f degre e
two) whil e th e orthogonalit y propert y i s onl y approximate .
In thi s sectio n w e dea l wit h th e Dirichle t character s t o variou s moduli . Le t
k ^ 1 and Q ^ 1 . W e conside r th e se t H(k,Q) o f character s x ( m o d kq) suc h tha t
X =£V > where £ is an y characte r t o modulu s k an d ip is an y primitiv e characte r t o
modulus q with 1 ^ q ^ Q , (q, k) = 1 . Throughou t w e assum e tha t T ^ 3 , iV ^ 1 ,
and w e denot e

(9.40) H = kQ 2T, C = lo g HN.

First w e establis h th e followin g extensio n o f Theore m 9.1 .

THEOREM 9.1 2 . For any complex numbers a n we have

lt 2 3
(9.41) ] Tf \ Yl anX(n)n \ dt^:G(N + H)C

where the implied constant is absolute.

 9. TH E DIRICHLE T POLYNOMIAL S 239

REMARKS. Ou r estimat e (9.41 ) i s not a s strong a s it coul d possibl y be , namel y

the logarithmi c facto r C? could b e remove d b y takin g a traditiona l approac h base d
on th e larg e siev e inequalit y fo r additiv e characters . Th e transitio n fro m multi -
plicative t o additiv e character s bein g performe d b y Gaus s sum s (se e Chapte r 3) ,
this approac h ha s the extr a advantag e o f working well for sequence s (a n) supporte d
in shor t segment s M < n < M + N. Nevertheless , w e have chose n a direc t metho d
simply t o sho w ne w ideas . Her e th e quantit y H = kQ 2T i s abou t th e numbe r o f
harmonics x( n)ntt bein g employe d (thin k o f t a s a discret e variabl e rangin g ove r a
well-spaced se t o f point s wit h |£ | < T) .
P R O O F . Fo r th e proof o f (9.41 ) w e firs t assum e tha t (a n ) i s supporte d i n a
dyadic segmen t X ^ n ^ 2X wit h X ^ 1 . B y th e dualit y principl e th e proble m
reduces t o th e estimatio n o f
2
(9.42) E | E / c x(t)X(n)n^dt

for an y comple x number s c x(t). W e enlarg e th e oute r summatio n b y introducin g a

smoothing functio n f(n) ^ 0 with f(n) ^ 1 for X ^ n ^ 2X. The n squarin g ou t
we se e tha t th e dua l for m (9.42 ) i s bounde d b y

(9.43) E E // \c xAh)cxAh)B(t1 -t2,XiX2)\dtidt2

i 0 i 0
Xi X2 ''

where

(9.44) 5 ( t , x ) = £ / ( n ) x ( n ) n".

Here \ = X1 X 2 is a character o f modulus £ = k[qi,q2] < kQ 2 an d \t\ = \t\ — ^1 ^ T.

By Poisson' s summatio n

s(«^E<K)
where G i s the Gaus s su m

x7
a(mo d t)
lt
and F i s the Fourie r transfor m o f f(x)x 1

•>o o
it
F(y) = / f(x)x e(xy)dx.
Jo
Have i n min d tha t F(y) i s als o a functio n o f t. W e choos e f(x) suc h tha t

(9.45) F ( 2 / ) « X e x p ( 1*
L1 l/yX\\-
- ( ^ r ) 5 ) ,i f y > 0.

For this , ther e ar e man y goo d choices , fo r example ,

(9.46) /(I ) = „„(»_£_!)

24 0 9. TH E DIRICHLE T POLYNOMIAL S

satisfies al l th e requirements . Fo r th e proo f o f (9.45) , w e mov e th e pat h o f in -

tegration i n th e Fourie r integra l o f f(x)x lt fro m R + t o e 2^lR+ wit h 0 < 0 < \
getting
f°° / 5\
\F(y)\ ^X exp(- + 0\t\ - (x + x~ l) cos6 - 2irxyX sm0)dx.

Since x~ l co s 0 + 2irxyX si n 0 ^ {nyX si n 20) *, thi s yield s

\F(y)\ ^ X e x p ( ^ + 0\t\ - (TH/ X sin 20) * ) (costf)" 1 .

We choose 0 = T~ l gettin g (9.45) .

By (9.45 ) w e deriv e tha t (b y th e trivia l boun d \G(h/t)\ < £)

4
(9.47) B(t, X) = S X^1 F(0) + 0(fTexp(| - ( * ) ))

where
/•OO

(9.48) F(0 ) = / f(x)x u

dx < C X(|t| + l ) " 2 ,

£x = 1 if x i s principal , an d 5\ — 0 otherwise . Not e tha t % = X1 X 2 i s principa l

if an d onl y i f \i = X 2 (thi s i s th e plac e wher e ou r hypothesi s tha t th e character s
^(mod q) ar e primitiv e i s essential) . Insertin g (9.47 ) an d (9.48 ) int o (9.43 ) w e
estimate (9.42 ) b y
T
{X + H 2eM-(X/H)l>)} £ [ T c ,~dt.
\ x(t)\
0
xen(k,Q)'

Since thi s hold s fo r an y comple x number s c x(t), i t follow s b y dualit y tha t fo r an y

complex number s a n ,
(9.49)
a
E/ | E a nx(n)n
u
\\t<£{X + H 2eM-(X/H)")} E l -| 2 '
xeH(fc,Q) ° X<n^2X X<n^2X

This resul t i s a s goo d a s (9.41 ) provide d X ^ H(logH) 2, bu t i t i s ver y poo r

for smalle r X. Not e tha t w e establishe d (9.49 ) onl y fo r sum s ove r n i n dyadi c
segments. T o deriv e a resul t fo r shor t sum s w e shal l increas e X artificially , whil e
preserving th e dyadi c shap e o f th e rang e o f th e characte r su m (thi s ide a i s du e t o
E. Bombieri) . W e begi n wit h th e inequalit y

] T logp ^ ^2 logp-log/e g ^ \P
P<p^2P P<p^2P
p\kq

which hold s fo r P ^ 9 log kq. Henc e th e lef t sid e o f (9.49 ) i s bounde d b y

2 lo u 2
~P E( SP) E / I E a nX(pn)(pn) dt.
tJU
p<P^2P xen(k,Q) x<n<:2X
 9. T H E D I R I C H L E T P O L Y N O M I A L S 241

Here we used th e multiplicativit y o f our harmonic s x( n)n%t t o replace n b y m = pn.

Now m range s ove r th e dyadi c segmen t pX < m ^ 2pX, s o applying (9.49 ) wit h X
replaced b y pX w e get a ne w estimat e
2
{PX + H 2 exp(-(PX/H)i)} ^ |a n|
X<n^2X

where P i s an y numbe r ^ 9logkQ. W e choos e P = ( 9 + HX~ 1 )(logH)2 showin g

that

u] 2
E/ | E *nX{n)n \dt^{X + H){\ogH) 2 £ |a n| .
xeH(k,Q)^° X<n^2X X<n^2X

Finally we replace the dyadic segment X < n ^ 2X b y the whole interval I ^ n ^ N

by subdividin g th e latte r an d applyin g Cauchy-Schwar z inequality . Thi s produce s
an extr a facto r log2iV , an d i t complete s th e proo f o f Theore m 9.1 2 . •

To ever y x G W ft Q) w e associate severa l point s (possibl y none )

(9.50) S r{x) = <Tr(x) + MX) w i t h 0 < <T r(X) ^ 1 , | * r ( * ) | ^ T.

Let S(k,Q,T) b e th e se t o f al l suc h point s counte d wit h relevan t multiplicity .

We sa y tha t <S(fc , Q,T) i s well-space d i f fo r an y s ri(xi), s r2(\2) h i 5(fc , Q,T) wit h
Oi,Xi) ¥= ( r2,X2) w e have either x i 7^X2 , or x i = X 2 = X and \t ri(x)-tr2(x)\ ^ 1 -
Clearly, i f 5(fe , Q,T) i s well-spaced, the n it s cardinalit y satisfie s

(9.51) R = \S{k, Q,T)| < 3A;Q 2r = 3ff .

Nowhere in the proof o f Theorem 9.1 2 did we use essentially th e continuous measur e
in t , therefor e th e sam e estimate s hol d tru e i f on e replace d th e integratio n b y a
summation ove r an y se t o f well-space d point s £ r (x)- The n on e ca n chang e n ltr^
into n~ Sr^ a t th e cos t o f a n extr a facto r (lo g 27V)2 i n (9.41 ) b y th e argument s
from th e proo f o f Propositio n 9.1 1 . Thi s wa y Theore m 9.1 2 become s
THEOREM 9.1 3 . Let S = S(k,Q,T) be a well-spaced set of points (9.50). For
any complex numbers a n we have

(9.52) Yl I E a nX(rc)rc~ ar(x) | 2 <G(7 V + # ) £ 5

sr(x)es I ^ J V
where the implied constant is absolute.
Theorem 9.1 3 i s a n extensio n o f Theore m 9.4 . Nex t w e establis h a n extensio n
of Theore m 9.6 , namel y
THEOREM 9.1 4 . Let S = S(k,Q,T) be a well-spaced set of points (9.50) of
cardinality R. For any complex numbers a n we have

3
(9.53) Yl I E anX(n)n- 'M\2 <&G(N + RHi)£*

where the implied constant is absolute.

242 9. T H E D I R I C H L E T P O L Y N O M I A L S

P R O O F . Thi s goe s b y modification s o f ou r argument s fro m th e proo f o f Theo -

rem 9. 6 an d Theore m 9.1 2 . Firs t w e assum e tha t cr r(x) — 0> i- e-> s r(x) = ^r(x)-
Then b y dualit y w e reduce th e proble m t o th e estimatio n o f

(9.54) V = ]T / ( „ ) |£ ]T cx(tr)X(n)n^\2
r
nX

for an y comple x number s c x (t r ), wher e f(n) is any non-negativ e functio n o n R +

such tha t / ( n ) ^ 1 if 1 ^ n ^T V (in wha t follow s w e assum eT V ^ 2 , otherwise
(9.53) i s trivial). Th e dua l for m satisfie s

(9.55) V < X i E E E h x . ^ i K ^ r J i ?^ -tr 2,XlX2)\

Xi Xi n r 2

where B(t, x ) i s given b y (9.44) . A t thi s poin t w e are goin g to refin e th e estimatio n
(9.47). W e writ e

L
(9.56) B(t, X) = TT: f (s- it , X)f(s)ds
27TZ J ( 2 )

where L(s,x ) i s the Dirichle t L-functio n an d f(s) is the Melli n transfor m o f /

/•OO

f(s) = / f(x)x 1 s
- dx.
Jo
Moving th e integratio n t o th e imaginar y lin e w e deriv e

(9.57) B(t, X) = Sx^f-f(l + zt) + 0(£H\t\ + l)*(logJW(|t| + l)) 2 ).

Here th e leadin g ter m come s fro m th e pol e o f L(s — it, x) a t s = 1 + it if x i s

principal (o f course , i t agrees wit h tha t i n (9.47)) an d th e erro r ter m i s obtaine d
from tw o estimate s

(9.58) L(s,x ) < ^ ( W + 1)* log*(|s| + 2)

(see Exercis e 9 in Sectio n 5.9 ) an d
ls|
(9.59) / ( 5 ) < e " logTV
on R e s = 0. T o se e th e latte r tak e
(9.60) f(x) = h[e-x'N - e~ x)
so f(s) = 5T(s)(TV s - 1 ) . Henc e w e als o ge t / ( l + it) < TVe"'*' . No w (9.57 ) show s
that th e dua l for m satisfie s

Xr

and thi s prove s (9.53 ) wit h th e logarithmi c facto r C 2 i n plac e o f £ 4 , bu t onl y i f al l

the point s s r(x) are imaginary . The n w e change n lt int o n~ a~lt a t th e cos t o f tw o
extra logarithm s completin g th e proof . •

There ar e tw o essentia l ingredient s i n ou r proo f o f Theore m 9.1 4 , namel y th e

reduction t o th e dua l for m (9.54 ) followe d b y the estimatio n (9.57 ) o f the characte r
 9. T H E D I R I C H L E T P O L Y N O M I A L S 24 3

sum B(x,t). However , thi s estimatio n i s no t th e bes t possible . Assumin g th e

Lindelof hypothesi s

(9.61) L(s,x)<&Ws\Y
for R e s = \ an d x(nio d £) one deduce s (b y movin g th e integratio n t o th e critica l
line rathe r tha n t o th e imaginar y one ) tha t
<p]f)
(9.62) B(t, x) = tfx^V/(I + it) + 0(N*e°(\t\ + l) £ ).

This yield s
SrM S
(9.63) J2 I Yl anX(n)n- \< C G(iV + RN*H )C4

in plac e o f (9.53) . Th e result s (9.53 ) an d (9.63 ) ar e du e t o H . Montgomer y [Mo2] .

9.6. Th e reflectio n method .

The Lindelo f hypothesi s ma y be an open problem fo r a long time. Nevertheless ,
M. N . Huxle y [Hu2 ] ha s succeede d t o establis h unconditionall y a resul t whic h i s
almost a s goo d a s (9.63 ) i n som e importan t ranges . Firs t h e di d i t fo r polynomial s
D(s) i n Corollar y 9. 9 b y th e subdivisio n trick . However , thi s tric k doe s no t mak e
sense fo r polynomial s D(s,x) twiste d b y characters . Fo r thes e polynomial s Huxle y
invented a completel y differen t brillian t tric k whic h h e call s th e reflectio n metho d
for reason s soo n t o b e clear . Usin g Huxley' s idea s w e ar e goin g t o prov e
THEOREM 9.1 5 . Let the conditions of Theorem 9.1 4 hold true. Then
sAx)
(9.64) ] TI Y, a nX(n)n- \< G(N + iJ§ H*N*)£ 6

where the implied constant is absolute.

For th e proo f w e can assum e tha t th e coefficient s a n ar e supporte d i n a dyadi c

segment N ^ n ^ 2N ', an d a s before , w e ca n als o assum e tha t al l th e point s s r (x)
are o n th e imaginar y line . Th e latte r simplificatio n cost s a los s of factor (lo g 27V)2,
so w e must sho w (9.64 ) wit h C 4 i n plac e o f £ 6 .
The proble m reduce s t o th e estimatio n o f th e dua l for m V. I n th e proo f o f
Theorem 9.1 4 w e have estimate d V fo r an y comple x coefficient s c x (t r ), however , i t
is sufficient t o dea l wit h
a
(9.65) c x(tr) = D(s r(X),x) = ^2 nX(n)n-s^K
n

Precisely, lettin g C be th e lef t sid e o f (9.64) , w e hav e

2
(9.66) C ^ VG
(see the derivatio n o f the dualit y principl e i n Chapte r 7) . No w we are goin g to tak e
advantage o f these specia l coefficient s (9.65) . B y (9.55 ) w e are agai n face d wit h th e
character su m B(x,t). W e shal l no t improv e th e erro r ter m i n (9.57 ) bu t rathe r
replace i t b y anothe r characte r sum ; i n othe r words , w e us e a kin d o f summatio n
24 4 9. T H E D I R I C H L E T P O L Y N O M I A L S

formula (se e Chapter 4 , i n particula r (4.26)) . W e derive suc h a formul a fro m a

functional equatio n fo r the corresponding L-function .
Without los s of generality, w e can assume tha t al l the characters i n C have the
same parit y (b y dividing C into tw o sums i f necessary), s o the character \ — X1 X 2
in B(x,t) i s alway s even . I f x ( m o d £) i s primitive , the n w e have th e functiona l
equation (4.73 )
s
(9.67) L(s, X)=ex"f(s)el- L(l-s,x)
where \e x\ = 1 an d

(9.68) 7(s )= ^ - l r ( ^ ) / r ( | ) .

Suppose x(mod £) is induced by the primitive character x*(mo d £*) with £*\£. Then

L( 5 , X ) = L ( 5 , x * ) n ( 1 - X * ( p ) P ~ s ) .
p\£

Hence th e functiona l equatio n (9.67 ) fo r L(s,x* ) yield s th e followin g functiona l

equation fo r L(s, x),
(9.69) L{s,x) = e x^(s)P(s)L(l-s,x)

where e x = s x* and
X
(9.70) P(s) = {t)^ s I t( - X*(P)P~*)( l ~ X'WP*- 1
1
)- .
p\£

Notice tha t 7(5) , P(s) ar e holomorphic i n Re s < 1 , that |P(s) | = 1 on Re s = \,

whence b y the convexity boun d o r Phragmen-Lindelof principl e
(9.71) \P(s)\ ^£^~ a i f cr = R e (s) ^ \.
Moreover, L(s , x) is holomorphic on C except fo r a simple pole at s = 1 with residu e
(f(£)/£ i f x ( m ° d £) i s principal. Therefor e b y movin g th e integration i n (9.56 ) t o
the lin e R e s = — e an d applying th e functional equatio n (9.69 ) w e obtain

(9.72) B(t, x) = S x^f-f(l + it) + exB*(t, X )

where

(9.73) B*(t, x)
= - L / " l{s~ it)P(s - it)L(l - s + it, x)f(s)ds.
2
™ J(-e)
Expanding L( l — s + zt, x) int o th e Dirichlet serie s and integrating termwis e w e get
00

(9.74) B*(t, X) = ^ 3Hx(m)m-5-1 (

where
s
(9.75) g(y) = -^/ 7 (s - it)P(s - it)f{s)y -^ds.

Notice w e moved th e integration fro m R e s = — e i n (9.73 ) bac k t o R e s = ^ in

(9.75) a s we can because 7(5 ) and P(s) hav e n o poles i n Re s < 1.
 9. T H E D I R I C H L E T P O L Y N O M I A L S 245

The formul a (9.72 ) i s the on e w e were talking about , bu t i t i s not ye t read y fo r

applications becaus e th e dua l functio n g(y) i s no t evaluate d explicitl y enough . T o
see it s propertie s w e tak e agai n th e majoran t functio n f(x) give n b y (9.46 ) wit h
X = iV , s o it s Melli n transfor m i s
l 5 2
(9.76) f{s) = J exp(^----y- dx = 2e / Ks(2)Ns

where K s(y) stand s fo r th e Bessel-Macdonal d function . Henc e f(s) ha s exponentia l

decay o n an y vertica l lin e s = a + it. Usin g analyti c propertie s o f / ( s ) , 7(5 ) an d
P(s) on e shows , b y movin g th e integratio n i n (9.75 ) sufficientl y fa r t o th e lef t (a s
in (1 0.64)) , tha t

(9.77 ) ^)«^M-K^T))' )
where th e implie d constan t i s absolut e (compar e thi s estimat e wit h (9.45)) . Henc e
^(ra) i s very smal l i f mN > £(\t\ + l)(logiW(|t | + l)) 3 .
In ou r cas e \t\ ^ T and £ < kQ2, s o th e term s o f B*(t,x ) wit h mN ^ HC 3

contribute ver y little , precisel y w e ge t b y (9.77 )

1
(9.78) B*(t, X)= ^ s H x H r n ^ - '+ Oir )

where
3
(9.79) MN = HC
and th e implie d constan t i s absolute .
an
To mak e (9.78 ) read y fo r application s i t i s desire d t o separat e m fro m \ dt
involved i n g(m). W e accomplis h thi s goa l quickl y b y introducin g (9.75 )

1
(9.80) B*(t, X) = ^-. I1 {s)P{s)f{s + it){ Y. x t m K ^ H O r ) .

Since \K 3(2)\ ^ 8|T(s)| < C e~^ fo r s = \ + iu, w e ge t

oo . .

/ +i(t
J2 x(m)m-* +u)le-^du + H- 1

where th e implie d constan t i s absolute .

This resul t i s mor e precis e tha n (9.57) . Indeed , estimatin g triviall y b y (9.81 )
one get s
(9.82) B * ( i )X)<ifi£§

which recover s (9.57) . Bu t on e ca n d o bette r b y puttin g th e su m

(9.83) Yl Xi&Mm-i+W-i-t-j+t. )

in th e origina l polynomia l
u
(9.84) Yl ^nXi(n)n ^
N<n^2N
246 9 . T H E DIRICHLE T POLYNOMIAL S

which appear s a s a coefficien t i n (9.55 ) (her e an d hereafte r w e use th e abbreviate d

notation t r = t r(x)). B y (9.55) , (9.72 ) an d (9.81 ) w e deriv e

Xr lr 2
oo
T> _ H
/ EE E E \ 1 , Xl
(^r2,X2)V(itr ! »*r 2 , X2 5 « ) | e d«
~°° X i X 2 r i r 2

+tf~1EEEEip^,xi)z>(^2,x2)i
Xi X 2r ir 2

where V(it ri, Xi ? ^r2 > X2; ^) denote s th e produc t o f the tw o sums (9.83 ) an d (9.84) .
For fixed X2,tr 2iu thi s produc t i s a ne w polynomia l o f typ e (9.39) . Th e lengt h o f
P ( i t r i , X i ) = V(it ri,xi'->ttr2iX2',u) bein g 2MN = 2HC S i s perfec t fo r a n applica -
tion o f Theore m 9.1 3 . W e obtai n

E E i m - ^ x i ) ! < ( ^ E E ip ( ^ ^ i ) i 2 ) " «(i?GF)i£ 6.

r
Xi ^ iX i i

After thi s bound , whic h doe s no t depen d o n X2^2,^ 5 w e estimat e

EEi^^'^i^^EEi25^'^)!2)^^)^
X2 r 2 X 2r 2

Integrating i n u, w e find that th e contributio n o f these polynomial s t o D i s at mos t

0(R(CGHN)zC5). Th e polynomial s fro m th e diagona l x i = X 2 = X contribute

r
xn 2
and th e remainin g one s contribut e
jff_1
EEEEi I) ^'^) I3 ^'^)i <<F " liiC<<c -
Xi X 2r ir 2

Adding u p thes e contribution s w e obtai n P < C NC + R(CGHN)z £ 6 . Henc e w e

derive b y (9.66 ) tha t C ^ GN + GRi{HN)^C A. Thi s complete s th e proo f o f
Theorem 9.1 5 .

9.7. Larg e value s o f D(s,x)>

The notatio n fro m previou s sectio n hold s here , i n particular , w e recal l tha t
H = kQ 2T an d C = lo g HN. I n thi s sectio n w e deduc e a fe w bound s fo r th e
number o f larg e value s o f th e polynomial s D(s,x)- First , a s i n Theore m 9.7 , w e
deduce fro m (9.52 ) th e followin g
na
THEOREM 9.1 6 . Suppose for any s r (x) ^ well-spaced set S(k, Q , T) we have
(9.85) \D(s r(X),x)\>V.

Then the cardinality of this set, say R = \S(k,Q,T)\, satisfies

(9.86) i ? « ( i f + iV)Gr 2£5
where the implied constant is absolute.
 9. T H E D I R I C H L E T P O L Y N O M I A L S 24 7

If V i s relatively large , the n (9.53 ) yield s a bette r boun d (fo r details se e th e

proof o f Theorem 9.8) , namel y th e following
THEOREM 9.1 7 . Let the conditions be as in Theorem 9.1 6 with
3
(9.87) V^G^H^C .
Then
2 4
(9.88) R < GNV~ £
where the implied constant is absolute.
The estimat e (9.88 ) i s idea l fo r applications , unfortunatel y i t i s establishe d
only for very larg e V relativ e to the "conductor" H. On e can remove the condition
(9.87) a t a cost o f introducing a n extra ter m i n the bound (9.88) . T o this en d w e
apply (9.64 ) an d get the following resul t
THEOREM 9.1 8 . Let the conditions be as in Theorem 9.1 6. Then
2 3
(9.89) R < (GNV~ + G NHV-6)C
1 S

where the implied constant is absolute.

The las t estimat e (9.89 ) i s the celebrated resul t o f Huxley [Hux] . I t is surpris-
ing t o see that i n the special cas e o f the Riemann zet a function , hi s subdivisio n
argument yield s a bound (9.28 ) whic h agree s wit h (9.89 ) obtaine d b y very differen t
arguments (reflectio n method) .
The assumption tha t th e ordinates t r(x) o f the points s r(x) — 0"r(x)+^r(x ) a r e
well-spaced ca n be relaxed, i t suffice s tha t the y d o not cluster to o much. Suppos e
that fo r any sr (x) £ S{k,Q,T) w e have
(9.90) \{n ; sri(X) G S(k,Q,T), \t ri - t r\ < 1 } | ^ L.
Then clearl y al l Theorem s 9.1 3 to 9.1 8 are true fo r th e set S(k,Q,T) wit h th e
relevant uppe r bound s multiplie d b y L (t o see this divid e th e set S(k, (J , T) int o
at mos t 2L subset s o f well-spaced point s an d apply th e results fo r eac h o f thes e
subsets separately) .
In particular , w e know tha t th e number o f zeroes p — ji + ry of L(s, x) f° r a
given characte r x ( m ° d kq) with | 7 — t\ ^ 1 is bounded b y 0(\ogkq(\t\ + 3) ) (see
Chapter 5) . Therefore , b y virtue o f the above remark , al l Theorems 9.1 3 to 9.18
hold fo r any subset <S(fc , Q, T) o f zeros of
L
(9.9i) n n n ^O
q^Q ^ ( r a o d q) £(mo d k)
(q,k) = l i\) primitiv e

in a rectangle a ^ a , \t\ ^ T (counte d wit h multiplicity ) wit h jus t on e extr a facto r

£, an d G replaced by
(9.92) G(a) = J2 K\ 2n~2a

in the relevant uppe r bounds .

In th e next chapte r w e apply thes e theorem s fo r selecte d point s s r(x) whic h
are zero s o f (9.91).
This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/11

CHAPTER 1 0

ZERO-DENSITY ESTIMATE S

10.1. Introduction .
In the absenc e of a proof o f the Gran d Rieman n Hypothesis , i t i s natural t o as k
how many zeros of a given L-function ca n lie off the critical line. Sinc e we know there
are n o zero s o n th e lin e R e (s) = 1 , a fe w suc h zero s woul d no t normall y influenc e
applications, bu t a larg e numbe r o f these , especiall y whe n the y ar e nea r thi s line ,
may distor t substantiall y th e result s (a n exampl e bein g th e distributio n o f prime s
in shor t intervals , se e Sectio n 1 0.5) . Therefore , w e shoul d as k ho w man y zero s
can possibl y b e i n a give n region , an d ou r estimate s shoul d revea l tha t th e furthe r
the regio n i s fro m th e critica l line , th e smalle r th e numbe r o f zero s i t contains . A
quantitative statemen t o f thi s kin d i s called a zero-densit y theorem .
The region s w e need mos t i n practic e ar e th e rectangle s
(10.1) R(a,T) = {s = a + it; a ^ a, \t\^T}
for \ ^ a ^ 1 and T ^ 3 . Le tT V (a, T) denot e th e numbe r o f zero s o f th e Rieman n
zeta function i n R(a, T) (counte d wit h corresponding multiplicity) , tha t i s we count
the zero s p = (3 + ij wit h
(10.2) p > a, | 7| < T.
In thi s notatio n th e Rieman n hypothesi s assert s tha t N(a, T) = 0 fo r an y a > \
and T ^ 3 , while the bes t know n zero-fre e regio n (du e to Vinogradov an d Korobov )
says tha t
2/3
(10.3) N(a,T) - 0 i fa > 1 - c(logT)~ (loglogT)1 /3

where c i s a n absolut e positiv e constan t (se e Chapte r 8) . Recal l tha t th e numbe r

of al l zero s p = ft - h 27 with \j\ ^ T satisfie s (se e Chapte r 5 )

(10.4) N(T) = - lo g -^ + O(logT) .

7r Zire
By analog y t o th e convexit y principl e concernin g th e orde r o f £(s ) o n vertica l line s
one ma y expec t th e followin g boun d fo rA T (a, T) t o b e tru e
DENSITY CONJECTURE . For \ ^ a ^ 1 and T ^ 3 we have
2(1
(10.5) 7V(a , T) < T - a ) lo g T
where the implied constant is absolute.
Of course , th e convexit y principl e fo r zero s i n rectangle s doe s no t hol d fo r gen -
eral analyti c functions ; nevertheless , th e densit y conjectur e fo r £(s ) ha s a reliabl e
249
25 0 10. ZERO-DENSIT Y ESTIMATE S

backup - th e Rieman n Hypothesis . Thi s conjectur e i s particularl y attractiv e be -

cause i t ha s a chance t o be proved b y th e technolog y whic h i s available today , whil e
it replace s th e Rieman n hypothesi s i n variou s application s t o th e distributio n o f
primes.
A grea t dea l o f effor t wa s mad e t o establis h estimate s o f typ e
c
(10.6) 7V(a , T) < T ^ ) (1 " a ) ( l o g T ) A
with c(a) a s smal l a s possibl e (th e logarithmi c facto r i s of secondar y concer n here ,
although i t wil l b e importan t t o eliminat e i t i n Chapte r 1 8) . Th e firs t zero-densit y
results wer e prove d b y Boh r an d Landa u [BL ] i n 1 91 4 . Thei r ide a wa s basi c i n
many work s ove r thre e decades , notabl y i n th e work s b y F . Carlso n [Car ] an d A .
E. Ingha m [Ing] . Tw o ne w device s wer e introduce d b y P . Tura n i n 1 949 , an d i n
collaboration wit h G . Halas z i n 1 958 . Grea t refinement s wer e mad e b y H . Mont -
gomery i n 1 969 . Th e othe r significan t contributor s t o thi s fascinatin g theor y ar e
M. N . Huxley , M . Jutil a an d D . R . Heath-Brown .
In this book we shall try to give a glimpse of the current stat e of knowledge. Ou r
results will no t b e a s stron g a s th e bes t know n ones , thoug h nearl y so . Moreover ,
we treat i n considerabl e detai l th e zero s o f Dirichlet L- functions. Le t N(a,T, \) b e
the numbe r o f zero s p x = f3 x + i^ x o f L(s , x) wit h fi x > a an d |7 X | ^ T (counte d
with multiplicity) . Th e ultimat e goa l woul d b e t o prov e th e followin g
GRAND DENSIT Y CONJECTURE . Let / c ^ l , Q ^ l , T ^ 3 and \ ^ a ^ 1 .
Then
a
(10-7) E E E N(a^^)^H^- \logH)A
q^Q -0(mo d q) £(mo d k)
(q,k) = l -^primitiv e

where H = kQ 2T and A is an absolute constant, the implied constant being also

absolute.
Our mai n result , th e Gran d Densit y Theore m 1 0.4 , will give (1 0.7 ) i n the rang e
| < a ^ 1 . Moreover , i t yield s (1 0.7 ) i n th e whol e segmen t \ < a ^ 1 , bu t wit h
the constan t c = 1 2/ 5 i n plac e o f 2 .
10.2. Zero-detectin g polynomials .
First w e outline th e ol d and ne w strategies. Give n a holomorphic functio n L(s)
in a regio n D C C ther e i s a variet y o f tool s fo r countin g it s zero s insid e D. A
typical on e i s th e integra l formul a o f Jense n
R
,f
/ A n , ,dr I f , \L(s)\ds
10-8 / n(r)— = — / l o g - ) - (-
Jo r 2<KI J lsl=R IL(0) Is

where n(r) denote s th e numbe r o f zero s o f L(s) i n th e dis c |s | ^ r (assumin g

L(s) ^ 0 fo r s = 0 an d fo r s o n |s | = R). Ther e ar e simila r formula s fo r othe r
domains. Not e tha t on e need s bot h uppe r an d lowe r bound s fo r \L(s)\ o n th e
boundary t o deduc e a n estimat e fo r th e numbe r o f zero s insid e th e region . O f th e
two, th e lowe r boun d i s th e hardes t t o find. I n vie w o f thes e requirement s an y
integral formul a i s practical onl y fo r a regio n sufficientl y wid e s o that it s boundar y
is distanced fro m th e area in which the zeros are concentrated (otherwis e one canno t
 10. Z E R O - D E N S I T Y E S T I M A T E S 251

claim a reasonable lowe r boun d fo r \L(s)\ o n dD). Bu t thi s i s not the cas e fo r the
narrow rectangle s R(a, T).
To avoi d th e abov e barrier s on e get s th e following idea : multipl y L(s) by an -
other holomorphi c function , sa y M(s), whic h i s suspected t o be large whe n L(s)
is small , an d coun t th e zero s of the produc t N(s) = L(s)M(s) rathe r tha n o f L(s)
alone. O f course , th e extr a zero s of M(s) will occur , bu t hopefully no t s o many as
to chang e th e orde r o f magnitude o f the origina l numbe r ( a positive proportio n o f
the extr a zero s is acceptable).
In realit y th e ide a work s onl y for counting "fictitious " zero s because N(s) doe s
vanish at the true zeros of L(s). Thu s the points at which we want th e mollifier M(s)
to be large ar e onl y hypothetical . Havin g n o possibilit y t o inspect th e distributio n
of thes e hypothetica l point s on e ma y tr y M(s) whic h mimic s 1 /L{s) a t almost all
points, s o N(s) is expected t o be close to one almos t everywher e i n a given region .
It woul d b e tautological t o tak e jus t 1 /L(s) fo r the mollifier M(s) , o r naiv e t o
deform 1 /L{s) smoothl y t o get a useful holomorphi c function . Bu t suppos e 1 /L(s)
is give n b y an absolutely convergen t serie s i n a certain region . The n th e chances
are that it s partial sum s d o well approximate t o \/L(s) uniforml y i n a wider regio n
which doe s no t contain th e true zero s o f L(s). Therefore , a finite bu t sufficientl y
long partia l su m o f such kin d offer s a good candidat e fo r M(s).
The abov e strateg y serve s wel l fo r counting fictitious zero s o f Dirichlet serie s
with multiplicativ e coefficients .
First t o illustrate th e ideas w e giv e a quick treatmen t o f L(s) = C( 5)- I* 1 this
case
1 /i(ra )
c(s) Alt ps) z ^
but onl y if Re s > 1. Nevertheless, on e expect s tha t th e partia l su m

(10..) M W . £ <g >

approximates t o l/((s) i f Re s = a > | . Indeed , th e Riemann hypothesi s ensure s
that
£
(10.10) - ^ = M(s) + 0{(\s\M) Mi-a).

REMARKS. Th e partia l su m M(s) was firs t use d i n th e contex t o f zero-density

estimates b y Carlson [Car ] whil e Boh r an d Landa u [BL ] wer e working earlie r wit h
the partia l produc t
i
(Ian) ^)=n( -i)-
Either devic e produce s non-trivia l results , yet M(s) is more efficien t tha n P(s).

It help s to keep hol d o n th e siz e of the points , s o we restrict ou r analysi s to the

region
(10.12) 5 = c r + zt, c r ^ a , T<\t\^2T,
25 2 10. Z E R O - D E N S I T Y E S T I M A T E S

and w e assume tha t T i s large. I n thi s regio n

n
(10.13) C ( s ) = E ~ S+ °( T
~a)

where th e implie d constan t i s absolute . Henc e

s 1a
(10.14) C(s)M(s)= Y^ a nn~ ^0(T~ M -a\og2M)
n^TM

where th e coefficient s ar e

(10.15) a n= ^ /x(m) , |a n|^r(n),

dm=n
m^M,d^T

and th e erro r ter m i s obtained b y th e trivia l estimatio n

1 _a
(10.16) M(s) < M lo g 2M.

Assuming H M ^ T w e hav e a\ = 1 and a n = 0 i f 1 < n ^ M o r n > MT. W e

cover th e interva l M < n ^ MT b y dyadi c segment s N < n ^ 2N wit hT V = 2 £M,
0 ^ £ < L = [lo g T/ lo g 2]. Fo r eac h o f thes e segment s denot e
a
(10.17) D £(s)= Yl -
N<n^2N

Thus w e hav e

(10.18) C(s)M(s) = l+ Yl De{s) + E(s)

O^KL

where th e erro r ter m satisfie s \E(s)\ < \ fo r s i n th e regio n (1 0.1 2) , provide d

a
Mi-a ^ T (logT)"2.
An importan t featur e o f th e identit y (1 0.1 8 ) i s th e absenc e o f term s wit h 1 <
n ^ M o n th e righ t side . Th e lef t sid e a t a zer o o f £(s) , sa y a t s = p , vanishes s o
we mus t hav e
D
(10.19) | J2 ^P) >k.

Hence
1
(10.20) \D e(P)\ > (2L)-

for som e 0 ^ £ < L. Thi s lowe r boun d i s rathe r larg e b y compariso n t o th e mean -
value o f \Di(s)\ 2 o n th e lin e R e s = a. Indee d w e hav e
2a
(10.21) G £(a) = J2 K\ 2
n~2a « A^- (log2A0 3 .
N<n^2N

For thi s reaso n w e cal l Di{s) zero-detectin g polynomials . Wha t w e showe d i s tha t
every zer o o f £(s ) i n th e regio n (1 0.1 2 ) i s detecte d b y a t leas t on e o f a fe w fixed
polynomials b y wa y o f producin g unusuall y larg e value .
 10. Z E R O - D E N S I T Y E S T I M A T E S 253

R E M A R K S . O f course, th e lower boun d (1 0.20 ) ma y not hold a t point s othe r

than th e zeros o f £(s). I n fact, on e should expec t tha t

(10.22) D < N* -ot+e

£(s)

for an y s in the region (1 0.1 2) . Comparin g (1 0.22 ) wit h (1 0.20 ) on e could ge t a

contradiction whic h prove s th e Riemann Hypothesis . Well , thi s argumen t i s noth-
ing bu t wishfu l thinkin g becaus e on e needs th e Riemann Hypothesi s t o establis h
(10.22). Thes e remark s revea l tha t th e concept o f a zero-detecting polynomia l is
somewhat superficial , i t will loos e it s meaning whe n th e GRH i s fully established .
In realit y Dirichle t polynomial s d o not assume larg e value s ver y often .

Now the theory of Dirichlet polynomial s as developed in Chapter 9 tells us that

(10.20) canno t happe n to o often, i n particular, producin g a n interesting estimat e
for th e number o f zeros o f the zeta function . Precisely , le t Ri be the number o f
zeros of ((s) i n (10.12) whic h are detected b y the polynomial Dt(s). The n Theore m
9.7 (o r Theorem 9.1 6 ) togethe r wit h th e closing remark s o f Chapter 9 yield
(10.23) R £ < ( T + A O A ^ - ^ l o g T )11
by (1 0.20 ) an d (10.21). Sinc e M ^N < MT, thi s give s
(10.24) R £ < {TM 1 -2* + ( T M ) 2 - 2 a } ( l o g T ) n
for an y 0 < £ < L. Choosin g M = T2^1 w e ge t
4a(1
(10.25) R e < T -a)(logT)n.
Adding u p the Ri we get an upper boun d fo r all zeros o f ((s) i n (1 0.1 2 ) (som e of
these ma y be detected b y several polynomial s D^(s)) , the n w e extend th e result t o
the zero s i n (10.2), an d conclude
THEOREM 1 0.1 . For \ ^ a ^ 1 and T > 2 we have

(10.26) N(a,T) < T 4 c * ( 1 - a )(logT) 1 3 .

This estimat e (apar t fro m th e logarithmic factor ) wa s obtained b y F. Carlso n

[Car] i n 1920 (se e also E. Landau [La3 ] of 1921). Twent y year s later A . E. Ingham
[Ing] prove d

(10.27) N(a,T) < T 3^(logT)5.

Either (1 0.26 ) o r (10.27) impl y tha t th e Riemann hypothesi s i s "almos t true " in
a statistica l sense , namel y tha t almos t al l zeros o f £(s) lie arbitrarily clos e t o the
critical line . I t follows b y (10.27) tha t th e number o f zeros p — / ? + ry with J7 J ^ T
and
+ 3
(10-28) ^ > i ^

is bounde d b y 0(T(logT) 5-4A) fo r A > 1.

Concerning th e zeros nea r th e critical lin e in 1942 A. Selberg [S4 ] prove d tha t
(10.29) N(a, T)<^{a-\)- lT, i f a > \,
25 4 10. Z E R O - D E N S I T Y E S T I M A T E S

where th e implie d constan t i s absolute . Hence , give n a functio n $(T )— > co a s

T—> • co, i t follow s tha t almos t al l zero s o f th e Rieman n zet a functio n li e i n th e
region

do.30 ) ^-il<^ -

10.3. Breakin g th e zero-densit y conjecture .

It i s amazin g ho w lon g th e estimat e (1 0.27 ) o f Ingha m resiste d improvements .
After sixt y year s i t i s still th e bes t know n resul t (apar t fro m th e logarithms ) i n th e
range | < a < | . Th e ne w methods , whic h emerge d fro m th e wor k o f Halas z an d
Turan [HaTu ] i n th e lat e sixtie s an d flourished i n th e seventies , produce d a ne w
estimate fo r th e numbe r o f larg e value s o f Dirichle t polynomials , namel y (9.28 ) i n
place o f (9.25) . Thi s ne w estimat e turne d ou t t o b e bette r tha n th e ol d on e fo r
treating zero s nea r th e lin e R e s = 1 . Firs t th e estimat e (1 0.27 ) wa s improve d i n
the rang e | < a < l b y H . L . Montgomer y [Mo3 ] i n 1 969 , an d shortl y afte r i n
the rang e | < a < l b y M . N . Huxle y [Hu2 ] i n 1 97 2 (se e a mor e genera l Theore m
10.4).
For th e zero s o f ((s) nea r th e lin e R e s = 1 on e i s als o helpe d b y havin g
shorter partia l sum s whic h approximat e t o £(s) . I n thi s sectio n w e us e a suitabl e
short polynomia l t o establis h a zero-densit y estimat e whic h i s even bette r tha n th e
zero-density conjectur e fo r a > | .
THEOREM 1 0. 2 ( H U X L E Y ) . For any a > | and T ^ 2 we have

(10.31) N(a, T) < T 3-^ (log T) A

where A = 300( a — | ) ~ 2 and the implied constant depends only on a.

REMARKS. Actuall y Huxle y prove d (1 0.31 ) fo r an y a ^ \ wit h A an d th e

implied constan t bein g absolute . W e coul d refin e ou r argument s t o ge t (1 0.31 ) i n
wider rectangle s an d wit h absolut e constants , bu t i t woul d obscur e th e presenta -
tion. Beside s provin g (1 0.31 ) ou r missio n i s also to introduc e neatl y som e technica l
novelties du e t o M . Jutil a [Ju4] . H e realize d tha t i t wa s sufficien t t o tak e th e
mollifier M(s) quit e short . Then , t o achiev e bes t results , h e make s a n adjustmen t
by raisin g th e zero-detectin g polynomial s t o appropriat e hig h powers . Thi s ide a of
Jutila i s somewha t simila r t o th e wor k o f D . R . Heath-Brow n o n hi s identit y fo r
primes (1 3.37) .

By WeyP s boun d (8.1 8 ) (th e va n de r Corpu t boun d (8.53 ) woul d als o d o th e

job) w e deriv e

(10.32)CO O = Yl n
~S + °(x*~aT* l o £ T )

for s = a + it wit h \ < a < 1 , T < \t\ < 2T , wher e 1 ^ X ^ T an d th e implie d

constant i s absolute . Henc e
s
(10.33) ((s)M(s) = 1 + ] Ta nn~ + E(s)
M<n^MX
 10. Z E R O - D E N S I T Y E S T I M A T E S 255

where 1 < M ^ X, \a n\ < r(n) (see (1 0.1 5 ) wit h T replaced b y X) an d

2
(10.34) E(s) < M^'Xi-'Ti (logT) .
We requir e

(10.35) |E( S )| < \

for 5 in th e rectangl e (1 0.1 2) , an d thi s hold s i f

(10.36) X a
~i ^ M 1
-^(logT)3
because T is assume d t o b e large .
Now th e situatio n i s th e sam e a s in Section 1 0. 2 excep t tha t w e emplo y fewe r
zero-detecting polynomial s (1 0.1 7) ; thei r lengt h N = 2lM is restricted b y M ^
TV < MX (i.e. , 0 < £ < L = [log Xj log2]) . Recal l tha t R t denote s th e numbe r
of zero s o f ((s) in (10.12) satisfyin g (1 0.20) . W e shal l estimat e Ri by an appeal
to (9.31 ) wit h a suitable k ^ 2 rather tha n (9.30 ) whic h wa s use d wit h k = 1 fo r
deriving (1 0.23) . I n other words , w e appl y (9.28 ) fo r th e polynomia l Di(s) k. W e
choose k ^ 2 depending o nT V suc h tha t P = N h fall s int o th e segmen t
2
(10.37) Z <P^ (MX) + Zi,
where Z is a given numbe r t o be specifie d late r subjec t t o Z ^ MX. T o se e tha t
(10.37) i s possible tak e k ^ 2 such tha t Z 1 ^ < M < Z 1 / ^ " 1 ) , an d i n case fc =2
use als o ] V ^ MX. Th e mean-valu e o f \Di(s)\2k o n th e lin e R e s = a is bounded
by
G 2
(10.38) fc=( Z ) l«n| n - 2 a T , ( n ) ) / C «1
P - 2 - ( l o g P ) 4 f c ^1+ )

where th e implie d constan t depend s onl y o n k. Moreover , th e lowe r boun d V =

(2L)" 1 o f \Dt(s)\ a t the zer o i s raised t o the powe r V k = (2L)~ k > (logT)" fc .
Therefore (9.31 ) yield s
(10.39) R £ < (P 2^-*) + TP4-6a)(log£)^
where Ak = 6(f c + l)(12fc + l) an d th e implie d constan t depend s o n k. Her e it would
be perfec t t o choos e P = T 1 ^2(^2OL~1 \ producin g th e bes t boun d
rp(l-a)/(2a-l) Q rp\ A k

but i t is not alway s possibl e becaus e P i s not precisel y localize d (sinc e k had t o b e
an integer!) . However , w e have reasonabl y goo d contro l o f P give n b y th e interva l
(10.37) wher e Z is at our disposal . Insertin g th e uppe r boun d fo r P t o the firs t
term an d th e lowe r boun d fo r P t o th e secon d ter m i n th e righ t sid e of (10.40) w e
get
4(1
(10.40) Ri < {(MI) " a ) + Z 3 < 1 - a ) +TZ 4 6a
- }(logT)Ak.

Now i t is clear tha t th e bes t choic e i s Z = T 1 ^ 3 * - 1 ) . W e als o tak e MX =

Zi(logT)9 whic h yield s

(10.41) Ri <T^^(logT) Afc+6.

256 10. ZERO-DENSIT Y ESTIMATE S

We stil l hav e som e freedo m t o choos e M an d X subjec t t o th e conditio n (1 0.36) .

For th e valu e MX = A' 3 / 4 ( 3 Q - 1 ) (logT) 9 thi s conditio n i s satisfie d wit h
(10.42) M = j 1 ( 6 a - 5 )/ 1 2 ( 3 o ; - 1 )
(7 3a)/6(3a 1 }
(10.43) X = T - - (logT)9.
Finally, summin g (1 0.41 ) w e get (1 0.31 ) wit h A = A k + 8 = 6( K + 1 )(1 2K + 1 ) + 8,
where K i s th e maxima l k whic h i s neede d t o pu t P = N k int o (1 0.37 ) fo r M ^
N < M X , i.e. , K = 1 + [lo g Z/log M] ^ 1 + 2( a - f ) " 1 . Henc e A < 300( a - | ) " 2 .
Having sufficientl y stron g estimate s fo r exponentia l sum s on e ca n mak e a goo d
approximation t o £(s ) b y partia l sum s whic h ar e ver y shor t fro m whic h on e ca n
derive relatively shar p zero-densit y estimates . Thus , using the Vinogradov estimat e
(8.92) i t wa s show n b y Halas z an d Tura n [HaTu ] tha t fo r an y a nea r on e
(1
(10.44) iV(a,T ) <cT -a)i|los^-a^3.

EXERCISE 1 . Prov e alon g th e abov e line s tha t th e Lindelo f Hypothesi s implie s

12 ( + £
(10.45) % T ) « T ^
for \ ^ a < 1 and an y e > 0, th e implie d constan t dependin g onl y o n s.
EXERCISE 2 . Assumin g th e Lindelo f Hypothesi s an d th e Montgomer y Conjec -
ture M (se e (9.23)) , prov e alon g th e abov e line s tha t
£
(10.46) N{a,T)<^T
if a > \ fo r an y e > 0, th e implie d constan t dependin g o n a an d e.
10.4. Gran d zero-densit y theorem .
Now we are goin g to exploi t th e idea s o f the previou s sectio n an d th e result s of
Chapter 9 for polynomials with characters in their ful l force . W e consider character s
x(mod kq) wit h (k,q) — 1 suc h tha t \ — ^ wit h £ an y characte r t o modulu s k
and I/J any primitiv e characte r o f conducto r q. Her e k i s give n whil e q varies ove r
the segmen t 1 ^ q ^ Q. Moreover , w e kee p th e point s s — a - f it i n th e rectangl e
a ^ a , \t\ < T , wher e T ^ 3 . Le t N(a, k, Q, T) denot e th e tota l numbe r o f zero s of
all I/(s , x) counte d wit h multiplicit y subjec t t o th e abov e restrictions , i.e. ,
N
N(a,k,Q,T)= J2 E m
Y, MW)-
q^Q V( ocl q) £(mo d k)
(q,k) = l ^primitiv e

Put D = kQT, H = kQ 2T an d £ = 2 log D. B y th e classica l boun d

N(a,T,x)<£Tlog(kqT)
(see Theore m 5.24 ) i t follow s tha t 7V(a , fc,Q, T) < C HC, whic h w e shal l refe r t o a s
the trivia l bound .
In thi s sectio n w e shal l establis h variou s non-trivia l bound s fo r N(a, fc, Q, T)
which ar e generalization s rathe r tha n improvement s o f th e result s b y Ingha m an d
Huxley (compar e (1 0.72 ) wit h (1 0.27 ) an d (1 0.31 )) .
 10. Z E R O - D E N S I T Y E S T I M A T E S 257

From the previous section it is clear that i f one has shorter Dirichlet polynomial s
approximating L(s,x) , the n a stronge r zero-densit y theore m ca n b e deduced . Th e
best unconditiona l approximatio n i s given b y th e approximat e functiona l equation ,
which, roughly speaking , represents L(s, x) a s a s u m o f two polynomials of length X
and Y respectively , wit h XY bein g th e conductor . Classica l formula s o f suc h kin d
(see e.g . [Lav] ) ar e quit e messy , s o w e rathe r develo p her e a differen t expressio n
for L(s , x) o f the sam e nature , bu t mor e friendl y fo r application s i n mind . Se e als o
Theorem 5. 3 fo r a genera l version .
The qualit y o f a n approximat e functiona l equatio n depend s o n th e choic e o f a
particular cut-of f function . W e hav e alread y experimente d wit h suc h approximat e
equations whe n studyin g th e reflectio n metho d i n Sectio n 9.6 . Ou r presen t settin g
is clos e t o tha t particula r one . On e ma y star t fro m an y smoot h functio n / o n R +
such tha t

(10.47) /(0 )- 1 and li m f(x) = 0

with f'(x) havin g rapi d deca y a s x tend s t o 0 o r oo. The n it s Melli n transfor m
/»0 0
1s
(10.48) f(s) = / f(x)x ~ dx
Jo

has a simpl e pol e a t s — 0 wit h residu e 1 and sf(s) i s a n entir e function . Indee d
we hav e
rOC
s
(10.49) sf(s) = - / f{x)x dx
Jo

which show s wha t w e claimed . Normall y w e woul d no t recommen d workin g wit h

any specific cut-of f function , bu t w e do so here for clarit y t o keep track of uniformit y
in severa l variable s sid e b y side . A nic e choic e i s

(10.50) /(* )= «/ exp(-y-- -\-H

^

where K is th e normalizin g constan t suc h tha t /(0 ) = 1 . W e shall se e fro m compu -

tations belo w tha t n~ l = 2K 0(2) = 2^2T F f(k)r-3(k) = 0.2277.. . . This functio n
satisfies
x
(10.51) 0 < f(x) < ne~

because e~ l/y < y, an d

1 /x
(10.52) 0 < 1 - f(x) < Ke~

because

(10.53) /(:r )+ / ( i ) = l .

By (1 0.49 ) fo r / give n b y (1 0.50 ) w e ge t

1s
(10.54) sf'(s) = Kf exp(-y--)y - dy = 2KK s(2)
25 8 10. Z E R O - D E N S I T Y E S T I M A T E S

where K s(z) i s th e Bessel-Macdonal d function . I n particular , b y th e powe r serie s

expansion o f K s(z) a t z = 2 we ge t

(10.55) sf(s) = - ^£ — \T(k=7) ~ f(kT7)J •

Hence w e see tha t f(s) i s od d

(10.56) f(s) = -f(-s).

Moreover, w e ge t th e unifor m boun d fo r s — a + it G C,
1|ff|
(10.57) / ( 5 ) < | s | - e-*|t|
where th e implie d constan t i s absolute .
Now w e ar e read y t o deriv e th e approximat e formul a fo r L(s,x) m question .
We assum e tha t s = a + it wit h \ < a < 1 and x i s a characte r o f modulu s £ ^ 1 .
To refres h you r memory , w e recommen d readin g agai n th e argument s fro m (9.67 )
to (9.75 ) i n th e previou s chapter . W e begi n b y evaluatin g th e su m
oo

(10-58) B(«,x ) = E x ( n ) » " ' / g ).

1

By contou r integratio n an d th e functiona l equatio n (9.69 ) w e ge t

u
B(s,x) = ^~ f L(s + u, X)X f(u)du

= S^Xi-ftl - s) + L(s, x ) + ^ / _ L(s + u, X )Xuf(u)du.

Then b y th e functiona l equatio n (9.69 ) an d b y (1 0.56 ) w e find tha t th e integra l o n

the lin e R e u — — 1 is equa l t o — exB*(s, x) wher e

u
£*(*, X) = ^ / l(s - u)P{s - u)L{\ -s + u, x)X- f(u)du.
2iri J {1 )

Here we expand L( l — s + u, x) m ^ ° Dirichle t serie s an d integrat e eac h ter m gettin g

oo

(10.59) B*(s, X) = Y1 xMm^gimX)

l

where
u
(10.60) g(y) = ^- [ 7 ( s - u)P(s - u)y- f(u)du.

Gathering th e abov e result s w e obtai n th e desire d expressio n

1
(10.61) L(s,x) = B(s,x)+e xB*(s,X) - 6 x^f(l - s)X ~s

where B(s, x) a n d -B*(s , x) a r e th e reduce d serie s given by (1 0.58 ) an d (1 0.59 ) wit h

arbitrary X > 0.
 10. Z E R O - D E N S I T Y E S T I M A T E S 259

It i s clear tha t B(s, \) r u n s essentiall y ove r n < C X, an d we show tha t B*(s, x)

runs essentiall y ove r m < 7 , wher e Y i s somewha t large r tha n -£|s|X _1 . T o thi s
end recal l th e estimate s
R e s
(10.62) | P ( s ) | ^ - ,
Res
(10.63) 7 ( s ) < N * -
which hol d uniforml y i n th e half-plan e R e s ^ | , an d th e estimat e (1 0.57 ) whic h
holds fo r al l s. Hence , movin g th e integratio n (1 0.60 ) sufficientl y fa r t o th e right ,
say t o th e vertical lin e
5
(10.64) R e W = max(l,i(^) )

one derive s

(10.65) 9 ( s ) «a e x p (_l(JL)*)

where th e implied constan t i s absolute. Henc e g(mX) i s very smal l i f mX i s some-

what large r tha n £\s\. Precisely , w e get by (1 0.65 )
1
(10.66) B*(s, X) = £ x(m)m'- g(mX) + 0 ( ^ )

provided
(10.67) XY ^£\s\(log£\s\f.
Next w e write (1 0.60 ) a s

(10.68) g(y) = — ( l(s - 2a + u)P{s - 2a + u)y u-2af(2a - u)du

by changin g u int o 2a — u an d movin g th e integration t o th e line R e u = 77 , where

1 < 7 7 < 2a. The n w e insert (1 0.68 ) int o (1 0.66 ) gettin g

(10.69) B*(8, X) =^/ ( E x W ^ S " + U _ 1 ) ^ W ^ + 0 ( l y)

where
W(u) = 7(5 - 2c r + u)P(s -2a + u)X u~2(Tf(2a - u).
By (1 0.63) , (1 0.62) , (1 0.57 ) w e get fo r u = rj + iv,

W{u) < ((|s | + |^|K)* +CT - r? X r? - 2<7 (2a - r y ) - 1 e - * H

«*-„)-(^"V-.-".
We assum e X 2 ^ £\s\ and7 7 = a + | , gettin g fo r a ^ a ,
(10.70) W(u )< ( a- iJ-^i^e-H .
Hence (1 0.69 ) yield s
oo
m s +<*~i+™ e d v +
/ 00 ' -
- OO m < -y\s
~ xy "
26 0 10. Z E R O - D E N S I T Y E S T I M A T E S

Before goin g furthe r le t u s stat e wha t w e hav e proved .

LEMMA 1 0.3 . Let x be a character modulo £ and s a complex number with
Re s ^ a > | . Choose X ^ (^H) ^ andY satisfying (1 0.67). Then we have (1 0.61 )
where B(s, x) i> s given by (1 0.58) with f as in (1 0.50) and B*(s, x) satisfies (1 0.71 )
with the implied constant being absolute.
Now w e proceed t o th e proo f o f ou r mai n resul t whic h i s
GRAND DENSIT Y THEORE M 1 0.4 . Let\ < a ^ 1 and e > 0 . Then we have

(10.72) N(a, fe, Q , T ) < ( £ > ( 2 + 0 ( l - a ) + H c(a)(l-a)^CA

where

(10.73) c(a )= n ^_i__J_).

Here D = kQT\ H = kQ 2T and C = 2\ogD. The exponent A in the power of

logarithm and the implied constant depend only on a and e.
REMARK. On e could remove e and asser t tha t bot h A an d th e implie d constan t
are absolute , however , suc h a refinemen t i s technicall y ver y complicate d t o mak e
along ou r line s (tr y th e origina l arrangement s b y H . L . Montgomer y [Mo2]) .

Note tha t (1 0.72 ) i s essentially a s stron g a s (1 0.7 ) i n th e rang e | < a < 1 , an d

considerably stronge r wit h respec t t o Q. I n particular , comparin g wit h th e numbe r
of terms i n th e sum , i t show s tha t fo r T fixe d an d an y a > | , ver y fe w L- functions
of primitiv e character s modul o q ^ Q hav e a zer o i n th e rectangl e (1 0.1 ) .
In wha t follow s % r u n s o v e r character s o f modulu s £ = kq < Q an d s = a +1 i s
a comple x variabl e restricte d t o th e rectangl e a ^ < r ^ l , | £ | ^ T . Furthermore , i f
X is principa l w e assum e tha t \t\ ^ £ 2 , becaus e otherwis e th e resul t i s trivial . Fo r
the proo f o f Theore m 1 0. 4 w e use (1 0.61 ) wit h

(10.74) X = D*C, Y = D X2C2.

The residua l ter m i n (1 0.61 ) i s smal l

5x^ff(l-s)X^«D-'
by virtu e o f (1 0.57) , an d th e serie s (1 0.58 ) ca n b e reduce d t o n ^ Y u p t o a
correction ter m 0(D~ 1 ) b y virtu e o f (1 0.51 ) . W e obtai n

(10.75) L ( S,X)= YJ X (n)n-s/(-|)

+ 0(x^- a / \Y^ x(n)n- s+a

-*+lv\e-^dv + D" 1 ).

Multiply thi s b y th e polynomia l

(10.76) M(s, X)= Y, M(™ M

 10. Z E R O - D E N S I T Y E S T I M A T E S 26 1

where 1 < M ^ D* gettin g

(10.77) L(s, X)M(slX)= X )a nX(n)n-s

n^MY
/•OO
/ f°° I I \
s
+ 0(C Y, a n(v)X(n)n- \e~Mdv + D^Mi)

where th e coefficient s ar e give n b y

(10.78) a n= Y. ^H/(^) >

dm=n
d<Y,m<M
d \<*-%+iv
m
(10.79) a n(v)= X ^ )(y)
dm=n
d^Y,m^M

so \a n\ ^ T(TI) an d |a n (^)| ^ r ( n )« Fo r n ^ M w e hav e mor e precis e estimate s

an= J2 M(m){ l + 0 ( e - x / d ) } = ^ M ( m ) + 0( J D- 2 ),
dm=n
v ot—±+iv

The first yield s a i = 1 + 0(£>~ 2 ) an d a n < D " 2

i f 1 < n < M , whil e th e latte r
gives
s r
IXa n(^)x(n)n- | < Y*" * ^ (n)n~* < Y ^ M * log2M .

We wan t thi s t o b e bounde d b y C~ 2 whic h i s th e cas e b y assumin g

a
(10.80) M ^D ~i£-6.

By th e abov e estimate s w e reduce (1 0.77 ) t o

s
(10.81) L(s,x)M(*, X) = l+X *nX(n)n-
M<n^MY

+0 C 1
( I E ^Wxtnln-'je-Wdnr ).
«'—OO -K/r^^s-Tk/r v
M<n^MY

We wan t t o trea t th e first su m an d th e integra l o f th e secon d su m i n on e way . T o

this en d w e combine th e su m an d th e integra l int o on e integra l wit h respec t t o th e
measure

(10.82) dp = \e-Wdv + \6(y)

where dv i s the Lebesqu e measur e o n R an d S(v) i s the poin t measur e a t v = 0 , th e

factor | bein g introduce d fo r th e normalizatio n

(10.83) / d/j, = 1 .
26 2 10. Z E R O - D E N S I T Y E S T I M A T E S

In thi s notatio n w e write (1 0.81 ) a s on e inequalit y

oo . .

/ s
I Y, a n(v)x(n)n- \dfi(v) + C-1
~°° M<n^MY
where fo r v = 0 w e redefin e a n (0) t o b e a n. Pro m no w o n w e onl y nee d t o kno w
that a n(v) doe s no t depen d o n s, x an d tha t |o n (v)| ^ T~{n). For convenienc e w e
also se t a n(v) = 0 i f n ^ M o r n> MY.
Recall that (1 0.84 ) holds for s in the rectangle R(a, T) an d a character x modul o
kq wit h q ^ Q, an d i n cas e \ i s principa l w e requir e \t\ > £ 2 , wher e C — 21ogD
and Z ) = kQT. I n particular , i f s = p is a zero of L(s, x) m thi s region, w e get fro m
(10.84)
roo .
p
J ] Ta n(v)x(n)n- \dn(v) > C' 1
J —c
M<n^MY
provided D is sufficientl y larg e (whic h conditio n ca n b e assumed , o r els e th e goa l
is trivial) . A s i n Sectio n 1 0. 3 w e brea k th e summatio n int o dyadi c segment s N <
n^2N wit hT V - 2 £M, 0^£<L= [lo g Y/ lo g 2]. Denot e
oo . .

/ s
IY a n{v)X{n)n- \dii{v).
"°° N<n^2N
For ever y zer o p bein g counte d ther e exist s £ such tha t
(10.86) Dt(p, X)>L-\

Let Ri denote th e numbe r o f zero s satisfyin g (1 0.86) . The n th e tota l numbe r o f

zeros i s
R^YRe^ LmaxR £.

Raising Dz{s,x) t o a suitable powe r 2k wit h k ^ 2 (dependin g o n N) w e obtai n

2
°° i i

/ s
Yb n(v)x(n)n- \ dp(v)
_0
° P<n<2* P
k
with P = N whic h fall s int o th e segmen t
(10.87) Z < P < ( M F ) 2 + Z*

where Z is an y fixed numbe r t o b e chose n later , subjec t t o MY ^ Z ^ H. A t thi s

point w e requir e M ^ D £/4, s o / c is bounde d i n term s o f e. Th e coefficient s b n(v)
are bounde d b y a power o f th e diviso r function , s o we hav e
2 2a
5 3 \b n(v)\ n- < pi-2"£^
fc
P<n^2 P

where A depends onl y o n k. Addin g u p thes e inequalitie s fo r th e point s (p,x)

satisfying (1 0.86 ) w e obtai n
oo . 2
dp(v).
/ -OO / _ \ , - or > ^ / n
(p,x)eSt P<n^2 k
tn
P
 10. Z E R O - D E N S I T Y E S T I M A T E S 263

Now we are ready t o us e two results fro m Chapte r 9 , namely Theore m 9.1 3 an d
Theorem 9.1 5 (se e als o the closin g remark s i n Chapte r 9 concerning th e spacin g o f
points p). Applyin g Theore m 9.1 3 w e get
(10.88) Ri < C P1~2a(P + H)C A
,
and b y Theore m 9.1 5

R£ ^ P 1 '2"^ + R]H^P*)£ A
,
the latte r givin g
(10.89) R £ < P 1 ~2a(P + H 3 4a
- )CA.
(here H — kQ 2T, an d A i s a constan t dependin g o n k whic h ma y b e differen t i n
each appearance) . A direc t compariso n o f (1 0.88 ) an d (1 0.89 ) show s tha t on e i s
stronger tha n th e othe r i f \ < a ^ | an d | ^ a < 1 respectively.
Suppose \ +s ^ a < | . The n introducin g P fro m (1 0.87 ) int o (1 0.88 ) on e get s

Rt < ((My) 4 *1 -"* + Z^ 1 -^ + HZ l 2a

~ )CA.
Clearly th e bes t choic e i s Z = #1 /(2-00 . W e als o assum e MY ^ Z± gettin g
A
(10.90) R t^iH^=^C .
It remain s t o verif y whethe r th e restrictio n MY ^ Z* doe s no t contradic t th e
conditions previousl y imposed . Fo r Y = D^C 2, an d sinc e D ^ H , thi s restrictio n
is verified wit h
2a-lo
2
M = H^^^C~ .
£/4
Note tha t M satisfie s (1 0.80) , M ^ £> an d obviousl y MY ^ Z . Th e exponen t A
and th e implied constan t i n (1 0.90 ) depen d onl y on e. Sinc e e is arbitrary, Theore m
10.4 is established i n the rang e \ < a ^ | (actuall y i n this rang e (1 0.72 ) hold s tru e
without th e first ter m becaus e i t i s absorbe d b y th e secon d one) .
Suppose | ^ a < 1 . The n introducin g P fro m (1 0.87 ) int o (1 0.89 ) on e get s
4 (1
Rt < ((M7) -Q) + Z 3(i-a) + HZ A Q(X
~ )CA.
Now th e bes t choic e i s Z — H1^3a~1^ givin g
2
Re « ( ( M ) 2 (1 _a)
+ H^^)C A
.
For M — D£^ thi s establishe s Theore m 1 0. 4 i n th e rang e | ^ a < 1 .

EXERCISE 3 . Sho w tha t fo r \ < a < 1 and e > 0,

(10.91) JV(a , jfe, Q,T) < (jD(3+e)(i-a ) + # 6 ( a ) ( i - a ) )£^

where
(10.92) 6(a )= n ^ _ ^ , _ i _ ) .

[Hint: Rais e th e polynomia l D^(s,x ) t o a powe r 2f c with k ^ 3 so tha t Z ^ P ^

(MY)3 + Z 4/3.}
26 4 10. Z E R O - D E N S I T Y E S T I M A T E S

EXERCISE 4 . Sho w tha t fo r \ < a ^ 1 and e > 0,

£){1
(10.93) N(a, k, Q, T) < (D^ ~a) + H *(*)^-*))cA
where
(10.94) a ( a ) = min (_l_ > _L_).
[Hint: Rais e th e polynomia l Di(s,x) t o a power 2/ c with k ^ 4 so tha t Z ^ P ^
( M 7 ) 4 + Z 5 / 4 .]
REMARKS. Fo r a nea r 1 , th e estimate s (1 0.91 ) an d (1 0.93 ) ar e stronge r tha n
(10.72) i n terms o f Q2 , bu t not in terms o f kT. I n particular, i f kT ^ Q £, the n
(10.93) yield s
Ar(a,fe,Q,T)<Q4(1-^+£
in th e rang e ~ ^ a ^ 1 , which i s essentially th e densit y conjecture . B y usin g hi s
Theorem 9.1 0 , M . Jutil a [Jut ] establishe d th e densit y conjectur e fo r zeros o f ((s)
in th e sam e range . Ever y exponen t a(a),fe(a),c(a ) take s it s maximu m a t a = | ,
being o ( | ) = f ,6(f ) = f, c(f ) = f . Fo r a = 1 we ge t o(l ) = f , 6(a ) = | ,
c(l) = |, an d fo r a = | al l thre e exponent s tak e valu e 2.
10.5. Th e gap s betwee n primes .
This sectio n i s written fo r the purpos e o f giving a n example o f how th e zero -
density theorem s ca n b e applied. W e hav e chose n a few question s abou t prime s
in shor t intervals , becaus e the y wer e inspirationa l fo r the developmen t o f density
theorems i n th e firs t place , an d fo r othe r statistica l studie s of zeros of £(s). B y n o
means d o w e attemp t t o show th e bes t result s fro m th e vas t literature .
By th e Prim e Numbe r Theorem ,

^{x) = ^2 A ( n ) = x + ^(x )'

where E(x) is a suitable erro r term , i t follows directl y tha t

(10.95) ^{x + y) - I/J(X) ~y
as x —> oo , provided y = y(x) i s somewhat large r tha n E(x). Henc e there i s a prime
number i n the shor t interva l (x , x + y] for al l sufficientl y larg e x.
Without improvin g th e existin g erro r ter m i n the Prim e Numbe r Theorem , G.
Hoheisel [Ho ] succeeded , nevertheless , i n showing i n 1930 tha t (1 0.95 ) hold s for
y = x6 with som e absolut e constan t 9 < 1. This i s an impressiv e achievemen t give n
the fac t tha t th e erro r ter m a s good a s E(x) = 0(x e) seem s t o be fa r beyond th e
current technology . Recal l that th e latter boun d translate s int o the non-vanishing of
£(s) i n R e s > 0. However , Hoheise l require d onl y tw o result s whic h wer e availabl e
in hi s time . Firs t i s a zero-free regio n of £(cr + it) o f th e typ e
l l0g T
(10.96) \t\^ T , G>\- °f > B
logT
for som e constan t B > 0 and al l T sufficientl y large . Th e secon d ingredien t i s the
zero-density estimat e o f type
c(1
(10.97) N(a,T) < T -a)(logT)A
 10. Z E R O - D E N S I T Y E S T I M A T E S 265

for al l \ ^ a ^ 1 with som e constant s c ^ 2 and A ^ 1 . Pro m thes e result s one

derives
THEOREM 1 0.5 . Put 0 = 1 - ( c + (A + I)/BY1. Then

(10.98) i/>(x + y) - i/>{x) = y + O ( ^ - )

V log x /
/or all y with x ^ l o g x ) 3 ^ y ^ x.
PROOF. On e use s the approximate "explici t formula " (se e Sectio n 5.9)
#*) = *- Y, f+og(logx) 2)
w i t h T = x1~e. Henc e

2/ ^ T w Vlogx /

Here the sum over th e zeros is bounded by

1 1
a
] T x^" ^2 / x ~ dN{a,T)

1
OL
<2aT*W(±,T) + 2(loga: ) f x - N(a,T)da
i
A c
< ^ - i T l o g T + ( l o g x ) ( l o g T ) f\T /xYda
J 7)

< x - i T l o g T + (TVx^OogT) ^
where7 7 = i?(lo g log T)/ lo g T. Sinc e
(c
(TV*)" - ( l o g T ) - ^ ) B = (\ogT)- A
-\
the su m over zero s i s 0(1/log x) completin g th e proof o f Theorem 1 0.5 . •

Hoheisel use d th e zero-free regio n (1 0.96 ) wit h a positive bu t very smal l con -
stant B du e to J. E . Littlewood [Lit] , an d the zero-density estimat e (1 0.97 ) wit h
exponent c = 4 due to F. Carlso n [Car ] (see (10.26)). Afte r Vinogradov' s widen -
ing th e regio n (1 0.96 ) (se e Corollary 8.28 ) on e can tak e B arbitraril y larg e s o
0 = 1 — c - 1 + e satisfie s th e conditions o f Theore m 1 0.5 . Pro m no w on the Ho -
heisel exponen t 0 depends onl y o n c in the zero-density estimate . Not e tha t on e
needs (1 0.97 ) t o hold throughou t th e segment \ ^ a ^ 1 with th e same c , so an
improvement i n a subrange doe s no t help. I n other words , al l one gets fro m (1 0.6 )
is c = maxc(a) . B y the Gran d Densit y Theore m 1 0. 4 for £(s ) one gets (1 0.97 )
with c = 1 2/ 5 (the maximum o f c(a) i s attained a t a = | ) , which yield s (1 0.98 )
for x e ^ y ^ x wit h 6 = ^ + e. Thi s i s the best resul t o f its kind obtaine d s o far
(in 1 97 2 by M. N. Huxley [Hu3]) . Th e density conjectur e c = 2 yields 0 = \ + e.
Various combination s o f the abov e analyti c argument s wit h siev e methods pro -
duced estimate s
(10.99) y < ip(x + y) - i/>(x) < y
26 6 10. Z E R O - D E N S I T Y E S T I M A T E S

in plac e o f th e asymptoti c formul a (1 0.95) , however , fo r shorte r intervals . Fo r

example, R . Bake r an d G . Harma n [BH ] go t (1 0.99 ) fo r y = x e wit h 0 = 0.534 .
From (1 0.99 ) wit h y = x 6 i t follow s tha t th e differenc e betwee n consecutiv e
primes satisfie s

(10.100) d n = P n + l - P n < P n -

Therefore w e hav e (1 0.1 00 ) wit h 6 = 0.53 4 a s a consequenc e o f th e Baker-Harma n

i
work. Recal l tha t th e Rieman n hypothesi s yield s d n < C Pn logpn , bu t eve n o n th e
Pair Correlatio n Conjectur e (se e Chapte r 25) , th e bes t tha t ha s bee n achieve d i s
(due t o D . Goldsto n an d D . R . Heath-Brow n [GHB] )

(10.101) d n < (p nlogpn)i.

Creating a probabilisti c mode l fo r prime s i n 1 93 7 H . Crame r [Cra ] wa s le d t o a

conjecture tha t
2
(10.102) d n < (logpj .

In th e othe r directio n R . Ranki n [Ra2 ] showe d b y constructin g specia l com -

posite number s tha t

(10.103) d n ^ (e 7 - £)(logp n )(loglogp n )(loglogloglogp n )(logloglogp n )~ 2

infinitely often , wher e 7 is the Eule r constant . Pau l Erdo s offered a price of $1 0,00 0
to anyon e wh o ca n replac e e 7 i n (1 0.1 03 ) b y a functio n increasin g t o infinit y (th e
largest pric e eve r offere d b y Erdo s fo r a solutio n t o a mathematica l problem) .
It i s conjecture d tha t th e normalize d gap s betwee n consecutiv e prime s d n* —
{pn+i — Pn)(logPn)- 1 ha s a Poisso n distribution , tha t i s fo r an y t > 0,

(10.104) li m -\{n ^ x; d n* ^ t}\ = 1 - e~K

x—>oo X

Many estimate s fo r ip(x + y) — ip(x) wer e establishe d o n averag e wit h respec t

to x.
EXERCISE 5 . Assumin g (1 0.97 ) wit h c ^ 2 prove tha t

2
(10.105) / (0( x + y) - ^J(X) - yfdx « y X{\ogX)~A
Jx

for X e ^ y < X wit h 0 — 1 — 2c _ 1 + e, fo r an y e > 0 an d an y A > 0 , th e implie d

constant dependin g onl y o n e and A. I n particular , b y taking c = 1 2/ 5 deduce tha t
(10.95) hold s tru e wit h y = x e fo r almos t al l x , wher e 0 > | i s a fixed number .
We close thi s sectio n b y pointin g ou t a forma l similarit y betwee n th e problem s
of prime s i n shor t interval s an d prime s i n arithmeti c progression s t o larg e moduli .
 10. Z E R O - D E N S I T Y E S T I M A T E S 26 7

EXERCISE 6 . Assum e th e followin g estimate s fo r th e zero s o f Dirichle t L-

functions wit h character s x ( m ° d q):
(1) L(s , x) 7 ^ 0 in th e regio n s = a + it wit h \t\ ^ T an d
l
(10.106) a * 1 - B ^ f
log qT
where B i s a positiv e constant , provide d qT i s sufficiently large ,
(2) Th e numbe r N(a,q,T) o f zero s o f al l L(s,x ) w ^ h x ( m ° d q) i n th e rec -
tangle \t\ < T , a ^ a satisfie s
(10.107) N(a,q,T) < (gT^-^log ^
for ^ ^ a < 1 , T ^ 3 , where c > 2 and A ^ 1 are suitabl e constants .
Prove tha t
(10.108) ^ ; , , 0 )= ^ ( 1+ 0 (_i_))

uniformly fo r x ^ ^ ( l o g g ) 3 wit h 6 — c -f ( A + l ) / £ . Se e Chapte r 1 7 for uncondi -

tional results .
This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/12

CHAPTER 1 1

SUMS OVE R FINIT E FIELD S

11.1. Introduction .
In thi s chapte r w e conside r a specia l typ e o f exponentia l an d characte r sums ,
called sometime s "complet e sums" , whic h ca n b e see n a s sums ove r th e element s of
a finite field. Althoug h th e method s o f Chapte r 8 can stil l b e applie d t o th e stud y
of suc h sums , disregardin g thi s specia l feature , th e deepes t understandin g an d th e
strongest result s ar e obtaine d whe n th e finite field aspec t i s taken int o accoun t an d
the powerfu l technique s o f algebrai c geometr y ar e brough t t o bear .
We have alread y encountere d i n th e previou s chapter s som e example s o f expo -
nential sum s whic h ca n b e interprete d a s sum s ove r finite fields, fo r example , th e
quadratic Gaus s sum s

<™= E (fMf)
x mo d p
or th e Kloosterma n sum s (1 .56 )

S(„,M= r . ^ ) .
x mo d p

In thi s chapte r w e wil l stud y thes e sum s i n particular . Th e culminatin g poin t o f

our presentatio n i s the elementar y metho d o f Stepano v whic h w e apply fo r provin g
Weil's boun d fo r Kloosterma n sum s
\S(a,b]P)\^2y/p
and Hasse' s boun d fo r th e numbe r o f point s o f a n ellipti c curv e ove r a finite field.
Then we survey briefly, withou t proofs , the powerful formalis m o f £-adic cohomolog y
developed b y Grothendieck , Deligne , Katz , Laumo n an d others , hopin g t o conve y
a flavor o f th e tool s involve d an d t o giv e th e reade r enoug h knowledg e t o mak e
at leas t a preliminar y analysi s o f an y exponentia l su m h e o r sh e ma y encounte r i n
analytic numbe r theory .
11.2. Finit e fields.
We first recal l briefl y som e fact s abou t finite fields, an d establis h th e notation s
used i n thi s chapter . Fo r ever y prim e p , th e finite rin g Z/pZ o f residu e classe s
modulo p i s a field, whic h w e denot e F p . Th e Galoi s theor y o f ¥ p i s ver y eas y t o
describe: fo r an y n ^ 1 , ther e exist s a uniqu e (u p t o isomorphism ) field extensio n
of F p o f degre e n , writte n ¥ pn. Conversely , an y finite field F wit h q element s i s
isomorphic (bu t no t canonically ) t o a unique field F pd, s o q = p d, an d F admit s als o
a uniqu e finite extensio n o f degre e n fo r an y n ^ 1 , namely ¥ pdn.
Let no w F = ¥ q b e a finite field wit h q—p d elements . I n mos t o f th e chapter ,
p is fixed an d w e chang e notatio n slightly , denotin g b y F a n algebrai c closur e o f F
269
270 11. SUM S OVE R FINIT E FIELD S

and b y F n C F th e uniqu e extensio n o f degre e n o f F fo r n ^ 1 . Th e contex t wil l

always indicat e clearl y tha t th e cardinalit y o f F n i s q n an d no t n .
The extensio n F n / F i s a Galoi s extension , wit h Galoi s grou p G n canonicall y
isomorphic to Z/nZ, th e isomorphism bein g the map Z/nZ—> • G n define d b y 1 H> a,
where a i s th e Frobeniu s automorphis m o f F n give n b y a(x) = x q.
Let F b e a give n algebrai c closur e o f F , s o by th e above ,

n^l

By Galoi s theory , fo r an y x G F, w e hav e X E F <<=4> <J(X) = x i f an d onl y i f

xq — x an d mor e generall y

(11.1) x £ ¥ n <=^> a n{x) = x ^ > x qn

= x.
Prom thi s w e can deduc e tha t F n i s the splittin g fiel d o f the polynomia l X q — XG
¥[X]. Mor e precisely , on e ca n stat e th e followin g resul t o f Gauss :
LEMMA 1 1 .1 . For any integer n ^ 1 , we have
qn
(11.2) Y[ Y[ P = X -X
d\n deg(P)= d

where the product ranges over all irreducible monic polynomials P of degree d di-
viding n.
P R O O F . Thi s i s a n immediat e consequenc e o f th e descriptio n o f finit e fields :
the roots of the polynomial o n the right sid e (i n an algebraic closure) ar e exactly th e
elements x G F n wit h multiplicit y on e and , conversely , ever y suc h x ha s a minima l
polynomial whic h mus t occur , exactl y once , amon g th e polynomial s P o n th e lef t
side. •

Associated t o the extension F n / F ar e the trace map and the norm map. Becaus e
of the abov e descriptio n o f the Galoi s grou p o f F n / F , th e trac e ma p T r = TvY n/w :
F n— » F i s give n b y

(11.3) Tt(a: )= V J <T \x)= £ x**

while th e nor m ma pT V = iV Fn / F : F *— > F * i s similarl y

a
(11.4) N(x) = Yl ^x) = Yl x q
= x 3
^ .

The equation s T r (x) = y an d N(x) — y, fo r a fixe d y G F ar e ver y important ,

Because th e extensio n F n / F i s separable , th e equatio n T r (x) = y alway s ha s a
solution. I f #o is a given solution, then all solutions are in one-to-one correspondenc e
with solution s o f T r (a) = 0 , b y x = Xo + a . Moreover , an y solutio n o f T r (a) = 0
is o f th e for m a = a(b) — b = b q — b for som e b £ F n , uniqu e u p t o additio n o f a n
element i n F .
Similarly, fo r an y y G F*, th e equatio n N(x) = y ha s a solution , an d i f xo i s a
given solution , th e se t o f solution s i s i n one-to-on e correspondenc e wit h solution s
of N(a) = 1 , whic h b y Hilbert' s Theore m 9 0 (o r b y direc t proof ) ar e al l give n b y
 11. SUM S O V E R F I N I T E F I E L D S 271

a = a(b)b l — b q x fo r some 6GF*, uniqu e u p to multiplication b y an element i n

F*.
As the additive grou p o f F is finite, the general theor y o f characters o f a finit e
abelian grou p (se e Chapter 3 ) can be applied. Character s o f F are called additiv e
characters, an d the y ar e all of the for m x—
i > > ip(ax) fo r some a G F, where tp is some
fixed non-trivia l additiv e character . Fo r instance, le t Tr : F —> TLjpZ be the trace
map t o the base-field, the n

(11.5) ip(x) = e(Tr(x)/p)

is a non-trivial additiv e characte r o f F. Fo r a given additiv e characte r ty and a G F,

we denote b y ipa the character x *-> I/J(CLX).
Applying th e general theor y o f characters o f finite abelia n groups , w e get the
orthogonality relation s

5>(*) = {q0 iotherwis

fx— 1,
e

(which i s used t o "solve " th e equation x = 0 in F) an d

f q i f ^ — 1 is the trivial character ,

E I/J(X) — <
1 0 otherwise .

The descriptio n o f characters o f the multiplicativ e grou p F * (als o calle d multi -

plicative character s of F) is not so explicit. Th e grou p structur e o f F* is well-known
(dating bac k t o Gauss) : i t i s a cycli c grou p o f order q — 1 . Generator s o f F* are
called primitiv e roots , an d there ar e <p(q — 1 ) of them, bu t no useful formul a fo r a
primitive roo t exists . Fixin g one , sa y z G F *, on e has an isomorphism

f F * ~ Z/(q - 1 ) Z
log : <
i >• n suc h tha t z n = x
I x—

and al l multiplicative character s o f F are expressed as

'alog(x)s
/alog(x)\
Q

for som e a £ Z/( g — 1 )Z , bu t suc h a descriptio n i s usually o f no use in analyti c

number theory .
As example s o f multiplicative characters , suppos e F = Z/p Z an d p^ 2 . The n
the Legendr e symbo l
'2N
X \->
\p, • )

is a non-trivial quadrati c character . I n general, i f 5 | (q—1), ther e i s a cyclic grou p

of orde r S consisting o f characters \ ° f ^* ° f order 5.
The orthogonalit y relation s becom e

^—^ y 0 otherwise ,
272 11. SUM S OVE R FINIT E FIELD S

(the su m ove r al l multiplicativ e characters) , an d

q—1 i f x = 1 is the trivia l character ,
£ *(*) = {0 otherwise .
xeF* ^
It i s usual to extend multiplicative character s to F by defining x(0 ) = 0 if x ^ 1 ,
and x(0 ) = 1 if x = 1 . Notic e the n tha t fo r an y b\q—\ th e formul a

(n.6) Y. x(* ) = \{ye F I ^ = x >i

(also a particula r cas e o f th e orthogonalit y relation s fo r th e grou p F*/(F*) d , a s

described i n Chapte r 3 ) i s tru e fo r al l x G F .
11.3. Exponentia l sums .
Let F = ¥ q b e a finite field wit h q = p m elements , p a prime . Exponentia l
sums ove r F ca n b e o f variou s kinds . Fo r th e simples t case , conside r a polynomia l
P G ¥[X] an d a n additiv e characte r -0 , and defin e th e su m

S(P) = £ > ( P ( * ) ) .
x£F

Slightly mor e generally , tak e a non-zer o rationa l functio n / = P/Q G ¥(X) an d

consider
S(f)= £ 1 >(f(x));
xeF
Q(x)^0
for instance , takin g q = p an d f(x) = ax + bx~ l, w e hav e S(f) = 5(a , 6;p) .
Multiplicative character s ca n als o b e used , gettin g sum s o f th e typ e

s*(/) = £**(/(*))
(where th e sta r i n J^ * mean s her e an d hencefort h tha t th e summatio n extend s
to al l x whic h ar e no t pole s o f / ) . Fo r g = p , X = ( " ) ( t n e Legendr e symbol )
and f(x) G Z[X] a cubi c polynomia l withou t multipl e root s modul o p , w e see tha t
—Sx(f) i s the p-th coefficien t a p o f the Hasse-Weil zet a function o f the ellipti c curv e
with equatio n y 2 — f(x) (se e Sectio n 1 4.4) .
Still mor e generally , on e ca n mi x additiv e an d multiplicativ e characters , an d
define sum s suc h a s
(11-7) S x(f,g) = ^2\(f(x))^(g(x)),
XEF

an exampl e o f whic h i s the Sali e su m T(a, b;p) define d b y

x N
mt , x v^ *( \ (ax ' ax+- f bx
bx\
r e
^p/ \ p
<".M= E Q x mo d p (^-)
which occurs in the Fourier expansio n of half-integral weigh t modula r forms ; se e [16]
for instance . I n contras t wit h th e seemingl y simple r Kloosterma n sum s 5(a , 6;p) ,
the Sali e sums T(a, b\p) ca n be explicitly compute d (se e Lemma 1 2.4 , and Corollar y
21.9 fo r th e unifor m distributio n o f th e "angles " o f th e Sali e sums) .
 11. SUM S OVE R FINIT E FIELD S 27 3

To end this list, we mention that al l these definitions ca n again be generalized t o

sums in more than on e variable, an d tha t th e summatio n variable s ca n be restricte d
to th e rationa l point s o f a n algebrai c variet y define d ove r F : som e example s wil l
appear i n th e surve y section s o f thi s chapter .
The exponentia l sum s whic h directl y aris e i n analyti c numbe r theor y ar e sum s
over th e prim e field Z/pZ . However , th e deepe r understandin g naturall y require s
considering sum s ove r th e extensio n fields F p n. Indeed , th e ver y reaso n fo r th e
success of algebraic method s lie s in the fact tha t a n exponential su m over F p doesn' t
really com e alone , bu t ha s natura l "companions " ove r al l th e extensio n fields ¥ pn,
and i t i s reall y th e whol e famil y whic h i s investigate d an d whic h i s th e natura l
object o f study . Thos e companio n sum s ar e easil y defined : tak e th e mos t genera l
sum S = S x(f,g) w e have introduced , the n fo r n ^ 1 let

(11.8) S n = Y? x(tfF n /F(/(aO)MltF n /F(s(z)))

x£Fn

where we use the multiplicativ e characte r x o iV and th e additiv e characte r i\> o Tr o f

F n . Al l th e sum s S n ar e incorporate d int o a singl e object , th e zet a functio n o f th e
exponential sum , whic h i s the forma l powe r serie s Z = Z x(f, g) G C[[T]] define d b y
the formul a

*—p(EM
n
n^l
Justification fo r th e introductio n o f th e zet a functio n come s fro m th e followin g
rationality theorem , conjecture d b y Weil , an d prove d b y Dwork .
T H E O R E M 1 1 . 2 ( D W O R K ) . The zeta function Z is the power series expansion
of a rational function; more precisely, there exist coprime polynomials P and Q in
C[T], with P(0) = Q(0 ) = 1 , such that Z = g .

As a corollary , denot e b y (c^ ) (resp . (/3j) ) th e invers e o f th e root s (wit h mul -

tiplicity) o f P (resp . Q) , s o

P = JJ(l-aiT), Q = JJ(l- i 8 i T).

Then usin g th e power-serie s expansio n

n^l

we find tha t th e formul a Z = P/Q i s equivalent t o th e formul a

for an y n ^ 1 , whic h show s ho w th e variou s sum s S n ar e related . I n particular ,

note tha t the y satisf y a linea r recurrenc e relatio n o f orde r d equa l t o th e numbe r
deg P + de g Q o f root s a^ , fij .
27 4 11. SUM S OVE R FINIT E FIELD S

COROLLARY 1 1 .3 . We have for any n ^ 1 the upper bound

3i

In particular,

(11.9) |S|<I>; l + I>i| .

3i

A commo n abus e o f language i s to spea k o f the c ^ an d (3j a s the root s o f the

exponential su m S. W e will describ e i n Section 1 1 .1 1 a number o f general facts
about thes e roots . I n the intervening sections , w e will prove Dwork' s Theore m and
estimate th e modulu s o f th e root s i n the importan t specia l case s o f Gauss sums ,
Kloosterman sum s an d for the local zet a functio n o f elliptic curves .
REMARKS. W e have calle d th e sums S n,77 - ^ 1 , "companions " o f the origina l
exponential su m S. However , on e can consider othe r companion s o f S as well. I f S
involves an additive character , it can also be very useful sometime s to consider 5 as
just on e element o f the family o f sums obtaine d b y varying th e additive character ,
specifically i f
s = £*x(/(*Ms(s))>
we als o introduc e fo r a G F,

Sa = £ * x ( / ( s ) W a ( s ( * )) = £ * x ( / ( * ) M a c , ( z ) ) .
x£F xe¥

Estimates o n average ove r a for the first fe w power moment s o f Sa are often easil y
derived b y elementar y means , an d they ca n be of great us e in estimatin g 5 , eve n
in additio n t o the method s o f algebraic geometry . Se e the proo f o f Weil' s boun d
for Kloosterma n sum s i n Section 1 1 . 7 and th e example s i n Section 1 1 .1 1 . Mor e
general type s o f families hav e bee n (an d still are ) extensively studie d b y Katz; see
for instanc e [Kl].

11.4. Th e Hasse-Davenpor t relation .

We consider genera l Gaus s sum s ove r a finite field. Le t F = ¥q be a finite field
with q — p m elements , an d le t ip be a n additiv e characte r an d \ a multiplicativ e
character o f F. Th e Gauss su m G(x, V J) ls

(11.10) G( Xi)=ExWW'
xeF
ls
(recall tha t \ extende d t o F by x(0) = 1 , 0 accordin g t o whethe r \ i s trivial o r
not). Whe n \ i s the Legendre symbol , on e recovers quadrati c Gaus s sums .
The associate d sum s ove r th e extensions fields are
G „ ( x , ^ ) = Y. x(N ¥n/¥(x))iP(Tr¥n/¥(x))
xe¥n
and th e zeta functio n is
qA Tn
(11.11) Z( X^)=e^(J2 ^ )-
n^l
 11. SUM S O V E R F I N I T E F I E L D S 275

In thi s case , Dwork' s Theore m wa s prove d b y Hass e an d Davenpor t an d i s

known a s th e Hasse-Davenpor t Relation .
T H E O R E M 1 1 . 4 (HASSE-DAVENPORT) . Assume x and ip are non-trivial. Then
we have for any n ^ 1 ,
-Gn(X,rl>) = {-G( X1, >))n
or equivalently the zeta function is a linear polynomial

Z(X,1>) = l + G{x,1 >)T.

Hence the only "root " fo r the Gauss sum is G{\, ip) itself. Thi s can be estimate d
elementarily, a s wa s don e fo r th e Gaus s sum s considere d i n Chapte r 3 .
PROPOSITION 1 1 .5 . We have

\G(x,1>)\ = y/Q
nor 7 s
if neither \ V ^ trivial, while
,/ N ,( 0 if tb non-trivial,
[q ifip = l
,, x ,f 0 if X non-trivial,
\G(XA)\=\ r ,
{ q ifx = 1.
nor ls
PROOF. Th e las t tw o statement s ar e immediate , s o assum e neithe r % ^
trivial. W e hav e

|G(x,^)l2= £ x(x)x(yMxM-y)
x,y£W*

= n writin z = x
^2^( z) Yl ^(( z ~ ^ (° § v~l)
— q (b y orthogonality , applie d twice. )
•
We no w tur n t o th e proo f o f th e Hasse-Davenpor t Relation . W e conside r th e
field F = ¥(X) o f rationa l function s o n F an d th e rin g R = ¥[X] o f polynomials .
Recall tha t R i s a principa l idea l domain . Fo r h £ R o f degre e d ^ 0 , w e define th e
norm
N(h) = q d.
The zet a functio n o f F i s th e Dirichle t serie s (analogou s t o th e Rieman n zet a
function)
CF(S)= £ N{h)->.
heR
h moni c
REMARK. Thi s coul d als o be writte n a s a su m ove r th e non-zer o ideal s a in R,

Ms) = X>(°)~s
a
whereA T (a) = \R/a\ = N(h) fo r an y polynomia l h suc h tha t a = (h). Bu t w e wil l
work wit h polynomial s t o emphasiz e th e elementar y spiri t here .
27 6 11. SUM S OVE R FINIT E FIELD S

The Dirichle t serie s CF(S) converges absolutel y fo r R e (s) > 1 . Indeed , puttin g

n(d) = {h G R | deg(h) — d, h i s monic} — q d

we obtai n immediatel y

cF(s) = J2n(d)i'ds = E<?(1~ s)d = c1 - ^y 1

-

On th e othe r hand , uniqu e factorizatio n int o irreducibl e polynomial s yield s a n

expression o f £ p a s a n Eule r produc t
1
<F(S)= n (i-A^rT
PeR
P moni c irreducibl e

which i s convergen t fo r R e (s) > 1 .

The firs t ste p i n th e proo f o f th e Hasse-Davenpor t Relatio n consist s o f writin g
the zet a functio n o f Gaus s sum s a s a n L- function fo r th e field F. Le t H C F * b e
the subgrou p o f rationa l function s whic h ar e quotient s o f moni c polynomials , an d
G C H a subgrou p wit h th e propert y

hih2 e G => hi, h 2 e G.

Then i f a : G - > C * i s a characte r o f th e grou p G, i t ca n b e extende d t o a totall y

multiplicative functio n o f th e se t o f moni c polynomial s h G R b y puttin g a(h) = 0
if h $L G. Th e correspondin g L- function i s define d analogousl y t o th e classica l
L-functions b y th e Dirichle t serie s
s 1s
L{s,a)= J2 a(h)N(h)- = l[(l-a(P)N(P)- )-
heR P
h moni c

for R e (s) > 1 .

For dealin g wit h Gaus s sum s w e conside r th e subgrou p G C H o f rationa l
functions / define d an d non-vanishin g a t 0 . Defin e a characte r A on G b y

X(h) = x ( a d ) ^ ( a i )

for h = X d — a\X d~l + • • • + (— l)ddd G R. Clearl y A is multiplicativ e o n moni c

polynomials, an d extend s t o a characte r o f G. I n thi s cas e w e ge t th e following :
LEMMA 1 1 .6 . We have L(s, A ) = 1 + G(x, *P)q~ s-
PROOF. W e arrang e th e Dirichle t serie s fo r L(s,A ) accordin g t o th e degre e o f
h:
£(*,A) = £ ( E A00)<r *
d^O deg(h)=d

and evaluat e eac h ter m i n turn . Fo r d — 0 , th e onl y moni c polynomia l occurrin g i s

h — 1, and A(l ) = 1 . Fo r d — 1, we hav e h — X — a so tha t

£ \(h) = Y,MX-a)=Y,x(aMa) = G(x,i>).

deg(fr) = l aE¥ a£F
 11. SUM S OVE R FINIT E FIELD S 27 7

For an y d ^ 2 we hav e
d
Y^ X(h)= Yl A(X -a
1 X1
d
" + -. . + ( - l ) d a d )
deg(h)=d ai,...,a d €F

d 2
=q ~ £ xMV-(ai)= 0
ai,a d GF

by orthogonality , becaus e a t leas t on e o f th e character s x > ^ i s non-trivial . •

On th e othe r hand , appealin g t o th e Eule r produc t w e will prove :

LEMMA 1 1 .7 . We have L(s , A) = Z(q~ s), where Z = Z(xii>) i> s
^e zeta func-
tion (1 1 .1 1 ) associated with Gauss sums.
Theorem 1 1 . 4 follow s fro m Lemma s 1 1 . 6 an d 1 1 .7 .
P R O O F O F LEMMA 1 1 .7 . Takin g th e logarithmi c derivativ e o f the Eule r prod -
uct, w e ge t

n ^ l rd=n p
deg(P)=d

while, o n th e othe r hand ,

1 Z'(q- S)
= ^G„(xi)f
logq Z(q-°) n>i

It therefor e suffice s t o prov e th e formul a

n d
(11.12) J2 d\(P) / = G n(X,iP)
P
d=deg(P)\n

for n > 1 , th e equalit y o f th e logarithmi c derivative s bein g sufficien t t o impl y

Lemma 1 1 . 7 since bot h side s ar e Dirichle t serie s wit h leadin g coefficien t 1 .
To prove (1 1 .1 2) , le t P b e on e o f the irreducibl e polynomial s appearin g o n th e
left side , o f degre e d \ n. It s roots , sa y a?i,... , x^ ar e i n F n . Fi x on e roo t x — Xj
and writ e
1
P = X d- a i X ^ " + • • • + (-l) d
ad.
We ge t

N(x) = (N Wd/F(x)r/d = a n /d
d ,

Tr(x) = - T r F d / F ( x ) = - a x

hence

HP)n/d = ( x ( a ^ ( a i ) ) " / d = xW*)*^) = x (N(x))^{Tr(x)).

27 8 11. SUM S OVE R FINIT E FIELD S

Summing ove r al l root s of P we deriv e

d

i=l
and summin g ove r al l P with deg(P ) | n, we get (1 1 .1 2 ) b y Lemma 1 1 . 1 since ever y
element i n ¥n wil l appea r exactl y onc e as one of the root s xi for some P. •

11.5. Th e zet a functio n fo r Kloosterman sums .

Next w e consider Kloosterma n sums . Le t F b e a finite field wit h q = p m
elements an d thi s time conside r additiv e character s ip and (p. We define th e Kloost -
erman su m associate d t o ip and <p by
1
(11.13) S^,ip) = -^2tl;(xMx- ),
xGF*

(the minu s facto r i s only fo r cosmetic reasons) . Whe n q — p is prime, an d I/J(X) =

e(ax/p), <p(x) — e(bx/p), w e hav e therefor e S(ip,ip) = —S(a,6;p) .
The companio n sum s ove r th e extensio n fields F n ar e
r1
S n f t M = - E ^(Tr(x)M rr(x- ))

and th e Kloosterma n zet a functio n is

We wil l prov e Dwork' s Theore m i n this case , whic h i s due to Carlitz.

THEOREM 1 1 .8 . Assume that ip and ip are both non-trivial Then

m<p) =
i-s(n,,l)T +^-
The proof i s very similar t o that o f Theorem 1 1 .4 . W e put R = F[X], F = ¥(X)
as before , an d conside r th e sam e grou p G C F* o f quotients o f monic polynomial s
defined an d non-vanishin g a t 0. We defin e a character7 7 : G -» C* by putting
77(h) = ip{ai)(p(a d-i/ad)

for a monic polynomia l h G G, where w e write (compar e th e previou s section )

ft = X d + a i X ^ " 1 + • • - + a d _xX + a d
(with ad 7^ 0 since h e G). Th e followin g computatio n verifie s tha t7 7 i s indeed a
character o f G: le t ft' = Xe + b\X&~x + • • • + be-\X + o e wit h 6 e 7^ 0, the n
ftft' = X d+e + (a x + ^ X ^ 6 " 1 + • • • + {a d-ibe + a A - i )* + a dbe
and
/,,A , ,. , N /a d-ibe + a dbe-i\
77(ftft ) = ^(a i + bi)(p{ — J
= ^{ai)ip{ad-ila d)^{bl)ip{be-i/be)

= n(h)r 1 (hf).
 11. SUM S O V E R F I N I T E F I E L D S 27 9

Recall tha t w e extend7 7 t o all h G R by putting 77 (h) = 0 for h $_ G.

LEMMA 1 1 .9 . For ip and ip non-trivial, the L-function associated to 77 is given
by
L(s,rj) = ^riWNih)-3 = 1 - S^^)q~ s + q 1 '28.
h

PROOF. B y arranging term s accordin g t o the degree o f h, we writ e

-ds
£ M = £( E »K>0
l\ ) )H y IL

d^O deg(h)=d

and evaluat e th e inner sums . Fo r d — 0, we have onl y h = 1 and 77(1) = 1 . For

d = 1, we have / i = X + a with a ^ 0 , hence

£ v(h) = E rt* + a) = E V-faM O = -s(V, *>)•

deg(/i) = l aGF * aGF *

For d = 2 , w e ge t
aX
E*? W = E ^ + + 6) = E ^Mafe" 1 )
d e g ( / i ) = 2 aG F aG F
6GF* 6GF *

= 9 - l + ( £ > ( « ) ) ( $ > ( & ) ) =9

aGF* &GF *

by applyin g twic e th e orthogonality o f characters, sinc e neithe r i\) no r ip is trivial.

Finally, fo r d ^ 3 , we get

d e g ( h ) = d aGF * ai,.. . ,a d_iGF

= / ~3 Y, V^iM«d-i O = 0
ai,ad_i€F
aGF*

since ther e i s free summatio n ove r 0 1 G F. •

LEMMA 1 1 .1 0 . For ip and p non-trivial, we have the identity

This lemm a complete s th e proof o f Theorem 1 1 .8 .

PROOF. Th e L-functio n ha s a n Euler produc t
1
L(sJr,) = l[(l-T,(P)N(P)-'r .
P

Taking th e logarithmic derivativ e w e get

1 L'{s,ri)
\ogq L(s,rj)
= Edeg(P)E 7 ?( P )V r d e s ( P ) s
r>l
d
E ( E E nipyy
n ^ l rd—n deg(P)= r
 28 0 11. SUM S OVE R FINIT E FIELD S

and a s befor e i t suffices t o prove th e formul a

n d
(11.14) Yl dn{P) l = -S ntyt<p)
d=deg(P)\n

for n ^ 1 . Le t
P = Xd + aiXd~l + • • • + ad^X + a d
be on e of the irreducibl e polynomial s o n th e lef t sid e of (11.14), of degree d | n, and
#i,..., Xd its roots , whic h li e i n F^ . W e hav e fo r eac h z,
_ n _n
T r ( ^ i ) = ^ T r F d / F ( Z i ) = - ^ ai

(since a ^ X ^ X " 1 ) = X d + ^ ^ d _ 1
+ • • • + a'1) an d

•&(*, ) = -Tr ¥d/¥(Xl ) = - 5 — •

Hence
,(P)»« = v ,(= 01 ) v (^)^(- I ,w-,-. )
and summin g ove r th e root s a^ , then ove r th e polynomial s P of degree d | n, w e
obtain (1 1 .1 4 ) b y Gauss' s Lemm a again . •

Theorem 1 1 . 8 allow s u s to factor th e Kloosterma n zet a functio n

Z(V,^) = ( l - S ( ^ ^ ) T + g r 2 ) - 1 = ( l - a T ) - 1 ( l - / 8 T ) - 1
where a an d / ? are comple x numbers , an d o f course a -f- /? = 5(^> , ip), a/3 = q. In
sharp contras t t o the cas e o f Gauss sums , however , th e root s a an d / ? cannot be
explicitly computed .
T H E O R E M 1 1 .1 1 ( W E I L ) . Assume that ip and <p are non-trivial and p ^ 2 .
Then the roots a and f3 for the Kloosterman sum 5( /0,^>) satisfy \a\ = |/3| = y/q,
and therefore we have

1
(11.15) 5 ( ^ ) ^ 2 ^ .

We will prov e Theore m 1 1 .1 1 in the nex t tw o sections .

COROLLARY 1 1 .1 2 . Let a, b, c be integers, c positive. We have
2
(11.16) \S(a,b]c)\ < r(c)(a,b,c)^ c^2.

P R O O F . B y the twisted multiplicativit y (1 .59 ) for Kloosterman sums , it suffices

to conside r c = pv wit h p prime an d v ^ 1 . If p j nra, w e hav e Ramanuja n sum s
for whic h th e resul t i s easy (se e (3.2) , (3.3)) . Otherwise , th e case v = 1 follows
from Theore m 1 1 .1 1 fo r p ^ 3 , and fo r p = 2 one check s immediatel y tha t the
Kloosterman sum s modul o 2 satisfy Theore m 1 1 .1 1 : w e hav e 5(1 ,1 ; 2) = 1 , an d
the associated zet a function i s therefore Z(T) = 1 +T+2T 2 , wit h roots (—l±iy/7)/4
of modulu s l / \ / 2 .
The cas e p \ nm and / ? > 2 can b e dealt wit h elementarily ; se e Exercis e 1 o f
Chapter 1 2 . •
 11. SUM S O V E R F I N I T E F I E L D S 281

EXERCISE 1 . Conside r a general Kloosterman-Sali e su m

and it s associate d companion s S n an d zet a functio n Z , wher e ip and <p ar e additiv e

characters o f F and \ ^ s multiplicative (s o \ = 1 is the cas e of Kloosterman sums) .
Show tha t
Z = (1 - 5( X ; ^ ^)<T S + x(-a) l 2s l
X{b)q - r
where

«,)-.(5M), *)-.(») .
11.6. Stepanov' s metho d fo r hyperelliptic curves .
We will prove Theore m 1 1 .1 1 by deducin g it from th e Rieman n Hypothesi s fo r
certain algebrai c curve s ove r finite fields. However , w e us e Stepanov' s elementar y
method (se e [Ste] , [Sch] , [Bo3] ) instea d o f Weil's arguments .
Let F be a finite field with q element, o f characteristic p. W e will only conside r
algebraic curve s Cf ove r F given b y equation s o f the typ e

(11.17) C f : y2 = f{x)
for som e polynomia l / G ¥[X] o f degree m ^ 3 . W e assum e moreove r th e followin g
condition

(11.18) Th e polynomia l Y 2 - f(X) G F[X, Y] i s absolutely irreducibl e

(i.e. i t is irreducible ove r th e algebrai c closur e o f F). Thi s i s a minimal regularit y

assumption o n th e curv e Cf. I t is easily see n to be equivalen t t o the conditio n tha t
/ i s not a square i n F[X], an d w e will us e it in this form .
REMARK. Stepanov' s metho d ha s been refined b y Schmidt [Sch ] an d Bombier i
[Bo3] an d i s capable o f handling th e genera l cas e o f the Rieman n Hypothesi s for
curves; th e cas e of curves wit h equatio n o f the typ e y d = f(x) i s not muc h harde r
than th e on e treate d here . W e limi t ourselve s t o the curve s Cf fo r simplicity, an d
because i t suffice s fo r the application t o Kloosterma n sum s an d elliptic curves .
Note tha t curve s o f the type y 2 = f(x) ar e instances o f so-called hyperellipti c
curves, which ar e quit e naturall y distinguishe d amon g algebrai c curve s (bu t no t all
hyperelliptic curves are of this form; se e for instanc e elliptic curves in characteristic s
2 an d 3) .

The proble m w e consider i s that o f estimating th e numbe r |C/(F) | o f F-rationa l

points of C/, i.e. , th e numbe r N of solutions (x , y) G F 2 t o the equatio n y 2 = / ( # ).
We ar e especially intereste d i n this questio n whe n q is large (typically , a s with
exponential sums, the polynomial / G ¥[X] i s fixed, and we consider th e F n -rational
points fo r all n ^ 1 ) , althoug h w e will obtai n completel y explici t inequalities .
THEOREM 1 1 .1 3 . Assume that f G ¥[X] satisfies (1 1 .1 8), and m = deg(/) ^
3. Ifq > 4m2, then N = \C f(¥)\ satisfies

\N - q\ < 8rriy/q.
282 11. SUM S OVE R FINIT E FIELD S

Clearly w e ca n assum e tha t p > 2 , a s otherwis e th e ma p y \-> y 2 i s a n auto -

morphism o f F an d N — q.
Stepanov's idea , whic h wa s inspire d b y result s o f Thu e [Thu ] i n diophantin e
approximation, i s to construct a n auxiliary polynomial o f degree r, say , having zero s
of hig h multiplicit y (a t leas t £, say) a t th e x-coordinate s o f points o f Cf (F) . Henc e
one get s easil y th e inequalit y
N ^2r£~\
the facto r tw o being the highes t possibl e multiplicit y o f a given x-coordinate amon g
points i n C/(F) . Thi s inequalit y turn s ou t t o b e s o stron g tha t i t give s th e upper -
bound o f th e theore m (certainl y a surprisin g fact!) . A tric k the n deduce s th e
lower-bound fro m this .
We firs t distinguis h amon g th e point s (x , y) thes e wit h y = 0 . Le t iV o be th e
number o f distinct zero s of / i n F, which is also the number o f points (x , 0) G C/(F).
If (x , y) i s a poin t o f Cf wit h y =fc 0 , i t follow s tha t f(x) i s a squar e i n F , whic h i s
true i f an d onl y i f g(x) = 1 where
g = f c wit h c = |(«7-1 ) .
Conversely, give n x G F wit h g(x) = 1 , ther e ar e exactl y tw o element s y G F*
with y 2 — f(x). Hence , writin g
(11.19) 7V ! = | { x G F | g(x) = l}\
it follow s tha t
(11.20) N = No + 2Nx.
We will estimate N\ b y following the strategy sketche d above , but i n order to handl e
the lowe r boun d later , w e generalize slightl y an d conside r fo r an y a G F th e se t
(11.21) S a = {x G F | f(x) = 0 or g(x) = a}.
To produc e polynomial s vanishin g t o a larg e order , w e wis h t o us e derivative s
to characteriz e whe n thi s occurs . I n characteristi c 0 , a polynomia l P ha s a zer o of
order £ at x o i f an d onl y i f al l th e derivative s P^ wit h 0 < i < £ vanish a t XQ. I n
characteristic p > 0 , however , thi s i s n o longe r tru e i f £ > p, a s th e exampl e o f th e
polynomial P = X p shows , sinc e p( k) = 0 for al l k ^ 1 , in particular , P ^ ( 0 ) = 0 .
A satisfactor y solutio n follow s b y considerin g othe r differentia l operators .
DEFINITION. Le t K b e an y field . Fo r an y k ^ 0 , th e fc-th Hass e derivativ e i s
the linea r operato r E k : K[X] -> K[X] define d b y
n
Ekxn = f \Xn-k

1
for al l n ^ 0 , an d extende d t o K[X] b y linearity . W e als o writ e E = E (bu t
beware tha t E k ^ E o E o • • • o E).
REMARK. Fro m th e binomia l expansio n

Xn = (X - a + a)n - Y^ (l) aU~k(X - ^) fc5

and b y linearity , w e se e tha t th e valu e o f E kP a t a poin t a G K, fo r P G if[X] ,

is simpl y th e coefficien t o f (X — a) k i n th e Taylo r expansio n o f P aroun d a. Thi s
 11. SUM S O V E R F I N I T E F I E L D S 283

explains th e properties o f the Hasse derivatives , bu t we cannot tak e thi s a s a defi -

nition, becaus e th e values of a polynomial ove r a finite field do not characterize the
polynomial.

Note tha t fo r K o f characteristi c p > 0 , w e ge t EX P = E 2XP = • • • =

E ~ Xp = 0 , bu t E PXP = 1 ^ 0 an d w e see that th e Hass e derivative s detec t
p l

the zer o o f X p o f order exactl y p at 0 . Thi s i s a general fact , a s Lemma 1 1 .1 6 will

show.
LEMMA 1 1 .1 4 . The Hasse derivatives satisfy
k
k EJ Ek 3
E {fg) = Y,^ ^ ~ d)
j=o

for all f,g€ K[X], and more generally,

k
(11.22) E {h---U)= Y. {E^h)---{E^f r)

/ o r A , ...,f r£K[X].

I t suffice s t o consider / = X m, g = X n, an d the first formul a follow s

PROOF.
from th e identity
m n
n + m\ ^ ( \(
k =
) f^\j]\k-j
which i s obvious fro m th e combinatorial interpretatio n o f the binomial coefficients .
Then th e second formul a follow s b y induction. •

COROLLARY 1 1 .1 5 . (1 ) For all k, r ^ 0, and all a e K, we have

Ek(X-a)r = ( r
\(X-a)r-k.

(2) For all k, r ^ 0 with k ^ r, and all f,gE K[X], we have

k r k
E (f9l = hg ~
for some polynomial h such that
deg(h)^deg(f) + kdeg(g)-k.

PROOF. Fo r (1), we apply (1 1 .22 ) t o /i = • • • = f r = X - a , gettin g

Ek(X - a) r = J2 EJl
(X-a)-- E Jr
(X - a)
j i H \-j r=k

and onl y term s wit h al l ji G {0,1}, 1 < i ^ r , giv e non-zer o contribution s sinc e
Ei(X — a) = 0 for j > 2 , from th e definition. Henc e (1 ) follows .
For (2) , we observe tha t i f k ^ r , w e have ji = 0 fo r a t leas t r — k indice s
in (1 1 .22) , whic h give s (2) . •

1 1 .1 6 . Let f e K[X] and a e K. Suppose that (E kf)(a) = 0 for all

LEMMA
k < L Then f has a zero of order ^ £ at a, i.e., is divisible by (X — a)*.
28 4 11. SUM S OVE R FINIT E FIELD S

P R O O F . Le t

be th e Taylo r expansio n o f / aroun d a. B y (1 ) o f Corollar y 1 1 .1 5 , we obtai n

Ek
f= E ^ V - " )4
and evaluatin g a t a we get a ^ = 0 for al l k < £, hence / i s divisible b y (X — a) e a s
claimed. •

We need anothe r technica l lemma .

LEMMA 1 1 .1 7 . Let K = ¥ be a finite field of characteristic p with q elements,
and let r = h(X, X q) G ¥[X], where h G F[X , Y]. Then
Ehr = (E k
xh)(X,X
q
)
for all k < q, where on the right side E xh denotes the Hasse derivative of h
performed with respect to X.
I t suffice s t o conside r h = X nYm, s
PROOF. o we must prov e that E kXn+mq —
k n rnq
(E X )X . Pro m Lemm a 1 1 .1 4 , we ge t

3=0
so i t suffice s t o sho w tha t E jXmq = 0 for 0 < j < q to prov e th e lemma . Bu t
^mq\ mq (mq — 1 N
v •1 1
3• /J •J \ J - • °

in characteristi c p , an d th e resul t follows . •

We com e t o th e hear t o f Stepanov' s method , th e constructio n o f th e auxiliar y

polynomial.
PROPOSITION 1 1 .1 8 . Assume that q > 8m, and let £ be an integer satisfying
m < £ ^ q/8. Then there exists a polynomial r G F[X] of degree
deg(r) < c£ + 2m£(£ - 1 ) + mq
which has a zero of order at least £ at all points x G S a (recall c= \{q— 1 )^ .
We wil l look , b y th e metho d o f indeterminat e coefficients , fo r a polynomia l r
of th e specia l for m
q
(11.23) r = f Yl (rj+s j9)yJ

for som e polynomial s r^ , Sj G F[X], t o b e constructed , eac h o f whic h ha s degre e

bounded b y c - m. Henc e suc h a polynomia l r ha s degre e bounde d b y
(11.24) deg(r ) ^ £m + c - m + cm + Jq ^ (J + m)q.

The nex t lemm a i s crucia l t o ensur e tha t r ^ 0 . Thi s i s wher e w e nee d th e

assumption (1 1 .1 8) .
 11. SUM S OVE R FINIT E FIELD S 285

LEMMA 1 1 .1 9 . We have r = 0 e ¥[X] if and only if Tj = Sj = 0 £ ¥[X] for

all j.
P R O O F . W e ca n assum e (b y a shif t X \-> X + a i f necessary ) tha t /(0 ) ^ 0 .
Suppose tha t r = 0 but no t al l r^, Sj ar e zero ; let k b e th e smalles t inde x fo r whic h
one o f r*fc , Sk is non-zero. Dividin g b y f £Xkq, w e get fro m (1 1 .22 ) th e identit y

k^j<J

which w e write i n th e for m ho + h\g = 0 , wher e

s
h0 = ] T rjXU-V*, hi = J2 3xU~k)q-

We squar e thi s equation , the n multipl y bot h side s b y / , gettin g

hlf = h\f«.
Since / 6 F[X] , we hav e
f{Xy = f(X") = /(0 ) (mo d X« )
hence
rlf = s 2kf(0) ( m o d i ' ) .
However, th e degre e s o f th e polynomial s i n thi s congruenc e ar e bounde d b y
2deg(r/e) + m ^ 2( c — m ) + m < q, an d 2deg(s/c ) < 2( c — m ) < g , respectively .
So there mus t b e equalit y r\f — s\f(f$), whic h contradict s th e assumptio n (1 1 .1 8 )
that / i s no t a squar e i n ¥[X]. •

We no w evaluat e th e Hass e derivative s o f r .

LEMMA 1 1 .20 . Let k ^ £. Then there exist polynomials r K- J , s^ each one of
degree ^ c — m + k(m — 1 ) such that

Ekr = f~k J2 {rW + sfg)**.

0^3 < J

PROOF. W e ca n writ e r = h(X, X q) wher e h G F[X, Y] i s the polynomia l

Hence b y Lemm a 1 1 .1 7 , we hav e

Ekr = (E kxh)(X,Xq) = Y\ (E k
(fr3) + E k i+c
(f sj))Xjq.
o^j<J

By (2 ) of Corollary 1 1 .1 5 there exist polynomial s r^ an d s : ' satisfyin g E k(f£rj) =

fi-kr(k) an d E k(f+cSj) = f-k+c s(k) wit h deg ( r W ) < ; deg(rj) + kdeg(/ ) - if e <
c — m + fc(m — 1 ) an d deg(s ^ ^ ) ^ c — ra+fc(m — 1 ) . Thi s i s th e desire d result . •

Recall tha t w e wis h r t o hav e zero s o f orde r ^ £ at point s i n <S a (se e (1 1 .20)) .
If f(x) = 0 , clearly thi s i s the case . S o let x E <Sa, with f(x) 7 ^ 0. Applyin g Lemm a
286 . SUM S O V E R F I N I T E F I E L D S

11.21, w e evaluat e E kr a t a poin t x G <Sa, using g(x) = a, an d mos t importantl y

xq = x:
Ekr{x) = f{xY~ k ] T (rf\x)+asf\x))xi

l k
= f(x) - *W(x)
where a^ G ¥[X] i s th e polynomia l

0<j<J

We ca n no w prov e Propositio n 1 1 .1 8 : i f <T^ ) = o fo r al l k < £, Lemm a 1 1 .1 6

shows tha t r ha s a zer o o f orde r ^ £ at al l point s i n S a- Th e syste m o f equation s
{k)
(11.25) a = 0 , fo r al l k < £
is a homogeneous syste m o f linear equations , th e unknown s bein g the coefficient s o f
the polynomial s r*j , Sj, the equation s correspondin g t o th e coefficient s o f th e cr^) .
We observ e tha t
deg(a (/c) ) <c-m + k(m - 1 ) + J ,
so th e numbe r o f equation s doe s no t excee d B = £(c — m - f J ) + \£{£ — l)(r a — 1 )
while, o n th e othe r hand , th e numbe r o f coefficient s o f th e rj an d Sj i s a t leas t
A = 2( c — TO)J.B y choosin g J larg e enough , w e ca n mak e A > B. The n th e
system (1 1 .25 ) ha s a non-trivial solution , an d b y Lemm a 1 1 .1 9 this produce s r ^ O
such tha t r ha s zero s o f orde r > £ at al l point s x G <Sa. Takin g

J = - ( c + 2m(*-l) )

one ca n chec k tha t A > B (recal l tha t 2c — q — 1 and S£ ^ q). Th e degre e o f r i s

bounded b y (1 1 .24) , whic h give s Propositio n 1 1 .1 8 .
We no w prov e Stepanov' s Theore m 1 1 .1 3 . First , le t a b e arbitrary , an d appl y
Proposition 1 1 .1 8 . Sinc e th e auxiliar y polynomia l r i s non-zer o an d vanishe s t o
order ^ £ a t point s i n <S a, w e hav e ^|<S a| ^ deg(r ) ^ c£ + 2m£(£ — 1 ) -f - mq s o
|<Sa| ^ c + 2m( ^ — 1 ) 4- mqt~x. W e choos e £ = 1 + [^/g/2] , whic h give s th e boun d
(11.26) \S a\ < c + 4mVg .
To prov e Theore m 1 1 .1 3 , tak e firs t a = 1 getting

7V0 + A ri = |<S a|< f + 4 m V g

hence th e uppe r boun d
(11.27) N = N 0 + 2/V i < 2(iV 0 + iVi ) < q + S T O ^ .
To ge t a lowe r bound , b y th e factorizatio n X q — X = X(X C — l)(X c + 1 ) w e hav e
f(x)(g(x) - l)( 5 (z) + 1 ) = 0
for al l x € F , henc e N 0 + Ni + N 2 = q wher e iV 2 = |{ x G F | g(x) = - 1 } | .
By (1 1 .26 ) applie d t o <S_i , we hav e

iV0 + iV 2 = |<S_i | < | + 4 m v ^

 11. SUM S O V E R F I N I T E F I E L D S 287

hence
Nx = q - N 0 - A/ i > | - 4mv^ ,
and finally ,

(11.28) N = N 0 + 2N 1 ^2N1 >q- Sm^/q.

Clearly (1 1 .27 ) an d (1 1 .28 ) prov e Theore m 1 1 .1 3 .

11.7. Proo f o f Weil' s boun d fo r Kloosterma n sums .
Let F b e a finit e field wit h q elements , o f characteristi c p ^ 2 . Le t ip be an y
fixed non-trivia l additiv e characte r o f F . Fo r an y additiv e characte r <p ther e exist s
a uniqu e a G F suc h tha t (p = ip a, henc e an y Kloosterma n su m S(ip,tp) i s o f th e
form

x£F*
for som e a , b G F. W e conside r a an d b as fixed an d writ e g = aX + bX -1 . W e
will prov e Weil' s boun d (1 1 .1 5 ) b y relatin g th e averag e o f th e Kloosterma n sum s
S(i/ja,^b) over ip t o th e numbe r o f point s o n a n hyperellipti c curve , wher e th e
contribution o f th e trivia l characte r ipo — 1 will b e th e mai n term .
LEMMA 1 1 .21 . For any n ^ 1 and any x G F n ; we have

(11.29) | { * G F n | y * - 2/ = * } | = ] [ > ( ! * ( * ) )

where the sum ranges over all additive characters of¥ and Tr is the trace F n— > F .
PROOF. I f T r (x) = 0 , the n th e equatio n y q — y = x ha s q solutions exactly , a s
recalled i n Sectio n 1 1 .2 , and i n thi s cas e w e have ^(T r (x)) = 1 for al l 0 , henc e th e
right sid e of (1 1 .29 ) i s also equal to q. O n the other hand , i f Tr (x) ^ 0 , the equatio n
yQ — y — x ha s n o solution , an d th e characte r su m i s zer o b y orthogonality . •

Prom thi s lemm a w e deduce tha t

ip ip xGF *

(11.30) =\{(x,y)eF* nx¥n \ y* - y = g(x)}\ = N n, say ,

for n ^ 1 . I f ip = ^o 5 th e trivia l character , w e hav e

For i/; ^ yyo , let a^ , fy b e th e "roots " o f th e Kloosterma n su m S ( ^ a , ^ ), s o b y

Theorem 1 1 . 8 we hav e a^f3^ — q an d

Sn(^a,^)=Q^+/^,
for al l n ^ 1 .
We ca n therefor e writ e
288 11. SUM S OVE R FINIT E FIELD S

The equatio n y q — y = g(x) doe s no t obviousl y describ e a curve , sinc e g i s no t a

polynomial, bu t multiplyin g b y x i t i s equivalent wit h
Ca}b : ax 2 -(y q-y)x + b=0
(note that x = 0 is not possibl e since b ^ 0) . Becaus e p ^ 2, th e number o f solutions
is equa l t o th e numbe r o f solution s o f th e discriminan t equatio n o f thi s quadrati c
equation
Da,b : (y q-y)2-4ab = v 2,
i.e., N n = \D a^{¥n)\. Thi s i s o f th e for m (1 1 .1 7 ) wit h deg(/ ) = 2q, an d becaus e
Aab ^ 0 it satisfie s (1 1 .1 8) . Henc e b y Theore m 1 1 .1 3 we hav e
1 +n/2
\Nn-qn\ 1
< 6g
if n i s larg e enough , s o that q n > 1 6q.
By (1 1 .30 ) w e ge t a shar p estimat e fo r th e root s ct^ , /J^ , o n averag e
n/2
(ii.3i) -\J2 q
K + 05) <16g"
</^V> o
for n larg e enough . Th e followin g simpl e lemm a show s tha t th e individua l root s
must b e o f modulu s ^ y/q:
LEMMA 1 1 .22 . Letuj\, ..., u r be complex numbers, A, B positive real numbers
and assume that
r
|Vw"| < AB n
holds for all integers n large enough. Then \UOJ\ ^ B for all j.
P R O O F . On e ca n d o thi s b y han d (usin g Dirichlet' s bo x principle) , bu t a nic e
trick give s th e resul t immediately : conside r th e comple x powe r serie s

/w-E (!>;)«"-Ei^; -
n^l j 3
The hypothesi s implie s tha t / converge s absolutel y i n th e dis c \z\ < B~ l, henc e /
is analyti c i n thi s region . I n particular , i t ha s n o pole s there , whic h mean s tha t w e
must hav e \WJ\~ 1 ^ B~ x fo r al l j . •

From thi s lemm a applie d wit h A — 16g, B = y/q, w e deduc e th e uppe r bound s
\ai>\ ^ y/Q^ I A/; | ^ \/q fo r al l ip = £ ipQ. Since a^/3^ — g, w e hav e i n fac t \a^\ =
\Pi/>\ — \/# > anc ^ s o Theore m 1 1 .1 1 is proved .
REMARKS. (1 ) We see here twice how crucial the introduction o f the companio n
sums K n is : first becaus e th e curv e D a^ ha s very hig h degree , s o Stepanov's boun d
\N — q\ < Srriy/q is trivial when applie d t o F itself, an d secondl y because onl y by th e
consideration o f all extension fields ca n w e determine th e exac t orde r o f magnitud e
of th e roots , an d obtai n Weil' s boun d S(ip,(p) ^ 2y/q wit h th e shar p constan t 2 .
(2) Th e constan t 2 is optima l i n Weil' s boun d fo r fixed a , b and q. Indee d w e
have

1
• a^z -fyz'
n^l
 11. SUM S O V E R F I N I T E F I E L D S 289

This i s a non-zer o rationa l functio n wit h pole s o n th e circl e \z\ = 1 /y/q, henc e thi s
is it s radiu s o f convergence . Therefor e

limsup|Snty0)V6)r1/n = ^ = .

This mean s tha t fo r an y e > 0 there exis t infinitel y man y n suc h tha t

\Sn(il>aM>(2-£)qn/2.

It is conjectured (thi s follows from th e Sato-Tate Conjectur e fo r the angles of Kloost-

erman sum s describe d i n Chapte r 21 ) tha t th e Wei l boun d i s als o optima l whe n
a, b are fixed, n = 1 , an d q = p - » +oo . However , thi s remain s ver y muc h open .
See the remar k a t th e en d o f the introductio n t o Sectio n 1 1 . 8 for th e cas e o f ellipti c
curves.
(3) Usin g th e extensio n o f Stepanov' s metho d t o curve s o f th e typ e y d = f(x)
and a n analysi s o f th e correspondin g zet a function , on e ca n prov e th e followin g
estimate fo r complet e characte r sums :
THEOREM 1 1 .23 . Let ¥ be a finite field with q elements and let \ be a non-
trivial multiplicative character of F* of order d > 1 . Suppose f G ¥[X] has m
distinct roots and f is not a d-th power. Then for n ^ 1 we have

\^2x(N(f(x)))\^(m-l)qn/2.
x£Fn

This i s Theore m 2 C , p . 43 , o f [Sch] . I n particular , w e ge t th e followin g

corollary whic h wil l be use d i n provin g th e Burges s boun d fo r shor t characte r sum s
(Theorem 1 2.6) .
COROLLARY 1 1 .24 . Let x (mo d p) be a non-principal multiplicative character.
If one of the classes b v (mo d p), v = 1 , . . . , 2r is different from the remaining ones
then

| ] Tx ((x + 6i ) • • • (x + b r))x((x + 6 r + 1 ) . . . (x + bar) )^ 2rp 2

x ( m o d p)

PROOF. Observ e tha t

X((x + &!).. . (x + b r))x{(x + 6 r + 1 ) . . . (x + b 2r) = x(f(x))

with
Hx)= n (x + bj) Yl (x + b.y- 2.
Prom th e assumption , on e o f th e bi is a roo t o f / o f orde r eithe r 1 or p — 2 , whic h
is coprim e wit h th e orde r d \ (p — 1 ) o f x, s o w e c a n a PPly Theore m 1 1 .23 . •
290 11. SUM S OVE R FINIT E FIELD S

11.8. Th e Rieman n Hypothesi s fo r ellipti c curve s ove r finite fields.

A particularl y importan t case o f th e Rieman n Hypothesi s i s tha t o f ellipti c
curves. Historicall y thi s wa s first establishe d b y Hass e usin g globa l methods . I n
the notatio n o f Section 1 1 .6 , this mean s that w e consider curve s Cf wit h degf = 3 ,
so the equatio n i s of th e for m

(11.32) C : y 2 = x 3 + a 2x2 + a 4x + a 6
(the numbering reflect s th e traditional notatio n fo r elliptic curves, see Section 1 4.4) .
In contrast wit h that section , w e emphasize that w e are considering the affin e curve ,
without th e poin t a t infinity . Th e cubi c polynomia l f(x) — x3 + CL2X 2 + a 4x + a$
cannot b e a square , s o thi s curv e satisfie s th e assumptio n (1 1 .1 8) . Moreover , w e
assume tha t / doe s not hav e a doubl e root ; thi s mean s tha t th e curv e C i s smoot h
(see Sectio n 1 1 .9) , an d i t i s a necessar y conditio n fo r wha t follows .
In thi s case , Theore m 1 1 .1 3 implie s tha t fo r q > 3 6 th e numbe r N = |C(F) |
satisfies
\N-q\< 2 4 ^ .
In Sectio n 1 1 .1 0 w e wil l prove , a s befor e fo r Kloosterma n sums , th e rationalit y
and th e functiona l equatio n o f the correspondin g zet a function , fro m whic h w e will
deduce:
THEOREM 1 1 .25 . Let C be an elliptic curve overF given by

C : y 2 = x 3 + a<ix 2 + a 4x + a§

with cti E F. Then for all n ^ 1 we have

(11.33) | | C ( F n )|-gn|^2^/2.

REMARKS. Theore m 1 1 .2 5 i s optimal. Indee d lettin g n — > +00 , thi s follow s a s

for Kloosterma n sum s fro m Lemm a 1 1 .22 . However , i t i s also true i n the horizonta l
sense a s th e followin g exampl e shows : le t E/Q b e th e ellipti c curv e wit h equatio n

E : y 2 = x3 - x

which ha s comple x multiplicatio n b y Z[i}. A s befor e w e conside r th e affin e points ,

not th e projectiv e ones . Th e discriminan t o f E i s 6 4 s o E ca n b e reduce d modul o
p to a n ellipti c curv e ove r Z/p Z fo r an y od d prim e p. On e show s (fo r instanc e b y
relating E t o th e curv e y 2 — x4 + 4 b y changin g (x,y ) i- > (yx~ 1 ,2x — y 2x~2) fo r
(x,y) ^ (0,0) , se e e.g . [1 4 ] o r [IR] ) tha t \E(Z/pZ)\ = p i f p = 3 (mo d 4 ) an d
\E(Z/pZ)\ =p-2a pifp = l (mo d 4) , wher e
2 2
n= n + h

with ix = a p-\ribp = 1 (mo d 2( 1 +i)) (thi s congruenc e determine s n u p t o conjuga -

tion). Fo r an y e > 0, Theorem 5.3 6 (generalize d slightl y t o ad d th e congruenc e con -
dition) show s that ther e exist infinitel y man y Gaussia n prime s TT = 1 (mo d 2( 1 + 2))
such tha t |arg7r | < e. Henc e |Im(7r) | ^ £|7r | an d
2
\p-\E{Z/pZ)\\ = <2\a v\>2{l-e )^
for infinitel y man y p.
 11. SUM S O V E R F I N I T E F I E L D S 291

Theorem 1 1 .2 5 will b e prove d i n Sectio n 1 1 .1 0 afte r som e geometri c an d alge -

braic preliminaries . Thi s goe s a bit furthe r awa y from th e hear t o f analytic numbe r
theory, ye t w e includ e ful l detail s becaus e th e Hass e boun d i s als o importan t a s
being the simples t cas e of the very important Delign e bound fo r Fourie r coefficient s
of modular forms . Th e reader will also certainly appreciat e th e elegance and beaut y
of th e geometr y involved .

11.9. Geometr y o f ellipti c curves .

In explainin g th e specia l geometri c feature s o f ellipti c curves , w e ma y a s wel l
consider a mor e genera l case . S o let k b e a n arbitrar y field , fc an algebrai c closure ,
and le t C b e th e curv e give n b y th e equatio n

C : y 2 = f{x), wit h / = X 3 + a 2X2 + a 4X + a 6 e k[X],

identified wit h th e se t o f solution s (x,y) G fc 2. W e assum e a s befor e tha t / doe s

not hav e a double root . I f k' jk i s any extension , w e let C(k f) b e the se t o f solution s
in (A/) 2.
The geometr y o f th e ellipti c curv e become s muc h cleare r i f w e wor k wit h th e
projective versio n o f the curv e C, namel y th e curv e E i n the projective plan e given ,
in homogeneou s coordinate s (x : y : z), b y th e equatio n

E : y z — x + a 2x z + a^xz + a§z .

Putting z — 1 gives bac k C ; o n th e othe r hand , "a t infinity" , w e ar e onl y addin g

one point : takin g z — 0 yield s x — 0, an d al l element s ( 0 : y : 0 ) (wit h y ^ 0 )
correspond t o a singl e poin t o o = ( 0 : 1 : 0) i n th e projectiv e plane . Notic e tha t
this point o o is rational ove r the bas e field fc, so that fo r al l extensions fc'jkw e have

£(fc') = C(fc')U{oo} .

The mai n propert y o f th e curv e E tha t w e will us e i s th e beautifu l fac t tha t it s

points for m a n abelia n group , wit h identit y elemen t oo. Throughout , p denote s
points o n E, no t th e characteristi c o f the field k. Th e grou p la w (denote d b y + ) i s
described b y th e geometri c conditio n tha t fo r an y thre e (distinct ) point s p\, p 2 an d
p 3 i n E, w e hav e p\ -j - p2 + Pz = 0 i f an d onl y i f th e thre e point s ar e collinea r (i n
the projectiv e plane) , an d th e opposit e o f a point (x : y : z) i s the poin t (x : — y : z)
(symmetry wit h respec t t o th e x-axis) . Thi s wa y on e ca n construc t th e su m o f an y
two distinct points , b y computing th e equatio n o f the lin e joining them , an d takin g
the opposit e (i n th e sens e above ) o f th e thir d intersectio n poin t wit h th e curve .
That ther e ar e exactl y thre e intersectio n point s follow s immediatel y fro m th e fac t
that th e polynomia l / i s o f degre e 3 . I n addition , t o comput e th e doubl e p + p o f
a poin t p , th e sam e constructio n i s don e wit h th e tangen t lin e a t p ; th e conditio n
that / ha s n o doubl e roo t ensure s tha t thi s tangen t lin e alway s exists .
Also, becaus e / G k[X], i t follow s easil y tha t th e ^'-rationa l point s E(k f), fo r
any extensio n k' /fc, for m a subgrou p o f E(k).
We d o no t prov e thos e fact s here ; completel y elementar y proofs , b y computin g
explicitly th e coordinate s o f th e su m p\ + P 2 o f tw o point s accordin g t o th e recip e
above an d checkin g th e abelia n grou p axiom s (associativit y i s the onl y difficulty) ,
are fairl y straightforwar d (se e fo r instanc e [IR] , ch. 1 8 , 1 9) .
292 11. SUM S OVE R FINIT E FIELD S

We no w introduc e som e furthe r geometri c object s relate d t o E or , mor e gen -

erally, t o an y smooth , projective , algebrai c curve 1 . Thu s conside r agai n a mor e
general case : le t k b e a n algebraicall y close d field , an d le t E b e a plan e algebrai c
curve ove r A; , i.e. give n b y a n equatio n
f(x,y,z) = 0
for som e homogeneou s / G k[X, Y , Z\. W e identif y E wit h th e se t o f point s i n th e
projective plane . W e assume that E i s smooth, whic h means here that fo r an y poin t
p = (x : y : z) o f E, no t al l partia l derivative s df/dX(p), df/dY(p), df/dZ(p),
are zero . I n thi s cas e th e lin e wit h equatio n

?L(p)(X-x) + ZL(p)(Y-y) + ^{p)(Z-z)=0

is well-defined an d i s the tangen t lin e t o E a t p. Fo r ellipti c curve s y 2 = / ( # ) , thi s
smoothness conditio n i s equivalent t o th e fac t tha t th e polynomia l / ha s n o doubl e
roots, b y a simpl e calculation .
Let C b e th e affin e curv e correspondin g t o E give n b y
C : f(x,y,l) = 0
in k\ Le t g(X, Y) = f(X, Y, 1 ) G k[X, Y]. W e define k[C] = k[X, Y]/(g). Element s
of k[C] can b e interprete d a s function s o n C. W e assum e tha t (g) i s a prim e idea l
(this i s easil y checke d i n th e cas e o f ellipti c curves ) s o tha t k[C] i s a n integra l
domain, an d w e le t k(C) o r k(E) b e it s quotien t field , calle d th e functio n field o f
C o r o f E. I t i s a finit e extensio n o f th e fiel d k(X) o f rationa l function s ove r k
(for ellipti c curve s y 2 = / ( # ) , i t i s a quadrati c extensio n k(X)(y/J)). W e interpre t
elements o f th e functio n fiel d a s rationa l function s o n E 1 , s o give n a poin t p G E
and a n elemen t p> G k(E), eithe r ip has a pol e a t p o r (p(p) G k i s defined .
Now th e importan t poin t i s tha t becaus e E i s smoot h i t i s possibl e t o defin e
the orde r o f ip at p for ever y p in E an d ever y non-zero rationa l functio n p G k(E)*.
As expected , thi s orde r behave s lik e it s analogu e fo r holomorphi c function s theor y
or rationa l functions . Precisely , fo r ever y p G R, ther e i s a discret e valuatio n
ord p : k(E) x - > Z ,
which give s th e orde r o f th e zer o (i f ^ 0 ) o r pol e (i f < 0 ) o f a rationa l functio n a t
p. A s a discret e valuation , i t satisfie s
ordp(c) = 0 , fo r c £k*,
ordp(<pV0 = ordp M + ord p(ip),
ordp((p + *p) ^ min(ord p ((/?),ordp(^)).
We sketc h a proo f (se e als o [Sil] , Prop . II . 1.1): conside r th e rin g O p = {p G
k(E) \ p i s define d a t p}. Thi s i s a noetheria n loca l domai n wit h maxima l idea l
vtip = {p E O p | p(p) = 0 } an d residu e fiel d O p/mp ~ k (b y evaluatio n a t p).
Because E i s smoot h a t p (ther e i s a tangen t line) , th e /c-vecto r spac e trip/m ^ i s of
dimension 1 (because th e curv e i s in th e plane , i t i s of dimension ^ 2 ; the equatio n
of th e tangen t lin e give s a relation , an d i t i s eas y t o se e tha t trip/m ^ ^ 0) . B y
Nakayama's Lemm a (se e for instanc e [AM] , p. 21 ) , it follow s tha t m p i s a principa l
x
The reade r ca n withou t damag e assum e tha t w e ar e jus t dealin g wit h th e ellipti c curve s
described before , wit h k = ¥. I n thi s cas e ever y incomplet e assertio n ca n b e checke d b y hand .
 11. SUM S O V E R F I N I T E F I E L D S 29 3

ideal. Le t IT be a generator ; the n m ^ is generated b y 7rd for any d ^ 1 . Th e order

ordp ca n be defined fo r p i n Op, ip ^ 0 , b y
ordp((p) = max{d ^ 0 | (p G m^ }
and extende d t o a homomorphism k(E)* — > Z. Th e propertie s abov e ar e then quit e
easy t o check.
For ellipti c curve s y 2 = /(#) , on e can easily se e that i f p =fi oc , and p = (x,y)
with y 7^ 0, it is possible t o take n = X — x. Fo r p = oo , one can take 7 r = X/Y,
and on e finds tha t ordoo(:r ) = —2 , ordoo(y) = —3 .
Every non-zer o elemen t p G k(E) ha s finitely man y zero s an d poles, an d to
package the m convenientl y w e define a diviso r o n E t o b e a forma l finit e linea r
combination wit h coefficient s i n Z of symbols [p] , one for each poin t p G E. Divi -
sors for m a fre e abelia n grou p Div(E) . Tw o homomorphism s ar e important. On e
associates t o a non-zer o rationa l functio n p> th e divisor (denote d (cp)) of its zeros
and poles :
( k(E) x ->Div(E )
] ¥>•->(¥> ) = ] T o r d p ( y > )
I p£E
and th e second give s the degree o f a divisor :
,/ Div(£ ) - + Z,
I \p] »- > 1 .
As suggeste d b y the notation, divisor s o f the typ e (ip) are called principa l divisors .
Ordinary rationa l function s hav e a s man y zero s a s poles , wit h multiplicity ,
and th e same hold s fo r (p G k(E)*: thi s mean s tha t fo r all p> G k(E)x, w e have
deg((y>)) = 0 (se e for instanc e [Sil] , II-3) . Fo r elliptic curve s y 2 = / ( # ) , an eas y
proof ca n b e derived b y observing tha t k(E) i s a quadratic extensio n o f k(X). Th e
non-trivial elemen t i n the Galois grou p i s p i- > (p defined b y (p(p) = <p(—p). I t is
clear tha t ord p((p) = ord_ p ((^), henc e deg(cp) = deg(</?) . No w cptp i s in k(X). On e
can chec k th e following compatibility : i f ip G k(X), wit h diviso r (ip)i = Yl nixi
(as a n ordinary rationa l function) , the n it s divisor a s an element o f k(E) i s (V>) =
^2^i(\pi] + [ _Pi])> where pf is any point o f E wit h ^-coordinat e Xi. In particular ,
0 = deg(('0)i ) = 2deg((^)) . Applyin g thi s t o p(p gives
0 = deg((p<p) = deg(<p ) + deg((p) = 2deg(p).

D E F I N I T I O N . 1 . Tw o divisor s D i an d D2 suc h tha t D\ — D2 = (p) fo r som e

(/? G k(E)x ar e called linearl y equivalent . Thi s i s denote d D\ ~ D 2l an d is an
equivalence relatio n o n Div(f7).
2. Th e group Div(Z? ) carrie s a partia l ordering , compatibl e wit h th e grou p
structure, define d b y D ^ 0 if and only i f al l the coefficients i n the formal sum
giving D ar e > 0. Suc h divisor s ar e called effectiv e divisors .
3. Fo r a divisor D G Div(£7), let
L(D) = {0 } U {p G k(E) | foO+ D ^ O } ;
this i s a ^-vecto r space . Le t ^(-D) = dimL(D) , a n integer o r +oo .
If D = ni\pi} + ...+nk[pk}-mi[qi]-...-mj[qj] withn^ , mi ^ 0 , then p G L(D),
</? 7^ 0, means simpl y tha t p has
29 4 11. SUM S OVE R FINIT E FIELD S

(1) Pole s of order a t most n ^ at pi, 1 ^ i ^ k;

(2) Zero s of order a t least rrti at g^, 1 ^ i ^ j.
It follow s immediately tha t i f D2 ^ £>i , then L(D 2 ) C L{D\). Also , if Di ~ D2 ,
then writin g D\ = D2 + (VO? the map<£> H > ^><£ induces a n isomorphism L{D\) — >
L(£) 2 ), i n particular, £(Di ) = ^(^2 ) only depend s o n the linear equivalenc e clas s
of th e divisor .
The followin g interpretatio n i s also clear : ther e i s a bijectio n
( P(L(D)) —> {Effectiv e divisor s linearl y equivalen t t o D},

between th e projective spac e P(L(D)) o f L(D) an d the effective divisor s linearl y

equivalent t o D (b y definition o f L(D), th e map has image i n the set of effectiv e
divisors). Thi s als o require s th e important fac t tha t
(11.35) L(0 ) - k, £(0) = 1.
In othe r words , a n everywhere define d rationa l functio n o n E i s constant: thi s is
obvious for elliptic curves , sinc e regularity o n C forces suc h a <p to be a polynomia l
g{X) + Yh(X), an d regularity a t 00 then force s h = 0 , g G k.
We first notic e th e following simpl e lemma :
LEMMA 1 1 .26 . Let D be a divisor on E such that £{D) > 0 . Then either
deg(D) > 0 or D ~ 0.
PROOF. lt£{D) > 0, there is a non-zero element tp G L(D), so that (ip)+D ^ 0.
Taking th e degree w e find tha t deg(D) ^ 0 . S o we need onl y sho w no w that i f
deg(D) = 0 , the n D ~ 0 . Bu t ((p) + D i s then a n effectiv e diviso r o f degre e 0;
clearly i t must b e = 0 , so D = — (p) = {p~ x) ~ 0 . D

To prov e th e rationality o f the loca l zet a functio n o f ellipti c curves , w e will

need t o kno w th e valu e o f £{D) fo r effectiv e divisors . Thi s i s compute d b y the
Riemann-Roch Theorem .
T H E O R E M 1 1 .27 . Let E be an elliptic curve over an algebraically closed field
k. For any divisor D on E, £{D) is finite and we have the formula
(11.36) £(D)-£(-D)=degD.
Equivalently, by Lemma 1 1 .26, we can compute £(D) for any D by:
1. Ifdeg(D) ^ 0 , and D / 0, then £(D) = deg(D) .
2. IfD~0, then£(D) = 1.
3. Ifdeg{D) < 0, then £(D) = 0.
This i s the Riemann-Roch theore m specialize d fo r elliptic curves . Se e the re -
mark below fo r the general case .
The simpl e proo f fo r th e Riemann-Roc h theore m hinge s o n th e remarkabl e
interaction o f the group structur e o n E wit h th e divisor group . Indeed , conside r
the map
a : Div(E ) - • E
defined b y <j(ni[pi ] + .. . + rik[pk}) — n\Pi + •• • + ^kPk, the + o n the right sid e
corresponding t o the group la w on E.
 11. SUM S O V E R F I N I T E F I E L D S 295

PROPOSITION 1 1 .28 . Let D be a divisor on E. Then we have

D~[a(D)] + (te%{D)-\)[oo].

P R O O F . I n essenc e thi s "is " th e grou p la w itself : b y induction , w e nee d onl y

consider D = [p] + [q] an d D — [p] — [q] fo r som e point s p an d q. I f p o r q i s th e
origin oo , the resul t i s obvious .
So conside r first th e cas e o f D = [p] + [q] wit h p , q ^ o o an d p ^ q. The n
the equatio n a X + 6 Y + c = 0 o f th e lin e joinin g p an d q define s a n elemen t
ip = aX + bY + c G &(£), whic h b y definitio n o f th e grou p la w satisfie s
(y>) - [p ] + [<? ] + [r ] - 3[oo ]
where —r — p + q = a(D). Sinc e (y? ) ~ 0 we ge t
D = [p ] + fo] ~ (y, ) - [ a(D)] + 3[oc ] - -[(7(D) ] + 3[oo] .
But similarl y fo r an y poin t p i n E, th e equatio n o f th e lin e joining p an d — p give s
a functio n wit h diviso r [p] + [—p ] — 2[oc] s o tha t fo r an y p
(11.37) -\p] + [oo ] - [-p ] - [oo ]
and henc e D ~ [c(^) ] + [oo] , as desired .
If p = g , th e equatio n o f th e tangen t lin e give s 2[p] - f [2p ] — 2[oo ] ~ 0 , an d th e
result agai n follows . Finall y i f D = [p ] — [g] , use (1 1 .37 ) t o reduc e t o th e previou s
case:

\P] ~ [q] = [P\ + {-[q] + [oo] ) - [oo ] - [p ] + [-q] - 2[oo ] - [ p + q] - [oo] .

D

P R O O F O F TH E RIEMANN-ROC H THEOREM . Notic e tha t (1 1 .36 ) fo r D o r -D

are equivalent . B y Propositio n 1 1 .28 , we hav e
e(D)=e{[v(D)] + (degD-l)[oo]),
£(-D) = £([-a(D)} + ( 1 - degZ?)[oo]) .
One of the two divisors on the right i s effective, s o we can assum e tha t D i s effectiv e
and o f th e for m D = [p ] -f n[oo ] wit h P G E an d n > 0 .
If p = oo , then D = ra[oo] wit h m > 1 . W e must prov e £(D) = m . An y elemen t
p of L(m[oo]) ha s n o poles on C , henc e i s a polynomial (p G fc[X,Y"], which satisfie s
3rdoo((^) ^ —m . Sinc e ordoo(X ) = — 2 and ordoo(F ) = —3 , w e hav e
ordoo(^) = max(-2deg(#) , - 3 - 2deg(ft) ) fo r ^ = g(X) + Yh(X).
Let V = {2deg(#), 3 + 2deg(h)} wher e g an d h ar e polynomial s i n X. On e
sees immediatel y tha t V i s th e se t o f al l positiv e integers , excep t 1 . No w usin g
monomials X a o r YX b a s basi s elements , w e chec k tha t
£(m[oo]) = \{n G V \ n ^ m} | = m .

Now we consider D = [p ] +n[oo] with n ^ 0 and p ^ oo . I f n = 0 , we can use the

automorphism g i-> g—p which sends p to o o to get a n isomorphism L([p\) ~ L([oo] )
which implie s £{[p}) — 1 .
29 6 11. SUM S OVE R FINIT E FIELD S

If n ^ 1 , we have D ^ n[oo ] hence L(n[oo]) C L(D). Thu s n < ^(£>) . Moreover ,

because onl y a simple pol e i s allowed a t p , we have £{D) ^ n + 1: i f 7r p is a functio n
with a simpl e zer o a t p , w e have a fc-linear ma p
L(D)/L(n[oo}) - > fc
ip*-± (7T p(f)(p)

which i s tautologicall y injective . Thu s w e mus t sho w tha t ther e i s i n L{D) on e

more fc-linear ly independen t elemen t no t i n L(n[oo]) .
Write p = (x,y ) i n affin e coordinates . W e hav e th e followin g divisors :
( X - * ) = [p ] + [-p]-2[oc] ,
(Y + y) = [-p] + \p'] + \p"]-3[oo]
for some p' and p". Le t< p = ( r + y ) / ( X - x ) , the n ((£ ) - -[p ] + [p'] + [p"]-[°°] ^ - # •
We claim tha t p ^ p' , p" . Indee d p ' an d p" , b y definition , ar e o f the for m (#i , — y),
(^2, —y). Thi s can be equal to p only if y — 0 , but fo r y = 0 by assumption ther e ar e
three distinc t root s o f f(x) = 0 . Henc e <p ha s a simpl e pol e a t p , s o (p $• L(n[oo]) ,
and w e obtain th e require d formul a £{D) — n + 1 = deg(D) . •

We mus t no w conside r som e rationalit y questions . W e assum e tha t w e hav e

an ellipti c curv e E give n b y y 2 — f(x) wit h / G k[X, Y]. Th e precedin g analysi s
applies t o a n algebrai c closur e fc of fc. There i s a natural actio n o f the Galoi s grou p
Gk o f k o n E (o n th e coordinates) , an d o n th e divisors . A poin t o r a diviso r i s
called fc-rational i f it i s G^-fixed. Notic e that fo r a divisor D — n\p\ +.. . -\-rtjPj thi s
does no t mea n tha t th e pi ar e i n E(k). Als o Gk act s o n k(E) an d th e fiel d fixe d b y
Gk i s k(E), th e fractio n fiel d o f k[X, Y]/(f). Th e diviso r o f a functio n (p G k(E) i s
obviously fc-rational.
If D i s define d ove r fc, we defin e
Lk(D) = {0 } U {ip G fc(£) | (y> ) + £ > > 0 }
and w e le t £ k{D) = dim fc L fe (D). I t i s clea r tha t £ k(D) ^ £(D) .
THEOREM 1 1 .29 . For any k-rational divisor D, we have
ek(D) = e(D).
In particular the Riemann-Roch formula holds with £k(D) instead of £(D).
P R O O F . Thi s i s a specia l cas e o f th e followin g theorem , whic h i s a formulatio n
of Hilbert' s Theore m 9 0 fo r GL(n): le t fc be a field , fc an algebrai c closure , V a fc-
vector space with an action of Gk • Then there is a basis of V mad e of elements which
are G^-fixe d (equivalently , le t Vk — VGk, the n V = Vk <S> fc,or dim/ - Vk = dim ^ V).
For a proof , se e fo r instanc e [Sil] , Lemm a II.5.8.1 . •

REMARK. Thi s theor y adapt s i n th e followin g wa y t o mor e genera l algebrai c

curves (se e fo r instanc e [Ha] , IV) : i f E i s smoot h an d projectiv e ove r fc, on e ca n
define a certai n diviso r clas s K (calle d th e canonica l class , an d relate d t o differen -
tials o n E). I t ha s degre e deg(K) = 2g — 2 for som e intege r g ^ 0 , calle d th e genu s
of E, an d th e Riemann-Roc h Theore m take s th e for m
(11.38) £(D) - £{K -D)= deg(D) + 1 - g.
 11. SUM S OVE R FINIT E FIELD S 297

Elliptic curve s correspon d t o g = 1; i n thi s cas e th e canonica l clas s i s trivial , an d

this reduce s t o Theore m 1 1 .27 . Th e cas e g — 0 corresponds t o th e projectiv e line ,
and i s als o ver y easy . Th e proo f o f (1 1 .38 ) i s muc h mor e involve d tha n th e on e fo r
elliptic curves , sinc e ther e i s n o grou p la w o n th e curv e whic h woul d help .

11.10. Th e loca l zet a functio n o f ellipti c curves .

Let C b e a n ellipti c curv e ove r F give n b y (1 1 .32) . I t i s more convenien t t o us e
here th e correspondin g projectiv e curv e E, a s describe d i n Sectio n 1 1 .9 . Th e zet a
function o f E i s define d a s th e forma l powe r serie s

(11.39) Z (E)=exp(^2l-^^Tn).
n^l

We firs t relat e Z(E) t o point s o n th e curve , b y givin g it s Eule r produc t (com -

pare Lemm a 1 1 .7) . T o d o thi s w e introduc e som e terminology , whic h come s fro m
the languag e o f schemes .
DEFINITION. Le t E be a n ellipti c curv e ove r a finite field F wit h q elements .
A close d poin t o f E i s the Galoi s orbi t o f a poin t x$ G E(¥). Th e degre e deg(x ) o f
a close d poin t x is th e cardinalit y (necessaril y finite) o f th e orbi t an d it s nor m i s
Nx = q de^x\ Th e se t o f close d point s o f E i s denote d \E\.
This notio n i s analogu e t o tha t o f a n irreducibl e polynomia l i n ¥[X] use d fo r
the zet a function s o f Gaus s sum s an d Kloosterma n sums . T o ever y close d poin t
x G \E\ i s associated a n F-rationa l diviso r whic h i s simply th e forma l su m o f all th e
elements i n th e orbit . Th e degre e o f thi s diviso r i s the degre e o f x. Moreover , i t i s
easy t o se e that th e grou p o f F-rational divisor s i s the fre e abelia n grou p generate d
by th e divisor s associate d t o close d points .
LEMMA 1 1 .30 . We have the Euler product expansion
d e g
(11.40) Z(E)= Yl ( l - T ^)-\
xe\E\
where the product is over all closed points of E.
P R O O F . Thi s i s very clos e to Lemma s 1 1 . 7 and 1 1 .9 . Firs t b y decomposin g th e
points i n E(¥ n) i n Galoi s orbit s w e obtai n

\E(¥n)\ = J2d E !'

d\n xe\E\
deg(cc)=d

which i s the analogu e o f Lemm a 1 1 .1 . The n w e hav e

^' n^ l

and, o n th e othe r hand , thi s operato r applie d t o th e righ t sid e o f (1 1 .40 ) yield s

^ d e g ^ T - ^ ^ r ^ d£ l )
x£\E\ n ^ l n^ l d\n x€\E\
deg{x)=d

hence th e result . •
298 11. SUM S OVE R FINIT E FIELD S

Using the Riemann-Roch Theorem , w e now prove the rationality an d functiona l

equation o f th e zet a function .
T H E O R E M 1 1 .31 . The zeta function Z(E) of an elliptic curve is a rational
function. More precisely, it is of the form

where a G Z is defined by the relations \E(¥)\ = q + l — a or, in terms of C, |C(F) | =

q — a. The zeta function satisfies the functional equation Z{E, (qT)~ 1 ) = Z(E, T).
LEMMA 1 1 .32 . Let d ^ 0 and let hd(C) be the set of linear equivalence classes
ofF-rational divisors of degree d. Then hd{C) is finite and \hd(C)\ = \ho(C)\ ^
\E(¥)\.
PROOF. Fo r an y rationa l diviso r D , w e hav e b y Propositio n 1 1 .2 8 th e equiv -
alence D ~ [a(D)\ + (degD — l)[oc] , thu s th e linea r equivalenc e clas s o f D only
depends o n o~(D). I f D is F-rational , i t follows tha t cr(D) G E(W), an d therefor e
the inequalit y \h d(C)\ < \E(¥)\ holds .
Moreover, i t is clear tha t th e ma p D i->> D-f - d[oo] wit h invers e D—
i > > D — d[oo]
induces a bijection betwee n hd(C) an d ho{C). D

P R O O F O F THEORE M 1 1 .31 . Sinc e F-rationa l divisor s o n E are simpl y com -

binations wit h intege r coefficient s o f divisor s associate d t o close d points , th e Eule r
product (1 1 .40 ) give s th e forma l powe r serie s expressio n

Z(E) = Y^ Tdeg(D)

where th e su m i s over al l effectiv e F-rationa l divisor s o n E.

Split th e su m accordin g t o the degre e d of D; fo r d = 0, the onl y effectiv e
divisor i s D = 0 so
Z(E) = l + J2Td £ 1 .
deg(D)=d
For eac h d, spli t furthe r th e su m ove r divisor s o f degree d in linear equivalenc e
classes. B y Lemm a 1 1 .32 , ther e ar e ho(C) equivalenc e classe s fo r eac h d. Fo r a
given clas s (tha t o f D say) , th e contributio n i s the numbe r o f effective (F-rational )
divisors linearl y equivalen t t o D. B y (1 1 .34 ) an d Theore m 1 1 .29 , thi s i s equa l t o

Since d > 1 , the Riemann-Roc h theore m implie s £(D) — deg(D ) = d , so the

computation o f Z(E) i s now straightforwar d

z{E) = \ + Y,Td E i =i + ^ ^ x > d - i ) r d

deg(D)=d

= h 0(C)/ T dT N = h 0(C)T
q-l\l-qT l-TJ (l-T)(l-qT)
l b T
(11 42) - - + «T2
1}
(l-T)(l-qT)
 11. SUM S OVE R FINIT E FIELD S 299

where b is define d b y ho(C) = q -f 1 — b.

This prove s th e rationality , an d give s th e precis e form , excep t tha t w e need t o
prove that a — 6 , where |i£(F) | = q + 1 — a, o r equivalently ho(C) = |£(F) | (actuall y
in Lemma 1 1 .32 , we have already shown \ho(C)\ ^ |-£(F)| , but w e do not nee d it an y
more). T o obtai n thi s equality , star t fro m th e origina l definitio n (1 1 .39 ) o f Z(E),
and compare with (1 1 .42) : th e latter is seen to imply that |i£(F) | = q+1 —b = HQ{C).
Finally th e functiona l equatio n o f Z(E) i s a forma l consequenc e o f (1 1 .42) . •

It i s worth recordin g separatel y on e o f th e las t step s o f th e proof .

PROPOSITION 1 1 .33 . Let E be an elliptic curve over a finite field F, let D be
a divisor on E. Then D is principal if and only if deg(D) = 0 and cr(D) = 0 G E.
More precisely, the map j : D i- > cr(D) is an isomorphism between the group of
divisor classes of degree 0 and E(¥).
PROOF. A diviso r D i s F n -rational fo r som e n ^ 1 ; lookin g a t E ove r F n , i t
suffices t o prov e th e isomorphis m betwee n classe s o f F-rational divisor s o f degre e 0
and F-rationa l points . Bu t j i s a surjectiv e (j([x\ — [oo] ) = x) ma p betwee n finit e
sets wit h th e sam e cardinalit y (\ho(C)\ = \E(¥)\). •

This i s th e specia l cas e o f th e so-calle d Abel-Jacob i Theorem , fo r a n ellipti c

curve ove r a finite field. I t actuall y hold s ove r an y field, an d a generalizatio n t o
all (smoot h projective ) curve s i s th e conten t o f th e theor y o f jacobia n varietie s
associated t o curves .
To conclud e th e proo f o f Theore m 1 1 .25 , we procee d a s i n th e cas e o f Kloost -
erman sums : fro m (1 1 .41 ) , w e deriv e
| E ( F n ) | - ( g n + l ) = a n + /3n
where 1 - aT + qT 2 = ( 1 - a T ) ( l - f3T). The n Lemm a 1 1 .22 , applie d wit h th e
input fro m Stepanov' s Theore m 1 1 .1 3 , show s tha t \a\ ^ y/q, |/? | ^ y /q1 an d sinc e
a/3 — q, this conclude s th e proof .
EXERCISE 2 . Assumin g th e genera l Riemann-Roc h formul a (1 1 .38) , prove tha t
for a smoot h projectiv e algebrai c curv e E o f genu s g ove r a finite field F wit h q
elements, th e zet a functio n

n^l

is a rationa l functio n o f th e for m

Z[E)
- (l-D(l-tfT )
for som e polynomia l P wit h integra l coefficient s an d degre e 2g.
[Hint: Th e questio n wil l arise whethe r ther e exis t F-rationa l diviso r classe s o n
E o f degre e 1 (which i s obvious fo r ellipti c curve s sinc e th e poin t o o i s F-rational) .
The imag e of the degre e map i s 5Z for som e 6 \ (2g — 2) (th e degre e o f the canonica l
class). Usin g thi s fact , find a preliminary for m o f the zet a functio n an d analyz e th e
poles t o sho w tha t actuall y 5 = 1 (see [Mor2] , 3.3). ]
30 0 11. SUM S OVE R FINIT E FIELD S

11.11. Surve y o f furthe r results : a cohomologica l primer .

The method s o f Stepano v ar e ver y usefu l and , i n certai n circumstance s the y
provide th e bes t tool s availabl e today , especiall y whe n th e genu s o f th e curv e i s
large compare d t o th e cardinalit y o f th e finite field (se e fo r instanc e th e proo f b y
Heath-Brown o f non-trivia l estimate s fo r Heilbron n sum s [HB2]) .
However, th e deepes t understandin g o f exponentia l sum s ove r finite fields an d
the greates t impac t o n classical problems o f analytic numbe r theor y come s from th e
sophisticated concept s o f algebrai c numbe r theory , especiall y th e £-adi c cohomol -
ogy theor y a s develope d b y Grothendiec k an d hi s collaborators , whic h giv e a ver y
powerful an d flexible framewor k fo r workin g wit h ver y genera l exponentia l sums .
The proo f o f the Rieman n Hypothesi s fo r varietie s b y Delign e [Del] , an d eve n
more hi s far-reachin g generalizatio n [De2] , ar e th e basi s fo r th e extensiv e wor k o f
Katz, Laumo n an d others . I t i s beyond th e scop e of this book to discuss this theor y
in great detail . Le t u s direct th e interested reade r t o the survey articles [Lau] , [K2].
Study o f th e foundationa l basi s o f th e £-adi c theor y ca n b e starte d i n [De3 ] an d
continued togethe r wit h application s i n the book s o f Katz, fo r instanc e [K3] , [K4].
We will limit thi s section to a short introductio n o f the basic vocabulary an d w e
will stat e a fe w o f th e mos t fundamenta l result s i n thi s language . W e the n includ e
examples t o sho w tha t suc h knowledg e ca n alread y b e ver y usefu l eve n whe n on e i s
not familia r wit h th e detail s an d backgroun d o f algebrai c geometry .
In Section s 1 1 . 4 an d 1 1 .5 , w e hav e show n tha t Gaus s sum s an d Kloosterma n
sums ca n b e relate d t o analogue s o f Dirichle t character s ove r finite fields. Th e
£-adic cohomologica l formalis m whic h w e no w discus s ca n b e though t a s relatin g
exponential sums , dually , t o object s whic h ar e Galois-theoreti c i n nature .
The exponentia l sum s S n define d b y (1 1 .8 ) ca n b e interpreted a s sums ove r th e
algebraic curv e Uf }9 consistin g o f ¥ p minu s the pole s of the rational function s / an d
g. Mor e generally , on e wishe s t o conside r exponentia l sum s no t onl y ove r curve s
but ove r mor e genera l varieties . W e wil l us e som e basi c vocabular y o f algebrai c
geometry t o describ e suc h situations , bu t wil l illustrate the m i n th e simple r cas e of
curves. Alread y th e cas e o f (1 1 .8 ) an d Uf, g ar e quit e interesting .
Let F b e a finite field an d U/¥ b e a smoot h algebrai c variet y o f dimensio n
d ^ 0 (technically , w e assum e a s par t o f th e smoothnes s assumptio n tha t U i s
geometrically connected , an d a s part o f being a variet y tha t U i s quasi-projective) .
The simples t example s i n dimensio n 1 ar e £//,$ , o r smoot h projectiv e curves . I n
dimension d > 1 , th e mos t importan t example s ar e th e affin e d-spac e A d , wit h se t
of point s A d(¥) = F d , an d th e projectiv e d-space . Th e exponentia l sum s ove r U
will b e o f th e typ e

(11.43) S n= J2 x(N{f(x)))4,(Tr(g(x)))
xeu(¥n)
where / an d g ar e F-rationa l function s define d o n U.
To U/¥ i s associate d th e so-calle d arithmeti c etal e fundamenta l grou p 7Ti(C7 )
which "classifies " etal e covering s V — > U o f U, an d i s th e analogu e bot h o f th e
Galois grou p o f a field, o r o f th e "ordinary " topologica l fundamenta l group . A
morphism o f algebrai c varietie s i s etal e i f i t i s flat an d unramified ; i f U i s a curve ,
this means V i s a curve, / i s non-constant an d unramified . Fo r the simpler purpose s
of exponential sums , the fundamenta l grou p ca n b e considered somewha t a s a black
 11. SUM S O V E R F I N I T E F I E L D S 301

box i n wha t follows , bu t on e shoul d kee p i n min d tha t th e element s o f V i n K\{U)

act a s automorphism s o f an y etal e coverin g TT : V —> U (i.e . TT^X) = n(x) fo r
any 7 G TTI(U) an d x G V), an d tha t i t i s a functor : an y ma p E /— » V betwee n
varieties induce s a continuou s grou p homomorphis m IT\(U) —> TTI(V). (On e shoul d
fix a base-poin t i n definin g 7Ti(£/) , bu t a mor e o r les s canonica l choic e exists , th e
so-called "generi c point " o f the schem e U.)

EXAMPLES. (1 ) Le t U be a singl e point {x} define d ove r F . The n 7ri(U) i s th e

Galois grou p Gal(F/F) .
(2) Le t U/¥ b e a smoot h curve , no t necessaril y projective . Ther e i s a a n
associated smoot h projectiv e curv e C/¥ suc h tha t U C C wit h complemen t a finite
set T o f points . I f U — F fo r instance , the n C = F 1 i s th e projectiv e line , an d
T = {0,oo} .
The fundamental grou p can be described concretel y a s follows: le t K = ¥(U) —
¥{C) b e th e functio n field o f £/ , i.e . th e field o f rationa l function s o n U o r C (i f
C = P 1 , the n K = ¥{t) i s th e usua l field o f rationa l fractions) . W e hav e th e
Galois grou p GK — G&\(K / K) o f K. Fo r ever y close d poin t x o f C , ther e i s
the correspondin g discret e valuatio n ord x o f K. Thi s extend s t o th e separabl e
closure K o f K, an d give s ris e t o a decompositio n grou p D x < GK an d a n inerti a
group I x < D x a s i n classica l algebrai c numbe r theory , wit h th e propert y tha t
Dx/Ix ~ Gal(F g /F 9 ), wher e ¥ q i s th e residu e field o f x , a finite field wit h q = Nx
elements. The n TTI(U) "is " th e quotien t o f GK b y th e smalles t close d norma l
containing al l inerti a group s I x fo r x a close d poin t o f U.

Fix a prim e numbe r £ ^ p. Th e object s use d t o interpre t exponentia l sum s

over U ar e th e so-calle d £-adi c sheave s o n [/ . I n th e simple r cases , thos e wil l b e
"lisse", i n which cas e there i s a simpler alternat e Galois-theoreti c descriptio n whic h
we take a s definition .

DEFINITION. Le t U/¥ b e a smoot h variet y ove r a finite field. A liss e £-adi c

sheaf o n U is a continuou s representatio n p : TT\(U) — > GL(V) wher e V i s a finite
dimensional Q^-vecto r space . Continuit y refer s t o th e profinit e topolog y o n TTI(U)
and th e f-adi c topolog y o n V.
Note th e similarit y wit h th e definitio n o f Galoi s representation s o f numbe r
fields (se e Sectio n 5.1 3) . Becaus e o f th e origina l definitio n o f a sheaf , on e usuall y
denotes ^-adi c sheave s b y curl y letter s J 7, <? , etc. Notic e tha t on e ca n obviousl y
speak o f direct sums , tensor product , symmetri c powers , etc., of lisse £-adic sheave s
by performin g th e correspondin g operation s o n th e representations . Als o on e ca n
speak o f irreducibl e sheaves , etc .
An importan t f-adi c sheaf , denote d Q^(l) , i s obtained b y considering th e natu -
ral actio n o f 7Ti (U) o n £-powe r root s o f unity , whic h arise s fro m th e etal e covering s
where on e simpl y extend s th e bas e field fro m F t o it s extensio n b y root s o f unity .
This actio n i s given b y a certai n characte r xe : ^i(U) — > Q}. Usin g thi s sheaf , on e
defines Tat e twists : i f J 7 i s a liss e £-adi c shea f an d i e Z , the n on e denote s T(i) (J 7
twisted i times ) th e shea f whic h correspond s t o th e actio n p' o f 7Ti (U) o n th e sam e
vector spac e bu t wit h
P/(7) = XK7)P(7) ;
in othe r words , .F(l ) = J(8)Q^(1 ) fo r instance .
302 11. SUM S OVE R FINIT E FIELD S

Exponential sum s aris e b y lookin g a t th e actio n o f th e Frobeniu s element s a t

points o f U. Le t x b e a close d poin t x o f 1 7 , which ca n b e see n a s a Galoi s orbi t
of point s i n U(¥). Th e fundamenta l grou p o f th e "point " x i s th e Galoi s grou p
Dx o f th e residu e field o f U a t x , isomorphi c t o F n wher e n i s th e degre e o f x. B y
functoriality ther e is a map D x— > TTI(U). W e have D x ~ Gal(F n /F n ) an d th e latte r
is generate d (topologically ) b y th e Frobeniu s morphis m ex , s o takin g th e imag e
we ge t i n TTI(U) a well-define d conjugac y class , calle d th e arithmeti c Frobeniu s
conjugacy clas s a t x. I n particular , fo r an y ^-adi c shea f on e ca n spea k o f th e trac e
Trp(a x ) withou t ambiguity . However , i t turn s ou t tha t i t i s th e invers e F o f a
(the so-calle d geometri c Frobenius ) whic h appear s naturall y i n th e cohomologica l
description o f exponentia l sums . W e denot e b y F x th e correspondin g conjugac y
class; i t i s called simpl y th e Frobeniu s conjugac y clas s a t x (omittin g th e adjectiv e
geometric).

THEOREM 1 1 .34 . Let U /¥ be a smooth variety, let f ^ 0 and g be ¥-rational

functions on U, let ip be an additive character and \ & multiplicative character of
¥. Let S n — Sn(U, / , #,X>V0 be the associated exponential sums over U(¥ n) as in
(11.43). Then there exists a lisse l-adic sheaf T on U of degree 1 with the property
that for all n ^ 1 we have

(11.44) S n= Y. Tr
(F*\F)
xeu(¥n)

where we denote Tr(g \ J7) = Tr(p(g) \ V), T corresponding to the representation

p : 7Ti(i7)->GX(V) .
To compar e wit h th e character s use d t o describ e Gaus s sum s an d Klooster -
man sums , on e shoul d thin k o f th e latte r a s analogue s o f Dirichle t character s o r
Hecke characters , wherea s th e ^-adi c sheave s give n b y thi s theore m ar e analogue s
of Galoi s characters . Th e correspondenc e betwee n th e tw o concept s i s a n instanc e
of reciprocit y o r class-fiel d theory .
We sketc h th e constructio n o f T i n th e cas e wher e X = 1 and g i s a non-zer o
rational functio n o n U = A 1 — {poles o f #} , over F , whic h make s i t clea r tha t thi s
is very closel y relate d t o th e argumen t i n Sectio n 1 1 .7 . Conside r th e curv e

(11.45) C : y q-y = g(x)

and notic e tha t ther e i s a surjectiv e ma p TT : (x,y) \-> x fro m C t o U. Fo r an y

a e F , th e equatio n y q — y — a = 0 i s separable , henc e i t ha s q distinc t root s i n
F. I n fac t th e additiv e grou p o f F act s o n th e root s b y translation : i f y i s a roo t
and z G F, the n (y + z) q — {y + z) = y q — y = a. Moreover , TT : C — > U i s a n
etale coverin g (we'v e just see n i t i s everywhere unramifie d an d surjective) . I n othe r
words, 7 r is a n etal e Galoi s coverin g wit h Galoi s grou p isomorphi c t o th e additiv e
group F (covering s give n b y suc h equation s ar e calle d Artin-Shreie r coverings) .
The fundamenta l grou p TTI(U) act s o n C b y automorphism s o f th e covering ,
which mean s a s translation s b y element s o f F a s above . Thi s define s a surjectiv e
map (p : TTI(U) - > F suc h tha t ^(7 ) = 77 / — y fo r an y y G C. (Thi s doesn' t depen d
on th e choic e o f y becaus e th e actio n o f 7 o n C mus t b e a morphis m o f curves. )
Consider th e trivia l £-adi c shea f Qi o n C , o r equivalentl y th e trivia l represen -
tation o f 7Ti(C) . B y th e abov e w e ca n construc t th e induce d representatio n p fro m
 11. SUM S O V E R F I N I T E F I E L D S 303

7Ti(C) t o 7Ti(L r), which ca n b e describe d a s th e spac e

V= { /: TTI(£ 0 -) • Q * | /(r 7) = / ( 7 ) for an y r e m(C)}

(where r G ^\{C) i s see n throug h th e ma p 7Ti(C )—> • ?ri(?> 0 comin g fro m 7r) , o n
which 7Ti(L r) act s b y translatio n o n th e righ t

p{l)fiT ) = /( T 7) -
The element s f £ V depen d onl y o n 7ri(C)\jri(U) ~ F (i.e . o n th e automorphism s
of the covering C —> U) , whic h implies that V ~ Q ^ is an ^-adic sheaf on C/ of degree
g. Th e representatio n spac e V ca n b e decompose d ove r th e additiv e character s ^
ofF,

where £ ^ i s th e ^-eigencomponen t o f V , namel y

£ ^ = { / € V | p(7) / = ^{f{l))f for al l 7 € *n([/)} .

It i s eas y t o se e tha t eac h £ ^ i s a n £-adi c shea f o n [7 , an d becaus e p i s induce d
from th e trivia l representation , eac h C^ i s o f degre e 1 .
Then fo r ever y additiv e characte r ^ , th e £-adi c shea f o n U corresponding t o C^
is the shea f satisfyin g (1 1 .44 ) fo r th e exponentia l sum s S n(U,g,ip).
Indeed, i f x G U(¥ n)1 an d y satisfie s y q — y = g(x), the n th e Frobeniu s o f x
acts o n y b y y qU = y + Tr ¥ri/¥(g(x)) sinc e

^ n - 1 / -y qn
- y ^ " 1 + y qn~l + yq - y
= (y q ~ y)^' 1 +'" + y q-y = Tr{y q-y)= Tvg(x).
Hence (f(cr x) — Trg(x ) an d b y definitio n o f C^ i t follow s tha t cr x act s o n C^
by multiplicatio n b y ^(Trp(x)) , henc e F x — a~l act s b y ij;(Trg(x)), whic h give s
(11.44).
In particular , not e tha t takin g th e trac e fo r Qi o n C w e deriv e

|C(F n )| = ^ 5 n ( f / , / , ^ ) ,

as i n (1 1 .30) .
EXERCISE 3 . (1 ) Le t S n b e th e characte r su m (1 1 .8 ) wit h g = 0 fo r som e
multiplicative characte r x oi F * an d som e non-zer o rationa l functio n / G F(x), o n
the variet y U = A 1 — {zeros an d pole s o f / } . Describ e a s abov e th e constructio n o f
the shea f C satisfyin g (1 1 .44 ) i n thi s case . [Hint : Us e th e cove r y d = / ( # ) , wher e
d i s the orde r o f th e multiplicativ e characte r x]
(2) Le t S n b e a s i n (1 1 .8) , U C A 1 th e complemen t o f th e zero s an d pole s o f
/ an d th e pole s o f g. I f C^ i s th e shea f satisfyin g (1 1 .44 ) fo r / = 1 and C x i s th e
sheaf satisfyin g (1 1 .44 ) fo r g — 0 , sho w tha t C = C^ ® Cx satisfie s (1 1 .44 ) fo r S n.
EXAMPLES. (1 ) Eve n th e case p — 1 is interesting whe n dealin g wit h a genera l
variety U. Thi s "trivial " ^-adi c shea f i s denote d Q^ , an d on e ha s S n = |L r (F n )|.
(2) Fo r th e shea f Q^(l) , notic e tha t a x act s b y £ \-> £ q fo r an y roo t o f unit y i f
Nx = q. Therefor e F x act s b y £ i—)• £ x / 9 an d i n particula r th e onl y eigenvalu e o f F x
is q~ l.
 30 4 11. SUM S OVE R FINIT E FIELD S

Now i n additio n t o U /¥w e conside r it s "extensio n o f scalars " U/¥ ove r th e

algebraic closur e o f F . Ther e i s a correspondin g geometri c fundamenta l grou p
7Ti([7), which sit s i n a n exac t sequenc e
1
(11.46) -> m(U) - > m(U) - > Gal(F/F ) - > 1 .
To ever y ^-adi c shea f T o n U ar e associate d th e £-adi c cohomolog y group s wit h
compact suppor t o f U wit h coefficient s i n T. Thos e ar e finite-dimensiona l Q r
vector spaces , denoted , H lc(U,F) fo r i > 0 . Th e ke y poin t i s that th e Galoi s grou p
of F act s naturall y o n H lc(U, J7 ), an d i n particular , s o d o th e Probeniu s a an d it s
inverse F , th e geometri c Probenius . Th e ke y to th e cohomologica l interpretatio n o f
exponential sum s i s th e
GROTHENDIECK-LEFSCHETZ T R A C E FORMULA . Let U/¥ be a smooth variety
of dimension d ^ 0 , T an i-adic sheaf on U. We have H %C{U\F) = 0 if i > 2d and
for any n ^ 1
Tr
(11.47) J2 ^F* I F) = Tr(F n | H°C(U,F)) - Tr(F n \ Hlc(U\F)) + .. .
xeu(¥n)
n
- Tr(F | H 2cd~\U,F)) + Tr(F n \ H 2d
(U,F)).

Therefore t o evaluat e th e exponentia l sum s (1 1 .43 ) usin g th e associate d sheaf ,

we nee d t o kno w th e traces , o r equivalentl y th e eigenvalues , o f F (equivalently , o f
a — F~ l) actin g o n H lc for 0 ^ i < 2d . I t turn s ou t tha t i n mos t case s H® and H 2d
are eas y t o compute :
PROPOSITION 1 1 .35 . Let T be a lisse £-adic sheaf on a smooth variety U/F,
corresponding to the representation p of n\(U) on the Q^-vector space V. We have
y*i(u) if U is projective,
(11.48) HZ(U,?)~
0 if U is not projective,
and
2 d
(11.49) H c (U,T)^VVl{0)(-2d)

where V G denotes the space of vectors invariant under the action of a group G on
an abelian group, and VQ denotes the space of co-invariants, the largest quotient
of V on which G acts trivially. In both cases, the isomorphisms are canonical
isomorphisms of vector spaces with an action of the Galois group of¥.
Since V i s a representatio n o f 7Ti(£/) , th e exac t sequenc e (1 1 .46 ) show s tha t
y*i(u) a n ( j V 7ri(jj^ ar e acte d o n b y Gal(F/F) , "through " th e give n representatio n
P-
This propositio n show s tha t fo r a curv e C//F , th e onl y "difficult " cohomolog y
group i s Hl(U,F).
EXAMPLE. Le t U — E/¥p b e a n ellipti c curve , T — Q^ th e trivia l sheaf . B y
the propositio n on e ha s
(1) H®{E, Qt) = Q£ , with trivia l actio n o f F (sinc e Qt i s the trivia l sheaf) .
(2) Hl(E,Qt) = Q * ( - l ) , s o b y definitio n o f th e twist , F act s b y multi -
plication b y p (o n root s o f unity , i.e . o n Q^(l) , o~ act s b y £ i- > £ p , henc e F b y
multiplication b y p _ 1 ) .
 11. SUM S O V E R F I N I T E F I E L D S 305

The Lefschet z trac e formul a (1 1 .47 ) give s

\E(Wn)\ =p n + l-Tr(F n\ H l
c{E,Q,))

(compare Theore m 1 1 .31 ) .

More generally , on e derive s th e rationalit y o f th e zet a functio n directl y fro m
the Trac e Formula .
COROLLARY 1 1 .36 . Let U, S n and T be as in Theorem 1 1 .34. For 0 < i < 2d,
let bi = di m if* (£7, T) and
bi

Pi{T) = det( l - FT | Hlc(U,F)) = JJ ( 1 - a^T).

i=i
We have

and for n^ 1 ,

(11.50) S n= J2 (-!)Xi -

Theorems 1 1 .4 , 1 1 . 8 an d th e resul t o f Exercis e 1 ar e al l specia l case s o f thi s

corollary, togethe r wit h suitabl e computation s o f cohomology groups . Th e number s
bi are calle d th e ^-adi c Bett i number s fo r T.
Of muc h greate r importance , however , i s Deligne' s vas t generalizatio n o f th e
Riemann Hypothesi s [De2] . On e start s wit h th e followin g "local " definition :
DEFINITION. Le t w e Z b e a n integer . A liss e ^-adi c shea f T o n U/¥ i s sai d
to b e pur e o f weigh t w i f fo r an y close d poin t x o f [/ , al l eigenvalue s o f F x actin g
on th e Q ^ vecto r spac e V associate d t o J- ar e algebrai c integer s al l conjugate s o f
which hav e th e sam e absolut e valu e equa l t o q w/2 wher e q = Nx i s the cardinalit y
of th e residu e field.
For instance , th e trivia l shea f Q ^ is pure o f weight 0 (all eigenvalues 1 ) . Fo r an y
i £ Z , Qi(i) i s pure o f weight —2z , and i f T i s pure o f weight w, the n F(i) i s pure of
weight w — 2i. Fo r an y exponentia l su m (1 1 .43) , th e associate d shea f T i s pur e o f
weight 0 because th e onl y eigenvalu e a t x i s the roo t o f unit y x(Nf(x))ip(Trg(x)).
THEOREM 1 1 .3 7 ( D E L I G N E ) . Let U/¥ be a smooth variety and J 7 a lisse l-adic
sheaf on U, pure of weight w. Let i ^ 0 and let £ be any eigenvalue of the geometric
Frobenius F acting on if*(£/ , J7 ). Then £ is an algebraic integer, and if a G C is a
conjugate of '£, we have
(11.51) |a | < </™+*)/ 2.

The conclusio n i s als o phrase d a s sayin g tha t H lc{U^!F) i s mixe d o f weight s

^i + w. I f ther e i s equality i n (1 1 .51 ) , the n H lc{U,T) i s said t o b e pur e (o f weigh t
w + i). I n certai n cases , on e ca n appl y dualit y theorem s (fo r instanc e Poincar e
duality) t o deduc e furthe r tha t (1 1 .51 ) i s a n equality .
30 6 11. SUM S OVE R FINIT E FIELD S

R E M A R K . Althoug h Deligne' s proof i s a monumental achievemen t o f very dee p

algebraic geometry, it is an interesting fact tha t a crucial use is made of a generaliza-
tion of the metho d o f Hadamard an d de la Vallee Poussin for proving non-vanishin g
of L-function s o n the line R e (s) = 1 (see Section 5.4) . Similarly , i n Deligne's firs t
proof [Del] , th e ideas o f the classical Rankin-Selber g metho d fo r modular form s
are essentia l (specifically , Delign e acknowledge s th e influence o f [Ra3]).
EXAMPLE. Le t C/¥ b e a smooth connecte d projectiv e curv e (fo r instance, an
elliptic curve) . B y Proposition 1 1 .3 5 as in the previous example , w e have easily :
(1) H®(C,Qt) = Qe, with F actin g trivially .
(2) Hf(C,Qi) = < M - 1), wit h F actin g b y multiplication b y p.
It i s more difficul t t o show tha t
(3) Hl(C,Qi) ~ Q ^ , as Q# vector space s (no t as Galois-modules!), wher e
g ^ 0 is the genus of C (for an elliptic curv e g = 1).
Moreover, ther e i s a Galois-invariant perfec t pairin g

It follow s tha t i f a is one of (the complex conjugate s of ) the eigenvalues o f F o n

iJi((7, Qg), the n p/a i s one also. Henc e fro m Theore m 1 1 .37 , sinc e Q# is pure of
weight 0 , one deduces tha t \a\ = yjp. Thu s

|C(Fn)|=p" + l - £ "?
2=1

where the c^ G Q are the eigenvalues of F on H\. Estimatin g triviall y now , w e get

|C(F„)|-(p n + l) ^2gp n/2

recovering the Riemann Hypothesis , an d in particular, Theore m 1 1 .2 5 for the case

<? = !•
In th e case of exponential sum s (1 1 .43) , the sheaf F i s pure o f weight 0 , hence
denoting
l
d(J-) = max{i | H c(U,F)^0},
we derive directl y fro m (1 1 .50 ) an d (11.51) th e bound

(11.52) \S n\^ Y, ^ ^

for n ^ 1 and, in particular,

nd 2
(11.53) \S n\^q ^ (j2bi)-
i

As in the case of Kioosterman sums , th e exponent d(F)/2 i s best possibl e i n this

inequality. Th e bound d{F) ^ 2d gives a trivia l estimat e (becaus e U is smooth
of dimensio n d , it has about q nd points , a s proved b y the Riemann Hypothesi s for
the trivia l shea f Qi). An y improvement o f this trivia l boun d i s equivalent wit h
Hcd(U,F) = 0 , and the square roo t cancellatio n ofte n expecte d fro m heuristi c
reasonings i s equivalent wit h H lc(U,F) = 0 for i > d. Althoug h no t always true ,
this turn s ou t to hol d "generically" , a s the analytic intuitio n suggest s (se e fo r
instance Theore m 1 1 .4 3 below).
 11. SUM S O V E R F I N I T E F I E L D S 30 7

For th e exponential sum s (1 1 .43 ) w e have d(.F ) < 2d, unless T i s the trivia l
sheaf Q^ , s o there i s always a non-trivia l bound . Thi s follow s fro m (1 1 .49) , sinc e
T i s of degree 1 so the space o f co-invariants i s either th e whole spac e (meanin g
the representatio n i s trivial) o r 0. However , thi s smal l gai n is usually insufficien t i n
applications.
Another surprisin g consequenc e o f Deligne's resul t an d the discreteness o f in-
tegers i s the following "self-improving " statement :
COROLLARY 1 1 .38 . Let S n be an exponential sum as in (1 1 .43) and T the
associated sheaf. Suppose w ^ 0 is an integer such that

\Sn\^qw'2+5
for some S G [0, \[ and n ^ 1 . Then we have d{F) ^ w, hence | S n | <C qw^2.

A secon d issu e i n applyin g th e estimates (1 1 .52 ) o r (1 1 .53 ) i n the context of

applications t o analytic numbe r theor y i s that w e usually hav e F = Z/pZ , wit h the
prime numbe r p varying. I n this case , wherea s th e variety U can be defined ove r
Q (o r Z) so that th e sum is, for all p, ove r th e F p -points o f the reduction U p of
U modul o p , the sheaves T v genuinel y depen d o n p (se e the equation (1 1 .45)) , i.e.
there i s no theory o f sheaves ove r U/Z givin g eac h T v b y "reductio n modul o p".
(Katz ha s asked a number o f times for such a theory o f "exponential sum s ove r Z";
see e.g. [K2], but it remain s elusive. ) Thus , th e Betti number s

bz(p)= dim W C(UP,TP)

of th e cohomolog y group s ca n depend o n p, an d the applicabilit y o f the result s
above woul d b e ruined , eve n wit h th e Riemann Hypothesis , i f thes e dimension s
were not bounded i n a reasonable wa y in terms o f p.
This i s in fac t th e case . Th e first genera l resul t i n thi s directio n i s due to
Bombieri [Bo4 ] fo r additive characte r sum s (1 1 .43 ) wher e / = 1 , and was general -
ized b y Adolphson an d Sperber [AS1 ] , [AS2 ] fo r general sum s (thei r method s are
p-adic, base d o n Dwork's origina l ideas) . I n general, thos e result s boun d th e Euler
characteristic
2d

Xc(-F) = E ( -1) i d i m f f c ( ^ ^ ) = £ (- 1
)'6*'
i=0 ( K ^ 2 d

of a shea f T o n (7/F , but furthe r argument s o f Kat z [K5 ] show ho w to deduc e

bounds for
2d
b
tTc(JO = 53dimff*(C7,J0= E *>
i=0 0^i^2d
(hence fo r bi < cr^T)) fro m thos e fo r XcP7 )-
T H E O R E M 1 1 .39 . Let U/Q be a smooth variety over Q, f and g functions
on U with f invertible. Let £ be a prime number and for all p ^ £ such that the
reduction U p ofU modulo p is smooth, let \ andip be any multiplicative and additive
characters of¥ p. Let Tv be an £-adic sheaf on Up such that

E x(Nf(x))i,(Trg(x))= £ Tr{F x\7p)

xEU(¥pn) x<EU{F pn)
30 8 11. SUM S O V E R F I N I T E FIELD S

for n ^ 1 . We have cr^Pp) ^ C where C is a constant depending only on U, f and

9-
A simpl e explici t boun d i s give n i n [AS3 ] i f f(x) = 1 , s o onl y th e additiv e
characters occur , an d g is a Laurent polynomia l o n [ / = ( Q — {0}) d . Th e sum s ove r
Z/_pZ i n questio n ar e therefor e sum s i n d variable s o f th e typ e

(11.54) S ftP= J2 W(^i,...,^) )

xi,... ,x de(i/pzy

where / G Q[#i, x± 1 , . . . , Xd, x^1} i s a non-zer o Lauren t polynomial . Writin g

/ = Y2 a3xJ

for som e (finite ) se t J C Z d , th e Newto n polyhedro n W(f) o f / i s defined t o b e th e

convex hul l i n R d o f J U {0}.
PROPOSITION 1 1 .40 . With the above assumptions, denoting by Tf, v the asso-
ciated sheaf for the sums Sf iP, we have

IXc(^/ l P )|^d!Vol(W(/)),
d
vdFf*)
1 < 0 dWol(W(f))
for any p not dividing the denominator of any coefficient of f, where Vol (W(f)) is
the volume of the Newton polyhedron in the subspace spanned by W(f) in R d, with
respect to Lebesgue measure.
Note tha t b y usin g exclusion-inclusio n an d detectin g polynomial s equation s b y
means o f multiplicativ e characters , on e ca n us e combination s o f sum s o f th e typ e
(11.54) t o describe muc h more general ones. Also , in many cases , one can sho w tha t
only all the odd (o r even) cohomology groups vanish, i n which case |xc(^ r )| — o~ c(F).
See als o Theorem s 1 1 and 1 2 of [K5 ] fo r explici t estimate s i n quit e genera l cases .
We no w giv e example s o f computation s usin g thes e fundamenta l results . Fo r
exponential sums arising in analytic number theory, one often need s nothing more, if
one uses skillfull y som e othe r simpl e trick s suc h a s averaging ove r extr a parameter s
to analyz e th e weigh t o f th e roots .
EXAMPLE 1 . Th e Kloosterma n sum s 5(a , b;p) fo r ab ^ 0 can b e treate d usin g
Proposition 1 1 .4 0 wit h d = 1 an d f(x) = ax + bx~ x. The n W(f) i s th e interva l
[—1,1]. B y Propositio n 1 1 .35 , w e hav e H® — H% = 0 i n thi s cas e sinc e U =
P 1 — {0,oo } i s no t projective , s o a c = ~Xc B y Theore m 1 1 .37 , H^ i s mixe d o f
weight ^ 1 . Henc e w e recove r th e Wei l bound :

\S(a,b;p)\^acp1/2^2p1/2.

(Of course , i n fac t w e hav e b\ = 2 and th e las t inequalit y i s a n equality) .

EXAMPLE 2 . Th e previou s exampl e generalize s t o th e multipl e Kloosterma n
sums define d b y

(11.55) K r(a,q)= E e ( l K * i + ;••+*-) )

 11. SUM S O V E R F I N I T E F I E L D S 309

for r ^ 2 and a j^ 0, so K 2(a,p) = S(a, l;p ) (se e [Bo4] , [Del]) . Withou t appealin g
to th e L- function, on e ca n nevertheles s ge t som e informatio n b y averagin g ove r a.
We ge t

2
(11.56) Yl \ K ^ «)l = I' ~ I1"' « -L

Hence \K r(a,q)\ < q r/2. T o improv e thi s elementar y boun d w e appea l t o the
following Lemm a
LEMMA 1 1 .41 . Given a finite set of distinct angles 6i modulo 2TT and complex
numbers a.j we have

] T \Y jale{nei)\ = N\\a\\ 2 + O(l)

where the implied constant does not depend on N. Hence

(11.57) limsu p \y^aie(n0j) > all .

n—>-+oo

PROOF. W e hav e

2
£ |^a,e(n^)| = iVEl^|2 + E Ea ^ E< n ^-^))-
The inner su m i s bounded b y a constant independen t o f TV, s o the first resul t follow s
and (1 1 .57 ) i s a n obviou s consequence . •

From (1 1 .56 ) an d (1 1 .57 ) it follows tha t amon g th e K r(a,p), ther e i s a t mos t

one roo t o f weigh t r, say fo r if r (ao,p), an d al l othe r root s ar e o f weigh t ^ r — 1 .
Notice that K r(ao,p) £ Q(/ip ), the cyclotomic field o f p-th root s of unity. Usin g
the Galoi s actio n o n Q(/x p), th e conjugate s o f K r(a0)p) ar e K r(aovr,p) fo r v e F*.
By th e Rieman n Hypothesis , thi s mean s tha t th e conjugat e o f th e roo t £ of weigh t
r i s stil l a root o f weigh t r for K r(aov2,p). Henc e v r = 1 for al l v £ Fp , whic h i s
only possibl e i f p — 1 | r. In particula r al l root s ar e o f weigh t < r — 1 if p > r + 1.
One therefor e get s b y Propositio n 1 1 .4 0

1
(11.58) K r(a,q)^q^ ^2

where th e implie d constan t depend s onl y o n r.

In th e cas e o f Kloosterma n su m th e Newto n polyhedro n i s th e simple x wit h
vertices ( 1 , 0 , . . . ,0) , . . . , ( 0 , . .. , 0,1), (—1 ,.. . , — 1) whos e volum e i s 1 /r! . More -
over, i t is know n tha t th e zet a functio n i s a polynomial s o \c — ~o~ c and w e ge t
the precis e estimat e
\Kr{a,q)\^rqW2.
This wa s firs t prove d b y Delign e [De3] , without an y assumptio n o n p an d r.
310 11. SUM S OVE R FINIT E FIELD S

EXAMPLE 3 . Her e i s anothe r higher-dimensiona l example . I n [CI1 ] , th e fol -

lowing exponentia l su m ove r finite fields appears :

(11.59) W( X,ip;p)= J2 X(xy(x + l)(y + l))i>(xy-l)

x,y mo d p

where p i s prime , x * s a non-trivia l quadrati c characte r modul o p an d ip i s an y

multiplicative characte r modul o p.
T H E O R E M 1 1 .4 2 (CONREY-IWANIEC) . There exists an absolute positive con-
stant C such that

(11.60) \W( X,t,p)\^Cp

for all p and all ip as above.

The first ste p o f th e proo f i s t o appl y Theore m 1 1 .3 4 t o sa y ther e exist s a n
£-adic shea f T o n th e algebrai c surfac e

U = {(x,y) | xy(x + l)(y + l)?0},

pure o f weigh t 0 for whic h w e have, fo r q — p n , n ^ 1 , the formul a

W(X,1>\q)= E x(N(xy(x + l)(y + l)))i>(N(xy-l))

x,y€U(¥q)

=£ T±(F {Xty)\f).
x,y£U{Fq)

The Lefschet z trac e formul a (1 1 .47 ) take s th e for m

4
W(X^;q) = Y.i-lYTriF" \ W C(U,T)).
2=0

By Theore m 1 1 .37 , eac h H lc(U ,F) i s mixe d o f weight s < i. Le t (0^,2^,1 ^ ) b e th e

family o f eigenvalue s o f F actin g o n th e whol e cohomolog y (wit h multiplicities) ,
together wit h thei r inde x an d weight . W e hav e \a u\ = p w»l2 an d

1/

By Theore m 1 1 .3 9 i n thi s case , th e tota l numbe r o f root s a v i s bounde d b y a

constant independen t o f p. Thus , w e gai n o n th e trivia l boun d W < C p2 i f w v ^ 3
(instead o f 4) , an d th e statemen t o f th e theore m i s that w v ^ 2 .
The second step is to show that ther e is at mos t on e root o f weight ^ 3 (actually ,
it mus t b e = 3 ) and , i f i t exists , the n ip = x i s th e non-trivia l quadrati c character .
This wil l b e derive d fro m th e followin g averag e formula :

(11.61) A = -±j ] T \W( X^q)\2 = q 2

-2q-2.
 11. SUM S O V E R F I N I T E F I E L D S 311

To prove thi s formula , ope n th e square an d sum over ip first gettin g b y the orthog-
onality o f character s

A= Y1 Y, X(U V (U
11 1 + 1)(V 1 + 1))X(U2V 2(U2 + 1)(V2 + 1))
UiVi=U2V2

= Yl J2^ Ul + 1
)(t'l + 1 ))X((W2 + l)(«i«i + U2))X{U2)
U±,Vi,U2

(where w e shorten th e notation fro m x ° N t o %) . Nex t th e summation ove r u\ i s

performed givin g

if u 2 = vi,
S(vi,ix 2 ) = > X(^ i + l ) x ( ^ i ^ i + ^ 2 j = < , . _, .
if u 2 y^ vi.
^o { -xM ~ x(u2)
Then th e sum over u 2 i s performed givin g

^T x(u2 + i)x(^2)s(vi, w 2) = gx(^i +1 ) + xOi)

and finall y th e sum over v\ give s

A
=Y1 x(vi + X ) («(Vl + X ) + xM + X
) = ^ - 2 ) -2.

From (1 1 .61 ) , usin g Lemm a 1 1 .41 , i t i s clea r tha t fo r al l ip and v w e hav e

wy ^ 3 and moreover, w v ^ 2 except fo r at mos t on e root, fo r one character if;. I f
this cas e occur s fo r ip, it happen s fo r ^ too , so the only possibilit y i s ip being a real
character. Sinc e
W(x,i;q) = (^2x(u(u + i)))2 = l,
u
we must hav e ip = X-
The las t ste p i s to trea t th e cas e ip = x separately . Precisely , on e can sho w
that
\w(x,x;q)\^4q
n
for an y p and q = p . Thi s is done in a purely elementary manne r withou t appealin g
to the Riemann Hypothesis , an d we refer t o [CI1 ] for the details. I n fact, W . Duke
showed tha t W(x^XiP) = 2Re(J 2 (x,£))> wher e J(x>£ ) i s th e Jacob i su m an d £
is a quarti c characte r modul o p = 1 (mo d 4). Als o W(XIX'->P) i s th e P_ ^ n Fourier
coefficient o f the modular for m 77(4z) 6 o f weight 3 and level 12.

When U is not a curve, numerous geometric subtleties can be involved in dealing

with th e non-trivia l cohomolog y group s H lc with i ^ 0 , 2d. Her e ar e two genera l
bounds, amon g many : th e first on e is due to Deligne [Del] , an d the second i s the
recent versio n i n [FK ] of a general "stratification " theore m o f Katz an d Laumon .
T H E O R E M 1 1 .43 . (1 ) Let f e Z[A"i,.. . , X m ] be a non-zero polynomial of
degree d such that the hyper surface Hf in P m _ 1 defined by the equation

Hf : f d(xu... , x m) = 0,
312 11. SUM S OVE R FINIT E FIELD S

where fd is the homogeneous component of degree d of f', is non-singular. For any

p \ d such that the reduction of Hf modulo p is smooth, any non-trivial additive
character i\> modulo p and any n ^ 1 we have
m nm 2
(11.62) | £ - " 1E >(Tr(f( Xl,...,xm)))\^(d-l) q ' .
Xi,... ^XmeWpTi

(2) Let d ^ 1 , n ^ 1 be integers, V a locally closed subscheme of A^ of dimen-

sion di m V(C) < d and f G Z [ X i , . .. , Xn] a polynomial
Then there exists C = C(n , d, V, f) and closed subschemes Xj C Ag of relative
dimension ^ n — j such that
xnc...ax2cx1ahl
and for any rational function g non-zero on V, any h G (Z/pZ)n — Xj(Z/pZ), any
prime p, any non-trivial additive character ip and multiplicative character \ rnodulo
p, we have

^ X{9(x))^(f(x) + ftixi H h nxn) < Cp*~

xe

Note that i n (1), subject t o a geometric conditio n o n Hf, w e obtain squar e roo t
cancellation i n the exponential sum . I n (2) the assumptions ar e much less stringent ,
but th e conclusion i s weaker: w e have a family o f exponential sum s parameterize d
by h e A n , an d roughl y speakin g w e hav e squar e roo t cancellatio n fo r "generic "
sums (fo r h outside a n exceptional subvariet y X\ o f codimension ^ 1 in A n ), while
worse an d wors e bound s ca n occu r onl y o n smalle r an d smalle r subvarieties . W e
will giv e a n applicatio n o f (2 ) in Chapte r 21.
The ^-adi c theor y an d formalism ar e much mor e develope d tha n wha t w e have
surveyed here . I t ca n also deal wit h sum s ove r singula r varieties , bu t the necessar y
algebraic notion s become rather formidable , an d we concede being unable to discuss
the pervers e sheave s tha t aris e i n more advance d situations .
We wish to emphasize, however , tha t thi s theor y i s also particularly well-suite d
to the study of families o f exponential sums . Th e parameters definin g those , say S x,
are mos t naturall y themselve s point s o n som e algebrai c variet y X/F. I n favorabl e
circumstances ther e exist s a lisse £-adi c shea f T o n X (correspondin g t o a n actio n
of 7Ti(X) o n V) suc h tha t

(11-63) S x,n = Tr(FZ\V)

for ever y valu e o f th e paramete r x G X an d an y n ^ 1 . Fo r exampl e a non-zer o
polynomial / o f degre e ^ d over F ca n be describe d a s an F-rationa l poin t o f the
affme paramete r spac e X = A^ + 1 — {0 } by
d
f = ^ciiX 1 h- > ( a0 , . . . , a d ).
2= 0

There i s an ^-adic shea f T o n X suc h tha t

£>(/(*)) = T*(i? i n
xe¥n
 11. SUM S O V E R F I N I T E F I E L D S 313

Purity o f a shea f satisfyin g (1 1 .63 ) depend s o n a firs t applicatio n o f th e Rieman n

Hypothesis. I f i t holds , th e applicatio n o f th e £-adi c theor y typicall y result s i n
equidistribution statement s (followin g fro m [De2] ) fo r th e argument s o f th e expo -
nential sums . Thi s equidistributio n i s i n som e spac e o f conjugac y classe s o f th e
"monodromy group " o f the situation . W e refer fo r instanc e t o [KS ] fo r a very luci d
introduction t o thes e profoun d aspects .
11.12. Comments .
In thi s closin g sectio n w e giv e som e impression s abou t ho w th e exponentia l
and characte r sum s ove r finite fields interac t wit h analyti c numbe r theory . Ther e
are mor e subtl e issue s betwee n th e tw o subject s tha n jus t application s o f result s
concerning the first on e for solvin g problems of the latter. W e could be quite specifi c
by covering completely a few representative examples , bu t i t woul d b e long and no t
transparent enough . Rathe r w e decide d t o discus s principles , idea s an d trick s i n
general term s an d guid e th e reade r t o particula r publications .
First o f al l som e exponentia l sum s appea r whe n on e use s Fourie r analysi s t o
get a hol d o n th e sequenc e unde r investigation . Ther e ar e n o finite fields i n th e
background, s o th e resultin g sum s ar e no t immediatel y relate d t o object s o f alge -
braic geometry . However , on e ca n complet e thes e sum s (b y anothe r us e o f Fourie r
analysis) an d the n facto r the m int o sum s o f prim e powe r moduli . Usuall y on e ca n
evaluate thes e loca l sum s explicitly , o r giv e stron g estimate s b y elementar y o r a d
hoc methods , excep t whe n th e modulu s i s prime . Bu t i n th e prim e cas e on e ma y
naturally conside r th e su m a s bein g ove r a finite field. Thi s schem e allow s u s t o
appeal t o th e powerfu l result s fro m algebrai c geometry . However , th e drawbac k o f
finishing b y estimates fo r every complete sum individually i s that on e cannot exploi t
a possibl e cancellatio n fro m extr a variable s offere d b y th e varyin g modul i (finit e
fields of different characteristic s d o no t interac t i n algebrai c geometry) . Sometime s
a kin d o f reciprocit y formul a ca n hel p tur n th e modulu s int o a variabl e (se e fo r
example [1 1 1 ] o r [M3]) . Anothe r scenari o i s tha t th e sum s ove r modulu s appea r
in th e spectra l resolutio n o f a differentia l operator , i n whic h cas e th e spectra l the -
ory produce s estimate s fa r stronge r tha n thos e derive d b y algebrai c geometry . Fo r
example, thi s i s th e cas e o f sum s o f Kloosterma n sums ; se e Chapte r 1 6 . On e ca n
also thin k thi s wa y abou t th e rea l characte r sum s wit h cubi c polynomials ; the y
are coefficient s o f a cus p for m associate d wit h a n ellipti c curves , s o th e modularit y
gives extr a cancellatio n i n summatio n ove r th e modulus .
As a rule the exponential/character su m o f a given modulus whic h comes out of
analytic numbe r theor y i s incomplete . Thi s itsel f i s not a proble m becaus e variou s
completing technique s ar e availabl e a s mentione d above . Completin g i s a natura l
step t o take , bu t i s i t usefu l o r wasteful ? A t thi s poin t on e shoul d realiz e tha t a
bound fo r a complet e su m hold s essentiall y th e sam e fo r th e origina l incomplet e
sum. Thi s mean s tha t th e resul t i s relativel y weake r fo r a shorte r sum . Stil l i t i s
non-trivial whe n th e lengt h o f the origina l su m i s larger tha n th e squar e roo t o f th e
modulus. Ver y shor t sum s canno t b e treated thi s way . W e do no t hav e a n absolut e
recommendation whe n t o complet e o r no t a give n incomplet e sum . Ou r experienc e
suggests executin g th e Fourie r metho d a s lon g a s th e resultin g summatio n ove r
the frequencie s i s shorte r tha n th e rang e o f th e origina l sum : a t leas t on e ca n fee l
that on e i s progressing . Bu t sometime s i t i s wort h actin g otherwise , acceptin g a
step backward s i n this respec t whil e opening a position fo r a stronger secon d move .
314 11. SUM S OVE R FINIT E FIELD S

For example , imagin e tha t th e amplitud e i n th e exponentia l su m i s no t a rationa l

function, bu t nevertheles s become s on e afte r a n applicatio n o f the stationar y phas e
method t o th e relevan t Fourie r transform . I n thi s cas e th e losse s fro m th e rang e
of summatio n ca n b e recovere d wit h extr a saving s b y application s o f algebrai c
arguments (se e Sectio n 8 o f [CI1 ] , wher e thi s gam e i s playe d i n severa l variable s
simultaneously).

Whatever th e argument s whic h lea d t o complet e sum s ma y be , i n the final ste p

one canno t bea t th e squar e roo t savin g factor . Therefor e t o receiv e a non-trivia l
result on e mus t first produc e someho w a su m wit h a numbe r o f term s large r tha n
the squar e roo t o f th e modulus . Ther e ar e severa l way s t o ge t started , dependin g
on th e shap e o f th e exponential/characte r sum . First , on e ca n tr y t o appl y a Wey l
shift wit h the effec t o f squaring th e numbe r o f terms. Similarl y on e may just squar e
the whol e sum , o r rais e i t t o a highe r powe r t o produc e eve n mor e points . Not e
that shiftin g th e variabl e an d squarin g th e su m ar e no t th e sam e things ; th e first
requires some additiv e feature s o f the variable while the latte r nothin g a t all . Thes e
operations see m t o b e quit e superficia l a t first glanc e (w e ar e takin g essentiall y
replicas o f the origina l sum) , ye t wit h ingenuit y on e can rearrang e th e point s s o th e
summation goe s i n a skewe d direction , th e consecutiv e term s repe l violentl y an d
randomly producin g a considerabl e cancellation . Thi s i s eas y t o say , muc h harde r
to execute . I n fac t on e need s man y othe r devices , suc h a s gluin g severa l variable s
with smal l multiplicit y i n orde r t o arriv e a t a singl e variabl e ove r a rang e large r
than th e squar e roo t o f th e modulus . On e als o mus t smoot h ou t thi s composit e
variable befor e applyin g algebrai c arguments . Usuall y a n applicatio n o f Cauchy' s
inequality an d enlargin g th e oute r summatio n (du e t o positivity ) doe s th e job . A
powerful exampl e ho w this work s i s given b y Burges s [Burl] . I n thi s pape r a shor t
character su m i s estimate d b y a n appea l t o th e Rieman n Hypothesi s fo r algebrai c
curves. A n interestin g poin t i s tha t afte r al l th e trick s on e comes t o a complet e
character su m fo r a curv e o f a larg e genus , althoug h th e origina l su m i s over a lin e
segment.

Different argument s fo r buildin g on e extr a larg e variabl e ar e applie d i n [FI4] ,

consequently th e final complete sum comes in three variables, or equivalently i n fou r
variables ove r a hypersurface . Her e th e Delign e theor y applie s (se e th e Appendi x
by Birc h an d Bombieri) , althoug h th e relate d variet y i s singular . On e shoul d no t
be surprised an d afrai d o f that singularity , because , afte r all , the process of creatin g
more point s o f summatio n a t th e star t i s quit e superficial . I n thi s gam e on e mus t
be experience d whe n mixin g th e point s t o b e sur e tha t i t i s quite random . Anothe r
interesting cas e o f creatin g an d estimatin g exponentia l sum s i n thre e variable s i s
given b y N . Pit t [Pi] .

Applications o f th e Rieman n Hypothesi s fo r curve s ove r finite fields ar e b y

now customary . Muc h les s successfu l ar e th e us e o f genuinel y higher-dimensiona l
varieties. Ther e ar e reason s fo r th e difficultie s involved . Firs t o f al l whe n mor e
variables appear , stronge r restriction s ar e impose d o n the m whic h ar e harde r t o
resolve ( a kin d o f uncertaint y principle) . Jus t imagin e havin g a n abundanc e o f
points t o wor k with , bu t whic h ar e no t fre e becaus e o f som e sid e conditions . Fo r
example ho w would yo u cop e with a requirement tha t th e determinant s o f a famil y
of ellipti c curve s matc h th e conductors ?
 11. SUM S O V E R F I N I T E FIELD S 315

Of course , ther e ar e als o direc t application s o f th e Rieman n Hypothesi s fo r

varieties t o traditiona l problem s o f solvabilit y o f diophantin e equation s b y mean s
of th e circl e metho d (se e example s i n Chapte r 20) . I f th e numbe r o f variable s i s
sufficiently large , on e need s nothin g t o manipulate , excep t fo r completin g th e su m
by th e standar d Fourie r method . Som e ingenuity , however , i s require d t o appl y
the circl e metho d ( a variant o f Kloosterman) t o treat diophantin e equation s wit h a
relatively smal l number o f variables, a n excellen t exampl e bein g th e wor k o f Heath -
Brown o n cubi c form s [HB6] .
It i s possibl e i n som e circumstance s t o bea t th e boun d fo r exponentia l sum s
which i s derive d b y th e Rieman n Hypothesis . Thi s i s becaus e th e angle s o f th e
roots o f th e //-functio n themselve s var y s o tha t additiona l cancellatio n ma y occur .
Deligne an d Kat z hav e establishe d suc h occurrence s fo r familie s parameterize d b y
points o n curve s o r varieties . I n othe r words , i n thei r case s on e i s actuall y con -
sidering exponentia l sum s i n mor e variables . However , th e cancellatio n o f root s
can als o occu r fo r familie s parameterize d b y point s ove r smal l irregula r sets . Mor e
important fo r analyti c numbe r theor y i s tha t thes e set s ca n b e quit e general , n o
structure o f a subvariet y i s needed , bu t instea d a kin d o f a bilinea r for m struc -
ture woul d suffice . I n practic e i t i s no t clea r ho w t o wor k wit h th e roots , s o on e
returns t o th e correspondin g exponentia l sum s wher e manipulation s wit h th e pa -
rameters (grouping , gluing , etc. ) ca n b e performed properl y accordin g t o th e shap e
of th e involve d rationa l functio n whic h i s see n wit h th e nake d eye . I n thi s proces s
one mus t no t destro y th e complet e variable s sinc e i n ou r min d th e correspondin g
summations ar e alread y execute d i n term s o f th e roots . Therefor e whe n applyin g
Cauchy's inequalit y t o smoot h th e on e variabl e compose d ou t o f th e parameters ,
we put al l th e complet e variable s t o th e inne r summation , sa y n o f them , togethe r
with som e remainin g parameter s whic h wher e no t use d i n th e composition . Thes e
inner parameter s ar e critica l fo r enlargin g th e diagonal . After squarin g ou t w e ge t
a complet e exponentia l su m i n 2 n + 1 variables whic h depend s o n th e inne r pa -
rameters. Excep t fo r a fe w configuration s o f those , th e complet e exponentia l su m
satisfies th e bes t possibl e boun d derive d fro m th e Riemann Hypothesis , thu s savin g
the facto r o f squar e roo t o f th e modulu s pe r eac h variable . Sinc e th e numbe r o f
variables i s large r tha n twic e th e original , w e wi n th e game . Th e abov e recip e i s
somewhat oversimplified, ye t i t reveal s th e sourc e o f extra saving . On e ca n se e how
it works in the particular cas e of [CI1 ] . Speakin g of [CI1 ] we would like to ad d tha t
here th e exponential/characte r sum s i n severa l variable s emerg e afte r application s
of harmoni c analysi s wit h respec t t o th e hyperboli c Laplac e operato r rathe r tha n
by th e traditiona l Fourie r analysis .

The profoun d theor y o f Delign e an d othe r geometer s i s bein g use d i n analyti c

number theor y wit h spectacula r effects , ye t mor e idea s nee d t o b e invente d t o
fully exploi t it s potential . Perhap s on e shoul d g o beyond borrowin g estimate s an d
penetrate deepl y insid e th e theory . Thi s i s a grea t subjec t fo r futur e research . P .
Michel [M3 ] mad e th e firs t significan t step s (se e als o [FK ] an d [KS1 ]) .
This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/13

CHAPTER 1 2

C H A R A C T E R SUM S

12.1. Introduction .
In analyti c numbe r theor y on e ofte n encounter s sum s o f typ e

(12.1) S="£F(x)
xev
n
where V C Z i s a finite se t an d F : V —> C i s a periodi c functio n o f perio d q.
Because V doe s no t matc h th e periodicity , S i s called a n incomplet e sum . Suppos e
(12.2) F(x) = X (f(x))iP(g(x))
where x , if; ar e multiplicativ e an d additiv e character s t o modulu s g , an d / , g ar e
rational function s wit h intege r coefficients . Precisely , w e assum e tha t

(12.3) / = /i//o , 9 = 9i/9o

where fo,fi,9o,9i€ %[x] an d

(12.4) (fo(x)go(x),q) = l if xeV.

Having fixed thes e polynomial s (whic h ar e no t unique ) w e defin e
(12.5) X (f(x)) = X (h(x)fo(x)),
(12.6) il>{g(x))= >(gi(x)g
1 0(x))

where, a s usual , a denote s th e multiplicativ e invers e o f a modul o q. The n th e

resulting su m

(12-7) S = Y?x(f(x))iP(g(x))
xeV

is calle d a n incomplet e characte r sum . Here , an d hereafter , th e sta r restrict s th e

summation t o the point s o f V satisfyin g (1 2.4) . Th e residue classes x(mod q) whic h
satisfy (1 2.4 ) wil l b e calle d admissible . A n importan t case , bu t b y n o mean s th e
only one , i s V bein g a box . Whe n th e bo x ha s siz e exactl y q we ge t

(12-8) 5 ( x , V ) = E * X(f(x))rl>(g(x))
x(mod q)

which i s calle d a complet e characte r sum .

In thi s chapte r w e giv e basi c technique s o f estimatin g incomplet e characte r
sums i n on e variabl e ove r a n interval . A s a n exercis e fo r th e reader , w e sugges t
generalizing som e o f th e forthcomin g result s t o severa l variables .

317
318 12. CHARACTE R SUM S

12.2. Completin g methods .

The mos t commo n treatmen t o f a n incomplet e su m

(12.9) S(M;N)= ]T * F(n )

M<n^M+N

goes b y expandin g i t int o complet e sum s (thi s i s calle d th e completin g technique) ,

(12 ,o) «(!) . r n*H-~)

x(mod q)

and treatin g th e latte r b y variou s arithmeti c means . T o d o s o we split th e summa -

tion int o th e *-admissibl e residu e classe s n = x(mo d q), an d detec t thes e classe s b y
the orthogonalit y o f additiv e character s gettin g

(12.11)
WW)-i £ *(§)5(! )
a(mod q)

where

S '(T)
(12.12) / an\
<$
M<n^M+N

Usually th e mai n contributio n t o S(M;N) come s fro m a — 0 i n (1 2.1 1 ) fo r whic h

A(0) = T V (we assum e tha t M an d N ^ 1 are integers) . Fo r 0 < \a\ ^ § , w e hav e
W f J K g l o l " 1 . Henc e
LEMMA 1 2.1 . Let F{x) be a complex-valued function defined on the residue
classes x (mo d q) which are •-admissible. Then the corresponding sums satisfy

1
(12.13) S(M;JV)--S(0)|< Yl N" ^-)
0<|a|^

Suppose tha t th e complet e sum s satisf y

(12.14) s(-)|<c(0)(a + M ) V
for som e b G Z an d 6 > | . Thi s boun d i s ofte n tru e wit h 0 = | + £ for an y £ > 0 .
Then (1 2.1 3 ) become s

e
(12.15) S(M]N) - —5(0) 1 < c(0)£(b,q)q

where
1
(12.16) e(b,q)= Yl H- (a + 6,^)i
0<|a|<§

Since £(b,q) i s usuall y smal l (fo r example , w e hav e £(0,q) ^ 2r(q)logq), th e in -

equality (1 2.1 5 ) show s tha t th e incomplet e su m S(M;N) satisfie s essentiall y th e
same boun d a s doe s th e correspondin g complet e sum s 5 ( - ) , u p t o th e mai n ter m
fS(0) .
 12. C H A R A C T E R SUM S 319

Clearly th e abov e metho d o f completin g th e su m o f a periodi c functio n F(x)

works fo r mor e genera l sum s o f typ e
(12.17) S = Y^F(n)G(n)
n

where G{x) i s a nic e functio n o f analyti c characte r whic h decay s t o zer o rapidl y a s
\x\— > oo . No w w e presen t anothe r method . Suppos e G(x) i s o f Schwart z clas s o n
M. The n on e ca n appl y th e Poisso n summatio n formula , givin g

where G(y) i s the Fourie r transfor m o f F(x). I n principle bot h method s ar e equiva -
lent, howeve r th e latte r ma y be preferable i n case G(x) propagate s som e vibrations .
If the vibration s ar e regular , G(-) ca n b e evaluate d asymptoticall y b y th e station -
ary phas e metho d (se e for exampl e Corollar y 8.1 5 ) givin g a n asymptoti c expansio n
for S i n term s o f S(^). O n th e othe r hand , i f G(x) i s wild , on e shoul d hav e som e
reservation fo r completin g S i n term s o f additiv e characters . On e shoul d tr y t o
use "harmonics " whic h ar e mor e adequat e fo r th e particula r case . Th e Fourie r
coefficients o f cus p forms , holomorphi c o r non-holomorphic , ma y serv e wel l a s th e
relevant harmonic s i n man y cases .
12.3. Complet e characte r sums .
Let Xq a n d i>q b e multiplicativ e an d additiv e character s respectivel y t o th e
modulus q. Th e complet e characte r su m

S{xq,1>q)= E * X qU(x))%{g(x))
a?(mod q)

is multiplicativ e wit h respec t t o q. Indeed , le t q = rs wit h (r , s) = 1 . The n Xq

factors uniquel y int o XrXs wher e ;\> > Xs a r e multiplicativ e character s t o modul i r , s
respectively. Th e additiv e characte r ip q is give n b y
'ax\
(12.19) ^ ( x ) = e(— )
Q>

for som e a G Z. B y th e "reciprocity " formul a

(12.20) - + - = -(mo d 1 )
rs q
where ss = l(mo d r ) an d rf = l(mo d 5) , th e additiv e characte r factor s int o i/j q =
ip*ips' Hence i t follow s tha t
(12.21) S( Xq,A) = S(Xr,r r)S(Xs,rs)-

Therefore th e problem o f evaluating a complete character su m o f modulus q reduces

to tha t o f prim e powe r moduli .
The complet e su m S(Xi V O °f modulu s q = p@ shouldn' t b e confuse d wit h th e
character su m over the finite field ¥ q whic h was considered i n Chapter 1 1 , except fo r
q = p i n which cas e 5(x ,VO i s indeed on e o f such sums . Th e cas e o f prime modulu s
belongs t o th e theor y o f L- functions fo r curve s ove r finite fields, a s describe d i n
Chapter 1 1 . Th e rationalit y o f the relevan t L-functio n togethe r wit h th e Rieman n
320 2 . CHARACTER SUM S

Hypothesis (bot h prove d b y A . Wei l [Wei ] i n this case ) yiel d algebrai c number s
9\ > • • • , 9r wit h \g v \ — p, p 2,1 suc h tha t
(12.22) S(x,il>)=9i + '-+9r,
The number r is bounded independentl y o f the characteristic p. Moreover , assumin g
some non-singularit y condition s fo r th e rationa l function s / , 9 wit h respec t t o the
characters x^-> ther e ar e n o root s g v with \g v\ = p, s o (1 2.22 ) yield s
(12.23) \S( X,n^rpi.

See Chapter 1 1 and i n particular Sectio n 1 1 .1 1 for mor e complet e discussio n o f thi s
situation.
There is no need to algebrai c geometry fo r th e complet e characte r sum s S(x,ip)
to modulu s q = p$ with (3 ^ 2 , because i n this cas e elementar y argument s are
available.
LEMMA 1 2.2 . Let q = p2a with a ^ 1 . Then we have

(12.24) S( x,ip) = pa £ * x(f(y))Hg(y))

y(mod p Q )
%)E0(modpQ)

where h(y) is the rational function given by

(12.25) h(y) = ag'(y) + b^(y)

with the integers a , b depending on the characters ip, \ which are determined by
(12.19) and (1 2.27) below.
REMARK. Sinc e th e character s \,ty hav e modulu s p 2a i t is required fo r cor -
rectness o f the summatio n (1 2.24 ) t o know tha t x(/(2/))V ;(5f(y)) doe s no t depen d
on th e choic e o f the representativ e o f y(mod p a) o n th e curv e h(y) = 0(mod p a).
This propert y follow s i n th e cours e o f th e proof .
PROOF. Writ e x = y + zpa, where y and z run independentl y ove r an y fixed
systems o f residue classe s modul o p a , an d y is restricted b y the condition p \
fo{y)9o{y)- W e hav e
(12.26) f{x) = f(y) + f'(y)zp° (mo d p 2«)
Indeed, th e congruence (1 2.26 ) i s easily see n fo r monomials x n by the binomia l
formula
(y + zpa)n = yn + ny n-1 zp°l + • • • .
Then i t extends t o arbitrary polynomial s fi(x),fo(x) wit h integra l coefficient s b y
linearity. Moreover , i f /o(y) = 0(mo d p), the n (1 2.26 ) i s verified fo r f(x)/fo(x) a s
follows:

A(*)//„(*) = (My) + Mv)zPa)fo(y)(i - fo(y)My)zPa)

= Mv)Hv) + (My)fo(y) - R{y)f'0{v)Mv))*Pa-
By (1 2.26 ) w e ge t
a
x(f(x)) = x(f(y))x(i + j(y)z P ).
 12. CHARACTE R SUM S 32 1

Clearly %( 1 - f zpa) i s an additiv e characte r t o modulu s p a , s o there exist s an intege r

b (uniquel y determine d modul o p a) suc h tha t
'bz\
(12-27) X X^
( l ++ zp-)
z ^ )== eey( ^ ) .

Hence
a
(12.28) *(/(*) ) = x(f(y))e(b^(y)zp- ).

Similarly w e get (1 2.26 ) fo r th e rationa l functio n g(x), whenc e

f
(12.29) ^(g(x)) = 4>(g(y))e(ag (y)zp-a).
a
Multiplying (1 2.28) , (1 2.29 ) an d summin g ove r th e residu e classe s y,z modul o p
we ge t
a
s(x,V0= E * x(f(y)M9(y)) £ e(h(y)z P- ).
y(mod p a) ^(mo d p a)
The inne r su m vanishe s unles s h(y) = 0(mo d p a) i n whic h cas e i t equal s p a. Thi s
completes th e proo f o f (1 2.24) . •

LEMMA 1 2.3 . Let q = p 2ot+l with a ^ 1 . Then we have

(12.30) 5(x , </0 = Pa E * x(f(y)Mg(y))G p(y)

y(mod p a)
h(y)=0(mod p a)

where G p(y) is the Gauss sum

2 a
(12.31) G p(y)= Y. e p(d(y)z + h(y)p- z).
z(mod p)

Here h(y) is the rational function (1 2.25) but with b given by (1 2.35) below, and
b f h
(12.32) d(y) = y\y) + - j{y) + (p - l) -(L{y)f.

REMARK. A S Z run s mo d p , s o does z 2/2p, therefor e th e Gaus s su m (1 2.31 ) i s

correctly defined .
PROOF. W e writ e x — y - f zp a wit h y runnin g modul o p a subjec t t o p \
fo(y)go(y), an d z runnin g modul o p a+1 . A s befor e w e argu e tha t

(12.33) f(x) = f(y) + f(y)zp a + ^f"(y)z 2p2a (mod p 2a+l

).

Note tha t th e rationa l functio n \f"{y) ha s integra l coefficients . Thi s i s clea r fo r

the monomia l y n becaus e n(n — 1) = 0(mo d 2) , the n fo r an y integra l polynomia l
by linearity , an d finally fo r a rationa l functio n f(y) = fi(y)/fo(y) b y th e identit y
1 £ti _ 1 f'f f-l f f f-2 I f £" f-2 | f( £f\2 £-3
— 1
2/ 2^ 1 1
^ 0 J J0J0 2l loJo T / U O J JO •

By (1 2.33 ) we get
2 a
(12.34) X (f(x)) = x(/(y))x(l + (£(y)z + \fr (y)z P°)P ).
32 2 12. CHARACTE R SUM S

Consider th e functio n
v2s
1 e 1
« + ^) = \p(^r + (P- 2pj
4)- a+1 v J

2a+1
This i s a characte r o f th e subgrou p o f residu e classe s x ( m o d p ) wit h x =
l ( m o d p a ) . Indee d
f ((1 + zp a){l + wp a)) = £(l + (z + w + zwp a
)pa)
2
(Z + W Z +W2\

= £( l + 2 p a ) £ ( l + w p a ) .
Since th e subgrou p ha s orde r p a + 1 an d £ 6 ar e al l differen t fo r distinc t b modul o
p a + 1 , i t follow s tha t ther e exist s a n intege r b (uniquely determine d modul o p a + 1 )
such tha t
bz bz \
(
Using (1 2.34) , (12.35) an d

(12.36) mx))=my))e (?mz+-m^

we deriv e tha t
2 a
S(X,1>)= E X{f(ym9{y)) £ e p(d(2/)z + h(y)p- z).
a 1a
y(mod p ) z ( m o d p + )

Here th e innermos t su m vanishe s unles s h(y) = 0 (mo d p a ) i n whic h cas e i t equal s

paGp(y). Thi s complete s th e proof o f (1 2.30) . •

The Gaus s sum s G p(y) wer e compute d i n Chapte r 3 . I f p\ 2d(y) , w e hav e

(12.37) G p ( ! ,)=£p!,s(^))e,(_Se(^))2).

The formulas (1 2.24 ) an d (1 2.30 ) represent trul y th e final stage of computation .

The onl y term s whic h ar e no t determine d o n th e righ t sid e ar e th e root s o f th e
congruence h(y) = 0(mo d p a). However , i n practic e ther e ar e no t man y root s
(essentially a bounde d number ) s o on e get s th e estimat e |5(x , VOI ^ cq 1 /2 wit h c
depending mildl y o n th e coefficient s o f th e rationa l function s / , g.

EXERCISE 1 . Usin g Lemm a 1 2. 2 an d Lemm a 1 2. 3 together wit h (1 2.37 ) eval -

uate th e Kloosterma n su m

(12.38) S(m,n;g) =^ e (J
x(mod q)

for q = pP wit h (3 ^ 2 . Suppos e p \ 2mn. Sho w tha t S(m,n\q) vanishes , unles s

( f ) = (^ ) i n whic h cas e
/ I \ i /2£\
(12.39) S(m,n;q) = 2 ( -)q*Re e qe( — J
 12. C H A R A C T E R SUM S 32 3

where £ 2 = ran (mod q).

The Kloosterma n sum s t o modulu s q = p@ wit h / 3 ^ 2 were compute d firs t b y

H. Sali e [Sal] . H e als o compute d th e so-calle d Sali e sum s
J
/ ^ ^ x r^ x v
(12.40) T(m,n;q)= t
£ ^(-Je
f%\ fmx
(J rnx\
x ( m o d q)

where (- ) i s th e Jacobi-Legendr e symbol , includin g th e cas e o f prim e modulus .

One ca n d o i t fo r composit e modul i a s well .

LEMMA 1 2.4 . Suppose (q,2n) = 1 . Then

(12.41) T(m,n;q) = eqq$(^) ] T e(^) .

v2=mn(mod q)

PROOF. Conside r th e function F{u) = T(ra , nu 2; q) defined fo r u (mo d q). Th e

Fourier transfor m o f F(u) i s

u ( m o d q)

nxu2 — m? N
= E (fMv ) E « (
x(mod g ) u(mo dq )

I /n\ v-^ * fmx — 4nxv 2\

£ q
o \-a) Sc(mod<q ) ; )
by th e formul a (3.21 ) fo r quadrati c Gaus s sums . Notic e tha t th e Jacobi-Legendr e
symbol cancele d out , therefor e th e las t su m i s th e Ramanuja n su m

5(0, m - 4nv 2 ; q) = ^ dfi(q/d).

d\(4mn — v 2,q)

Hence b y Fourie r inversio n

m-\ E ^KfW(=)S> © E
f (mo d q ) d| q i>(mo
«(=) •
dq)
•L> 2 =4ran(mod d)

If (u , #) = 1 , this simplifie s t o

v2=?nn(mod g )

In particular , fo r u — 1 we ge t (1 2.41 ) . •

Suppose (q,2mn) = 1 . The n T(m,n;q) vanishe s unles s ther e exist s £ with

2
(12.42) £ = mn(mod q).
324 2 . CHARACTE R SUM S

Given £, al l th e solution s t o v 2 = run (mod q) ca n b e writte n explicitl y a s v =

(rf — ss)£, wher e r , s ru n ove r th e factorization s rs = q with (r , s) = 1. Henc e th e
formula (1 2.41 ) ca n b e writte n mor e explicitl y a s follows :

(12.43) T(m,n ]q)=eqqiQ £ c (^(j-j)).

(r,s) = l

The Kloosterma n sum s t o prim e modulu s canno t b e compute d i n elementar y

terms. Thi s shouldn' t b e surprisin g i n view of the result s (an d conjecture ) concern -
ing th e distributio n o f angle s o f Kloosterma n sum s (se e Sectio n 21 .2 , i n particula r
Theorem 21 . 7 an d th e Sato-Tat e Conjecture) .
12.4. Shor t characte r sums .
In thi s sectio n w e ar e dealin g wit h characte r sum s ove r a shor t interva l

(12-44) S X(N)= Yl *(n)>

M<n^M+N
where \ i s a non-principa l Dirichle t characte r o f modulu s q. B y Lemm a 1 2. 1 we
obtain
a
(12.45) \S X(N)\^2 Y. ~l\9x(")\

where

(12.46) g x(a)= £ x(*)e(y )

;r(mod q)

are th e classica l Gaus s sums . Le t x*(mo d q*) b e th e primitiv e characte r whic h

induces x- W e hav e

(12.47) g x(a)=gx.{l) £ d?(ald)n{q/dq*)

d\(a,q/q*)

by Lemm a 3.2 . Henc e

(12.48) \g x(a)\<cr((a,q/q*))V¥-

Inserting thi s boun d int o (1 2.45 ) w e deriv e

(12.49) \S x(N)\^3r(q/q*)V¥logq.

Since r(m) ^ 2y/m, thi s yield s

THEOREM 1 2.5 . For any non-principal character \ (mo d q) we have

(12.50) | J^ x(n)\<6y/Qlogq.
M<n^M+N

This inequalit y wa s prove d i n 1 91 8 independently b y G . Poly a [Pol ] an d I . M .

Vinogradov [V4 ] excep t fo r th e constan t 6 . B y th e Gran d Rieman n Hypothesi s fo r
Dirichlet //-function s on e ca n deriv e a slightl y stronge r estimat e
SX(N) <^ y/qloglogSq,
 12. CHARACTE R SUM S 325

however, nothin g bette r tha n th e Polya-Vinogrado v inequalit y i s know n i n gen -

eral excep t fo r th e constan t factor . Fo r recen t improvement s o f thi s constan t se e
A. Hildebran d [Hi2] .
Next w e giv e anothe r derivatio n o f th e Polya-Vinogrado v inequalit y withou t
completing S X(N). W e hav e

where f(x) = min( x - M , 1 , M + N + 1 - x) i f M < x ^ M + N + 1 , and /(a: ) = 0 ,

otherwise. Thi s redundan t cut-of f functio n will hel p t o separat e th e variable s i n
the forthcomin g expressions . Fo r an y intege r a we hav e

Sx(N) = J2x(n + a)f(n + a).

n

Hence
ASX(N)= J2 E X(n + a)f(n + a).
0<a^A M-A<n^M+N
Now separat e th e variable s i n f(n + a) b y th e Fourie r transfor m
oo
f(t)e((n + a)t)dt
/ -oo

getting
OO

|/(t)|B(t)d*
/ -oo
where B(t) i s th e su m i n tw o independen t variable s

B
W= I E I E X( n + a)e(ni )
0<a^A M-A<n^M+i V

By Cauchy' s inequalit y
1
1
cc(mod q) n

^G + J)EE | E X(ni+x) x(n2+x)

n\ n 2 x(mo d g )

Suppose x(mo d q) i s primitive . The n

(12.51) ^2 X(ni + ^)x(^ 2 + a ) = 5(0 , ni - n 2 ; g)

x(mod qr )

is th e Ramanuja n sum . Henc e w e derive tha t

m2^(j + 1 . 1 \/A + N
] ) (^ + l)(A + N)R(q)
q)\ q
where
S
(12.52) R(q)= Y, \ (0>V><1 )
y(mod q)
32 6 12. C H A R A C T E R SUM S

Assuming tha t N ^ | (a s w e ca n b y virtu e o f periodicity) , an d choosin g A = N

the abov e boun d simplifie s t o 6R(q). Henc e
oo
\f(t)\dt.
/ -oo

By |/(0 I — ( T T 0 - 2 | s m ( 7 r 0 s m ( 7 r ^ ) l w e m
^ e r tha t th e integra l i s bounde d b y
3 log 2iV, an d conclud e tha t
|5x(JV)K8fl(g)hogg.
Note tha t i?(g ) = 2 W(«V(<7) ^ T
(?)9- therefor e
(12.53) |5 x (JV)|<8(T((7) 9 )ilogg.

For q prime this estimate give s the Polya-Vinogradov inequalit y whil e for composit e
q it i s only slightl y weaker .
REMARK. Th e facto r log g emerge d i n th e proces s o f separatio n o f variables .
Since thi s facto r canno t b e entirel y ignored , w e mak e a poin t tha t i n genera l th e
losses from separatio n o f variables ar e not exclusivel y o f technical nature , bu t occu r
inevitably fo r arithmeti c reason s a s well. Thi s comment sound s eve n stronger i n th e
context o f bilinea r forms , wher e factorizatio n b y superficia l separatio n o f variable s
at n o cos t woul d b e fata l (thin k o f Corollar y 7.3) .

The Polya-Vinogrado v inequalit y (1 2.50 ) i s trivial fo r characte r sum s S X(N) o f

length N < C #2, whil e th e tru e boun d i s expecte d t o b e
e
(12.54) S x(N)<&N*q ,
which i s non-trivia l i f N ^ > q 3e. I n a serie s o f paper s D . Burges s [Burl] , [Bur2 ]
established severa l result s fo r relativel y shor t characte r sums , on e o f the m bein g
THEOREM 1 2. 6 (D . BURGESS) . Let x be a primitive character of conductor
q > 1 . Then
1
(12.55) 5 x(A^)<7V -^S*+£
for r = 2 , 3, and for any r ^ 1 if q is cubefree, the implied constant depending only
on e and r.
The hypothesi s tha t q i s cubefre e ca n b e remove d a t th e cos t o f weakenin g
the result . Le t q = k£, wher e £ is th e maxima l cubefre e factor . B y factorin g th e
character x(mo d q) accordingl y an d splittin g th e summatio n int o classe s modul o k
one derive s fro m Theore m 1 2.6 ,

(12.56) S X(N) < Ni-rkh^+e

(if £ = 1 , then fc = q and (1 2.56 ) follow s b y th e periodicit y o f x) .
Burgess's boun d (1 2.55 ) fo r q cubefree i s non-trivia l i f i V ^> g 4 + 4?+e? an d b y
taking r sufficientl y large , on e derive s

(12.57) S X(N) < N 1 -6, ifN^ q + +vrs

for an y 0 < 5 ^ | , th e implie d constan t dependin g onl y o n 6.
 12. C H A R A C T E R SUM S 32 7

We shal l giv e a proo f o f a slightl y stronge r estimat e

l
(12.58) \S X(N)\ ^ cN -rpri$(\ogp)r

but onl y fo r character s x ( m ° d p) t o prim e modulus . Her e c is an absolut e constan t

(c = 3 0 i s fine). Ou r argument s g o b y inductio n wit h respec t t o N. Firs t notic e
that th e assertio n (1 2.58 ) i s eithe r trivia l o r i t follow s fro m (1 2.50) , unles s
r
(12.59) c v\ + h log p ^ T V ^ p i + i i Q gp

which conditio n w e hencefort h assume . Applyin g a shif t n \-> n + h wit h 1 ^ h ^

H < N w e obtai n

(12.60) S X{N)= Yl x(n + h) + 26E(H)

M<m^.M+N

where |0 | < 1 and

(12.61) E(H) = ctf-rp&ilogp)*

by th e inductio n hypothesi s (1 2.58 ) fo r th e tw o characte r sum s o f lengt h h whic h

do no t overla p wit h th e origina l segment . Le t H = AB, wher e A,B ar e positiv e
integers. W e us e th e shift s o f typ e

(12.62) h = ab wit h 1 ^ a < A , 1 O< B.

Averaging (1 2.60 ) ove r a , b we get

s
* w = jj E* E *( n afc
+ )+ 20E w
l^a^A M<n^M+N

where \0\ ^ 1 . B y x( n + a fy = x( a )x(& n + &) > where a i s the multiplicativ e invers e

of a modulo p (her e we employ tw o properties o f characters, th e periodicity an d th e
multiplicativity) w e deduc e tha t

(12.63) \S X(N)\ < H^V + 2£(# )

where

^ = E " H E ^+ 6 )|
z(modp)
1 ^6^ £
and v(x) i s th e numbe r o f representation s o f x a s an ( mo d p ) wit h 1 ^ a ^ A an d
M < n ^ M -\- N. W e shal l estimat e V withou t usin g th e inductio n hypothesis , s o
the implie d constan t wil l b e independen t o f c .
As a and n vary over short segment s relative to q, many residue classes x(mod q)
are no t represente d b y an; i n othe r words , v(x) i s ofte n zero . Moreover , i/(x) i s
essentially bounded , bu t i t i s impossibl e t o analyz e it s rando m change s o f size .
Therefore w e relax i/(x) i n V b y applyin g th e Holde r inequalit y
328 12. C H A R A C T E R SUM S

where

v1= Y, "(*) • ^ = E ^
x ( m o d p) a;(mo d p)

w
= E | E x( * + &)f.
x ( m o d p) l^b^B

We could hav e restricted th e outer summatio n i n W t o the residu e classe s x(mod p)

for whic h u(x) ^ 0 , bu t w e are no t abl e t o tak e advantag e o f such a condition . Th e
extension t o th e complet e su m i s massive, nevertheles s i t i s nominal relativ e t o th e
length o f the characte r su m raise d t o power 2r. Thi s explain s ou r choic e of Holder' s
exponents.
Clearly, w e hav e V\ — AN. W e shal l sho w tha t V 2 i s als o o f thi s magnitud e
(essentially).
LEMMA 1 2.7 . We have
(12.64) V 2 < SANiANp- 1 + log3A) .

P R O O F . V 2 i s th e numbe r o f quadruple s {ai,a 2,ni,n2) wit h 1 ^ a\,a 2 < A

and M < ni,n 2 ^ M + N suc h tha t a\n 2 = a2ni(modp) . Fi x a\,a 2 an d pu t
a\n2 — a 2ri\ — kp. W e hav e
AN
K - (fl l — ^ 2 )
P V
and (ai , a2)|/c. Give n ai ,G&2 an d k as above we find that th e number o f pairs (ni,n 2)
satisfying th e equatio n a\n 2 — a 2ri\ = kp i s bounde d b y 2N(ai,a 2)/max(ai,a2).
Therefore
(01,02) / 2AN
iax(ai,a ) \(ai,a );
i ^ 2 a m a x ( f l l ' G 2 )2l ( a i ' f l 22) p 7
Hence (1 2.64 ) follow s afte r simpl e estimates . •

LEMMA 1 2.8 . We have

(12.65) W ^ (2rB) rp + 2rB 2r
p^.

This estimat e fo r W lie s a t th e hear t o f Burgess ' method . Assumin g (1 2.65 )

we complet e th e proo f o f (1 2.58 ) a s follows . W e choos e B = [rp^] givin g
2r
W < (2r) p*.

Then w e choos e A = [N/9rp^\. Not e tha t A ^ 1 by th e lef t sid e o f (1 2.59) , an d

AiV ^ N 2/9rp^ ^ p(logp) 2 b y th e righ t sid e o f (1 2.59) . Therefor e Lemm a 1 2. 7
gives
V2^AN(4\ogp)2.
Recalling tha t V\ — AN w e deriv e
V ^ 2r(AN) 1 ~^(4p* logp) ^ ^ N 2'^^2^ logp)- .
 12. C H A R A C T E R SUM S 329

Next not e tha t H = AB < N/9 an d H > JV/1 0 . Therefor e w e deriv e b y (1 2.61 )
and (1 2.63 ) tha t
\SX(N)\ < (lO+^cJ^-ip^aogpji .
Taking c = 3 0 we obtai n (1 2.58) .
It remain s t o prov e Lemm a 1 2.8 , fo r whic h w e ca n assum e B < p. A s i n
Burgess' origina l proo f w e appeal t o th e Rieman n Hypothesi s fo r curve s ove r finite
fields. B y Corollar y 1 1 .24 , fo r an y \ modul o p whic h i s non-trivial , w e hav e

(12.66) | ] T X((x + h)--(x + b r))x{{x + b r+1 ).--(x + b 2r)) ^ 2rp 2

cc(mod p)

if one o f th e classe s b v (mod _p) , v — 1, . .. , 2r i s differen t fro m an y othe r one .

We hav e

W= ^2---Yl Yl X({x + b 1 )--(x + b r))x{{x + b r+1 )".{x + b 2r)).

l^bi,... for^B x(mod p)

The complete sum satisfies (1 2.66 ) except possibly for the (&i,.. . , b2r) whic h can b e
arranged i n r equa l pairs. Thu s the number o f exceptions i s bounded b y r ( 2 r )B r ^
(2rB)r. Henc e (1 2.65 ) follows .
REMARK. Choosin g th e involve d parameter s i n th e abov e proo f mor e econom -
ically on e ca n ge t (1 2.58 ) wit h th e facto r c(logp) ^ replace d b y c'(\ogp)^ fo r som e
absolute constan t d ^ 1 (fo r othe r comment s se e [Fri]) .

EXERCISE 2 . Prov e alon g th e abov e line s th e estimat e (1 2.55 ) wit h r = 2 fo r

any modulu s q > 1 .
The Burges s bound (1 2.55 ) with r = 2 implies a subconvexity boun d fo r Dirich -
let L- functions i n conducto r aspec t (compar e Sectio n 5.9) .
T H E O R E M 1 2.9 . Let x be a primitive Dirichlet character modulo q ^ 2, and
s = | + it. We have
+£
(12.67) L(s,x)<£\s\q^

for any e > 0, the implied constant depending only on e.

PROOF. B y Theore m 5. 3 w e hav e

where V s(y) i s a rapidl y decayin g functio n give n b y (5.1 3) . Precisel y i t satisfie s

Hence
33 0 12. C H A R A C T E R SUM S

By (1 2.55 ) wit h r = 2 we hav e

13
s
S(x) = 22x(n) <min(<7,x2gi6+ ).

Then b y partia l summatio n w e ge t

\L(s,X)\ ^ J" \s(x)(x-V a(x/y/q)) dx

and th e resul t follow s b y applyin g th e relevan t estimates .

•
12.5. Ver y shor t characte r sum s t o trul y composit e modulus .
It i s natura l tha t Burgess' s metho d fail s t o produc e a non-trivia l estimat e fo r
the characte r su m S X(N) o f length N < C q*, wher e q is the conducto r o f \. O n th e
other hand , w e do have effective treatment s o f exponential sum s of Weyl's type (se e
Chapter 8 ) which ar e very short relativ e t o the siz e of the amplitud e function . Her e
we appl y th e differencin g proces s t o reduc e th e amplitud e functio n severa l time s
until i t i s small enoug h (howeve r unbounded ) s o that a non-trivia l estimat e fo r th e
resulting su m i n th e fina l ste p i s availabl e fro m specia l sources . I n thi s respec t th e
case o f th e characte r su m S X(N) i s quit e differen t becaus e on e normall y canno t
reduce th e conducto r b y shifting th e argument ; absolutel y not , i f the conducto r i s a
prime number . Ther e is , however, a possibility t o imitat e th e Weyl-Va n de r Corpu t
differencing idea s fo r characte r sum s o f modulu s whic h facto r int o relativel y smal l
numbers; w e sa y tha t suc h a modulu s i s trul y composite . A n importan t estimat e
for characte r sum s o f specia l composit e modul i wa s give n b y S . W. Graha m an d J .
Ringrose [GRi ] (se e als o [H-B ] an d [FI3]) . I n thi s sectio n w e presen t ou r versio n
of th e Graham-Ringros e magnificen t result .
We shal l procee d b y inductio n wit h respec t t o th e numbe r o f factor s (no t nec -
essarily primes ) i n th e modulu s o f th e character . Fo r thi s reason , i n orde r t o mee t
the inductio n hypothesis , w e trea t mor e genera l sum s tha n necessar y fo r th e fina l
result.
Let £(mo d k) b e a multiplicativ e character . W e conside r

(12.68) St(N)= ] T £(/(n) )

M<m^M+N

where f(x) i s a rationa l functio n wit h al l zero s an d pole s bein g integers . W e writ e
771

(12.69) f(x) = Y[(x-a 1 /)d"

z/=l

where d v ar e non-zer o integer s an d a v ar e an y integers . W e do no t assum e tha t ai ,

. . . , CLfYi a r e distinct , s o th e representatio n (1 2.69 ) doe s no t defin e th e exponent s
d i , . . . , dm uniquely . Recal l that fo r a given representation / = fi/fo wit h / i , // Q G
Z[x] w e define d £(/(#) ) — £,(fi(x))£(fo(x))i therefor e th e su m (1 2.68 ) i s actuall y
restricted b y
(12.70) ( ( n - a i ) - . - ( n - a m),fc) = l.
 12. C H A R A C T E R SUM S 331

THEOREM 1 2.1 0 . Let £ = xi'"Xr-iX> where \t ( m ° d qi) for 1 ^ £ < r

are any characters and x (mo d q) is a primitive character of conductor q > 1 , q
squarefree. Let f be given by (1 2.69) with

(12.71) ( d i , . . . ,dm,h ) = l
J
where h is the order of x- Put NQ — max{gi,... , qr-i, q }q*, and

(12.72) A=niI(S-S) -

Then for any N ^ NQ we have

1
(12.73) IS^N^ ^4N((Aq 1 ---qr.l,q)q- r(qy2Tm(q)2n2~r•

The powe r o f proo f stem s fro m estimatio n o f th e complet e characte r su m

(12.74) S( X)= £ x(/(*)) -

:r(mod q)

If q = p i s prime w e are dealin g wit h a sum ove r th e field ¥ p. I n thi s cas e w e deriv e
from [Sch ] th e followin g
PROPOSITION 1 2.1 1 . Let x (mo d p) be a non-principal character of order h
which satisfies (1 2.71 ). Then

(12.75) | S ( X)|<(m-l)(A,p)M-

P R O O F . Firs t not e tha t p ^ 2 becaus e ther e i s onl y th e trivia l characte r o f

modulus two . Fo r m = 1 we hav e f(x) = (x — a\) dl wit h (di , ft)=0 an d
dl
S(x)= E x 0*) = o,
a:(mod p)

so (1 2.75 ) hold s (i n thi s cas e A - 1 ) . Le t m ^ 2. I f A = 0(mod p) , the n (1 2.75 )

is trivial . No w suppos e al l th e number s a i , . . . , am ar e distinc t modul o p . I f / i s
a polynomia l (i.e. , al l th e exponent s c?i,.. . , dm ar e positive) , the n th e polynomia l
F(X, Y) — Yh — f(X) i s absolutel y irreducibl e (se e Lemm a 2C o f [Sch] ) an d ou r
assertion follow s fro m Theore m 1 1 .23 . I f f{x) — fi(x)/fo(x) wit h fi(x),fo(x) £
Z[x] having all distinct zeros , then we consider the polynomial g(x) = fi{x)fo(x) p~2
in place of the rationa l functio n f(x). Not e that x(f( x)) — x(d( x)) a n d (p~2 , ft)=
1, becaus e h\(p — 1). Therefor e th e conditio n (1 2.71 ) hold s fo r g(x), s o (1 2.75 ) i s
established. •

Proposition 1 2.1 1 extends b y multiplicativit y a s follow s

COROLLARY 1 2.1 2 . Let x (mo d q) be a primitive character of squarefree con-
ductor q and order ftwhich satisfies (1 2.71 ). Then

(12.76) |S( X )K(m-ir^(A,g)M

where uo{q) is the number of prime divisors of q.
332 12. CHARACTE R SUM S

PROOF. I f q = p\ • • -p t, the n x — Xi ' * * Xt> wher e Xi > • • • >Xt ar e primitiv e

characters o f orde r / i 1 ? . . . , ^ respectivel y whic h ar e divisor s o f h. Therefor e th e
condition (1 2.71 ) i s satisfied fo r eac h character Xs(mo d p s). Multiplyin g (1 2.75 ) fo r
X i , . . . ,X t w e ge t (1 2.76) . •

Now w e ar e read y t o prov e tha t fo r N ^ TVQ ,

1
(12.77) \S Xl...Xr.lX(N)\ < c r(m)((Aqi • • • 9 r_ 1} q)q~ )2'rN
by inductio n wit h respec t t o r . Th e factor s c r(m) ^ 1 will b e determine d i n th e
induction process ; the y depen d o n g , but i t i s not necessar y t o displa y thi s becaus e
q wil l no t chang e throughou t th e process . Moreover , w e d o no t chang e th e se t
of exponent s { d i , . . . ,<i m} i n th e representatio n (1 2.69) , althoug h ever y inductio n
step wil l doubl e th e numbe r o f zero s an d pole s o f th e resultin g rationa l functions .
an
For r = 1 we hav e £ = x d

W =E x(f(n)) = ^s(x) + eq
where 5(x ) i s th e complet e su m (1 2.74 ) an d |0 | ^ 1 . Henc e w e obtai n b y (1 2.76 )
that (becaus e q? ^ N)

w (,9/) (A,g)^
(12.78) | S x(A0|<m (^^)T./V.

Now le t r ^ 2 . W e appl y th e shif t n »-> • n + /M7 1 and averag e ove r 1 ^ h ^ H

getting
n
"%W = £ E *i(/( )) E l(/( n + ^i)) + 2 ^9 l
M<n^M+N l^h^H

where £ = X 2 ' " X r - i X a n d |0 | ^ 1 - Her e th e erro r ter m i s obtaine d b y trivia l

estimation o f the shor t characte r sum s whic h d o not overla p wit h th e origina l sum .
Hence \S^{N)\ ^ T + 2Hq 1 , wher e

iE | E «/( n + ^i))
M<n^M+N l^h^H
By Cauchy' s inequalit y
T 2
<^ E *| E £(/( " + Mi)//(n + Mi))
l^InM^H M<n^M+N
Here th e inne r su m i s o f typ e (1 2.68 ) fo r th e reduce d characte r £ and th e quotien t
function /(# ) = / ( # + h\qi)/f(x + / ^ l ) - Thi s ha s 2 m zero s an d poles . Fo r f(x)
written a s (1 2.69 ) w e introduc e th e polynomia l
A
0 ) = 1 1 I I ^+ a
^ ~ a ^ ) > degA(x ) =m(m-1 ) ,

so A = A(0) . The n th e correspondin g polynomia l fo r f(x) i s

A(x) = A 2(x)A2(x + y)(x + y) 2rn
 12. C H A R A C T E R SUM S 33 3

where y = (hi - h 2)qi- Henc e A = A(0 ) = A 2A2(y)y2m an d

(Aq2 - - • q r-i,q) < (A«?i • • • q r.1 ,q)(yA(y),q/(q, qi)).

By th e induction hypothesi s w e ge t

| £ l ( / ( " ) ) | < c r _ 1 (2m) 5 (ft 1 - h 2)((Aqi • • • qr_uq)q-1f~' >

n

where g(h) = (hqiA(hqi)^q/(q Jqi))2 . W e choose H = q and estimate a s follow s

ft
jp X ) * 9(hi-h 2) = - J2 ( 9iA(ftgi),g/( 9 ,gi)) 2 l " r
l ^ / u , / i 2 ^ # ft (mo d <? )

2/(mod g )

5
i^ (J/A(J/), 9))

2/(mod 9 )

For th e inner su m we have

- J2 (j/A(ft),g)<^;|{y(modd);yA(j/)=0(modd)} |
y(mod q) d\q

< ] T ( 1 + d e g A )w ( d ) = ( 2 + d e g A ) " ^ .

We hav e deg A = m(m - 1 ) , so 2 + degA ^ 2m 2 and (2 + deg A )^ ^ r(g)r^(g) .

Gathering th e above estimate s w e obtain

\St(N)\ s C cl1(2m)(r(q)Tl(q)(Aq1 • • • qr_1,q)/q)2'"N + 2q Ql.

Since qiq% < A/" , this yield s (1 2.77 ) fo r r ^ 2 , provided

C-(m) > ci 1 (2m)(r( g )T^(g)) 2 "" + 2 .

By (1 2.78 ) w e get (1 2.77 ) fo r r = 1 , provided ci(ra ) ^ r m (g). On e can check tha t
the number s
cr(m) = A(T(qY 2Tm{q)2r)2~r
satisfy th e above (recurrence ) inequalit y (us e r2rn(q) = r(q)r rn(q)) givin g (1 2.73) .
Our mai n interes t i n Theorem 1 2.1 0 is for / ( x) = x (s o m = 1 , A = 1 ) in which
case i t yield s
THEOREM 1 2.1 3 . Let \i (mo d qi) for 1 ^ £ < r be any characters and
let x (mo d q) be a primitive character of conductor q > 1 , q squarefree with
(q^qi ' * *^r-i) = 1 - Then for N ^ No = max{qi,... ,q r-i,q±}q± we have

T((?r 2
(12.79) Yl Xi'--Xr-ix(rc ) <4iV| ~
M<m<^M+N
33 4 12. C H A R A C T E R SUM S

COROLLARY 1 2.1 4 . Let x be a primitive character of conductor q > 1 , q

squarefree. Suppose all prime factors of q are $J NQ . Then
r/2
1 /r2r
(12.80) £ x(n)k4JM*) V
M<n^M+N

where r is any integer with N r ^ q 3.

PROOF. Le t p b e th e larges t prim e facto r o f q. W e ca n writ e q = qi - — q r wit h
qi < pq* fo r 1 ^ £ < r (som e o f th e factor s ca n b e one) . W e hav e qi < TV's .
Moreover, on e o f th e factors , sa y q r, satisfie s q r ^ qr. W e ca n als o requir e tha t
uj{qr) ^ cj(q)/r. T o mee t thes e condition s defin e q r a s th e smalles t produc t o f
the larges t prim e factor s o f q suc h tha t q r ^ qr. The n defin e qg successively fo r
1 ^ £ < r a s th e larges t produc t o f th e larges t prim e factor s o f q/q r • • • qi+\ suc h
that qi ^ pqr (i f q = q v • • • #g+i, the n pu t ^ = • • • = q\ = 1 ) . Clearly , ever y prim e
factor o f q is used onc e i n thes e r steps , s o q = q\ • • • q r. Accordingl y x — X i ' '' Xr,
where Xi a r e character s modul o qt an d Xr i s primitive . Therefor e (1 2.80 ) follow s
from (1 2.79) . •

The diviso r functio n i n th e estimate s (1 2.79) , (1 2.80 ) i s a nuisance . I n genera l

we hav e r(q) = (a x + 1 ) • • • (a n + 1 ) ^ 2 a i + - " + a - = 2 QM i f q = p? 1 • • - p ^, wher e
H(g) i s th e numbe r o f prim e factor s o f q counte d wit h multiplicity . Thi s ca n b e
quite large . W e estimat e Q(q) a s follows ,

CZ

^-—' lnc
logdr r\ o g 2l n a :7.f^Vlog
^ l ^-^ ^-—' V lno *d/7lo l nga 2/ /
7. log lnczr log
? 2
z
d\q d\q d<z °

where z > 1 and c is a n absolut e constant . Choosin g z = logq fo r q ^ 3 we ge t

(1181 ) ( . ( . K ^ f ! . <
log log q V lo g log q
Hence
1 +C//loglog(? log(? loglog(?
(12.82) T(Q ) ^ 2( ) / .

EXERCISE 3 . Prov e tha t r(q) ^ g V log log 3? for al l q ^ 1

Using thi s crud e estimat e w e deriv e fro m (1 2.80 )

COROLLARY 1 2.1 5 . Let x be a primitive character of conductor q ^ 3,q square-
free. Let p be the largest prime factor of q. Then, for

(12.83)T V > q £{q) + p9 with e(q) = 4(loglogg)" 2

we have

(12.84) Yl X(n) < iVexp(-V /lo^)

M<n<^M+N

where the implied constant is absolute.

 12. CHARACTE R SUM S 335

EXERCISE 4 . Le t th e condition s b e a s i n Corollar y 1 2.1 5 . Sho w tha t

(12.85) L(l , x) < e(q) log q + 9 logp + b
where b is a n absolut e constant .
12.6. Character s t o powerfu l modulus .
The trul y composit e modul i (not e tha t th e number s havin g n o larg e prim e
divisors, th e so-calle d "smoot h numbers" , o r "entier s friables " i n French , ar e trul y
composite) appea r i n practic e no t a s ofte n a s th e prim e moduli . Anothe r extrem e
are th e modul i whic h ar e hig h power s o f a fixed prime . Thes e ar e example s o f
powerful numbers . W e sa y tha t q is powerfu l i f it s kerne l

(12.86) * = n ^
p\q

is smal l relativ e t o q in th e logarithmi c scale .

Let x b e a primitiv e characte r o f conducto r q which i s a powerfu l number . I n
this cas e a larg e par t o f characte r value s ca n b e describe d a s a n exponentia l o f a
polynomial, s o the characte r su m S X(N) ca n b e reduced t o Wey l sums. T o this en d
consider th e followin g polynomial :

(12.87) L(X) = X~Y + -..±^- .

One ca n sho w tha t (se e [Ga2] , p . 1 92 )

a b
(12.88) L(x + y + xy) = L(x) + L(y) + ] T * c(a,b)x y
d<a+b^2d

with som e rationa l coefficient s e(a , b) such tha t c(a, b) G Z[a, 6].
Suppose 4 { q (fo r a technica l reason) . Le t d be sufficientl y larg e s o tha t q 2\kd.
Let D b e th e produc t o f al l m ^ d , (m,q) = 1 . Clearl y DL(kx) G Z[x], an d i t
follows fro m (1 2.88 ) tha t
DL(kx + ky + k 2xy) = DL(kx) + DL(ky) (mod q).
Hence th e functio n £( 1 + kx) = e(DL(kx)/q) i s a characte r o f th e multiplicativ e
group o f residu e classe s mo d q congruen t t o 1 mod k. I t i s eas y t o se e tha t th e
order o f £ is q/k, whic h i s the orde r o f th e group . Therefor e ther e exist s a n intege r
B suc h tha t
(12.89) x ( l + kx) = e(BDL(kx)/q).
Since q is the conducto r o f x i t follow s tha t B i s prime t o q/k. O n th e othe r hand ,
b is determine d onl y mo d q/k s o we ca n assum e tha t B i s prim e t o q.
By (1 2.89 ) w e arrang e th e characte r su m (1 2.44 ) a s follows :

SX(N)= J2 X(a)j2x(l + kan)= £ X («) £ > ( F ( a n ) / 9 )

0<a</c n 0<a<k n

where n run s ove r th e interva l (M — a)k_1 < n ^ (M- f N — a)/c_1, ad = l(mo d g) ,

and F(x) = BDL(kx) i s a polynomia l o f degre e d wit h intege r coefficients . Th e
33 6 12. C H A R A C T E R SUM S

last su m i s o f Weyl' s typ e whic h ca n b e estimate d b y Vinogradov' s method . Thi s

has bee n don e i n detai l i n [1 2] , gettin g
THEOREM 1 2.1 6 . Let x be a primitive character of conductor q, 2 \ q. Let k be
the product of prime factors of q. Then, for k 1 00 < N < N ; ^ 2N we have
1 6
(12-90) | Yl * ( ^ jN '
N<n^N'

where

7 = exp(200rlog 2 (1 200r)), S = (1 800r)- 2 (log3600r)~ 1 , r = | ^ | .

REMARKS. Th e first resul t o f thi s typ e fo r q = p £ wa s establishe d b y A . G .

Postnikov [Pos] . Ou r generalizatio n o f th e Postniko v formul a (1 2.89 ) fo r th e pow -
erful modulu s i s base d o n th e formul a (1 2.88 ) whic h i s du e t o P . X . Gallaghe r
[Ga2]. Ther e ar e application s o f th e result s t o boundin g th e L(s,x) , t o widenin g
the zero-fre e region , an d t o estimatin g th e leas t prim e p = a (mod q) (se e [Ga2 ]
and [1 2]) .
 http://dx.doi.org/10.1090/coll/053/14

CHAPTER 1 3

SUMS OVE R P R I M E S

13.1. Genera l principles .

Given / : N—> C w e ar e intereste d i n estimatio n o f th e su m

(13.1) V = J2fiP)
v
where p run s ove r prim e numbers , o r o f th e closel y relate d su m

(13.2) S = $>(n)/(n )
n

where A(n ) i s the vo n Mangold t function . Not e tha t S i s obtained fro m V fo r / ( n )

replaced b y f(n) logn u p t o a smal l contributio n o f term s n — p v wit h v ^ 2 . O f
course, th e problem s o f estimatin g S an d V ar e equivalen t b y partia l summation ,
but i t i s ofte n simple r i n practic e t o dea l wit h S.
If / i s a nice , smoot h functio n wit h compac t suppor t o n R + on e ca n expres s S
by the zero s of ((s) b y integrating —f(s)( ,(s)/((s)1 wher e / i s the Melli n transfor m
of / . On e get s

(13-3) S = /(l) - J2 /(~2n) - J2 fiP)-

n>0 p

However, eve n assumin g th e Rieman n Hypothesis , thi s expressio n (th e so-calle d

explicit formula , se e Exercis e 4 in Chapte r 5 ) i s no t alway s useful . Fo r example , i f
/ : R+— > C i s a n oscillator y function , suc h a s

(13.4,) f{x) = e{ax)g(x)

(13.5) f[x) = e{a^c)g{x)

with a e R* an d g(x) a bum p function , the n th e Melli n transfor m f(s) i s sprea d

over a large range of zeros so that th e expansion (1 3.3 ) fail s to produce goo d results .
As wit h an y summatio n typ e formul a (se e Chapte r 4 ) th e decisio n t o appl y i t o r
not ca n b e reasonabl y guide d accordin g t o whethe r th e origina l su m ha s mor e
terms ove r prime s tha n th e resultin g su m ove r th e zeros . Recently , equippe d wit h
powerful computers , researcher s d o no t hesitat e t o downloa d fro m th e internet 1
a lis t o f numerica l value s o f th e zeros , an d ar e no t afrai d o f usin g the m t o tes t
sums ove r prime s indirectl y vi a the explici t formula . Prime s an d zer o ar e n o longe r
different beasts , i t i s a matte r o f choic e t o wor k wit h on e o r th e othe r sets .

See Odlyzko's site : http://www.dtc.umn.edu/~odlyzko/zeta_tables /

337
33 8 13. SUM S OVE R PRIME S

Many interestin g arithmeti c function s / : N —> C ar e no t eve n define d o n th e

real numbers , i n whic h cas e th e explici t formul a (1 3.3 ) i s no t applicabl e (true , on e
can tak e smoot h extensio n o f / t o R + , bu t i t doe s no t wor k efficientl y i f th e se t o f
values f(n) fo r n G N is erratic). I n such a case we prefer t o us e a sequence notatio n
A = (a n) i n plac e o f th e functio n f(n) = a n. Sinc e A i s infinite , w e conside r th e
finite sum s

(13.6) S(A,x) = ^A(n)a n

n^.x

in plac e o f (1 3.2 ) wit h th e ai m o f estimatin g thes e fo r al l sufficientl y larg e x.

It i s not difficul t t o predic t th e asymptoti c behavio r o f S(A, x) b y appealin g t o
a randomnes s o f th e Mobiu s function :
MOBIUS RANDOMNES S LAW . The Mobius function n(m) changes sign ran-
domly so that for any "reasonable" sequence of complex numbers A — (a m ) the
twisted sum

(13.7) M(A,x) = ] P / i(m)am

m^x

is relatively small due to the cancellation of its terms.

Of cours e a reasonabl e sequenc e mean s chose n wit h n o bias . I n practic e i t i s

often th e case , excep t whe n playin g wit h siev e weight s whic h d o conspir e wit h th e
Mobius function .
To appl y thi s la w t o prim e number s w e write A = \i * L, o r

(13.8) A(n ) = Y, ^ d) lo S \ = ~ E M< 0 log d,

d\n d\n

and inser t th e latte r int o (1 3.6 ) gettin g

(13.9) S(A,x) = -^2fi(d)(logd)A d(x),

d

where
a
(13.10) A d(x)= Yl -
n=0(mod d)

The quantitie s Ad(x), s o also A d(x) logd , ar e "reasonable " i f the origina l arithmeti c
terms a n ar e themselve s "reasonable" . Therefore , accordin g t o th e principle , w e
believe tha t th e contributio n o f larg e d' s i n (1 3.9 ) i s negligible , s o 5(^4 , x) i s wel l
approximated b y

(13.11) S(A,D,x) = - J ] ) fjL(d)(logd)A d(x)

d^D

for som e D whic h i s relativel y small . No w fo r d ^ D th e quantitie s A d(x) ma y

satisfy (du e t o larg e averaging ) a simple , ye t stron g approximat e formul a
(13.12) Ad(x)=g(d)A(x) + r d(x)
 13. SUM S O V E R P R I M E S 339

where g(d) i s a nic e multiplicativ e function , A(x) = A\{x), an d rd(x) i s a smal l

error term . Ignorin g th e erro r ter m w e arriv e a t S(A, J9 , x) ~ H(D)A(x), wher e

ff(D) = - ] T > ( d ) ( l o g d ) 5 ( d ) .

Usually g is regularly distribute d ove r prime s s o there i s a limit o f H(D) a s D tend s

to oo , sa y H(D) ~ H(oo) — H wit h

(13.13) H = ->T>(d)(logd)0(d ) = I I ( 1 ~5(P))( 1 " ^ ) _ 1 -

Therefore, th e abov e transformation s lea d u s t o th e followin g asymptoti c formul a

(13.14) S(A, x) ~ HA(x), a s x - • oo .

Of course , provin g thi s rigorousl y i s a differen t matter . Wha t need s t o b e said ,

however, i s tha t th e heuristi c formul a (1 3.1 4 ) ha s neve r faile d t o hol d i n ever y
natural cas e i n whic h S(A, x) i s evaluated rigorousl y b y specia l means .
In 1 934 , I . M.Vinogrado v (se e [V5] , [V6] ) gav e a ver y goo d estimat e fo r th e
exponential su m

(13.15) V(a\x) = ^2e(ap)

p^x

which was needed fo r solvin g the Goldbac h proble m wit h three primes (se e Chapte r
19). I n th e absenc e o f th e Rieman n Hypothesis , thi s wa s a surprisin g accomplish -
ment. Hi s metho d applie s t o th e su m S(A,x) fo r a grea t variet y o f oscillatin g se -
quences A = (a n) whic h ar e no t multiplicative , suc h a s a n = e(an), a n = e(ay / n),
an — \( n + 1 ) ) wher e x i s a multiplicativ e non-principa l character .
Here i s how on e ca n explai n th e principle s o f Vinogradov' s metho d i n abstrac t
settings. On e start s b y repeate d application s o f th e exclusion-inclusio n procedur e
to arrang e S(A,x) int o sum s o f two type s

(13.16) Si = 2 J z J a dadn,
d^.D dn^x

(13.17) S 2
mn^x
M<m^N

where a^ , /3 m ,7 n ar e independen t rea l number s (essentiall y bounded ) an d D, M , N

are suitabl e parameter s (neithe r smal l no r larg e relativ e t o x). Th e su m o f type S\
can b e writte n a s

(13.18) S! = J2 a dAd(x)

where Ad(x) i s given b y (1 3.1 0) . Her e th e term s a n o f Ad{x) ar e cleare d o f compli -

cated coefficient s excep t fo r th e conditio n n = 0 (mo d d) whic h i s eas y t o handl e
for smal l d. Assumin g som e regularit y i n th e variatio n o f th e argumen t o f a n on e
shows tha t Ad(x) i s smal l fo r ever y d ^ D du e t o th e cancellatio n o f terms .
340 13. SUM S OVE R PRIME S

The su m o f type S 2 is considered a s a bilinear for m i n the coefficient s /3 m , 7n

(see Chapte r 7 for mor e genera l considerations) . Fo r estimatio n i t is convenient to
split S 2 int o bilinea r form s ove r dyadi c segment s

(13.19) B{M,N)= Y, P™ J2 W™-

M<m^2M N<n^2N

Applying Cauchy' s inequalit y on e remove s th e coefficient s /3 m ,7 n a s follows :

2 2
(13.20) \B(M,N)\ ^ (Y,\Pm\ )Y,J2\^ln2\C(nun2)
n\ ri2

where
(13.21) C(n un2) = V a ,mn\ Uj mri2 •

If ni = n 2 w e ca n d o nothing bette r tha n th e trivia l bound . However , sinc e N

is quit e larg e th e number o f such case s (th e diagona l terms ) i s relatively small .
On th e othe r hand , fo r a majority o f n\ / n 2 th e su m C{rii,n 2) i s small becaus e
of cancellatio n o f terms (th e variatio n o f the argumen t o f amrll i s not completel y
annihilated b y tha t o f am n 2 ) . I n this wa y on e arrive s a t a non-trivial estimat e for
S(M, N) an d als o fo r th e sum s o f typ e S 2.
Combining th e results fo r sums o f type Si an d £2 one deduce s a boun d for
S(A, x) whic h i s often remarkabl y strong . I n the forthcomin g section s w e provid e
details of the Vinogradov metho d i n particular contexts . W e shall als o present som e
other methods .
13.2. A variant o f Vinogradov's method .
In thi s sectio n w e develo p a formula fo r sums ove r prime s usin g som e modifi -
cations o f the origina l idea s o f Vinogradov. Th e argument s ar e of combinatorial
nature, the y hav e muc h i n common wit h siev e methods .
For 1 < n ^ x we put f3 n(x) = ( 1 +a;(n, y/x)) -1 , wher e cj(n, y/x) i s the numbe r
of prime s p\n wit h p ^ y/x. If n is squarefree, w e hav e
Y^ n , x j -Li f n has a prime facto r ^ y/x,
Pn/P(x) = J Q0 i f n is prime wit h y/x < n ^ x.

This nice formula i s due to O . Ramare. Henc e for an y sequenc e o f complex number s
A = (a n) we hav e

Yl a P = ^2 an
~ ^2^2 P™(x)amp
^/x<p^x Kn^x p^v ^ mp^x
(rn-,p) = l

where ^ restrict s th e summatio n t o squarefree numbers . Le t P(z) be the prod -

uct o f all prime s p ^ z wit h 2 ^ z ^ y/x. W e appl y th e above identit y t o th e
subsequence o f numbers a n wit h (n , P(z)) = 1 gettin g

Yl a P= ^2 an
~ ^2^2 Pm(x)am P.
^<p^x Kn^x Z<P^T/X mp^x
(n,P(z)) = l p\m,(m,P(z)) = l
 13. SUM S OVE R PRIME S 341

Now w e remov e th e restrictio n t o squarefre e number s n , an d th e conditio n p \ m.

The correctio n quantit y doe s no t excee d

E W < ( Ep>zk l f f E ^ 'f < \\A(x)\\xh-$

n^x n^x
Vp>z,p \n

where ||^4(x)| | i s define d b y

2
(13.22) \\A(x)\\=(j2\a n\ y.
n^.x

To thi s w e ad d th e term s a\ and a p wit h p ^ ^fx an d estimat e thei r contributio n

by
E \a n\ ^ \\A(x)\\x l<.
n^y/x
Next w e remov e th e conditio n (n,P(z)) = 1 by the exac t siev e (th e Legendr e
formula)
an=
^2 Yl ^( d)Ad(x).
n^x d\P{z)
(n,P(z)) = l
We kee p th e term s Ad(x) wit h d ^ yfx an d estimat e th e remainin g ones . Firs t b y
Cauchy's inequalit y w e get

T( dn
E \M*)\<
d\P(z) dn^x
E Edn^x
K«KM(*)II( E E - )f
d>^/x~ d\P( Z),d>y^C d\P( Z),d>y^

1
^\\A(x)\\(xlog3x)±( ] T ^(djd- )*.
d\P{z),d>V^

To th e las t su m w e appl y Rankin' s tric k a s follow s

1
J2 r(d)d~ ^ x~ £ Y, r{d)d 2e 1
~ ^ x~ £ Y[(l + 2p 2£ 1
-)
d\P(z) d\P(z) p^z
d>y/x

< x~ £ J ] ( l + p - 1 " 5 ) 2 2 ' 6 ^ x" e C(l + e) 2z3e

^+ if"=eXp(-g£)(1 + 1 o g ^
by choosin g e = 1/logz. Collectin g th e abov e estimate s w e get
PROPOSITION 1 3.1 . Let 2 ^ z ^ yfx. For any sequence of complex numbers
A — (an) we have
(13.23)
Y a P =Y K d A
) d(x) ~ Y Prn(x) ^ a
mp + s(x, z)y/x\\A(x)\\
p^x d\P(z) (m,P(z)) = l mp^x
d^y/x z<p^y/x
342 13. SUM S O V E R P R I M E S

where P(z) is the product of all primes p ^ z, ^2 restricts the summation to square-
free numbers, A^{x) is given by (1 3.1 0), \\A(x)\\ is given by (1 3.22), \fi m(x)\ ^ 1
and e(x,z) (which also depends on A) satisfies
21
(13.24) \e(x,z)\ < -^+exp(-i^)(log3x) .
y/z V 2 log z J
REMARKS. Mor e precisel y w e hav e (1 3.23 ) wit h f3 m(x) = ( 1 + uj(m, y/x)) 1- ,
where cj(ra , y/x) i s th e numbe r o f prime s p\m,p ^ y/x, bu t w e didn' t specif y thi s
to emphasiz e tha t on e doesn' t nee d anythin g explici t abou t thes e coefficients . Th e
first su m o n th e righ t sid e o f (1 3.23 ) i s o f typ e Si wit h D = y/x (se e (1 3.1 6 ) an d
(13.18)). On e ca n als o spli t thi s int o tw o sum s Si -f - S 2 accordin g t o d ^ z an d
z < d ^ y/x. The n Si i s o f typ e (1 3.1 6 ) wit h D = z whil e S 2 i s o f typ e (1 3.1 7 )
with M = z an dT V = y/x. Th e doubl e su m o n th e righ t sid e o f (1 3.23 ) i s als o o f
type (1 3.1 7 ) wit h M = z an d N = y/x. Th e estimatio n (1 3.24 ) i s fine fo r

(13.25) z = exp(V'logx) .
For thi s choic e o f z w e hav e
/
(13.26) Hx,z)\ ^ 3exp(-| v i o i ^ ) ( l o g 3 a : ) 21 .
If A — (a n) i s an oscillatory sequenc e (s o the mai n ter m i n (1 3.1 2 ) i s expected t o b e
small), the n withou t grea t los s on e ca n simplif y (1 3.23 ) b y writin g th e inequalit y
a
(13.27) \J2 p\^ E \M*)\ + Y1\ E
z)y/x\\A(x)\\.
Z<p^.y/x
We shal l retur n t o th e abov e construction s i n Sectio n 1 3.6 .

13.3. Linnik' s identity .

In 1 96 0 Yu.V . Linni k [Li2 ] gav e a marvelou s expressio n fo r
fn
A(n) _ f a _ 1 i = P aia > °»
(13.28) A'(n )
lognn (~ \ 00 otherwis e
in term s o f the stric t diviso r function s
(13.29) T' k(n) = \{n 1 ,...ink>2; m...n k = n}\
To this en d h e compute s th e logarith m o f

c(*) = E n " f l = 1 ft 1 -*>"')"

1P

in tw o way s usin g th e powe r serie s expansio n

00 / j

Iog(l-x) = - X ; y -
fe=l

First b y th e Eule r produc t fo r ((s) on e get s

(13.30) logC(s ) = 5]A ,(n)n-fi

n^2
 13. SUM S O V E R P R I M E S 34 3

On th e othe r hand , b y th e Dirichle t serie s fo r C( 5)>

(13.31)
logC(s) = l o g ( l - ( l - C ( s ) ) )

fc=l n^2 k=l

Comparing th e coefficient s i n bot h expansion s Linni k get s th e identit y

(13.32) A '(n) = - £ L l ^ ( n ) .

Note tha t r k(n) = 0 i f 2 k > n , s o (1 3.32 ) i s a finit e su m ove r 1 < k ^ log njlog2.
In application s Linni k use s (1 3.32 ) fo r number s n havin g n o smal l prim e factors ,
say fo r (n,P(z)) = 1 with z ^ 2 . The n (1 3.32 ) run s ove r k ^ lo g nj lo g z. Thi s
yields th e followin g
PROPOSITION 1 3.2 . Let 2 ^ z ^ y/x. For any sequence of complex numbers
A = (a n) we have

a
(13.33) E < V = - £ ^X ) «r^n)
p ^ r r k^K n^x
P>z (n,P(z) )=l

where K = lo g #/ lo g z and r fk(n) is the strict divisor function 1( 3.29).

If on e prefer s th e exac t diviso r function s r^{n) on e ca n replac e th e stric t one s
rk{n) i n th e Linni k formul a b y

(13.34) r' k{n)= T ( -1 ) * " ' ( * W ) -

O^i^k

Moreover, th e conditio n (n , P(z)) = 1 can b e relaxe d a t smal l cos t b y applyin g th e

exact siev e i n th e sam e wa y a s i n th e Vinogrado v formula .
On th e right-han d sid e o f (1 3.33 ) th e firs t ter m (k = 1 ) i s a su m o f typ e Si ,
precisely

(13.35) 5 i = ] Ta n,

(n,P(*)) = l

while an y othe r su m

(13.36) S k= J2
n^x
(n,P(z)) = l

for k ^ 2 can b e regarde d a s a bilinea r for m ( a su m o f typ e S2) . However , Linni k

gave a specia l treatmen t fo r Sk fo r a fe w smal l k (i n th e solutio n o f th e Hardy -
Littlewood equatio n p + x 2 + y 2 — iV) , an d onl y fo r large r k h e arrange d Sk int o a
suitable bilinea r form . B y extendin g th e clas s o f special form s involvin g th e diviso r
functions o f degre e k > 1 one gain s mor e flexibilit y i n th e remainin g part s o f th e
identity (1 3.33 ) fro m whic h t o arrang e bilinea r forms . O f course, no t ever y for m Sk
can be treated b y special means. Usuall y on e treats S i b y classical Fourier analysis ,
34 4 13. SUM S OVE R PRIME S

and on e ca n ofte n appl y t o # 2 th e spectra l theor y o f automorphi c forms . On e ma y

hope tha t S k wit h k ^ 3 can b e treate d b y Fourie r analysi s i n GL^ , bu t s o fa r thi s
has no t bee n successful .
In 1 98 1 D . R . Heath-Brow n [HB3 ] develope d a formula fo r A(n ) whic h i s
reminiscent o f tha t o f Linnik . Le t

M(s) = Y, M Mm

Then s
as)M(s) = l + J2an(z)n- .
n>z
Consider th e identit y
k1
£ ( S ) ( 1 _ C ( S ) M ( S ) ) * = £ ( * )+ J2 (-l) (Kk)C(s)Ck- (s)Mk(s).

Comparing th e Dirichle t coefficient s o n bot h side s w e obtai n

PROPOSITION 1 3.3 . Let K ^ 1 , z ^ 1 . Then for any n < 2z K we have

(13.37) A(n ) = - ] T (-^ k(Kk) J2'"H l i{ml)...^mk)\ognk.

k
v 7
l^k^K rn 1 ...m
1 kn ...nk=n
mi,... ,m k^z
This identity o f Heath-Brown ha s a few advantages ove r that o f Linnik. Mos t of
all, the numbe r o f terms K ca n b e chose n smalle r b y regulating wit h th e paramete r
z mor e efficiently .

EXERCISE 1 . Sho w tha t fo r n ^ zK we hav e

(13.38) M W = -E ( " W * ) Y.---Y. Kmi)...Kmk).

l^k^K ^ ' m ...mkn1
11 ...nk- =n
mi,... ,m k^.z

13.4. Vaughan' s identity .

The mos t popula r formul a fo r A(n ) i s du e t o R . C . Vaugha n [Va] . W e deriv e
it fro m
A(n) = £ > ( & ) lo g £ .
b\n

Here w e keep th e term s wit h b ^ y an d transfor m th e remainin g su m a s follow s

£M(fr)log^ = £ & ( & ) A ( c ) .

b\n bc\n
b>y b>y

Next w e keep th e term s wit h c > z and transfor m th e remainin g su m a s follow s

A
£ £A*(6)A(c ) =Y,H M&)A(c) - J]E^ (°)-
6c|n 6c| n 6c| n
b>y,c^iz c^.z b^.y,c^.z
Here th e complet e su m ove r al l b dividing n/c vanishe s unles s c = n , whic h i s no t
possible i f n > z. Addin g u p th e abov e expression s on e obtain s
 13. SUM S OVE R PRIME S 345

PROPOSITION 1 3.4 . Let y, z ^ 1 . Then for any n > z we have

lo A
(13.39) A(n ) = J2 »(*>) §\ " £ £ ^) (C)+ £ £ ^ 6 A C
) ( )'
6|n 6c| n 6c| n
b^y b^y,c^z b>y,c>z

Similarly w e deriv e a formul a fo r th e Mobiu s function .

PROPOSITION 1 3.5 . Let y, z ^ 1 . Then for any m > max(y, z) we have

(13.40) n(m) = - £ E M*>)M< 0 + E £ M%(c) .

6c|m frc|ra
b^.y,c^.z b>y,c>z

PROOF. W e star t fro m

fi(m) = ^^/i(6)/x(c )
bc\m

and spli t th e summatio n int o fou r range s accordin g t o b ^ y, b > y, c ^ z, c > z.

Then w e transfor m th e tw o middl e sum s int o

£ X>(% M = - £ £ M%(c),
bc\m bc\m
b^y,c>z b^.y,c^z

££/i(6)/x(c) = -££/x(%(c).
bc\m bc\m
b>y,c^z b^.y,c^z

Adding u p th e abov e expression s on e obtain s (1 3.40) . •

Vaughan's identit y (1 3.39 ) is not a s flexible fo r arrangin g bilinear form s a s those

of Linni k o r Heath-Brown , bu t i t prove d t o b e sufficien t fo r basi c application s an d
it i s the simples t o f al l (read y t o us e withou t re-groupin g it s terms) .
EXERCISE. Deriv e a formul a a l a Vaugha n fo r Ak(n).
13.5. Exponentia l sum s ove r primes .
As a n exampl e o f numerou s application s o f th e method s describe d i n th e pre -
vious section s w e prov e Vinogradov' s estimat e fo r th e exponentia l su m (1 3.1 5) .
Assuming

(13.41) | a - - | ^ — with(a,g )= l ,
z
I q\ q
Vinogradov [V6 ] showe d tha t

(13.42) V(a;x) < qi x% + q~^x + xexp(-^^/loga;) .

Using Vaughan' s identit y (1 3.39 ) w e deriv e practicall y th e sam e boun d fo r

(13.43) 5(a ; x) = ] T e(cm)A(n) .

346 3 . SUM S OVE R PRIME S

THEOREM 1 3.6 . Suppose a satisfies (1 3.41 )- Then for x ^ 2 we have

(13.44) S(a,x) < (qix* + q~^x + x^)(logx) 3

where the implied constant is absolute.
We begin th e proo f o f (1 3.44 ) b y estimating specia l exponential sums . W e have

e
| 52 ( an )|< min ( 7V -2M)-
Hence fo r an y number s x(m) > 0 we ge t

min iV a;(TO)
52 | 52 e(«™«)| < 52 ( ' '^u^)-
n^x(m)

As m varie s ove r a segmen t o f lengt h q/2, th e point s 1 1 am| | ar e al l distinc t an d

spaced b y l/2q a t least . B y thi s observatio n w e deriv e

£ min ( iV '^)^(1 +4M

^ ( i V + E «< _1 )
<€i(M + N + MNq- 1 + q) log 2g.
Similarly fo r x(m) = x/ m w e derive (conside r separatel y th e rang e 1 ^ m ^ q/2)

Y\ m i n f - ' ^ i\) <:(M + xq- 1

+q)log2qx.
^—' \m 2 \\am /
From thes e estimate s w e ge t
LEMMA 1 3.7 = For any numbers x{m) > 0 we have

e
(13.45) Y^ I ]L ( a m n ) | < ( M + N + M A ^ - 1 + g)log2g ,
n^rr(m)

(13.46) ^ I ] P e(amn)\ < ( M + x g - 1 +g)log2gx .

Next w e deriv e a boun d fo r genera l bilinea r form s

(13.47) B(x;N)= ^ | £7 m e(amn)

N<n^2N mn^x

where 7 m ar e comple x number s wit h | 7 m | ^ 1 . Applyin g Cauchy' s inequalit y an d

(13.45) w e ge t

B2(x;N)^2N J^Jl | E <a(™i - m 2)n)

< ( ^ + A ^ - h - + g+
)xlog2g .
9
From thi s boun d w e deduc e
 13. SUM S OVE R PRIME S 34 7

LEMMA 1 3.8 . For any complex numbers a m,/3n with \a m\ ^ l,|/3 n | ^ 1 we

have

(13.48)
1 1 Y Y amPne(amn) <z: ( ^ + ^ + - ^ q) 2 x^ (lo g x) 2.
mn^x
m>M,n>N

P R O O F O F THEORE M 1 3.6 . B y th e identit y (1 3.39 ) w e hav e

S(a;x) = / J /] n(m)(log£)e(alm) — V V V /i(m ) A(n)e(afmn)

£m^.x Imn^x

+
H Y1 Y1 M(^)A(n)e(a^mn ) + O(iV) .
imn^x
m^M,n^N

We choose M = N = x$. T o the first an d th e secon d sum s we apply (1 3.46 ) gettin g

0((xs + xq~ l + q)(\ogx) 2. I n th e las t su m w e consider in = fc as one variable wit h
coefficient
c(k) = ^ A(n ) ^ lo g A;.
n\k,n^N

Then w e appl y (1 3.48 ) gettin g 0((xs + x g - 1 - f <?)^ 2 (logo:) 3. Addin g u p thes e

estimates w e get (1 3.44) . •

By th e sam e arguments , bu t usin g th e identit y (1 3.40 ) i n plac e o f (1 3.39 ) (als o

apply Cauchy' s inequalit y alon g th e wa y t o remov e th e correspondin g coefficient s
c(k) whic h ar e bounde d b y r(k) J no t b y lo g k a s above) , on e show s
THEOREM 1 3.9 . Suppose a satisfies (1 3.41 )- Then for x ^ 2 we have
m
(13.49) ^^ ^( )e( a m
) < ( < p £ 2 _| _ q-\ x + x5)^2(logx) 4
.
m^.x

If a i s a rationa l numbe r wit h smal l denominator , the n th e su m (1 3.49 ) ca n b e

well estimate d b y a n appea l t o th e zero-fre e regio n o f L-functions . Indeed , fo r an y
Dirichlet characte r %(mo d k) w e have (se e (5.80) )

(13.50) ] P /i(m)x(m ) < k^x(logx)"

for an y x ^ 2 an d A ^ 0 , th e implie d constan t dependin g onl y o n A. Henc e w e

derive b y mean s o f Gaus s sum s tha t
A
(13.51) Yl A*Me ( —) < qx(logx)~ .
m^.x

Now le t a b e an y rea l number . Give n Q ^ 1 there exist s a rationa l numbe r a/q

with (a , g) = 1 , 1 ^ q ^ Q suc h tha t
a 1
(13.52) a -- — < —7Z
<7 qQ
34 8 13. SUM S OVE R PRIME S

Hence w e deriv e b y partia l summatio n tha t

/\ / J ( 27T£\| V^ /am\
E /i(m)e(am)| ^ (^ 1 + —J| ^ /^Weh r)
for som e 1 ^ y ^ x. Usin g (1 3.51 ) on e get s
5A
(13.53) Yl ^{rn)e{am) < C (q + ^ ) x ( l o g x ) ~
m^x

for an y x ^ 2 an d A ^ 0 , th e implie d constan t dependin g onl y o n A W e appl y

(13.53) i f q < x Q - 1 an d (1 3.49 ) i f x Q " 1 <q^Q gettin g
_1
y ^ fjb(m)e(am) < Q ^2(logx)-5A + (5^4(logx) 4 + x^(logx) 4.

Finally, choosin g ( J = x(logx) _ 4 A , w e conclud e th e followin g

THEOREM 1 3.1 0 . Fo r any real number a and x ^ 2 we have

(13.54) ^ fjb(m)e(am) < C x(logx)"A

ra^x

/or am / A ^ 0 ; where the implied constant depends only on A.

This estimat e wa s firs t prove d i n 1 93 7 b y H . Davenpor t [Dal ] b y applyin g
Vinogradov's method . Th e complet e uniformit y i n a o f thi s boun d i s it s pleasan t
feature. W e shal l exploi t (1 3.54 ) i n ou r treatmen t o f th e Goldbac h problem s i n
Chapter 1 9 .

EXERCISE 2 . Usin g th e identit y (1 3.39 ) prov e tha t

(13.55) ^ e(oiy/n)A{n) < x§ (log x)4

n^x

for an y rea l numbe r a ^ 0 , wher e th e implie d constan t depend s o n a .

13.6. Bac k t o th e sieve .
We elaborat e furthe r th e idea s o f Sectio n 1 3. 2 t o dra w result s whic h ca n b e
useful i n mor e delicat e situation s tha n tha t i n th e las t section . Th e goa l i s t o
produce th e boun d
a £7r x
(13.56) \52 p\ < ()
p^x

for an y e > 0 an d x ^ XQ(S) subject t o hypothese s o n sum s o f typ e (1 3.1 8 ) an d

(13.19) whic h ca n b e verifie d b y existin g technology . On e o f suc h technolog y wil l
be th e spectra l theor y o f automorphi c form s whic h w e presen t i n Chapte r 2 1 fo r
the proble m o f equidistributio n o f quadrati c root s t o prim e moduli .
Recall tha t P(z) denote s th e produc t o f al l prime s p < z. Give n x w e conside r
the sum s
(13.57) Q(A,z)= Y,
n^x
(n,P(*)) = l
 13. SUM S OVE R PRIME S 349

as a function o f z for th e sequenc e A = (an) an d it s subsequence s Ad = {cbmd)-

(See als o Chapte r 6 for a n elementar y introductio n t o siev e methods) . Suppos e
that \a n\ < r(n).
Let x s < z ^ x*. The n ou r su m over prime s i s approximate d b y Q(A,z) a s
follows:

Y a P = Q(^ z ) -^2^2apq + Ylap + ai

p^x pq^% p<z
z^p^q

= Q(A,z) + Ol log 1 2
v logx1
\2 ogz y (logx) y
One can reduce th e "sievin g level " z by applying th e Buchstab identity , namel y for
any sequenc e A and any w < z we have
A
(13.58) Q(A,z) = Q(A,w)- Y Q( P^P)-
w^.p<z

Applying thi s identit y twic e w e get

(13.59) Q(A,z) = Q(A,w)- Y ^p^)+EE^^) -

w^.p<z w^~q<p<z

Opening Q(A, w) an d Q(A P, w) b y the exact siev e o f Legendre w e get

Q(A,w)- Y Q(Ap,w) = Y^ KtfMx)

w^p<z d
where th e das h restrict s th e summatio n t o th e divisor s d of P(z) whic h hav e at
most on e prime facto r p^ w. W e fixy^ z and estimate th e partial su m over d > y
by applyin g Rankin' s tric k a s follows :

\Y fJ>{d)A d(x)\ ^Y z L T( m c 0 <• x(logx)Y^ r 1

d d
()~
d>y d>y md^.x d>y

and
£'r(d)d- 1 <!/- e £ , ' r ( d ) (f " 1
d>y d

<y~s(1+ E ^br 1 )Il( 1 + r ^" 1 )-

w^p<z p<w
Choosing e = (logw)" 1 w e get

E^>^(j)'n(-rM;P'w
d>y y p<z F
^ y/

Hence
8
Y M M ^ W « (- ) x(logx) .
yJ
d>y ^

The remainin g su m over d ^ y is estimated b y

(13.60) R{x) = Y\Mx)\-

d^y
35 0 13. SUM S OVE R PRIME S

Now w e procee d t o th e doubl e su m o f Q(A pq, q) ove r prime s p, q with w ^ q <

p < z. Le t x± < v ^ xz. W e estimat e th e partia l su m ove r q ^ v trivially as
follows:
/~ \2
A ( 1 +
E Q( P^ I)«T^ZlogEx ^-^ i'\lo gz

x , / log x \ x
< - logo;
lo 1\3 gogi
—^ — +
y (logx) 2

It remain s t o evaluate th e su m o f Q(Apq,q) ove r prime s p , q with w ^ q < v

and q < p < z. W e writ e thi s i n th e for m

(i3.6i) EE«A>„9 )= E E mp
w^q<v mq^x 7m a.
q<p<z w^q<v
q<Pm<z
where p m is th e leas t prim e facto r o f m and 7 m i s th e numbe r o f prim e divisor s o f
m whic h ar e < z, so 7 m ^ uj(m). W e want t o mak e it a bilinear form . Thi s require s
us to relax th e relatio n q < pm. T o thi s en d w e us e th e followin g lemm a whic h is
convenient fo r separatio n o f variables :
LEMMA 1 3.1 1 . There exists a function h(t) depending only on z such that
oo
\h{t)\dt <\ogQz
/ -oo
and for all integers 1 ^ a , b ^ z,
f°° f ,(^\ lt f 1 ifa^b,
(1363) (,)
L" ( i ) * = ( , ,/„> ,
PROOF. Pu t g{u) = mm{uz, 1 , 1 + (1 - u)z} fo r 0 < u < 1 + z_ 1 an d g{u) = 0
otherwise. The n fo r positiv e integer s a , b ^ z we hav e g(a/b) = 1 if a ^ 6 , and
g(a/b) = 0 if a > 6. W e als o hav e

^-el/Kir* /(0 )

where f(s) i s th e Melli n transfor m o f g{u), i.e. ,

f(8)=Jo 9 (U)U° du= - (-^? .

For s = it with t > 0 w e hav e three estimates \f(s)\ ^ min{log62; , 2t_ 1, Azt~ 2} fro m
which it follows tha t th e functio n h(t) = ^f(-it) ha s th e asserte d properties . •

Applying (1 3.63 ) t o (13.61 ) an d changin g th e notatio n q to p we obtai n th e

integral o f bilinea r form s
oo
lt
] T ] P lmCLmp{Pmlp) h{t)dt
/ -oo _mp^x
„^ „
w^.p<v
 13. SUM S OVE R PRIME S 351

which w e estimat e b y B(x)(logx) 2 wit h

a
(13.64) B(x)= ] T| Yl mPPlt
(m,P(w)) = l mp^.x
w^~p<v

for som e rea l numbe r t (us e u(m) < log 2m ^ logx , (1 3.62 ) an d 6z ^ x).
Collecting th e abov e estimate s w e obtai n
a
(13.65) \Yl p\ ^ R(x) + B(x)(\ogx) 2 + E(x)
p^x

where
1/ lo g w
(13.66) E(x) « ^ - lo g ( ( l o g * } ) + — ^ + (-) x(logxr
2
logx \6(\ogv)(logz)J (logx) \yj
and th e implie d constan t i s absolute .
Let 2 ^ A ^ #1 2 . W e choos e z — A~ 2x?, y = A _ 1 £ 2 5 v = A _ 1 £ 3 7 an d w =
^l/ioiogioga: gettin g E(x ) < C x(logx)- 2 lo g A. Henc e w e conclud e (tak e A = x £)
THEOREM 1 3.1 2 . Let A = (a n) be a sequence of complex numbers with \a n\ ^
r(n). Let x ^ e 1 2 an d ( l o g x ) - 1 < e ^ ^ . Suppose that

(13.67) V lV a d m k-^-,
2
^ ^ I ^- ^ I (logx)
X
(13.68) ] T IE? 'i4a.pra ^ (logx) 4
(m,P(iu)) = l pm^x
w^.p<v

/or an? / t G l, where the parameters y, w, v are given in terms of x and e by

(13.69) y = x \ '\ v = x\~ £, w =x ^/^^^x

and P(w) is the product of primes < w. Then

(13-70) E«,«i ^
p^x

where the implied constant is absolute.

REMARKS. Th e conditio n (1 3.68 ) i s regarde d a s a n estimat e fo r genera l bi -
linear forms . Th e facto r p lt (whic h comes o n ou r wa y o f separatio n o f variables )
contaminates th e coefficients , bu t i t ha s n o effec t i n practice . On e coul d remov e
this factor , however , a precis e statemen t woul d involv e anothe r parameter , whic h
compromises th e clarity . Som e o f th e restriction s i n th e bilinea r for m (1 3.68 ) ca n
also b e modified . Fo r example , on e ca n remov e th e coprimalit y conditio n i n th e
outer sum , bu t ou r experienc e show s tha t i t i s convenien t t o hav e i t (not e tha t
(m, P(w)) = 1 implies (m,p ) = 1 ) . A concept o f "prim e producin g exponents " wa s
discussed i n [DFI5] ; Theore m 1 3.1 2 assert s tha t th e pai r (^ , ^ ) i s prim e produc -
ing. Th e presenc e o f e i n th e parameter s (1 3.69 ) i s ver y important ; withou t i t th e
requirement (1 3.68 ) woul d b e impossibl e t o verif y fo r mos t interestin g sequence s
A = (a n ). A well balanced choic e i s e — e(x) — (log log x )- 1 givin g a bound fo r th e
352 13. SUM S O V E R P R I M E S

sum ove r prime s whic h show s onl y littl e cancellation , nevertheles s i t i s a satisfac -
tory result , particularl y i n application s t o th e equidistributio n problem s o f degre e
two object s relate d t o th e GL2 theor y (se e Chapte r 21 ) .
 http://dx.doi.org/10.1090/coll/053/15

CHAPTER 1 4

HOLOMORPHIC MODULA R FORM S

14.1. Quotient s o f th e uppe r half-plan e an d modula r forms .

Holomorphic modula r form s wer e historicall y discovere d an d studie d fo r pur -
poses o f complex analysi s an d algebrai c geometry , i n particular, i n connectio n wit h
elliptic functions . The n the y wer e introduce d i n algebrai c numbe r theory , particu -
larly for th e developmen t o f class field theory, Galoi s representations an d arithmeti c
of ellipti c curves . Mor e recentl y modula r form s hav e entere d th e territor y o f ana -
lytic numbe r theory . Beside s providin g ne w tools , the y ar e als o a sourc e o f dee p
problems wit h numerou s connection s t o arithmeti c geometry , notabl y throug h th e
theory o f //-functions .
See als o Chapte r 1 5 fo r a surve y o f th e theor y o f non-holomorphi c forms , a s
well a s Chapte r 1 6 for application s t o Kloosterma n sum s an d Chapte r 2 6 fo r som e
results abou t L- functions. Ther e ar e man y fine book s o n variou s aspect s o f thi s
vast subject , fo r instanc e [Bu] , [BG] , [Mi] , [Sh3] , [1 4] , [Sel] .
We star t wit h H , th e Poincar e uppe r half-plan e
(14.1) M = {zeC | y = I m (z) > 0}
r
w hich i s a n ope n subse t o f C (i n th e nex t chapter , i t wil l b e see n a s a riemannia n
manifold o f constan t negativ e curvatur e wit h th e Poincar e metric) .
There i s a n actio n o f th e grou p G = SX(2,R ) o n H I by linea r fractiona l trans -
formations
riAn\ az + b fa b
14.2) gz = - , fo r g = ,
^J * cz + d' u \c d
because
Im(z)
(14.3) lm(gz)
tcz + d\ 2
which i s positive i f I m (z) > 0.
The cente r o f G , namel y {±1 } , act s trivially , an d on e show s tha t th e grou p
of holomorphi c automorphism s o f H i s th e quotien t PSX(2,R ) = SL(2,R)/{±1 } .
Note that thi s actio n i s also isometric fo r th e hyperboli c Poincar e metric ; se e Chap -
ter 1 5 .
The actio n o f thi s larg e grou p give s E I a larg e amoun t o f symmetry . A usefu l
manifestation o f thi s i s th e followin g property : defin e first th e actio n o f G o n th e
extended uppe r half-plan e M * = l U l U o o , b y th e sam e formul a (1 4.2 ) i n additio n
to th e rule s go o = a/c i f c ^ 0 an d goo = o o i f c = 0 , the n fo r an y thre e distinc t
points v, w an d z i n R U o o , ther e exist s g G G suc h tha t
gv = 0 , gw = 1 , gz = oo.
353
35 4 14. H O L O M O R P H I C MODULA R FORM S

Associated t o a Riemannia n metri c i s a natura l invarian t measur e <i/i , which i n

the cas e o f H I i s th e hyperboli c measur e

(14.4) dfi(z) = y~ 2dxdy i f z = x + iy

which i s invarian t b y G , i.e . fo r an y / integrabl e o n H I we hav e

/ f{z)d^z) = f f{gz)dn(z).

Arithmetic come s whe n w e conside r discret e subgroup s o f G actin g o n H I and

the correspondin g quotien t spaces . On e shoul d kee p i n min d th e analog y o f Z c l
(or Z[i] C C) . Thi s i s a ver y ric h theory ; ther e ar e man y "different " discret e
subgroups o f G , an d t o classif y the m i s almos t th e sam e a s classifyin g al l Rie -
mann surfaces . Fo r arithmeti c applications , th e mos t importan t example s ar e th e
congruence subgroups , particularl y thos e whic h contai n on e o f th e so-calle d Heck e
congruence subgroup s To(q) define d b y

The group s To(q) hav e finite inde x i n SX(2,Z) , precisel y

[To(l):r 0 ( g )]= g n( 1 +P" 1 )-

p\q

For q varying on e obtain s ver y interestin g quotient s r 0 (g)\H[ (thei r genu s in -

creases) .
One prefer s t o dea l wit h compac t Rieman n surfaces , however , quotient s o f HI
are no t alway s compact . I t i s eve n possibl e t o hav e Vo l (F\M) = +oo , wher e th e
volume refer s t o th e natura l measur e induce d fro m (1 4.4) . Fo r example , th e grou p
of translation s b y integer s

r„-{(J \) ]k.z}
has th e vertica l stri p 0 < R e (z) ^ 1 "representing " th e quotien t r ^ H I , an d i t ha s
infinite area .
More generally , a fundamenta l domai n fo r th e actio n o f a discret e subgrou p
T C G o n H I is a nic e subse t F C H I which "almost " contain s on e poin t pe r orbit .
Precisely, F mus t satisf y th e followin g conditions :
(1) F i s a n ope n subse t i n HI.
(2) Th e closur e F o f F i n C intersect s ever y orbi t o f T , i.e . fo r al l z G H I
there exist s 7 G T wit h 72 : G F.
(3) N o tw o point s i n F ar e T-equivalent .
A fundamenta l domai n alway s exists , bu t i t i s certainl y no t unique . On e ca n
even choos e F t o b e a polygo n wit h (hyperbolic ) geodesi c sides .
For th e congruenc e subgroup s To(^) , on e ca n first conside r th e usua l funda -
mental domai n fo r SX(2,Z) ,

F 1 = {^GH I I | R e ( z ) | < | , \z\>\]

 14. H O L O M O R P H I C M O D U L A R F O R M S 355

and the n tak e th e unio n o f it s image s unde r cose t representative s i n 5L(2,Z) :

Fq= | J 7* 1
7er 0 (q)\r 0 (i)

is a fundamenta l domai n fo r T 0(q). (I t i s no t necessaril y connected ; fo r othe r

choices, se e [1 5] , Section 2.2. )
On a picture of F\, th e non-compactness appear s because Fi touche s the bound -
ary MUo o at infinity , s o (z , 2i, 3z,...) i s a sequence without convergin g subsequence .
However, b y the Gauss-Bonne t formul a o r a simple direc t computation , th e volum e
of SX(2,Z)\H I (whic h i s th e volum e o f an y fundamenta l domain ) remain s finite,
namely Vo l (SX(2, Z)\Bf) = | , an d th e sam e i s tru e fo r To(q) wit h

Vol(r 0 (, ? )\H) = | [ r o ( l ) : r 0 ( g ) ] .

For r suc h tha t Vo l ( r \ H) < +oo , th e point s a t whic h F touche s th e boundar y

are calle d cusp s of T (thi s intuitive descriptio n wil l be made mor e precise i n Sectio n
15.7 whe n w e discus s th e classificatio n o f element s i n G) , an d the y ar e i n finite
number (becaus e t o each cus p on e can attac h non-overlappin g smal l neighborhood s
in M , whic h hav e fixed positiv e volume) . Fro m th e above , o o i s th e onl y cus p
for SX(2,Z) . S o fo r To(q), the y ar e easil y classifie d b y considerin g th e actio n o f
r 0 ( g ) \ r o ( l ) o n oo. On e finds tha t ever y cus p ha s a uniqu e representativ e ^ wit h
(a, c) = 1 , c ^ 1 , c | g, an d a is modulo (c , | ). Thei r numbe r i s therefor e

h=Y,<P((c,d)).
cd—q

This i s equal t o r(q) i f g is squarefree .

For an y cus p a , it s stabilit y grou p T a = { 7 G T \ 7 a = a } i s infinit e cycli c (u p
to ±1 ) , generate d b y 7 a say , an d ther e i s a a a £ G , calle d a scalin g matrix , suc h
that cr a 00 = a an d
1
(14.5) ^ 7a^ a =fJ \Y
Using a a, mos t computation s an d propertie s relate d t o th e variou s cusp s o f
T ca n b e studie d fo r th e cus p a t infinity , wit h stabilit y grou p generate d b y th e
translation z y-> z - f 1 , fo r th e conjugat e grou p G~ lTcra (i t i s a n advantag e o f th e
Poincare mode l ove r othe r model s o f th e hyperboli c plane , fo r instanc e th e uni t
disc, tha t i t ha s suc h a distinguishe d cusp) .
We can no w defin e modula r form s a s holomorphic function s o n HI transforming
in a simpl e wa y unde r th e actio n o f a discret e subgroup . Le t k ^ 1 be a n integer .
Then w e hav e a n actio n (th e "weigh t k action" ) o f G = SX(2,IR ) o n function s
/ : H -> C by

ab
(14.6) ( /| fc g)(z) = j(g, z)~ kf(gz), wher e j(g, z) - cz + d for g =
cd
a
Let q ^ 1 be a n integer , an d \ Dirichle t characte r modul o q (no t necessaril y
primitive). Clearl y \ induce s a character o f r 0 (g) b y x(d) = x(d) f° r 9 —
356 14. HOLOMORPHI C MODULA R FORM S

A modula r for m o f weight k, level q, and nebentypu s (o r character ) \ i s a holomor-

phic functio n / o n H I which satisfie s

(14.7) / I * 7 = X(7)/, for all 7 G r 0 (g)

and i s holomorphi c a t all cusp s o f To(^) . Th e holomorph y a t cusps i s explained as
follows. First , fo r a cusp a of To(g), the functio n f a = f \k cra i s seen b y modularit y
to b e periodic o f period one , f a(z + 1 ) = f a(z). Thi s implie s no w tha t f a i s a
function o f th e paramete r q — e(z), namel y f a(z) = g a(q), wher e g a i s holomorphi c
in a punctured dis c { z G C | 0 < | 2 : | < r } . W e the n sa y tha t / i s meromorphic at
a if ga i s meromorphic a t 0. Therefor e f a ha s a Laurent serie s expansio n

AW = Yl fa(n)e(nz)
n^na

for som e intege r n a , and f a(na) ^ 0 . W e sa y tha t / i s holomorphic a t a if it i s

meromorphic an d n a ^ 0 . If , moreover, w e hav e n a > 0 , w e sa y tha t / vanishe s at
a. I f / vanishe s a t all cusps , it is called a cusp form .
EXERCISE 1 . Sho w fro m th e definitio n tha t a modular for m / i s a cusp for m
if an d onl y if the T-invarian t functio n
k/2
(14.8) 9(z) = y \m\
is bounde d o n H I (this criterio n i s very convenien t i n practice).

EXERCISE 2 . Sho w tha t i f we tak e th e sam e definitio n bu t k = 0 , th e corre -

sponding spac e o f modula r form s i s reduce d t o 0.

Clearly modula r form s of fixed level, weigh t an d characte r for m a vector space ,
which is denoted Mk(q, x) ? a n d th e cus p form s a subspace denote d S&(# , x)- Takin g
7 = — 1, w e see tha t / i s identicall y zer o unles s x ( —1) — (~l)fc? s o we assume tha t
X satisfies thi s consistenc y condition .
The firs t fundamenta l resul t i s that M^(g,x ) i s a finite dimensiona l vecto r
space. Thi s ca n b e proved i n general usin g th e Riemann-Roc h theore m fo r compac t
Riemann surfaces . I n the case o f To(g), a simpler proo f i s possible. Fo r k ^ 2,
the dimensio n o f Mk(q,x) c a n ^e compute d exactl y (se e e.g . [Sh3]) , an d the n
estimated ver y precisel y (th e cas e o f forms o f weight on e i s a tantalizing proble m
of deep arithmetica l significanc e becaus e o f links with Arti n L- functions, se e [Se2] ,
[Dul]). On e show s fo r instanc e tha t

dim M k(q, x ) = " i 2 " ^ + 0(y/qk)

where v q = [T0 (l) : T 0(q)].

For an y modula r for m / w e denot e b y df(n) its Fourier coefficient s a t oo, so
that
(14.9) f(z) = Y,af(n)e(nz).

Those coefficient s ar e ver y importan t object s (the y ar e arithmeti c harmonics) ,

and thei r orde r o f magnitude, i n particular, i s the subjec t o f many inquiries . For
 14. H O L O M O R P H I C M O D U L A R F O R M S 35 7

cusp forms , th e criterio n tha t y k^2\f(z)\ i s bounded o n HI implies quickl y th e trivia l

bound
k/2
(14.10) a f(n)<^n

which i s goo d enoug h t o star t wit h (se e below (1 4.54 ) fo r th e muc h deepe r Delign e
bound).
One als o define s th e Petersso n inne r produc t o n Mk{q, x) b y

(14.11) {f,g) = J f{z)W)y kd^(z).

This i s well-define d becaus e th e integran d i s ro(g)-invariant , a s a simpl e compu -
tation reveal s (compar e (1 4.8)) , an d th e integra l i s finite a s soo n a s on e o f / o r g
is a cus p form . I n particular , Sk{q,x) 1 S a (finit e dimensional ) Hilber t spac e wit h
this inne r product , an d thi s analyti c fac t i s very importan t i n what follow s (se e th e
definition o f "newforms " i n Sectio n 1 4.7) .
Notice als o tha t i t i s possibl e t o multipl y modula r form s o f th e sam e level : i f
/ e Mk(q,x), 9 £ M k>{q,Xf), w e n a v e fg € ^4+/c'(<2SXX') - I n particular , th e direc t
sum o f al l Mk(q), k ^ 0 , i s a grade d C-algebra . Thi s importan t algebrai c fac t wil l
not b e o f muc h us e i n thi s book , however .
It shoul d b e emphasize d tha t i n provin g tha t a functio n i s modular , th e mod -
ularity formul a (1 4.7 ) nee d onl y b e prove d fo r 7 i n a famil y o f generator s o f ro(g) .
Such group s ar e alway s finitely generate d s o thi s i s a finite se t o f condition s o n / .
For instanc e 5L(2 , Z) i s generate d b y th e tw o element s

(01 ) an d
( 1 "o 1 )
acting a s th e translatio n z \-> z + 1 and a s inversio n z M > — z~l respectively , an d
the tw o condition s fo r / t o b e modula r becom e
k
f(z) = f(z + l) an d f(^.)=z f(z).

We no w giv e example s o f modula r form s an d i n particula r o f cus p forms .

14.2. Eisenstei n an d Poincar e series .
The basic examples of modular form s ar e the Eisenstein an d the Poincare series .
We onl y defin e thos e serie s associate d t o th e cus p 0 0 fo r simplicity , bu t simila r
constructions ca n be done at ever y cusp. Le t k > 2 (to ensure absolute convergence )
and pu t
fc
Ek(z)= J2 X(7)j(7,*)- .
Clearly, Ek i s i n Mk(q,x)i a s ^ i s constructe d b y th e averagin g technique ; holo -
morphy a t th e cusp s i s checke d b y explici t computatio n o f th e Fourie r expansion ,
which als o show s tha t Ek i s no t a cus p form . I t i s calle d a n Eisenstei n series .
More generally let m ^ 0 . Th e ra-th Poincar e series is obtained b y the averagin g
technique applie d t o e(mz):
p k
m{z) = Yl X(7)j(7 , z)- e(mjz)
ler^xroiq)
358 14. HOLOMORPHI C MODULA R FORM S

for k ^ 2 (the convergenc e fo r k — 2 is not absolut e bu t ca n b e see n fo r m ^ 1 from

the Fourie r expansio n (1 4.1 2 ) below , usin g th e Wei l boun d fo r Kloosterma n sum s
(1.60)). W e d o no t conside r m < 0 as th e serie s diverge s then .
The importan t fac t i s that fo r m ^ 1 , we get cus p forms .
PROPOSITION 1 4.1 . (1 ) J/r a = 0 , then P 0 = E k.
(2) If m ^ 1 ; then P m is a cusp form in Sk(q, x)-
We sketc h par t o f th e proof , namel y th e followin g lemma .
LEMMA 1 4.2 . The Poincare series P rn has Fourier expansion
mn r
(14.12) P m(z) = 6(m,0) + J2p( ' Wnz

with coefficients
/?ir\k n 1
k
~ * .^
P ( 0 , n ) = ( - ) Ytkj E c-*5 x (0,n;c)
c>0
c=0 (mod q)

p ( m 5 n ) = f — j 2 f(5(m,n) + 27r i fc ^ c 1
S' x (m, n; c) Jfc _i f J J
c>0
c=0 (mod q)

if m ^ 1 . ifer e S x (ra, n; c) is the Kloosterman sum with character x,

(14.13) S x (m,n;c) = ^ *(d)e (

d (mod c)

and J/c- i « 5 the Bessel function of order k — 1 .

PROOF. W e firs t mak e a convenien t parametrizatio n o f th e coset s r ^ r o ^ ) ,
namely usin g th e actio n o f a n elemen t o f To o o n th e righ t an d o n th e left , w e se e
that To(q) i s the disjoin t unio n

(14.14) r 0(g) = ro o | j ( | J( J roof " ^ W )

0 0 d(modc )
c=0(modq) ( c,d) = l

(where fo r give n c an d d wit h (c , d) = 1 , a an d 6 are an y tw o integer s suc h tha t

ad — be = 1 ) .
Thus w e hav e correspondingl y

d J c
P m (z) = e(raz ) + 5 Z X T ^( ) ( '* *)
c>0 d (mod c)
c=0 (mo d g)

where
I(c,d;z) = J2(c(z + n) + dy ke(m(«c-c{c{z+1 n) + d) )).
 14. H O L O M O R P H I C M O D U L A R F O R M S 359

Next, b y th e Poisso n summatio n formul a

i z
^ ^]2jm(^+v)+dyH~-c{c{zZ)+d)-nv)dv
/am nd\
ii /
-oo+iy
Here w e observ e tha t th e inne r integra l vanishe s fo r n < 0 (mov e th e lin e
of integratio n upwards ; th e cas e n = 0 i s th e mos t important , s o th e reade r i s
encouraged t o chec k carefully) . The n puttin g togethe r th e differen t terms , an d
recognizing th e Besse l functio n i n th e integra l (se e [GR , 8.31 5.1 , 8.41 2.2]) , w e
obtain th e lemma . •

EXERCISE 3 . Prov e th e decompositio n (1 4.1 4) .

Notice tha t ther e ar e infinitel y man y Poincar e serie s P m , bu t the y al l liv e i n

the finite dimensiona l spac e Sk(q,x)- Thi s mean s tha t ther e mus t b e man y linea r
relations betwee n th e P m 's. O n th e othe r hand , suc h abundanc e tend s t o sugges t
that th e Poincar e serie s shoul d spa n th e whol e space . Thi s i s indeed th e case .
LEMMA 1 4.3 . Let f e Mk{q,x) be a modular form with expansion 1( \.9).
Then for any m > 1 ;

COROLLARY 1 4.4 . The Poincare series with ra ^ 0 span Sk(q,x)-

Indeed, an y cus p for m / orthogona l t o ever y P m mus t hav e al l it s Fourie r
coefficients a t o o equal to zero , an d henc e mus t b e zero . Becaus e th e spac e spanne d
by Poincare series is closed (thi s uses crucially the finite-dimensionality o f the space )
the resul t follows .
P R O O F O F LEMM A 1 4.3 . B y definitio n o f P m w e hav e

k k
(f,Pm) = [ /(*) { J2 x{l)Wi7z)~ ^m^z)}y d^z)

k
= Yl Im(<yz) f(-yz)e(m~fz)dfi{z)
jF
ieroo\r0(q)

(by modularit y o f / an d I m (jz) = \j(g, z ) | - 2 I m (z))

pi p + OO
k 2
=/ / y ~ f(z)e(-mz)dxdy
Jo Jo
r+oc pi
k 2
Taf{n) / y - e-2^n+m^dy e(( (n
r — m)x)dx
„ ^ n JO JO

r(fc-i) a m
(A.ixm)xkf e _ T / ( )

•
360 14. HOLOMORPHI C MODULA R FORM S

Taken together , th e tw o previou s lemma s impl y a formul a whic h show s ho w

the sequence s o f Fourie r coefficient s o f a basi s o f Sk(q,x) a r e almos t orthogonal ,
thereby givin g anothe r decompositio n i n arithmeti c harmonic s o f the diagona l sym -
bol (5(m , n). Le t T b e an y orthonorma l basi s o f Sk(q,x) (distinguishe d basi s wil l
be describe d i n Sectio n 1 4.7) . Sinc e P m i s in Sk(q,x)i w e c a n expan d i t a s a linea r
combination o f th e / G J , namely w e hav e

Pm=Yt(f,Pm)f

so that b y taking the n-th Fourie r coefficien t o f both side s (o n the right wit h Lemm a
14.3, o n th e lef t wit h Lemm a 1 4.2 ) w e obtain th e Petersso n formul a
PROPOSITION 1 4. 5 ( P E T E R S S O N ) . For any n ^ 1 and ra ^ 1 ,

(1415) a n
7 I ^ f e ^ E / ( ) ^ M= 5(m,n )

1
+ 2m * 2 ^c 5x(m,n;c)J/c_if— J .
c>0
c=0 (mod q)

EXAMPLE. Th e "first " cus p for m occur s fo r leve l 1 an d weigh t 1 2 ; i t i s th e

famous Ramanuja n A function , whic h ha s th e elegan t produc t expansio n

A(*) = QY[0--024> w i t h Q = < z)

(see e.g . [Sel ] fo r on e o f man y direc t proof s o f th e modularity) .

The existenc e o f a non-zer o / G £1 2(1 ) ca n b e predicte d b y considerin g th e
function / = E% — E%, where E± an d EQ are th e Eisenstei n serie s o f leve l 1 an d
weight 4 and 6 . Thes e hav e th e Fourie r expansion s (a t th e onl y cus p oo )

EA(z) = 1 + 24 0 Y^ °z(n)e(nz)
n^l

E6(z) = 1 - 50 4 ^2 °s{n)e(nz)

so tha t a direc t computatio n o f th e smal l orde r term s give s

El-El = 1 72Sq-{-0{q 2)
whence / i s non-zer o an d i s a cus p form . Sinc e di m 5i2(l ) = 1 , th e equalit y
/ = 1 728 A mus t follo w onc e we know that A i s indeed modular . Th e coefficient s i n
the expansio n o f A ar e denoted r(n) (no t t o be mistaken wit h th e diviso r function) .
These ar e integer s havin g man y fascinatin g properties , fo r example ,
m o d 61
9
r(n) = Y^ dU -
d\n

This kin d o f congruence s belong s properl y t o th e theor y o f Galoi s representation s

(see e.g . [Se3]) .
Evidently, A must als o be proportional to the non-zero Poincare series in 612(1),
but i t doe s no t aris e naturall y i n thi s way . Also , th e constructio n o f / exploit s th e
 14. H O L O M O R P H I C M O D U L A R F O R M S 361

fact tha t holomorphi c modula r form s ca n be multiplied. Actually , i t i s not ver y

difficult t o show tha t th e graded rin g 0fc^o-^fc(l ) i s isomorphic t o the polynomial
ring C[E4,EQ], an d E± and EQ ar e algebraically independent .
14.3. T h e t a functions .
The Eisenstei n an d the Poincare serie s ar e constructed b y averaging ove r the
group cosets . Thet a function s provid e anothe r importan t clas s of examples of mod-
ular form s whic h com e ou t in quite a different way . Th e most basi c thet a serie s is
Jacobi's 0 function an d equation (1 .53 ) i s indeed a modularit y propert y o f weight
\. However , w e hav e not defined half-integra l weigh t form s (se e for instance [Shi] ,
[14] for this theory ) s o we square 6 and obtain a weight 1 form

02(z) = J2r(n)e(nz).

More generall y w e consider firs t thet a serie s attache d t o binar y quadrati c forms .
Definite form s ar e related t o holomorphic thet a series , whil e indefinit e one s woul d
give non-holomorphi c forms , whic h wer e firs t constructe d b y Maass [Ma] .
Recall the notatio n o f Section 3.8 : K = Q(\/~D) is an imaginary quadrati c field,
D < 0 being a fundamental discriminant , 0 = Z © LOZ is the ring o f integers o f K,
w
~ 2
\{D + V~D), w is the number o f units o f O and H i s the ideal clas s group .
Let A E % be an ideal class . Th e theta serie s attache d t o A i s defined t o be

(14.16) 0 A(z) = - + y e{zNa)

aeA

where a runs ove r the non-zero integral ideals belonging to the clas s A. A s shown in
Section 22.2 , there i s a one-to-one correspondenc e betwee n idea l classe s and classes
of definit e binar y quadrati c form s o f discriminan t D. Le t $A — [a,b,c] b e the
quadratic for m associate d t o A. I f b E A, w e have b = (a) a for som e a E a - 1 ,
a = m + riz^. Then Nb = ^ ^ ( m , n) so taking th e units int o accoun t w e obtain the
formula

(14.17) 0A{Z) =-T e($ A(m,nj).

772,Tl G Z

Poisson summatio n (i n two variables) show s again tha t Qj± is modular, precisel y
9A £ MI(\D\,XD)I wher e XD i s th e Kronecke r symbo l (th e Dirichlet characte r
associated t o the field K). I t als o follow s tha t

OA(Z) =

From (1 4.1 6 ) o r (1 4.1 7 ) w e see that 6A is not a cus p form . Fo r a class grou p
character x E ?^ we define th e theta serie s

/x(Z) = ^ X (o)e(*)
a

where th e sum is over al l integral ideal s o f O thi s time . Splittin g int o classe s we
derive
/*(*) = E x(A)d A(z)
A<EH
362 14. H O L O M O R P H I C MODULA R FORM S

for x 7 ^ 1 , s o / x E MI(\D\,XD) a l s o - Computin g th e Fourie r expansio n a t othe r

cusps, one proves that f x i s a cusp form i f y2 ^ 1 . Recal l that th e characters of order
2 o f H ar e calle d genu s characters . The y wer e determine d (essentially ) b y Gauss ,
who showe d tha t th e subgrou p o f genu s character s i s isomorphi c t o (Z/2Z) t _ 1 ,
where t = u(\D\) i s th e numbe r o f distinc t prim e divisor s o f \D\.
For x n ° t a genu s character , w e hav e f x G SI(\D\,XD), a n d i t als o satisfie s

MIDI J
All clas s grou p character s ar e particula r case s o f Heck e character s o f K (se e
Section 3.8) , s o the y hav e a n L-functio n (ove r K) attache d t o the m

£*(*,*) = J>(a)(Wa)-*
a

and b y Melli n inversio n th e complete d L-functio n

AK(s,x)=(^)Sr(s)LK(s,X)
is expressed b y th e formul a
1
A*(5,x) = J^ (sT +iT 8)/x(-^L=)dy.
A fe w othe r representation s o f A^(s , x) a r e give n i n Sectio n 22.3 . Fro m th e abov e
integral representatio n w e derive, a s fo r th e Rieman n zet a functio n o r th e Dirichle t
L-function, tha t LK(S,X) n a s meromorphi c continuatio n t o th e whol e comple x
plane an d i s entire i f x ¥" 1 - Moreover , w e have th e functiona l equatio n
A*(s,x) = A i r ( l - s , x ) .

Much more general theta series are obtained b y considering any integral positiv e
definite quadrati c form , an d associate d harmoni c polynomial s i n n variable s instea d
of only binary quadrati c forms . However , ther e is no interpretation i n terms o f ideal
classes of number fields in the genera l cas e (fo r 4 variables, o f course, ther e ar e link s
with quaternio n algebras) .
Let A = (dij) b e a real , symmetric , positiv e definit e matri x o f ran k r =
0 (mod 2), an d
J\\X\ — XJ~±X — y CLi jX^Xj

the correspondin g quadrati c for m (th e theor y o f form s i n od d numbe r o f variable s

is somewhat harde r becaus e o f a complicated multiplie r syste m fo r th e correspond -
ing modula r forms) . A homogeneou s polynomia l P G C[X] i s calle d a harmoni c
polynomial i f A ^ P = 0 , wher e

1
dxidxj
i,3

is the Laplac e operato r associate d wit h th e matri x A 1 — (a* •). Wit h thi s dat a w e
define th e thet a serie s
I9(2) = Y, P(.m)e(^A[m]).
meZr
 14. H O L O M O R P H I C M O D U L A R F O R M S 363

Suppose A an d NA~ l hav e integra l coefficients , an d eve n diagonal , fo r som e

suitable intege r T V ^ 1 . The n 0(z) G Mk{N,XD) wher e k = \ + d e g P an d x D{d) =
( ^ ) i s the Kronecke r symbo l wit h D = (—1 )%\A\. Moreover , i f d e gP > 0, then
0(z) = 6(z; A, P) i s a cusp for m whic h satisfie s

k
0(z;A1P) = z%\A\-^z- 0(--;A-\l

where P*(x) = P{A~ lx) i s a harmonic polynomia l associate d wit h A~ l (se e als o
Section 20. 4 and , fo r instance , [1 4] , Chapter s 9 and 1 0 fo r complet e proofs) .

14.4. M o d u l a r form s associate d t o ellipti c curves .

Cusp form s associate d t o elliptic curve s hav e b o t h arithmeti c an d analyti c

flavors. W e begi n b y introducin g th e basi c invariant s o f elliptic curve s ove r an
arbitrary fiel d (se e [Sil ] an d als o Sectio n 1 1 .9) . An y ellipti c curv e E, defined ove r
a fiel d fc, has a plane mode l give n b y a Weierstrass equatio n

(14.18) E : y2 + a\xy + a^y = x3 + <i2X2 + a^x + CIQ

with coefficient s a,j G k. Associate d wit h th e a-coefficient s i s th e se t o f coefficient s

62, 64, be, 6 8 give n b y b 2 = a\ + 4a 2 , 6 4 = a xa^ + 2a 4 , 6 6 = a 2 + 4a 6 , b 8 = a2a6 -
aia^a^^ Aa 2a§ + a 2al — a\. Th e 6-coefficient s ar e relate d b y 46 g = b 2bQ — b2. Havin g
these quantitie s w e expres s th e discriminan t o f (1 4.1 8 ) b y
3
(14.19) A = -b\bs - Sb 4 - 21 b\ + 96 2 6 4 6 6 .

The cubi c equatio n (1 4.1 8 ) define s a n elliptic curv e ove r k i f an d onl y i f it i s

non-singular, whic h mean s t h a t it s discriminant A doe s no t vanish . T o sim-
plify equation s w e associat e wit h (1 4.1 8 ) anothe r se t o f coefficients C4,C 6 give n
by C 4 = b\ — 24&4, CQ = — b3 + 366264 — 21 665 . I n thes e term s w e hav e
3
(14.20) 2 A = c\-c\.

R E M A R K S . T h e firs t se t o f 6-coefficients arise s fro m completin g th e square .

Assuming th e characteristi c o f k is no t 2 , w e ca n chang e th e coordinate s (x,y) i n
(14.18) int o (x,y + ^x + ^) gettin g

(14.21) E :y2=x* + ^x2 + ^ x + ^ .

7
4 2 4
The secon d se t o f c-coefficients arise s fro m completin g th e cube . I f further th e
characteristic i s not 3 , we ca n chang e th e coordinate s (x , y) i n (1 4.21 ) int o (x— j | , y)
getting
(14.22) E : y 2 = x 3 _ | x _ | L = x 3+ a x +6 5

say. Thi s ha s th e discriminan t A = I2~ 3(c\ - c§ ) = - 1 6 ( 4 a 3 + 276 2 ). I n this

equation w e chang e furthe r (x , y) int o (x , |) an d multipl y throug h b y 4 getting

(14.23) E : y2 = Ax3 + Ax + B

with A = - j | = 4a an d i ? = — ^ = 46. No w th e discriminan t o f (1 4.23 ) i s simpl y

given b y

(14.24) A = -(A3 + 27B2) = (c3 - c 2

6)/1 728
36 4 14. H O L O M O R P H I C MODULA R FORM S

(notice tha t 1 72 8 = 1 2 3 ).

We retur n t o th e genera l equatio n (1 4.1 8 ) ove r an y field k. W e define th e

so-called j-invarian t by
(14.25) j = 4A-1 = (12c 4)3(c34 - eg)" 1 ;
this make s sens e becaus e w e assume A / 0 . Th e j-invariant play s a n importan t
role i n classification o f elliptic curves . Firs t w e single ou t two special cases .
CASE j = 0 . Thi s i s equivalent t o c4 = 0 , therefore i f E i s given b y (14.23), the n
A = 0 and A = - 2 7 £ 2 .
CASE j = 1 728 . Thi s i s equivalent t o c 6 = 0 , therefore i f E i s given b y (1 4.23) ,
then B = 0 and A = -A 3,
EXAMPLE 1 . Suppos e j ^ 0,1 728 . The n th e equation

(14.26) E
vJ
: y2 + xy = x3 ^ - x - -—^— —
y y - _ 1 7 2 8 • _ 1 7 2 8

gives an elliptic curv e ove r any field k which contain s j ; it s j-invariant is equal to j ,
the c-coefficient s ar e C4 = j(j — 1 728) _1 , CQ = —j(j — 1 728) _1 an d the discriminant
is A = j 2(j — 1 728) _3 . Als o the cubic equatio n

/1 A 0 7 \ T? 2 , • 3 J U~ 1 )2 ^ 33
(14.27) Ej : y + ixy = x x : T^T^X : —- —
vJ
3 y J y 4 J _ 1 7 2 g3 _ 1 7 2 g

(which is the quadratic twis t o f (14.26) b y j) give s an elliptic curv e ove r the field fc;
its invariants ar e c 4 = j 3 ( j - 1 7 2 8 ) " \ c 6 = - j 4 ( j - 1 7 2 8 ) _ 1 an d A = j 8 ( j - 1 7 2 8 ) ~ 3 .
EXAMPLE 2 . Suppos e j ^ 0,1 72 8 an d char.k ^ 2,3 . The n th e cubic equatio n

(14.28)
v yE : y2 = 4x 3 - 2
\ { x- , 2 ^{no
' j - 1 72 8 j - 1 72 8
with j G k give s a n ellipti c curv e ove r k whos e j-invarian t i s equal t o j an d the
discriminant i s A = 2 6 • 3 1 2 • j2{j - 1 728) - 3 .
From no w on we consider ellipti c curve s E define d ove r Q. W e can assum e
that E i s given b y the Weierstrass equatio n (1 4.1 8 ) wit h intege r coefficient s aj (if
necessary, chang e th e coordinates (x , y) int o (c _ 2 x, c~ 3y) and multiply throug h by
c6 t o kil l th e denominator s o f aj b y choosin g a suitabl e c G N). Ever y ellipti c
curve E/Q ha s a minima l model , tha t i s E i s given b y the equation (1 4.1 8 ) wit h
integer coefficient s suc h that ord p A is minimal for every p among all such equation s
defining E (thi s specia l equatio n i s possibl e t o arrang e b y a rationa l chang e of
variables becaus e Z has unique factorizatio n property) . Th e following conditio n
(14.29) min{ord p A, 3ord p c 4 } < 12
for al l p is necessary an d sufficient fo r the model (1 4.1 8 ) wit h aj G Z to be minimal.
Given a minimal equatio n (1 4.1 8 ) fo r E w e look a t its reduction modul o p. If
p \ A, then th e reduced cubi c i s nonsingular, s o it defines a n elliptic curv e ove r the
finite field F p . I n thi s cas e i t i s sai d E ha s good reductio n a t p . I f p \ A, the n
the reduce d curv e i s singular, an d it i s said E ha s bad reduction a t p. Th e ba d
reduction divide s int o tw o types:
 14. H O L O M O R P H I C M O D U L A R F O R M S 365

(1) Multiplicativ e reductio n (o r semistable) i f p | A and p \ c^ (thi s mean s

E/Fp ha s a node).
(2) Additiv e reductio n (o r unstable ) i f p | A and p | C4 (this mean s E/¥ p has
a cusp) .
The multiplicativ e reductio n i s said t o be split o r non-split i f the slope s of the
tangent line s a t the nod e ar e rationa l o r not, respectively .
For ever y p we defin e th e conducto r exponen t f p a s follows:
(1) f p = 0 i f E ha s goo d reductio n a t p.
(2) f p = 1 i f E ha s multiplicativ e reductio n a t p.
(3) f P = 2 i f E ha s additiv e reductio n a t p ^ 2 , 3.
When £ ha s additiv e reductio n a t p = 2 or p = 3 , then th e definitio n o f fp i s
quite involved (se e [Sil]) , but w e always have f p^2 an d f p — 2 is a certain "measur e
of wil d ramification" . I n these case s on e ca n sho w tha t $1 ^ 8 and / 3 ^ 5 . Having
the exponent s / p fo r all p the conducto r o f E/Q i s set to be

(14.30) N = Y[p fp-

v
Note tha t N i s squarefree i f and onl y i f E ha s n o additive reductio n a t all places.
In thi s cas e E is called semistable .
Hasse define d th e L-functio n o f E as an Euler produc t o f local factor s define d
by looking at the reductio n E p of E modulo p for al l primes p, whic h i s a curve ove r
the finite field ¥ p. Th e loca l zet a functio n i s defined b y the forma l powe r serie s

Zp(E) = exp{j2\E(^)\~
n^O

where |J£(F pn)| is the numbe r o f points of Ep in a finite field of cardinality p n (wit h
the poin t a t infinity included) .
Let a{p) b e the intege r define d b y | ^ ( Fp ) | = p + 1 — a(p). Notic e tha t i f E is
given by a simpler Weierstras s equatio n

y2 — x 3 + ax + 1

(to which one can reduce for p ^ 2 or 3), then a(p) ca n be expressed b y the quadrati c
character su m

<p) = - xL
mo d p
(— ~v—J-
For p I A , the coefficien t a(p) i s given by

0i f E ha s additiv e reduction ,
a(p) = 1i f E ha s spli t multiplicativ e reduction ,
— 1 i f E ha s non-spli t multiplicativ e reduction .
Hasse prove d tha t Z P(E) is a rational functio n fo r every p , an d fo r p \ A (th e
reduced curv e i s then smooth )

l a{p)X+pX2
(14.31) Z p{E)- -
(l-X)(l- PX)-
36 6 14. H O L O M O R P H I C MODULA R FORM S

Moreover, Hass e proved the Riemann Hypothesis for Ev i n the form o f the inequalit y
(14.32) \a(p)\^2^p
which implie s tha t bot h root s of the numerato r o f Zp hav e modulu s equa l to 1 /y/p.
See Sectio n 1 1 . 8 for elementary proof s o f (14.31 ) an d (14.32).
The Hasse-Weil zeta function o f E, already mentioned in Section 5.1 4 , is defined
by th e Euler produc t

(14.33) L(E, s) = I J ( 1 - a(p)p-T 1 1 ^1 " a

(P)P~° + P1" 2 *)" 1-
p\A pf A

Define a(n) for all n ^ 1 by expanding (1 4.33 ) int o the Dirichlet serie s
oo
s
(14.34) L(E, s) = Y^ a{n)n- .
l

This L- function i s holomorphic i n the region R e (s) > | b y the Riemann hy-
pothesis (1 4.32) . Hass e furthe r conjecture d tha t th e L-function thu s define d fro m
local data coul d be extended to an entire functio n satisfyin g th e functiona l equatio n
A(E,s) = wA(E,2-s)
where w = ± 1 an d
S
/\/N\
A ( E , S ) = ( ^ - ) T(s)L(E,s),
N bein g th e conductor o f E.
This functiona l equatio n i s the same as that satisfie d b y the Hecke L-function s
of cus p form s o f weight 2 and level N wit h trivia l characte r (se e Section 1 4.7) .
Around 1 950 , Shimur a an d Taniyama mad e th e conjecture tha t fo r any E/Q,
there i s a cus p for m whos e L- function i s equa l t o L(E,s). Late r Wei l refine d
the conjectur e b y postulating tha t th e level o f this for m shoul d equa l exactl y the
conductor o f E. Thi s refine d conjectur e ca n be tested computationall y sinc e the
space o f cusp form s o f a given level , weigh t an d character ha s finite dimension .
This dee p modularit y conjectur e wa s proved i n 1 99 5 by Wile s [W] , Wiles-
Taylor [TW ] for semistable curve s (wit h Fermat's Grea t Theore m as a consequence,
as ha d been show n befor e b y Ribet), an d is now known fo r all elliptic curve s ove r
Q, th e final work bein g du e to Breuil, Conrad , Diamon d an d Taylor [BDCT] .
T H E O R E M 1 4.6 . Let E/Q be any elliptic curve of conductor N. There exists
a primitive cusp form
oo

(14.35) f(z) = ^\{n)n*e{nz) G S 2(T0(N))

l

such that a(n) — X(n)n^, that is

oo

(14.36) L(E, s + \) = L(f, s) = J2 A W™" 5 -

l

Since the cusp for m / associate d t o E is primitive o f level A ^ it is an eigenfunc -

tion o f the Fricke involutio n (Wf)(z) = N'lz~2 f(-l/Nz), i.e. ,
(14.37) Wf = -wf, wit h w = ± 1.
 14. H O L O M O R P H I C M O D U L A R F O R M S 36 7

Hence th e functiona l equatio n o f th e complete d L-functio n is

(14.38) A ( / , s ) = w A ( / , l - s ) .
The sig n w — ±1 in this equatio n i s also calle d th e roo t numbe r o f E. I t i s
not difficul t t o comput e w computers fo r an y concret e curve , ye t ther e i s no simpl e
formula fo r w in general. I n the importan t specia l cas e o f semistable curves , tha t
is whenT V i s squarefree , w e obtain b y th e theor y o f Heck e operator s

(14.39) w = -n(N)a(N) = -n(N)\(N)N* = - ( - l ) m

where m is the numbe r o f place s wit h spli t multiplicativ e reduction .
The celebrate d Birc h an d Swinnerton-Dyer Conjectur e assert s tha t L(E, s)
(resp. L(f,s)) vanishe s a t s = 1 (resp. s = | ) t o order exactl y equa l t o th e
rank o f th e grou p E(Q) o f rational point s o n E. Thoug h r = rankE(Q) i s not eas y
to compute , it s parit y i s conjecture d t o b e determine d b y
(14.40) w= (-l) r,
and thi s ha s bee n prove d b y Nekovar [Ne ] unde r th e assumptio n tha t th e Tate -
Shafarevitch grou p o f E is finite (whic h is known i f the order o f vanishing o f L(E, s)
at s = 1 is ^ 1 ) .
Proofs o f the modularit y o f special ellipti c curve s existe d before , notabl y for
the curve s wit h comple x multiplication , du e to Deuring (fo r a self-contained proo f
in the cas e of the "congruen t number " curves , se e for instanc e [1 4] , Chapter 8) . A n
early exampl e (du e t o Shimura ) o f a modular curv e whic h i s not a CM-curve i s th e
following ellipti c curv e
E : y2 + y = x 3 - x 2
whose j-invariant i s j — —4096/11 an d discriminan t i s A = —1 1 .
It follow s fro m th e propertie s o f the conducto r tha t th e leve l o f the weigh t 2
cusp for m / correspondin g t o thi s particula r E must b e 1 1 . Bu t i t is easy to show
that 52(1 1 ) i s a 1 -dimensiona l space , an d actuall y i t is not to o har d t o construct
an elemen t i n this spac e fro m th e knowledg e o f the Ramanuja n A function, namel y

f(z) = {A(z)A(llz))1/12 = ? n ( l - <Z") 2(1 " <? Un ) 2

(q = e(z) as before). I n other words , th e modularit y o f E is equivalent wit h th e

identity

(14.41) g J I a - 9") 2(1 - Q lln)2 = £ a(n)q n.

No elementary proo f o f this seems to be known. On e proceeds in two steps. Th e

first on e consists o f showing tha t Yo(ll ) = r 0 (ll)\IH[ (which , suitabl y compactified ,
is a compac t Rieman n surface ) i s an algebraic curv e isomorphi c t o the (affine )
elliptic curv e E; thi s i s done b y constructing meromorphi c function s g\ an d gi
on X 0 ( l l ) (a s quotient s o f modular form s o f the sam e weight ) whic h satisf y th e
equation o f E, an d goes bac k t o Fricke an d Klein . Th e secon d step , whic h is
arithmetic an d muc h les s elementary , i s to deduce fro m thi s th e require d identit y
of //-functions . Thi s las t argumen t i s an exampl e o f the so-calle d Eichler-Shimur a
theory.
36 8 14. H O L O M O R P H I C MODULA R FORM S

Another interestin g exampl e i s the Gross-Zagie r curv e give n b y th e twiste d

Weierstrass equatio n
-139y 2 = x3 + 10x2 - 20 x + 8.
This curv e ha s conducto r N = 37 -1392 an d the sign of the functiona l equatio n
is e = — 1. Moreover , th e rank o f the group o f rational point s o n this curv e i s 3
while on e can show usin g the Gross-Zagier formul a tha t th e Hasse-Weil L-functio n
vanishes a t s = 1 to orde r exactl y 3 (see Appendix t o Chapte r 2 3 for a sketch).
Hence thi s exampl e confirm s th e conjectur e o f Birch an d Swinnerton-Dyer . I n
combination wit h a n earlier wor k o f Goldfeld i t yields a non-trivial effectiv e lower -
bound fo r the class numbe r o f imaginary quadrati c fields , a s will b e explained i n
Chapter 23.
14.5. Heck e L-functions .
Hecke showe d ho w t o produce Dirichle t serie s ou t of modular form s satisfyin g
many o f the properties o f Dirichlet L-functions . Conside r / G M/c(g,x), s o / ha s
the Fourie r expansio n a t oo

/(*) = ^2 af(n)e(nz)-

One defines it s Hecke L-function t o be the Dirichle t generatin g serie s for the Fourie r
coefficients

(14.42) L(f,s) = Y,af(n)n- s.

This L-functio n an d the completed L-functio n

(14-43) A ( S ,/)=(^)Sr(S)L(/,S)

are holomorphi c i n the region R e (s) > 1 + | b y the trivial boun d (1 4.1 0) .
We als o introduc e a n operator W (sometimes calle d th e Fricke involution ) o n
Mk{q,x) b y
k/2
(14.44) Wf{z) = q- z-kf(—)
V qz J

which, becaus e th e elemen t H q1

=( normalizes To(g) , i.e . H qTo(q) —
M
T0(q)Hq, induce s map s W : M k(q,x) ~+ k{q,X) an d W : S k(q,\) ~ >S k(q,x)-

T H E O R E M 1 4. 7 ( H E C K E ) . With notation as above, the L-function of f has a

meromorphic continuation to the whole complex plane and the completed L-function
satisfies the functional equation
k
(14.45) A(f,s) = i A(Wf,k-s).

Moreover, L(f,s) is entire if f is a cusp form, and otherwise it has only a simple
pole at s = k.
 14. H O L O M O R P H I C M O D U L A R F O R M S 369

P R O O F . Th e proof is very much like Riemann's (second ) proo f o f the functiona l

equation o f C( 5)- B y th e definitio n o f th e gamm a functio n w e hav e fo r al l n ^ 1 ,

so i n th e regio n o f absolut e convergenc e w e have th e representatio n

Here we split th e integra l int o th e par t fro m 1 to +o o an d th e par t fro m 0 to 1 and

we transform th e latte r a s follow s

by applyin g (1 4.44) . Addin g bot h part s w e obtai n th e integra l representatio n

Since f — cto(f) decay s exponentiall y fas t a t infinity , th e meromorphi c continuatio n

follows, an d th e functio n equatio n i s then clear . •

If / = Ek i s a n Eisenstei n series , on e ca n se e fro m th e Fourie r expansio n tha t

L(Ek,s) i s a produc t o f tw o Dirichie t L-functions , s o w e d o no t ge t an y reall y
new L-functions . Bu t whe n / i s a cus p form , th e associate d Heck e L-functio n i s a
genuine GL(2) object .
REMARK. Lookin g bac k a t th e thet a serie s i n Sectio n 1 4.3 , on e ca n observ e
that i t provide s a n equalit y
LK(S,X) = L(f x,s)
where o n th e lef t sid e i s th e L-functio n attache d t o a clas s grou p characte r o f th e
quadratic extensio n K, whic h ha s a n Eule r produc t
1
s
LK(s,x) = \[(l-x{p)Np- )-
p

of degre e 1 over prim e ideal s i n (9 , wherea s o n th e righ t sid e w e hav e th e Heck e

L-function o f th e modula r for m f x G M I ( | D | , X D) whic h ha s a n Eule r produc t o f
degree 2 over rationa l primes . Thi s i s a n earl y sig n o f th e existenc e o f a n instanc e
of Langland s functoriality . Similarly , L-function s define d ove r numbe r fields ar e
expected t o b e identica l wit h particula r higher-degre e L-function s ove r Q . Anothe r
case o f thi s phenomeno n appear s i n th e theor y o f ellipti c curves , wher e th e Hasse -
Weil zet a functio n o f a n ellipti c curv e E/Q wit h comple x multiplicatio n wa s als o
shown b y Deurin g t o coincid e wit h th e L-functio n o f a Hecke characte r o f weight 2
of a n imaginar y quadrati c field.
37 0 14. HOLOMORPHI C MODULA R FORM S

14.6. Heck e operator s an d automorphi c L-functions .

We no w sketc h th e theor y o f Heck e operator s an d th e Atkin-Lehne r theor y o f
primitive forms , whic h establishes a link between Dirichle t character s an d automor -
phic forms , showin g clearl y ho w th e latte r generaliz e th e former .
The Heck e operator s ar e certai n linea r operator s actin g o n space s o f modula r
forms, whos e existenc e an d propertie s ca n b e see n t o b e intimatel y relate d t o th e
arithmeticity o f the subgroups To(g ) o f 5L(2, M), although thi s will not b e apparen t
in thi s surve y (se e [Mar]) .
Fix integer s k ^ 1 , q > 1 and a character \ modul o q satisfying th e consistenc y
property x{~ 1 ) = ( — l)fe- Th e operato r T{n) i s define d b y th e formul a

(14.46) T(n)f(z) = \ £ X (a)ak £ /(^ )

ad—n 0^b<d

for n ^ 1 an intege r (notic e tha t T(n) als o depend s o n \ an d s o implicitl y o n q\

this ca n b e sometime s significant) . Th e followin g propositio n i s fundamental .
PROPOSITION 1 4.8 . For any n ^ 1 , T(n) acts on modular forms and on cusp
forms: in other words, it induces linear maps

T(n) : M fc (g,x)->M fc ((7,x),

T{n) : S k(q,x) ~ > S k(q,x)-

The proof o f this proposition require s identifying th e formula (1 4.46 ) a s anothe r

instance o f a n averagin g operator , thi s tim e ove r th e (finite ) orbi t spac e A n =
r 0 ( g ) \ r n , wher e T n i s the se t o f integral matrice s o f order 2 and determinan t n , o n
which 5L(2 , Z) act s naturally .
From th e definition , w e comput e th e actio n o f T(n) o n th e Fourie r expansio n
(at oo ) o f a modula r for m (o r eve n o f an y 1 -periodi c function) , namel y

T k
(n)f(z) = \ £ X(a)a £ £ a / ( m ) e (m££±*)
ad—n 0^.b<d m

•iE-<(-)E«c^)E'(?)
m ad=n 0^b<d
= £ ( £ X{a)a k-la}{di))e{mz).
771
ad=n
a£=m

Therefore

(14.47) T(n)f(z) = ^2( £ X (d)dk-'af{^))e(mz).

m d\(n,m)

This formul a coul d als o be take n a s a definitio n o f T(n), bu t th e modularit y o f

T(n)f woul d b e harde r t o prove . However , sinc e th e Fourie r expansio n determine s
a modula r form , i t ca n b e used t o giv e quick proof s o f many propertie s o f the Heck e
operators.
 14. HOLOMORPHI C MODULA R FORM S 371

PROPOSITION 1 4.9 . The Hecke operators commute. More precisely, for any
n ^ 1 and m ^ 1 we have

T{m)T(n)= £ x (d)^r(^)
d\(n,m)

or equivalently
T(mn)= £ M ( d ) X ( d ) d f1
e
- T(^)T(^).
d|(n,ra)

In particular , th e T(n) ar e multiplicative : T(mn) = T(m)T(n) i f m an d n ar e

coprime. Therefor e an y T(n) i s a product o f Hecke operators T(p u) fo r som e prime s
p an d integer s v. Thos e operator s obe y a secon d orde r linea r recurrenc e relatio n
(showing tha t the y ar e determine d b y T(p)):
T(p»+l) = T(p)T( Pn - X
1
(p)P k 1
- T(pv- )
which ca n b e summarize d b y th e identit y o f generatin g function s
v
£ T{p )Xv = ( 1 - T(p)X + X (p)pk-lX2)-\

so the Dirichle t generatin g serie s o f all the T{n) ha s a n Eule r produc t (o f degree 2 )

J2 T(n)n- S = J ] ( 1 - T (P)P~° + X O V " 1 - 2 5 ) - 1 -

n^>l p

For p prime , w e have th e formul a

T(p)f(z)=X(p)pk-1f(pz) + l£ /(—) •

If p divide s q (th e level) , thi s simplifie s t o

W1-;EO
The operator s T(p) fo r p no t dividin g q turn ou t t o hav e ver y differen t prop -
erties tha n thos e wit h p dividin g q\ th e latte r ar e ofte n referre d t o a s th e "bad "
Hecke operators, an d th e correspondin g p a s th e ba d o r ramifie d primes . Th e mai n
difference betwee n those two kinds of operators, fro m th e analyti c point o f view, lie s
in their relatio n wit h th e inne r product . Indeed , th e nex t lemm a fail s fo r (n , q) > 1 .
LEMMA 1 4.1 0 . Let n be coprime with q. Then the operator T(n) acting on the
space of cusp forms Sk(q, x) 2 5 normal with respect to the Petersson inner product;
more precisely, its adjoint is
(14.48) T(nY =x(n)T(n).

The lemm a mean s tha t fo r / an d g tw o cus p form s w e hav e

(T(n)f,g) = X (n){f,T(n)g);
note tha t indee d thi s formul a canno t hol d fo r n no t coprim e wit h g , i.e . fo r n wit h
x(n) — 0 ? because i t woul d impl y tha t T(n) = 0 , which i s certainly no t alway s true .
37 2 14. H O L O M O R P H I C MODULA R FORM S

The proof of this lemma (du e to Hecke) is quite involved (se e [1 4] , pages 1 04-1 0 6
for instance) .
From the normality of the T(n) wit h (n , q) = 1 and the fact tha t the y commute ,
we deduc e b y standard linea r algebr a th e following
PROPOSITION 1 4.1 1 . There is an orthonormal basis of the space Sk(q,x) °f
cusp forms which consists of eigenfunctions of all the Hecke operators T(n) for
(n,q) = 1.
Any non-zer o modula r for m / G Mk(q, x ) f° r which ther e exis t comple x num -
bers X(n) such tha t
T(n)f = X(n)f
for al l n coprim e wit h q is called a Hecke form . I f / i s a cuspidal Heck e form , the
adjointness formul a (1 4.48 ) show s tha t X(n) = x(^)A(n) , fo r (n,q) = 1 .
What doe s i t mea n tha t / i s a Hecke for m fo r the Fourier coefficient s aj(n) o f
/ ? Fro m (1 4.47 ) applie d t o / w e obtain b y taking th e first Fourie r coefficien t o f
both side s tha t
(14.49) X(n)a f(l) = a f(n), fo r all n wit h (n,q) = 1
so i f a / ( l) 7^ 0, the Hecke eigenvalue s an d the Fourier coefficient s ar e the same up
to a constant factor ! Thi s alread y prove s tha t th e Ramanujan r functio n i n Section
14.2 is multiplicative, a s already conjecture d b y Ramanujan himsel f (an d prove d by
Mordell). Indeed , sinc e $1 2(1 ) i s one-dimensional an d spanned b y the A function ,
perforce A is an eigenfunction o f all Hecke operator s (ther e ar e no ramified prime s
here), an d since r ( l ) = 1 , the eigenvalue i s r(n) an d the multiplicativity follow s
from Propositio n 1 4.9 .
14,7, P r i m i t i v e form s a n d specia l basis ,
Proposition 1 4.1 1 i s not exactl y wha t w e wish t o be true. I t woul d b e mor e
interesting t o hav e a basi s o f eigenfunction s o f al l th e Heck e operator s withou t
exceptions. The n (1 4.49 ) hold s for all n, and so a / ( l) must b e non-zero or otherwise
/ = 0 . Afte r normalization , th e Fourier coefficient s wil l be multiplicative, an d more
importantly th e Hecke L-functio n o f / wil l hav e a n Euler produc t

L(f, s) = fj (1 " af(p)p-s + X {p)pk-l-2srl-

P

Such a wonderful worl d doe s no t exist, however . Indeed , le t %* modulo q* be

the primitiv e characte r whic h induce s x - Tak e any q' and d with q* \ q', dq' \ q an d
let x' b e the character modul o q' induced b y x*- No w if / G Sk(q'', x')i th e functio n
/ | d (z) = f(dz) i s in Sk(q,x)i a s follow s fro m th e matrix identit y
d 0 \ fa p\ _ f a pd\ fd 0 \
0 l) [ 7 6)-\ild 6 J \01 )'
It i s also eas y t o show tha t th e Hecke operator s T(n) wit h (n , q) — 1 commute
with th e operatio n / \-> f |^ , so / bein g a Heck e for m implie s tha t / | d i s one .
However, sinc e
/ I d (z) = ^2a f(n)e(dnz),
 14. H O L O M O R P H I C M O D U L A R F O R M S 373

the firs t Fourie r coefficien t o f / \d vanishes i f d > 1 .

The theor y o f primitiv e form s (o r "newforms" ) remedie s t o thi s defec t b y con -
sidering tha t cus p form s suc h a s / \d in Sk{q, X) a r e no t reall y o f leve l q but com e
from lowe r levels , an d tha t a bette r theor y o f Heck e operator s shoul d hol d whe n
such "oldforms " ar e pu t aside . Thi s wa s develope d b y Atki n an d Lehne r [AL] .
Therefore, le t £ b (g, x) D e th e subspac e o f Sk(q, x) spanne d b y al l cusp form s o f
the typ e / \d where / < G Sk(q' ', x') w ^ h q' < q and dq' \ q. The n le t S*(q, x) b e th e
orthogonal complemen t o f S\(q,x) wit h respec t t o th e Petersso n inne r product , s o
b
Sk(q,x) = S (q,x)®S*(q,X)
(note tha t i f x i s primitive , w e hav e S%(q,x) = Sk{q,x)i a n d tha t ^ X i s trivial , q
is prim e an d fc ^ 1 0 or k = 1 4 , w e als o hav e S%(q) = Sk(q), becaus e 5^(1 ) = 0 fo r
those k).
Since (T{n)f) \d= T(n)(f \d) for (n,q) = 1 , i t i s immediat e tha t T(n) act s o n
Sfr(q,x) f° r ( n5 q) — 1 - Sinc e thos e Heck e operator s ar e normal , the y ac t als o o n
the orthogona l complemen t S%(q, x)- I n particular , bot h space s stil l hav e a basis of
Hecke forms. B y definition, a Hecke form / whic h i s in S%(q, x) ls calle d a primitiv e
form (o f weigh t fc, level q, an d characte r %) • Th e mai n resul t i s the following :
THEOREM 1 4.1 2 (MULTIPLICIT Y ON E PRINCIPLE) . Given a sequence (A(n) )
of complex numbers, the subspace of S^(q,x) spanned by Hecke forms with eigen-
values A(n ) for all (n , q) = 1 is at most one-dimensional. In other words, such a
Hecke form f, if it exists, is unique up to multiplication by a scalar.

EXERCISE 4 . Prov e tha t th e multiplicity-on e principl e i s equivalen t t o th e as -

sertion tha t an y primitiv e for m / G S%(q,x) n a s non-zer o firs t Fourie r coefficient .

Suppose no w that T i s a linear operato r o n th e spac e S%(q, x) whic h commute s

with al l T(n) fo r (n , q) = 1 . The n i t follow s tha t T act s o n the commo n eigenspace s
of thos e T(n). B y th e multiplicity-on e principle , thos e ar e one-dimensional , eac h
spanned b y a primitiv e for m / , an d therefor e / i s als o a n eigenfunctio n o f T . I n
particular applyin g thi s propert y t o th e "bad " T(n), i.e . fo r (n,q) ^ 1 , w e obtai n
the desire d result .
PROPOSITION 1 4. 1 3 . Let f e S%(q,x) be a primitive form. Then f is an
eigenfunction of all Hecke operators,
T(n)f = \ f(n)f, for all n^ I
and for all n^l, a/(n ) = A/(n)a/(l) . Hence the first Fourier coefficient a / ( l ) is
non-zero.
It i s natura l t o normaliz e a primitiv e form , s o tha t i t become s a distinguishe d
basis o f it s commo n eigenspace ; the n th e basi s o f S%(q,x) ° f normalize d primitiv e
forms i s also unique. Tw o natural normalization s exist : w e will say that / i s Hecke-
normalized i f its first Fourie r coefficien t i s a/(1) = 1 , so that th e Fourie r coefficient s
are the same as the Hecke eigenvalues, an d that / i s Petersson-normalized i f ||/| | = 1
(such normalizatio n i s ofte n mor e usefu l i n th e contex t o f th e Petersso n formula) .
If / i s Hecke-normalized, the n o f course | | / | | - 1 / i s Petersson-normalized. Th e nor m
ll/H i s relate d t o th e valu e o f th e adjoin t squar e L-functio n o f / a t 1 , an d play s a
37 4 14. HOLOMORPHI C MODULA R FORM S

role i n som e importan t applications . Se e Sectio n 5.1 2 , an d i n particula r Corollar y

5.45 fo r th e compariso n betwee n th e tw o normalizations .
When n o othe r indicatio n i s present , "primitive " wil l alway s mea n "Hecke -
normalized primitive" . I n an y case , onc e a normalizatio n i s chosen , th e basi s o f
primitive form s o f S%(q,x) i s unique . I n particular , i t i s possibl e t o spea k o f th e
set o f primitiv e form s withou t ambiguity .
The Heck e L-functio n o f a (Hecke-normalized ) primitiv e for m / , fro m th e pre -
vious arguments , ha s a n Eule r produc t

L(f, s) = J] ( 1 - A/(p ) + XWP"- 1 -2*)-1 -

P

This i s equivalen t wit h th e formul a

(14.50) A /(mn)= £ x (d)d"-lXf(^).

d\(m,n)

This ca n b e als o writte n b y Mobiu s inversio n a s

k
(14.51) A /(m)A/(n)= £ ^d) X{d)d -'\f(^)\}Q.
d\(m,n)

Moreover, th e functiona l equatio n o f Theore m 1 4. 7 is als o simple r becaus e th e

primitivity o f / wil l sho w tha t Wf i s expressed i n term s o f / . Th e linea r operato r
W i s a n isometry , an d i t i s eas y t o sho w tha t W commute s wit h th e goo d Heck e
operators:
WTx(n) = x(n)TX(n)W, fo r (n , q) = 1
(where th e superscrip t x indicate s fo r whic h characte r th e T(n) i s defined) .
To com e bac k fro m Sk(q,x) t o th e origina l spac e 5fc(g,x) > w e us e anothe r
operator, th e if-operato r define d b y

KM = 7FI).
Note tha t K act s o n Fourie r expansion s b y conjugatin g th e coefficients ,
K z
f( ) = ^~^(n)e(nz) = 70)> say.

One mus t b e carefu l becaus e K i s no t C-linear , namel y K(af) = aKf. Bu t

from th e definitio n i t i s clea r tha t K 2 = Id, tha t K i s a n isometry , an d als o tha t
K commute s wit h al l T(n). Conside r no w th e compositio n (als o onl y R-linear )

W = KW : S k(qlX)^Sk(q,x)

which satisfie s th e commutatio n relatio n

T(n)W = x(n)WT(n) , fo r (n , q) = 1 .
Let / b e a primitiv e form . The n fo r (n , q) = 1 we hav e
T{n)Wf = X (n)WT(n)f = X (n)W(Xf(n)f) = X {n)Y^)W{f) = X f(n)W(f)

by th e formul a fo r th e adjoin t o f T(n). B y th e multiplicit y on e principle , i t follow s

that Wf mus t b e a multipl e o f / , henc e
 14. H O L O M O R P H I C M O D U L A R F O R M S 37 5

PROPOSITION 1 4.1 4 . If f is a primitive form, then there exists rj E C, with

\r)\ = 1 , such that Wf — r\f.

That1 77 1 = 1 follow s fro m th e fac t tha t W i s a n involution , W = Id. Th e

eigenvalue rj is another ver y intricat e invarian t o f / . I t i s determined b y / bu t onl y
in specia l case s d o we have a formul a fo r it s valu e i n term s o f more accessibl e data .
PROPOSITION 1 4.1 5 . Let \ be a primitive character of conductor q. Then rj is
given by
k/2
ri = r(x)Xf(q)q- .

PROPOSITION 1 4.1 6 . Let q be squarefree, and \ trivial. Then 7 7 is given by

1
77 = fi(q)Xf(q)q -^.

From Propositio n 1 4.1 4 an d Theore m 1 4. 7 w e deduc e th e mai n resul t abou t

automorphic L-function s o f primitiv e forms .
T H E O R E M 1 4.1 7 . Let f be a primitive form. Then the Hecke L-function of f
has an Euler product expansion

L(f,s) = Y[(i-\ f(p) + x(p)p k-1 -2sr1

p

and has analytic continuation to an entire function. The completed L-function

S
A(f,s)=(^) T(s)L(f,s)

satisfies the functional equation

k
A(f,s) = i vA(7,k-s).

Notice that thi s theore m i s in perfect analog y wit h th e correspondin g resul t fo r

primitive Dirichle t characters . Eve n deepe r analogie s ar e reveale d b y th e insigh t o f
the Langland s program , whic h use s representation theor y t o unify th e whol e theor y
of L-function s (cf . [BG]) .
We conclude thi s sectio n b y statin g a non-trivia l generalizatio n o f the multipli -
city-one principl e whic h i s very usefu l i n applications .
T H E O R E M 1 4.1 8 (TH E STRON G MULTIPLICIT Y ON E PRINCIPLE) . Given a
sequence (A(n) ) of complex numbers and a fixed positive integer M, there exists at
most one primitive form f E S%(q,x) such that
Xf(n) = A(n) , for all n prime to M.

The mai n poin t i s tha t M ha s nothin g t o d o wit h th e leve l q. I n effect , thi s

shows that a primitive for m i s known whe n on e knows its eigenvalues fo r al l primes,
except finitely many . I n fac t on e ca n sho w usin g Rankin-Selber g L- functions tha t
finitely man y prime s suffic e (dependin g o n th e weigh t an d level) ; se e Propositio n
5.22 an d th e remar k following .
376 14. HOLOMORPHI C MODULA R FORM S

EXERCISE 5 . Sho w tha t i f / G Sk(q,x) i s a Heck e form , the n ther e exist s

a uniqu e primitiv e for m g G Sk(q',x') wit h q' \ q suc h tha t x'( n) — x{n) a n d
Ap(n) = A/(n ) for al l n prim e t o q.
14.8. Twistin g modula r forms .
We hav e alread y mentione d tha t w e intend t o exploit th e eigenvalues A/(n ) of
primitive modula r form s a s arithmetic harmonics ; however , befor e the y ca n be used
effectively fo r such a purpose, i t i s necessary t o get information o n the eigenvalue s
themselves. Fo r this , followin g ou r strategy , i t i s natura l t o prob e thei r behavio r
when twiste d b y othe r harmonics . Th e cas e o f primitiv e Dirichle t character s i s
basic an d the results ar e nice.
PROPOSITION 1 4.1 9 . Let f e Mk{q,x) be a modular form with Fourier co-
efficients a,f(n), not necessarily a Hecke form, q* the conductor of the Dirichlet
character x an d let IJJ be a primitive Dirichlet character modulo r. Let f ^ip be the
function on HI given by the Fourier expansion
(f®*p){z) = ^2^(n)a f(n)e(nz).

Then f <8>ip is also a modular form, more precisely f <8>ip G Mk(N, x^ 2 )? where the
level N is the least common multiple of q, g* r and r2 . If f is a cusp form, then so
is f ® ip.
PROOF. Sinc e ip is primitive , w e hav e fo r an y n th e expressio n fo r ip(ri) in
terms o f additive character s
u e
r(i/j)ip(n) = J2 ^( ) yy)
u (mo d r)

where r(yj) ^ 0 is the Gauss sum , whence

2
(14.52) {f®iP){z) = T®)-1 J2 /( +
u (modr )

It i s no w a matte r o f writing , fo r 7 = 1 , I i n To(N), th e matri x identit y i n

SL(2,R),

,0 l ) \ c d) \ c d -cj^ ) { 0 x
to obtai n

{{f®^)\kl){z)=x{d)T{^)-1 ] T 4{u)f(z + ^p)

u (mod r )

= x(#M 2 (/®^)W
which show s the modularity. A s for /<8>^ being a cusp for m whe n / is , the criterion
of Exercis e 1 applies immediatel y fro m (1 4.52) . •

By multiplicativit y o f tp and th e formul a (1 4.47) , w e se e tha t th e operatio n

of twistin g b y ip takes Heck e form s t o Heck e forms . O n the othe r hand , i t migh t
 14. H O L O M O R P H I C M O D U L A R F O R M S 37 7

not tak e a primitiv e for m t o a primitiv e form , becaus e th e leve l N a s describe d i n

Proposition 1 4.1 9 migh t no t b e optimal . Thi s happens , fo r instance , whe n / = f x
is on e o f th e form s o f weight 1 associated t o a n ideal-clas s characte r \ m Sectio n
14.3. Indee d sinc e th e Fourie r coefficient s o f f x ar e supporte d o n integer s whic h
are norm s o f ideal s o f th e field K = Q(y/D)^ s o twistin g b y th e field characte r
I/J = XD has n o effec t a t all , sinc e i t i s identically 1 on norm s o f ideals. B y Exercis e
5, however , ther e i s a uniqu e primitiv e for m g, wit h leve l dividin g TV , which i s suc h
that
ip(n)\f(n) = A p (n), fo r al l n coprim e wit h q.
In th e specia l cas e when q and r ar e coprime , however , ther e i s no problem an d
f ® ijj is primitiv e wheneve r / i s (a s migh t indee d b e expected) . I n thi s cas e th e
functional equatio n o f th e twiste d L-functio n take s th e followin g form .
PROPOSITION 1 4.20 . Let f E S^(q^x) be a primitive form andip be a primitive
Dirichlet character modulo r with (r , q) = 1 . Then f 0 i\) is primitive of level
N = qr 2 and the Hecke L-function of f 0 ij; is entire and polynomially bounded in
vertical strips. Moreover, the completed L-function (l^.Jf.3) satisfies the functional
equation
A(s, f ®ip) = i kwfjfA(k - s , 7 ® ^>)
where rjf is the eigenvalue of f for the operator W, and the root number w depends
only on x an d i>, namely
w = X\TW{q) .
r
The fact tha t ther e is no pole for an y of the twists of a cusp form show s that th e
sign o f th e Fourie r coefficient s ar e changin g randoml y an d independentl y o f thos e
of an y Dirichle t character . Th e lac k o f periodicit y o f th e Fourie r coefficient s i s a
welcome featur e i n analyti c numbe r theory . Se e [DuI2] , [DuI3] , [DuI4 ] fo r som e
applications o f twistin g a s a replacemen t o f positivity .
Twists ar e importan t als o a s a simpl e cas e o f Rankin-Selber g convolutio n
GL(2) 0 GL(1 ) (se e Chapte r 5) , an d the y occu r i n so-calle d convers e theorem s
which characteriz e modula r form s b y mean s o f analyti c propertie s o f L-functions .
The first convers e theore m usin g twist s i s du e t o Wei l [We3] . Deepe r variant s ar e
important i n proving instances o f Langlands functoriality , se e the survey b y Cogdel l
in [BG] . We quot e Weil' s result :
THEOREM 1 4.21 . Let
S
Li(s) = ] T a{n)n~ s, L 2(s) = J2 H™)™~

be two Dirichlet series absolutely convergent for R e (s) > C for some C > 0 .
Assume that there exists integers k ^ 1 , q ^ 1 and M > 0 such that for any ip
primitive modulo m, with (m,Mq) = 1 , the Dirichlet series
_s s
L(f ® ij), s) = ^2 a ( n ) ^ ( n ) n , L(g 0 T/> , S) = ^ 6(n)^(n)n~

admit analytic continuation to entire functions bounded in vertical strips such that
the functions
A ( / ® V>, s) = (27T)- sT(s)L(f ® V, s), A(g ®^,s) = (2w)~ sT(s)L(g ® </>, s)
378 14. H O L O M O R P H I C MODULA R FORM S

are entire and satisfy the functional equation

A(/<g ) i/;, s) = w^(qm 2)k/2-sA(g 0 ^ , k - s),

with
w^p = i*x(m)^(g)T(^) 2 r _1 .
or
T/ien t/ier e ezzst e a cws p /orr a / G Sk(q,x) f some \ modulo q such that
L(f1s) = L 1 (s) andL(Wf,s) = L 2(s).
14.9. Estimate s fo r th e Fourie r coefficient s o f cus p forms .
The questio n o f th e siz e o f th e Fourie r coefficient s (o r Heck e eigenvalues ) o f
cusp form s i s on e o f th e classica l problem s i n th e stud y o f automorphi c form s fo r
themselves. Conside r / e Sk(q,x)- B y th e criterio n fo r cus p form s o f Exercis e 1
and th e Parseva l formul a w e obtain, fro m th e Fourie r expansio n o f / ,

$>/(n)| 2 e - 4 ™"= f\f{z)\ 2

dx^y-k

for an y y > 0 , s o tha t fo r an y N ^ 1 ,

Y,\af(n)\2«y-ke^y

and b y choosin g y = TV -1 w e derive a boun d fo r th e second-momen t o f th e Fourie r

coefficients
2
(14.53) V J | a/(n)| «iVfc.

Hence th e trivia l boun d (1 4.1 0) , namel y a>f(ri) < C n^/ 2 , follow s b y positivity .
This las t ste p i s rather crud e an d i t i s therefore expecte d tha t th e resultin g boun d
for th e individua l coefficien t cif(n) ca n b e improved. O n th e othe r hand , th e boun d
obtained o n averag e turn s ou t t o hav e alread y th e correc t orde r o f magnitude , a s
shown b y th e propertie s o f th e Rankin-Selber g L-functions . Thus , i t show s tha t
af(n)1 o n average , i s of orde r n~^~ .
The statemen t tha t a/(n ) < C n~^~+£ fo r an y e > 0 holds individuall y i s know n
as the Ramanujan-Petersso n Conjecture . Fo r primitiv e cus p forms , thi s was prove d
by P . Delign e [Del ] fo r k ^ 2 a s a consequenc e o f th e Rieman n Hypothesi s fo r
varieties ove r finite fields (th e Wei l conjectures) , i n th e ver y shar p for m

(14.54) |A/(n) | ^ r ^ n ^

for an y n ^ 1 , where r i s the diviso r function . Th e correspondin g formul a fo r k = 1

holds an d i s due t o Deligne an d Serr e [DeSe] . Thi s i s an extremel y powerfu l result ,
in particular , becaus e i t i s completel y uniform . I f / i s no t primitive , b y writin g i t
as a linea r combinatio n o f Heck e forms , w e deriv e fro m (1 4.54 )

a
/ ( n ) ^ r(n)n~2 _ ,

but th e implie d constan t no w depend s o n / quit e badly .

 14. H O L O M O R P H I C M O D U L A R F O R M S 37 9

It i s often convenien t t o normalize th e Fourier coefficient s s o as to extract th e

expected orde r o f magnitude, writin g

(14.55) f(z) = y^a/(n)n~2~e(nz )

instead of (14.9), and similarly for primitive form s wit h A/(n) . Thi s has the effect of
putting the critical line of the L- function a t Re (s) — |, an d the functional equatio n
then relate s A(/ , s) an d A(/ , 1 — 5). Thi s normalizatio n i s mos t convenien t i n
analytic numbe r theor y an d will b e in effect fro m no w on, except wher e explicitl y
stated. I t i s als o th e normalizatio n use d i n Chapte r 5 , se e Section 5.1 1 for the
necessary adjustments . I n this context , Deligne' s boun d take s th e form |A/(n) | ^
r(n) fo r / primitive , an d the Hecke L- function become s

£ A/(n)n- s = n a - xf(p)p~s + x{p)p- 2srl-

n^lp

In recen t application s (fo r instance, o f the "amplificatio n method") , th e need

for lowe r bound s o n the Fourie r coefficient s ha s also arisen . Sinc e a n individua l
lower boun d i s impossible, a s there exis t form s / (primitive ) an d n (coprim e wit h
the level ) wit h Ay(n ) = 0 , this ca n only b e expected t o hold o n average. Indeed ,
from th e Rankin-Selber g metho d w e know tha t th e upper-boun d (1 4.53 ) ca n be
sharpened t o the asymptotic (remembe r tha t w e now normalize af(n) a s in (14.55))

(14.56) £ | a/ (n)|2= C/ iV + 0(iVf) ,

with cj > 0 the implie d constan t dependin g o n / . I n particular , fo r a fixe d / ,

the Fourie r coefficient s ar e not too small to o often. However , th e dependence of
the constan t o n / i n thi s formul a i s no t explici t enoug h an d so the uniformit y
which i s often require d fo r applications i s not present. I n this contex t a very usefu l
observation i s that i f / i s a primitive for m

\f{p2)-xf{p)2 = x(p).
From thi s simpl e formul a w e see that fo r p not dividing the level q it is not possible
for Xf(p) an d A/(p 2 ) t o be simultaneously ver y small ! Thi s i s very usefu l becaus e
it i s completely unifor m i n all parameters involved . Her e i s an example o f the us e
of thi s formula .
P R O P O S I T I O N 1 4.22 . Let f e S%(q,x) be a primitive form, \f(n) its Hecke
eigenvalues. There exists a sequence of complex numbers c(n) such that
-1
(14.57) ] T c(n)A/(n ) x V ^ l o g T V ) ,

(14.58) ] T |c(n)| 2 x y/NilogN)- 1

if N ^ > (log2g)2 , where the implied constants are absolute.

380 14. H O L O M O R P H I C MODULA R FORM S

I t is enough t o take c(n) — \{p) it n = p2 ^ N, c(n) = ~x{p)^f(p) i f

PROOF.
n = p ^ y/N, an d c(n) = 0 otherwise. The n th e left sid e o f (14.57) i s equal to

2
£ x(p)(A/(p )-A,(p)2) = J2 l^v/iVOogiV)- 1

by Tchebyshev' s estimat e fo r ir(yJW) an d u;(g) < C (log2g)(loglog3g)_1. O n the

other hand , c{n) satisfies (1 4.58 ) b y Deligne's boun d (1 4.54) . •

R E M A R K . Pro m the proof on e sees that th e numbers c(n) ca n be chosen to have

support o n primes o r squares of primes an d that \c(n)\ ^ 2. Thes e extr a propertie s
are no t important bu t could b e convenient fo r technical reasons .
14.10. Average s o f Fourier coefficients .
The Petersso n formul a (Propositio n 1 4.5 ) i s very usefu l i n applications; it per-
mits averagin g ove r natura l familie s o f automorphic forms , whic h ar e so to spea k
"complete". I n doing so, it shows that th e Fourier coefficient s o f automorphic form s
behave a s "harmonics", the y ar e nearly orthogona l (se e als o Sectio n 7. 3 for the un-
derlying philosophy). Thi s is also very true with the similar effect s o f the Kuznetso v
formula (se e Chapter 16).
It i s especially importan t t o deal wit h coefficient s o f primitiv e forms , whic h
benefit als o fro m multiplicativ e propertie s (amon g othe r things) . Th e issue arise s
that i n general ther e i s no basis o f Sk(q,x) ° f primitive forms . However , takin g a
basis o f Hecke form s (se e Propositio n 1 4.1 1 ) , say Hk(q,x)i a n d denotin g

their Fourie r expansion , Propositio n 1 4. 5 yield s

COROLLARY 1 4.23 . Let q ^ 1 , k ^ 2 . Let x be a character modulo q and
Hk(q,x) an y Heche basis of Sk(q,x)- F° r an V m, n ^ 1 , we have
(14.59)
V^ Xf(n)Xf(m) = <5(ra,n) + 2ni~k \ J c~ 1
Sx(m, n ; c)Jk-i ( ) •
feHk(q,x) 0 0
c=0 (mo d q)

where the superscript h means that f has been spectrally normalized to be of Peters-
son norm 1 . Without normalization this means the terms of summation are appro-
priately weighted, namely

<"*» E * . , - $ j^ E Z-
feHk{q,x) v J
feH UJ
"
k(q,x)

It i s important t o know ho w goo d the approximate orthogonalit y ca n really be,

that is , to estimate the series of Kloosterman sums . A simple but effective treatmen t
yields th e following corollar y (se e Theore m 1 6. 7 for a more refine d estimate) .
 14. H O L O M O R P H I C M O D U L A R F O R M S 381

COROLLARY 1 4.24 . With notation as above, we have for any m, n ^ 1 ,

k
(14.61) Yl A / (n)A / (m) = <5(ra,n )
feHk(q,x)
i.

+ 0^r3((m,n))(m,n,g)2(mn)4^1 og^ l + ^ - — ^ j
qy/k ^ Vqk

where the implied constant is absolute.

PROOF. Us e th e estimat e Jk-i(x) < C min(l,:r//c ) an d th e Wei l boun d fo r

Kloosterman sum s (Corollar y 1 1 .1 2 )

(14.62) |5(m , n- c)\ ^ (m , n, c) 1 / 2 c 1 / 2 r(c).

I I I
For c = qr, say , thi s yield s the boun d (m , n, g) 2 (m, n, r) 2 (gr) 2 r(g)r(r). Henc e th e
sum o f Kloosterma n sum s i n (1 4.59 ) i s estimate d b y

(m,n,q)2-^- > - — / ( m , n , r ) 2 m m 1 ,- ^ .
V^ , v ^ \* kqr J
Hence th e resul t follow s b y th e elementar y estimat e

r^

which hold s fo r an y positiv e X. •

If th e spac e o f old form s vanishes , the n th e se t o f primitive form s Sfc(g , x)* i s a

distinguished basi s o f Sk(q, x)- I n particular , thi s i s so if x i s primitive fo r instance .
However, w e wil l discus s furthe r th e specia l cas e wher e q is prime , \ i s trivia l
and k < 1 2 , becaus e thi s wil l b e use d i n Chapte r 2 6 fo r studyin g non-vanishin g o f
L-functions. I n tha t cas e S%(q) = Sk(q) becaus e 5^(1 ) = 0 (th e firs t cus p for m fo r
SL(2,Z) i s o f weigh t 1 2) . I f / G Sk{q)* i s a primitiv e for m i t ha s rea l eigenvalues ,
and sinc e q i s squarefre e th e sig n ej = ± 1 o f th e functiona l equatio n i s give n b y
sj — —Xf(q)q1^2 (Propositio n 1 4.1 6 ; note we have changed normalizatio n o f Fourier
coefficients slightly) . I t i s ofte n desirabl e t o b e abl e t o averag e ove r form s wit h a
fixed sign , sa y e — ± 1 . Thi s ca n b e don e b y insertin g a facto r ( 1 -f ££/) , an d
applying th e summatio n formul a twice :

= Y? A / ( m ) A 7 R - sq 1 ^2 ] r " X f(q)Xf(m)Xj^j.
ff

Because / i s primitive an d q is the level , Xf(q)Xf(n) = Xf(qn) fo r al l n. W e deduc e

38 2 14. HOLOMORPHI C MODULA R FORM S

PROPOSITION 1 4.25 . Let q be prime, 1 < k < 1 2 , e = ± 1. For any m , n ^ 1,

we /iai>e
(14.63) 2 ^ A / (m)A / (n) = 6(m, n) - eq 1/25(m1 nq)

+ 2m~ k Y^ c~ 1
(S(m, n ; c) - eq1/2S(m, nq; c))Jk-i ( 7Tx/mn
\.
c>0
c=0 (mo d q )

Here als o w e ca n estimat e som e o r all Kloosterma n sum s gettin g a n approxi-

mation t o th e diagona l symbo l i n term s o f th e Heck e eigenvalue s o f primitiv e cus p
forms.
COROLLARY 1 4.26 . Let q be prime, 1 < k < 1 2 and e = ± 1. For any m , n ^ 1
with (m , q) = 1 we have

(14.64) 2 ^ X f(m)Xf(n) = 6(m,n)

£f=£

2eiri~k v ^ 1 ^, _ N T /47 r /m? v

^ (r, * ^) =—i
l V "
9
1I1
/ , ,u (km(ran)
n ) 4 4/ / ((mn)
mn)4\ 4 \

and

(14.65) 2j : ' A / ( m ) ^ 0 = ,(m,„) + O(«)(^)* V , 1(l + (™)i

T/ie implied constant in both estimates is absolute.

P R O O F . Sinc e q does no t divid e m we hav e <5(ra , gn) = 0. Moreover , th e firs t
term involvin g Kloosterma n sum s o n th e righ t sid e of (14.63) i s alread y estimate d
in Corollar y 1 4.24 . Fo r th e second , w e write c = qr an d estimat e th e contributio n
of th e term s wit h q | r, by th e sam e metho d gettin g a bound whic h i s of smaller
order o f magnitude .
When (r,q) = 1 , the Kloosterma n su m factorize s
S(ra, n; qr) = S(mq, n ; r)S(0, m; q) = — S(mq, n ; r).
This give s (1 4.64) . Th e remainin g su m o f Kloosterma n sum s i n (1 4.64 ) ca n als o b e
estimated directl y b y usin g Weil' s bound , an d th e resul t i s give n i n th e erro r ter m
of (1 4.65 ) (thi s i s worse tha n th e erro r ter m i n (1 4.64 ) i n term s o f q). •

The erro r ter m o f (14.65) i s large r tha n th e erro r ter m i n (14.64) an d wil l no t
be sufficien t (just ) fo r application s i n Chapte r 26 . Bette r estimate s wil l b e derive d
from (1 4.64 ) b y exploitin g a cancellation i n sums o f Kloosterman sum s ove r th e
variables m and n.
 http://dx.doi.org/10.1090/coll/053/16

CHAPTER 1 5

SPECTRAL THEOR Y O F
AUTOMORPHIC FORM S

15.1. Motivatio n an d geometri c preliminaries .

Abelian harmoni c analysis , a s developed i n the previous chapter s (se e especially

Chapter 4) , give s powerfu l mean s o f transformin g o r estimatin g summation s ove r
integers, o r lattic e point s i n variou s geometrica l domains . However , i n analyti c
number theory , i t i s desire d t o hav e suc h tool s fo r sum s ove r othe r subset s o f
integers whic h ar e no t define d b y analyti c condition s alone . Commo n amon g thes e
are sum s ove r (a , 6, c, d) G Z4 restricte d b y th e determinan t equatio n

where h i s fixed. Fo r example , on e encounter s th e determinan t equatio n whe n

investigating th e powe r moment s (wit h amplification ) o f L-functions o n th e critica l
line t o handl e off-diagona l contributions .
Although th e classica l Fourie r analysi s applie s t o th e determinan t equatio n
quite well (especiall y whe n enhance d b y estimates fo r Kloosterma n sums ) th e spec -
tral theor y o n th e hyperboli c plan e produce s muc h stronge r results . Th e ke y poin t
is that fo r th e analysi s o f the determinant equatio n th e automorphi c form s ar e mor e
natural harmonic s tha n th e exponentia l function s o r Dirichle t characters . The y en -
ter th e moder n analyti c numbe r theor y throug h man y othe r doors . Fo r example ,
the ol d problem s concernin g th e clas s grou p o f quadrati c fields (Lagrange , Gauss ,
Dirichlet) canno t b e properl y understoo d withou t speakin g o f automorphi c theory .
Another incentiv e fo r ne w analyti c tool s i s to stud y automorphi c L-functions .
References fo r thi s chapte r ar e fo r instanc e [Bu] , [1 5] . Th e basi c theor y wa s
constructed b y Maas s [Ma ] an d Selber g [S6] .

As i n th e previou s chapte r w e conside r first th e grou p G — SX(2,R). Jus t

as Z i s a discret e subgrou p o f R , fro m whic h th e theor y o f periodi c function s an d
Fourier serie s arises , Y = To(l ) = SX(2,Z ) i s a discret e subgrou p o f G. Studyin g
the determinan t equatio n wit h h = 1 by harmoni c analysi s mean s takin g a kerne l
(smooth, compactl y supported ) k : G —> • C an d finding a summatio n formul a a l a
Poisson fo r th e functio n K : G —> C give n b y

K{g) = YJKl9)

which i s T-periodic ; JC{^g) — IC(g) for 7 G I\

383
38 4 15. S P E C T R A L THEOR Y O F AUTOMORPHI C FORM S

However, a usefu l simplificatio n i s possibl e b y performin g firs t a n ordinar y

Fourier expansio n wit h respec t t o th e compac t abelia n subgrou p

(which i s isomorphi c t o th e circl e R/27rZ) . An y functio n k ca n b e expande d a s a

series

where k m satisfie s a simple transformation rul e km(gr) = e(mO)k(g) fo r al l r = TQ G

K. Therefore , instea d o f function s o n G , w e can loo k a t function s o n th e quotien t
GjK. Thi s quotien t i s diffeomorphi c t o th e Poincar e uppe r half-plan e b y

/A : - > M
(15.1) a b \ ai +b
c d) ci +d
Instead o f doin g thi s reductio n explicitly , w e work o n H I and hav e G ac t o n it .
Recall fro m Chapte r 1 4 tha t th e Poincar e uppe r half-plan e i s H I = {z G C |
Im (z) > 0} . W e wil l se e i t no w a s a mode l o f th e hyperboli c plane , a complet e
Riemannian manifol d o f dimensio n 2 wit h constan t negativ e curvatur e — 1 when
equipped wit h th e Poincar e metri c
2
(15.2) ds = y~ 2{dx2 + dy 2) = y~ 2
\dz\2.
The geometr y o f H I with thi s metri c i s accessibl e becaus e H I has a larg e isom -
etry group . Indeed , th e actio n (1 4.2 ) o f G = 5L(2,R ) o n H I by linea r fractiona l
transformations
^ ^ az + b fa c
(lb.6)
vJ gz
u = - , ro ra =7
cz + du \c d
is a n actio n b y isometries : thi s follow s directl y fro m th e formul a abov e an d fro m
(15.4) d{gz) = (cz + d)~ 2dz.
Hence G C Isom(HI) , th e isometr y grou p o f HI , wit h th e cente r o f G actin g
trivially. Actually , PSL(2,R ) = SX(2,R)/{±1 } i s th e ful l grou p o f orientation -
preserving isometrie s o f HI , an d Isom(HI ) i s generate d b y G an d th e reflectio n a :
z ^ > —~z.
Using thi s larg e isometr y group , man y geometri c propertie s o f H ca n b e estab -
lished fairl y quickly :
(1) Angle s i n H ar e th e sam e a s euclidea n angle s (simpl y becaus e th e metri c
is conforma l t o th e euclidea n metri c |dz| 2 ).
(2) Th e hyperboli c geodesie s i n HI are the half-circle s (i n the euclidea n sense )
orthogonal t o th e "boundary " MUoo , an d th e vertica l line s R e (z) = constant . I n
particular, suc h a circl e o r lin e i s transforme d int o anothe r b y g G G.
(3) A geodesi c circl e o f cente r w, {z G HI | d(z,w) = r } , is als o a euclidea n
circle (wit h differen t cente r an d radius) .
(4) Th e distanc e (i n th e hyperboli c metric ) i s explicitly give n b y
\z-w\ + \z-w\
d(z, w) = lo g = ^ -,
 15. SPECTRA L THEOR Y O F AUTOMORPHI C FORM S 385

but i t i s actuall y ofte n mor e convenien t t o wor k wit h th e functio n u(z, w) give n b y

(15.5) u(z,w) =»Mz,v,)) = 4 ^-)£H

where u : R +— > R + i s th e functio n o f th e distanc e

(15.6) u{d) = i(cosh d - 1 ) = 2(sin h - ) .

REMARKS. (1 ) On e ha s t o b e carefu l tha t th e topolog y o f H I induced b y th e

Riemannian metri c i s differen t fro m it s (euclidean ) topolog y a s a subse t o f C . Fo r
instance, H I is complet e wit h th e Riemannia n metric , bu t o f cours e no t wit h th e
euclidean one .
(2) Notic e tha t th e compac t subgrou p K i s th e stabilize r o f i G HI, hence th e
isomorphism G/K ~ H I described abov e i s simpl y th e ma p g i-> gi o n G.

Recall tha t th e invarian t measur e associate d t o th e Poincar e metri c i s

(15.7) dfi{z) — y~ 2dxdy
(see (1 4.4)) .
When w e com e t o harmoni c analysi s o n quotient s o f H I by discret e subgroups ,
the congruenc e subgrou p r 0 (#)\HI i n particular , se e Sectio n 1 4.1 , thinking o f th e
abelian cas e o f R an d it s ow n discret e subgroups , man y ne w phenomen a occur .
First, an y discret e subgrou p o f R i s necessaril y o f th e for m aL wit h a 6 R , an d
they ar e therefor e basicall y "al l th e same" . Thi s similarit y allow s us , fo r instance ,
to deduc e quickl y a Poisso n summatio n formul a fo r a n arithmeti c progressio n fro m
the on e fo r Z (se e Sectio n 4.3) . O n th e othe r hand , ther e ar e ver y subtl e ques -
tions concernin g harmoni c analysi s o n ro(#)\H I (simila r t o thei r complex-analyti c
differences mentione d i n Chapte r 1 4) .
15.2. Th e laplacia n o n H .
Harmonic analysi s o n H I is associate d wit h th e hyperboli c Laplac e operator ,
which i s give n b y
/ B 2 ft 2
\

This i s a G-invarian t differentia l operator , i.e . fo r an y functio n / whic h i s twic e

differentiable w e hav e
A ( / | g) = A / .
Here an d hereafte r th e actio n o n th e righ t o f G o n function s i s o f cours e give n b y

(/ I g)(z) = f(gz).
There ar e man y way s t o analyz e A o n M , depending o n whic h coordinate s ar e
chosen. Usin g rectangula r coordinate s z = x + iy an d workin g b y separatio n o f
variables, on e consider s function s / o n H I which ar e periodi c o f perio d on e i n th e
variable x\ the n / ha s a Fourie r expansio n

(15.9) /(* ) = £/n(2/)e( r

38 6 15. S P E C T R A L THEOR Y O F AUTOMORPHI C FORM S

with coefficient s

(15.10) fn(y)= f f{x + iy)e(-nx)dx.

Jo
Applying th e Laplac e operato r on e get s
2 2
Af(z) =y*J2 (4n n fn(y) - %{y))e(nx).
nez

Of specia l concer n ar e th e eigenfunction s o f A , th e function s / suc h tha t A / = Xf

for som e A E C . I t i s bes t t o writ e th e eigenvalu e A in term s o f a paramete r s
such tha t A = s( l — s). Becaus e A i s a n ellipti c differentia l operator , an y suc h / i s
automatically real-analyti c o n H .
Continuing wit h a functio n / periodi c i n x , w e se e tha t i t i s a n eigenfunctio n
with eigenvalu e A i f and onl y i f f n i s a solution t o th e ordinar y differentia l equatio n

r + ( A _ w )/ = o.
For n = 0 all solutions ar e linea r combination s o f y s an d y l~s (replac e y l~s b y
l 2
y / \ogy i f s = 1 /2) . The n fo r n ^ 1 , tw o linearl y independen t solution s t o th e
equation ar e give n b y standar d Besse l function s
1 /2
fn(y) = 2y K1
s_ /2(27r\n\y)

and
fn(y) = 2y 1 '2I13_ /2(27r\n\y)
which ar e distinguishe d b y thei r asymptoti c behavio r a s y — > -foo . Thi s lead s t o
LEMMA 1 5.1 . Let f : M—>> C be an eigenfunction of A which is l-periodic in
x, with eigenvalue X = s( l — s), and suppose f satisfies the growth condition
f(z) = o(e 27ry ), as y -+ +oo.
Then there exist complex numbers a 7 b, and a n for n ^ 0 , such that

f(z) = ay s + by 1 '8 + y 1 '2 ^ a nKs_ /2(2n\n\y)e(nx)

1

(if s — 1/2, replace y l~s by y l 2

/ \ogy).
In particular , thi s explici t Fourie r expansio n implie s tha t / i s in fact o f at mos t
polynomial growth ; precisely , w e hav e
/(2)«j/ff + j /
i
-
for y —> +oo , wher e a = R e (s) a s usual .
15.3. Automorphi c function s an d forms .
We no w conside r a discret e subgrou p o f G fo r whic h Vo l ( r \ H) < +o c (suc h
subgroups ar e calle d Fuchsia n group s o f th e first kind) . O f particula r importanc e
are th e Heck e group s To(q) fo r q ^ 1 .
We defin e differen t space s o f T-periodi c function s o n H :

^ ( r \ M ) = { / : H - + C | / | 7 = / , for al l 7 eT}
 15. S P E C T R A L T H E O R Y O F A U T O M O R P H I C F O R M S 38 7

(the spac e o f automorphi c functions) ,

L2(r\H) = {/ G A(r\u) | ll/ H = J \f(z)\ 2dn(z) < +^}

(the spac e o f square-integrabl e automorphi c functions) ,

AS(T\M) = { / e A(T\M) I A/ = 5( 1 - s)f}

(the spac e o f automorphi c forms , o r Maas s forms , wit h eigenvalu e A = s( l — s)).

Let a be any cusp of T and a a a scaling matrix as in (1 4.5) . Fo r any / G L 2 ( r \ H )
the functio n f a = f | a a i s thu s 1 -periodi c i n x, an d ha s a Fourie r expansio n a s
in (1 5.9) . W e defin e furthe r
2
Lg(r\H) = {fe L (T\M) | / fli0 - 0 fo r an y cus p a }
(the spac e o f cuspida l automorphi c functions) ,

A°8(T\E) = L 20(T\M) n A S(T\U)

(the spac e o f cus p form s wit h eigenvalu e 5( 1 — 5)) .

Note, i n particular , tha t an y automorphi c for m i n L 2 ( r \ H ) (o r wit h moderat e
growth) admit s a Fourie r expansio n a t th e cus p 0 0 a s i n Lemm a 1 5.1 .
Our firs t goa l wil l b e t o describ e ho w t o decompos e element s o f L 2 (r\M) i n
"spectral" terms , simila r t o th e Fourie r decompositio n (serie s o r integrals ) o n R/ Z
or R . Thi s wil l b e don e i n term s o f th e eigenvalue s o f th e laplacia n actin g o n
automorphic functions .
We defin e th e domai n V o f A t o b e

V = { / G L2 ( r \ H ) I / an d A / ar e o f clas s C°° an d bounded} ,

thereby turnin g A int o a n unbounde d linea r operato r o n L 2 ( r \ H ) , define d o n th e

dense domai n V. Usin g Stoke' s formula , i t i s show n tha t A i s symmetri c an d
positive, an d b y general functiona l analysis , i t therefor e ha s a self-adjoint extensio n
to al l o f L 2 (r\M).
If the quotien t r \ H wer e compact , the n th e genera l Hilber t spac e theory woul d
imply tha t A ha s pur e poin t spectrum , an d tha t L 2 ( r \ H ) i s spanned b y eigenfunc -
tions o f A , a s i s th e cas e wit h R/Z . Du e t o th e lac k o f compactness , thing s ar e
more complicate d becaus e a continuou s spectru m exist s but , o n th e othe r hand ,
the presenc e o f the cusp s give s rise to Fourie r expansions , whic h ar e powerfu l tools .
15.4. Th e continuou s spectrum .
We start th e spectra l decompositio n b y constructing a space o f "obvious " auto -
morphic functions , namel y thos e constructe d b y th e familia r averagin g technique .
More precisely , fo r a tes t functio n ip : R +— > C whic h i s smoot h wit h compac t
support an d a cus p a , w e defin e th e incomplet e Eisenstei n serie s

(15.11) E a(z^)= ^2 ^(Ima-Sz )

7er a \r

which i s obviousl y i n L 2 (r\M) sinc e ip has compac t support . O f course , fo r th e

same reason , E a(',ip) ca n neve r b e a n eigenfunctio n o f A . However , usin g Melli n
388 15. SPECTRA L THEOR Y O F AUTOMORPHI C FORM S

inversion, w e find tha t

(15.12) E a(z, V ) = ^- J E a(z, s)^(s)ds

where a > 1 , - 0 is th e Melli n transfor m o f ip and E a(-,s) i s the Eisenstei n serie s

Ea(z,s) = Yl (IHKT-S^) -
7er a \r

the serie s bein g absolutel y convergent , uniforml y o n compac t sets , fo r R e (s) > 1 .
Because A(I m z) s = s( l — s)Im (z) , the Eisenstein serie s E a(z, s) i s an eigenfunctio n
of A , bu t unfortunatel y i t i s not squar e integrabl e s o it canno t b e use d directl y fo r
the spectra l decomposition .
We define a new space £ ( r \ H) t o be the closure in L 2 (T\H) o f the space spanne d
by th e incomplet e Eisenstei n series . Clearly , A act s o n £ ( r \ H ) . Th e solutio n t o
the spectra l decompositio n i n thi s subspac e lie s i n th e meromorphi c continuatio n
of the Eisenstei n serie s (wit h respec t t o th e s- variable) t o th e whol e complex plane ,
which wa s prove d b y Selberg . I n th e cas e o f To(q), th e analyti c continuatio n ca n
be see n quit e easil y b y computin g explicitl y it s Fourie r expansio n a t th e cus p oo .
Let a , b b e tw o (no t necessaril y distinct ) cusps . The n E a(-,s) ha s a Fourie r
expansion a t th e cus p b , which w e writ e

Ea(abz, s) = S aibys + <p a,b?/1 _s + y 1 /2 Yl <Pa,b{n, s)K s_1 /2{27r\n\y)e(nx).

The squar e matri x o f constan t term s <£(s ) = ((f a,b(s))a,b i s calle d th e scatterin g
matrix o f T .
In th e cas e o f T = 5L(2 , Z), the Fourie r coefficient s ar e (w e drop th e subscript s
a, b since ther e i s onl y on e cusp )

r ( s - i/2 ) c(2 s - i )
<p(s) = \J-K- T(s) C(2a ) '
-1/2

ab=\n\

and th e meromorphi c continuatio n ca n indee d b e see n fro m tha t o f th e Rieman n

zeta functio n ((s). I n thi s cas e th e ful l Fourie r expansio n o f E(z, s) i s
(15.13)
oo

E(z, s) = 0(s)y s + 0( 1 - s)y 1 -s + V^ r s_ /2(n)K1

1 s_ /2(27rny) cos(2^nx) .
l

For To(q) i n general , simila r expression s hol d i n term s o f Dirichle t L-functions .

Let R e (s) ^ 1 /2 . I t i s easil y see n tha t i n thi s regio n E(z, s) ha s onl y a simpl e
pole at s = 1 with constant residu e V~ l, wher e V = Vo l ( r \ H ). Fo r general fuchsia n
groups o f th e first kind , ther e migh t b e othe r pole s i n th e half-plan e R e (s) ^ 1 /2 ,
although onl y a finite number , al l i n th e interva l 1 / 2 < s ^ 1 , non e o n th e lin e
Re (5) = 1 /2 . Al l residue s ar e i n L 2 (r\EI) an d on e denote s 7^(T\H ) th e subspac e
that the y span ; thi s i s calle d th e spac e o f residua l spectrum .
 15. S P E C T R A L T H E O R Y O F A U T O M O R P H I C F O R M S 38 9

For To(q), on e furthe r see s that s = 1 is the onl y pol e o f E(z, s) , with constan t
residue
Vq = Vol(r 0 (g)\H) = | g j ] ( 1 +p' 1 ).
p\q

Using th e self-adjointnes s o f A , on e ca n prov e th e functiona l equatio n o f the Eisen -

stein serie s
E(z,s) = $(s)E(z,l-s)
where E(z,s) i s th e colum n vecto r (E a(z,s))a; correspondingly , fo r th e scatterin g
matrix w e hav e
$ ( s ) $ ( l - s ) =Id .
One show s tha t th e orthogona l complemen t t o 7£(r\H ) i n £ (r\M) ha s continuou s
spectrum o f multiplicity equa l to the numbe r o f cusps, describe d b y the eigenpacke t
of Eisenstei n serie s fo r R e (s) = 1 /2 . Fo r To(g) , w e therefore derive :

THEOREM 1 5.2 . Let f e £(T 0(q)\E). Then f has the expansion

E
(15.14) f(z) = ±- I f(z)dfi(z) + J2 J- I </ > ^ \ + it))E a(z, \ + it)dt

where V q — Vo l (ro(g)\lH). This formula is valid in L 2 sense, and the integrals are
absolutely convergent point-wise and uniformly on compact sets if f G V.
The mai n poin t i n th e proo f i s that , whil e E a(-,s) i s no t square-integrabl e
(because o f ter m S a^ys + ip a^(s)y1 ~s i n th e Fourie r expansio n a t th e cus p b) , i t
is almos t s o whe n R e (s) = 1 /2 . Indeed , on e ca n sho w tha t E a(a^z,s) < C y/y
when y —> • oo , an d tha t th e extr a integratio n i n t i n (1 5.1 4 ) save s a facto r logy,
which i s sufficien t t o tur n th e right-han d sid e int o a n L 2 function . O n th e othe r
hand, on e migh t b e tempte d t o sa y tha t (1 5.1 2 ) o n th e lin e R e (s) = 1 / 2 give s th e
spectral decomposition , bu t i t i s no t tru e becaus e th e coefficien t ip(s) i s no t th e
inner produc t (£^ a(-,,0), Ea(-, s)) a s i t shoul d b e (sinc e (E a(-,s),Ea(-,s)) diverges ,
multiplying b y th e Eisenstei n serie s an d integratin g i s no t permitte d i n (1 5.1 2) . I t
is necessar y t o appea l t o th e functiona l equatio n t o ge t th e prope r formul a whic h
(despite what (1 5.1 2 ) migh t suggest ) show s that i t i s necessary t o us e the Eisenstei n
series a t al l cusp s b to ge t th e expressio n o f the incomplet e Eisenstei n serie s a t th e
cusp a (se e [1 5] , 22.1-22.3).
In th e cas e o f a genera l subgrou p T , th e Eisenstei n serie s remai n i n man y
respects rathe r mysteriou s objects , bu t the y ar e comparativel y les s s o fo r To(q),
where th e explici t Fourie r expansio n yield s crucia l estimate s i n a straightforwar d
way, wit h th e consequenc e that , fo r th e application s i n thi s book , th e continuou s
spectrum wil l b e easil y manageable .
15.5. Th e discret e spectrum .
We no w tur n ou r attentio n toward s th e complemen t o f th e spac e £ ( r \ H ) .
LEMMA 1 5.3 . The orthogonal complement of £(T\E) is the space L\ ( r \ H ) of
cuspidal automorphic functions.
This follow s b y a simpl e computation , whic h w e write dow n completel y t o giv e
a feelin g o f th e whol e theory : w e simpl y comput e th e scala r produc t (/ , E a(-, ip))
390 15. SPECTRA L THEOR Y O F AUTOMORPHI C FORM S

of a n L 2 -automorphic functio n agains t a n incomplet e Eisenstei n serie s

(f,Ea(.^)) = [ f(z)E a(z,'ip)dfi(z)

JF

= I f(z) X ) ^(Ima-S*)^* )
F
7^r a \r
1
Ef Ff{^{lma-
1 z)d^z)
JF
ieTa\T

which b y automorph y o f / transform s t o

(f,Ea(-,il>))= Yl I f{<r azj$(z)dn{z)

1/ >+ 00
2
/ f((J az)^(z)y~ dydx
/o Jo
since th e set s o~ x^F ar e disjoin t an d mak e a partitio n o f th e stri p 0 < R e (z) < 1
(see (1 4.5 ) o r thin k o f a = oc) .
This w e can no w writ e a s
2
(15.15) (f,E a(.,r(>))= / faAy)Hv)y~ dy
Jo
where f a,o(y) i s the constant ter m i n the Fourier expansio n o f / a t a , an d the lemm a
is no w clear .
Again, A act s o n the spac e LQ(T\IH[) . Bu t no w on e ca n sho w that th e resolven t
of th e laplacia n i s a compac t operato r o n Lo(r\H) . Fro m thi s i t readil y follow s
that A ha s pur e poin t spectru m o n Lo(r\H ) an d thu s thi s spac e i s spanne d b y
eigenfunctions o f A , whic h ar e calle d Maas s cus p forms .
Because A i s self-adjoint, th e eigenvalue s o f A mus t als o b e non-negative , an d
indeed the y ar e positiv e sinc e th e eigenvalu e 0 correspond s t o bounde d harmoni c
functions o n H , whic h ar e necessaril y constant . I n term s o f s thi s mean s tha t
either 1 / 2 < s ^ 1 , o r R e (s) = 1 /2 . Ther e ca n b e onl y finitel y man y eigenvalue s
A = 5( 1 — 5 ) wit h 1 / 2 < s ^ 1 , an d the y ar e calle d exceptiona l eigenvalue s fo r
T. Althoug h i t i s fairl y eas y t o construc t fuchsia n group s havin g arbitraril y man y
exceptional eigenvalues , Selber g conjecture d tha t the y d o no t exis t fo r congruenc e
(arithmetic) groups .
SELBERG'S EIGENVALU E CONJECTURE . For any q > 1 , we have

where Xi is the smallest eigenvalue of A acting on LQ(r 0(q)\lHI).

Further, Selber g prove d
THEOREM 1 5.4 . For any q ^ 1 , we have A i ^ ^_ .
This ha s bee n improved , a s mentione d i n Sectio n 5.1 1 , the bes t resul t t o dat e
being A i ^ 975/409 6 b y Ki m an d Sarna k [KiS] .
As wil l appea r clearl y i n Section s 1 5. 8 an d 1 5.9 , exceptiona l eigenvalue s ar e
somewhat analogou s t o exceptiona l zero s o f Dirichle t L-function s (se e (5.51 )) , an d
 15. S P E C T R A L T H E O R Y O F A U T O M O R P H I C F O R M S 391

too man y o f the m ca n compromis e analyti c estimates . However , Theore m 1 5. 4 i s

analogous t o a zero-free stri p an d i s thus muc h stronge r tha n wha t i s known i n thi s
case. Th e mor e fruitfu l analog y i s wit h th e Ramanujan-Petersso n conjecture , an d
in representatio n theoreti c term s ther e i s a commo n formulation .
For genera l T , cus p form s ar e shroude d i n mystery ; indeed , i t i s b y n o mean s
clear tha t the y mus t exis t (se e [PS]) , excep t i n certai n case s wher e thi s i s ensure d
by symmetr y (fo r instance , i f th e reflectio n z —
i > • — ~z descend s t o th e quotient , od d
functions mus t com e fro m od d cus p form s sinc e al l Eisenstei n serie s ar e even) . Fo r
congruence subgroups , however , i t wa s show n b y Selber g usin g th e trac e formul a
that no t onl y d o cus p form s exist , bu t the y ar e i n a certai n sens e "dominating "
over th e Eisenstei n series . Thei r presenc e i s measure d b y th e countin g functio n

Nroiq)(T) = \{j | M<T} |

which i s shown t o satisf y th e Wey l La w

iVr„M(T) = v ° " r 4 f w ^ + 0(^rio S,r)

for T ^ 2 , wher e th e constan t i s absolute . Henc e ther e ar e man y eigenvalues ,
on averag e abou t qT betwee n T an d T - f 1 , bu t th e proo f i s indirec t an d doesn' t
produce an y o f the correspondin g cus p forms !
15.6. Spectra l decompositio n an d automorphi c kernels .
From th e tw o previou s section s w e ca n deduc e th e ful l spectra l decompositio n
theorem fo r th e automorphi c laplacia n o n ro(tf)\HI .
THEOREM 1 5.5 . Let u 0, ui, U2, ... be an orthonormal basis of the residual
_i
and cuspidal spaces, i.e. u^ — constant = Yol(To(q)\M) 2 G 1 Z(To(q)\H.) with
eigenvalue A o = 0 and Uj G .4^.(ro(g)\IH) ; with eigenvalue Xj = Sj(l — Sj) for
j = 1 , 2, . . . . Then any f G L2(To(q)\H.) has the spectral decomposition

(15.16) f(z) = Y </ , u3)u3(z) + Y 7- j </ > E *('> \ + **)>£<*(* , \ + tydt

valid in L 2 sense, and also converging absolutely and uniformly on compact sets if
f G V. Moreover, the Parseval formula holds

(15.17) ll/ f = ]T |(/,^-)| 2 + Y, T I K/> E*('> \ + it ))\2dt

We wil l no w appl y Theore m 1 5. 5 t o ge t a kin d o f Poisso n summatio n formul a

for quotient s o f th e hyperboli c plane . Fo r thi s w e conside r a kerne l functio n i n
two variables 1 fc : H I x H I -> C , sa y continuou s an d compactl y supported , whic h
is a "point-pai r invariant" , i.e . k(gz,gw) — k(z,w) fo r al l g G G. Thi s mean s
that k(z^w) depend s onl y o n th e distanc e d(z,w). W e expres s thi s propert y b y
saying tha t ther e exist s anothe r functio n i n on e variabl e k : R +— > C suc h tha t
k(z,w) = k(u(z,w)) wher e u i s the functio n i n (1 5.5) .

The secon d variabl e i s usefu l becaus e ther e i s n o privilege d basepoint .

392 15. S P E C T R A L THEOR Y O F AUTOMORPHI C FORM S

Associated t o k is an invariant integra l operato r L acting o n functions o n HI by

Lf(z)= [ f(w)k(z,w)dn(w)

and, fro m th e assumption tha t k depends only on the distance, it follows that L is G-
invariant. I n particular, L map s automorphi c function s t o automorphic functions ,
and indee d fo r / G «4(ro(^)\H) whic h i s bounded w e have anothe r formul a

Lf(z) = J f(w)K(z,w)dn(w).

where th e kernel K i s defined by

(15.18) K(z,w)= Yl k(z^w).

jer0(q)/{±i}

The ne w kerne l K i s automorphic i n both variables : indeed , i n w by the averaging

construction, an d in z becaus e k is symmetric.
Applying th e spectral decompositio n i n the z-variable give s th e expression

K(z,w) = ^2(K(^w),Uj)uj(z)

+ Y, J " / W" > W )>E*('> \ + it))Ea(z, \ + it)dt

where th e coefficients (K(-,w),Uj) (respectivel y (K(-,w),E a(-^ + it))) ar e again

automorphic an d can be also expande d similarly . However , th e next lemm a show s
that the y ar e simply proportiona l t o the cusp form s Uj (respectivel y th e Eisenstein
series).

LEMMA 1 5.6 . Let k : R +— > C be smooth and compactly supported, k(z )w) =
k(u(z,w)) the associated kernel and L the integral operator as above. If f : H—> C
(not necessarily automorphic) is any eigenfunction of the laplacian with eigenvalue
A = 5( 1 — 5) , 5 = 1 / 2 + it, t G C, then f is also an eigenfunction of the integral
operator L, with an eigenvalue which depends on A and L, but not on f. More
precisely, we have

where A = h(t), the transform k \-> h being obtained by the following steps:

r+oc
q(v) = / k(u)(u — v)~ l'2du,
JV

g(r)=2q[(Smhr-)2),
irt
h(t) = f g(r)e dt.
 15. S P E C T R A L T H E O R Y O F A U T O M O R P H I C F O R M S 393

For a proof, se e e.g. [1 5 , Th. 1 .1 6] . Th e transfor m abov e is called th e Harish -

Chandra/Selberg transfor m fo r 5L(2,R). I t can b e inverted by

irt
9(r) = - ^ / h(t)e~ dt,
1

i r+oc
1 /2
k(u) = / (v-u)- dq(v),

sufficient condition s fo r h being

( h{t) = h(-t),
(15.19) I h is holomorphic i n |Im (t)\ < \ + 5,
2 s
[h(t)^(\t\ + ir -
for som e 5 > 0 . Thes e condition s ar e in fac t sufficien t fo r th e analysis o f the
automorphic kerne l K(z, w), i.e . th e restrictio n tha t k has compac t suppor t ca n be
relaxed.
We deduc e fro m thi s th e spectral decompositio n whic h wa s our goal.
T H E O R E M 1 5.7 . Let h be a function satisfying the conditions (1 5.1 9), k the
Harish-Chandra/Selberg transform of h and K the automorphic kernel on r 0 ( g )
associated to k. Then the following expansion holds

K(z,w) = Y^ k{u(z,~/w)) = ^h{tj)u j{z)uj{w)

ier0(q)/{±i} j7*o

+ E — - / h(t)E a{z, \ + it)Ea(w, \ + it)dt

where the convergence is absolute and uniform on compact sets.

It i s importan t t o hav e som e contro l o n the growth o f the eigenvalues and
eigenfunctions i n this formula . Th e followin g i s often sufficien t fo r first estimations .
PROPOSITION 1 5.8 . Let T ^ 1 and z e H . We have

E M*)i 2
+ E^ / T
\ E*(z> \ + «)i2 * « T2
+ T y(z)
J
\t3\<T a * ~T

where the implied constant depends on V alone and

y{z) = max maxim (a" 1 72:).
a jer
For th e proof, se e [15], Section 7.2.
15.7. Th e Selberg trac e formula .
In Sectio n 1 5.6 , w e have obtaine d th e spectral expansio n o f an automorphi c
kernel, an d in doin g so , we have see n tha t i t i s usefu l t o conside r th e invarian t
integral operato r L associate d t o it, namely

Lf(z) = J f(w)K(z,w)d fi(w)

394 15. S P E C T R A L THEOR Y O F AUTOMORPHI C FORM S

where K i s give n b y (1 5.1 8) . Anothe r formula , whic h i s als o a generalizatio n o f

the Poisso n summatio n formula , arise s whe n on e compute s th e trac e o f L , usin g
the spectra l expansio n o f K o n th e on e hand , an d it s definitio n o n th e othe r hand :
this i s the celebrate d Selber g trac e formula . O f course , th e correc t definitio n o f th e
trace o f an operato r o n a n infinit e dimensiona l spac e i s already a problem, bu t her e
we take a pragmatic approac h an d defin e th e trac e o f L (o r o f K) t o b e the integra l
of th e kerne l o n th e diagona l (whic h i s well-defined becaus e ou r kernel s ar e alway s
at leas t continuous) :

(15.20) TrK= f K(z,z)dfi(z).

If there were no Eisenstein serie s in Theorem 1 5.7 , the spectral expansion woul d
yield immediatel y

since (UJ) is a n orthonorma l basis , an d w e would obtai n th e pre-trac e formul a

jF
j ier

The right-han d sid e is still no t ver y explicit. T o go further, Selber g partitione d

the su m ove r 7 i n such a way that th e partia l sum s ar e computable. Thi s i s done by
dividing th e grou p int o it s conjugac y classe s (tha t thos e ar e non-trivia l i s a featur e
of th e non-commutativit y o f T). Fo r on e conjugac y clas s C , sa y o f 7 G T, w e hav e
C = {r~ V \ reT}
which i s i n bijectio n wit h Z(7)\T , wher e Z("Y) is the centralize r o f 7 i n T ,
Z(j) = { r G 7 I 7T = rj}
so th e contributio n o f th e clas s C t o th e su m i s equa l t o
l
E k(u(r- ^TZ,z)) = E k{u(p(Tz,Tz))
rGZ(7)\r rGZ( 7 )\r

which, b y invariance , give s th e contributio n

T r c X = / k{u{nz,z))dii(z)
JZ(~f)\M

to the trac e o f K. Th e advantag e w e gain in performing thi s summation ove r 7 G C

is tha t th e abov e expressio n fo r T r c K no w depend s o n th e conjugac y clas s o f 7
in ST(2,R) , no t i n I \ Henc e i n furthe r computation s w e can replac e 7 b y it s mos t
standard representativ e T~ 1 JT fo r suitabl e r i n G. Suc h classe s hav e geometri c
meaning, an d th e fac t tha t 7 i s a n elemen t o f T wil l translat e thi s meanin g int o
information abou t geometri c dat a o f th e quotien t spac e r \ H .
We first describ e th e classificatio n o f th e conjugac y classe s i n G = 5L(2,R )
using th e geometri c actio n o f G o n I , namel y b y considerin g th e fixed point s o f
the isometr y o f H I induced b y a n elemen t g G G. Thi s give s informatio n abou t
conjugacy classes , becaus e i f z i s fixed b y g, the n TZ i s fixed b y th e conjugat e
r~1gr o f g. Becaus e th e grou p i s ver y symmetric , th e fixed point s wil l b e enoug h
to classif y g up t o conjugacy .
 15. S P E C T R A L T H E O R Y O F A U T O M O R P H I C F O R M S 395

Let g e G, 9 = [ , I , # 7 ^ ± 1 , actin g non-triviall y o n i U Q U 00. Th e

equation gz — z i s a quadrati c equatio n cz 2 - f (d — a)z — b = 0 , wit h discriminan t
A = (d — a) 2 + Abe = (Tig) 2 — 4 , an d ther e ar e thre e cases .
• Firs t case : \Tr g\ > 2 (g is called hyperbolic) . The n A > 0 , an d ther e ar e tw o
distinct fixed point s i n R (th e cas e c — 0, a ^ 1 also belong s t o thi s case , bu t on e
of the fixed point s i s 00) . Ther e exist s r G G which send s on e o f the fixed point s t o
0 an d th e othe r t o 00 ; the n b y th e observatio n abov e th e conjugat e r~ xgr fixes 0
and 00 , an d clearl y (changin g r int o r _ 1 i f necessary) , w e find tha t g i s conjugat e
to a uniqu e elemen t o f th e for m

(15-21) a(p ) = ±^ 1 ()/2 p _1

° /2)

for som e p > 1 , which act s geometricall y a s a dilatio n z ^ pz. W e cal l p th e nor m
of g an d writ e Ng = p; th e nor m classifie s exactl y a hyperboli c conjugac y class .
Notice tha t p ca n b e recovere d fro m g b y solvin g th e equatio n z + z~ x = Tr g an d
taking th e squar e roo t o f th e solutio n whic h i s large r tha n 1 .
Associated t o a hyperboli c motio n i s th e geodesi c 7 ^ i n H whic h i s th e half -
circle perpendicula r t o R a t th e tw o fixed points , o r i n th e cas e c = 0 , th e vertica l
line fro m th e rea l fixed poin t t o 00 . Thi s geodesi c i s preserved b y g (thi s i s obviou s
for a(p) , an d follow s b y conjugatio n i n th e genera l case) . Not e tha t th e nor m o f
g ca n b e recovere d b y log Ng = d(z,gz) fo r z G jg (g act s a s a translatio n o n th e
geodesic).
Conversely, t o ever y geodesi c 7 i n H , on e ca n associat e th e isotrop y subgrou p
T 7 i n G, containin g element s which preserv e 7. Thi s is generated b y ± 1, an elemen t
of orde r 2 whic h interchange s th e end-point s o f 7 i n R U {00}, an d th e subgrou p
of thos e element s fixing th e end-point s (whic h i s isomorphi c t o th e multiplicativ e
group o f positiv e rea l numbers , an d s o to R b y th e logarith m map) . Al l these fact s
follow b y conjugatio n t o th e cas e whe n th e geodesi c i s th e imaginar y axis , an d T 7
is generate d b y th e involutio n z1— > — z~x an d th e grou p o f element s a(p) above .
Let agai n g be hyperbolic . Th e centralize r Z(g) o f g in G is a subgroup o f T 7g ;
one check s tha t th e involutio n i s no t i n it , bu t th e othe r subgrou p is , s o Z(g) i s
isomorphic t o R (time s ± 1 , of course) .
• Secon d case : \Trg\ = 2 (g i s calle d a paraboli c motion) . The n A = 0 , s o g
has a uniqu e fixed poin t a , a G R i f c ^ 0 , an d a = 0 0 i f c = 0 . A s before , som e r
sends a t o 0 0 an d on e see s thi s wa y tha t g i s conjugat e t o a uniqu e elemen t

1x
(15.22) n(x) = ±
01

for som e x G R, whic h act s a s a horizonta l translatio n z i- > z + x. Th e paramete r

x classifie s th e paraboli c conjugac y classes . Th e grou p N = {n(x) \ x G R} i s th e
isotropy grou p o f 0 0 i n G. an d i t i s th e centralize r i n G o f an y o f it s elements . B y
conjugation, th e centralize r o f a paraboli c elemen t i s thu s isomorphi c t o R .
• Thir d case : |Trg | < 2 (g i s calle d a n ellipti c motion) . The n A < 0 , s o th e
quadratic equatio n ha s tw o conjugat e comple x roots , on e o f which , sa y z g, i s i n
M. Movin g z g t o i b y a n elemen t r G G , a shor t computatio n reveal s tha t g i s
396 15. SPECTRA L THEOR Y O F AUTOMORPHI C FORM S

conjugate t o a uniqu e elemen t

(15.23) r{0) = ±( COS.ea S i n ^

v v
' ' \ - s m 0 co s 0 J
with 0 G R/27rZ , whic h act s b y hyperboli c rotatio n o f angl e 0 around i. Th e grou p
K = {r{0) | 0 G R/27rZ}, isomorphi c t o R/27rZ , i s th e centralize r o f an y o f it s
(non-trivial) elements , an d th e centralize r o f an y ellipti c elemen t i s isomorphi c t o
K.
Now conside r a discret e subgrou p r o f G an d a n elemen t 7 G T. Le t Zri'j) b e
the centralize r o f 7 i n T ; i t i s o f cours e a subgrou p o f th e centralize r Z(j) o f 7 i n
G, an d i t i s a discret e subgroup . Whe n 7 i s hyperboli c o r paraboli c an d ^(7 ) i s
isomorphic t o M , thi s implie s tha t Zr(~f) i s infinit e cyclic , an d whe n 7 i s elliptic ,
this implie s tha t ^ ( 7 ) i s finite cyclic . I n al l cases , w e say tha t 7 (o r it s conjugac y
class i n T ) i s primitiv e whe n 7 i s itsel f a generato r o f th e centralize r Zri^f)- I t i s
clear tha t an y elemen t 7 i s o f th e for m ±7 ™ fo r som e intege r m ^ 0 and a uniqu e
primitive elemen t 70 . B y conjugatin g th e generato r 7 0 to on e of the standar d form s
described above , w e se e tha t th e actio n o f ^ r ( l ) o n H ha s th e followin g simpl e
fundamental domain , dependin g o n th e typ e o f 7 :
(1) I f 7 i s hyperbolic , 7 0 i s conjugat e t o (1 5.21 ) fo r p = N^ 0, actin g b y
z ^ pz, an d a fundamenta l domai n i s th e horizonta l stri p
Z r ( 7 ) \ H = {z = x + iy G M | 1 < y < p}.
(2) If 7 i s parabolic, 7 0 is conjugate t o (1 5.22 ) fo r som e x, an d a fundamenta l
domain i s th e vertica l stri p
Z r (7)\EI = {z = x + iy I 0 < x < 1}.
(3) I f 7 i s elliptic , 7 0 i s conjugat e t o (1 5.23 ) fo r som e 6. Le t w b e th e fixed
point o f 70 (th e center o f the rotation) ; a fundamental domai n fo r Zr(j) i s obtaine d
by takin g th e domai n betwee n tw o geodesie s goin g throug h w wit h a n angl e equa l
to 0 (by discreteness, notic e that 0 must b e of the for m 2TT/£ fo r som e integer £ ^ 1 ) .
In th e specia l cas e o f T — SX(2, Z), the paraboli c an d ellipti c conjugac y classe s
are ver y eas y t o describe . First , th e fixed poin t o f a paraboli c elemen t mus t b e
in Q U {00}. Bu t al l suc h point s ar e T-equivalent , b y Bezout' s theorem , an d sinc e
no = n(l ) clearl y generate s th e isotrop y grou p o f 0 0 i n 5X(2,Z) , w e find tha t it s
conjugacy clas s i s th e onl y paraboli c conjugac y clas s i n SX(2,Z) . Th e centralize r
ZY(JIQ) i s {n(x) I x G Z}, an d a fundamenta l domai n fo r Zr(no) i s th e vertica l
strip {z — x 4 - iy \ 0 < x < 1}.
By simpl e computations , on e finds als o tha t SX(2 , Z) contain s onl y tw o primi -
tive conjugac y classe s o f ellipti c elements , wit h representative s ( ~ x J , ( _ 1 0 ) o f
order 2 (a s a n isometry , bu t 4 as a matrix) , fixing i an d o f orde r 3 (a s a n isometry ,
but 6 a s a matrix) , fixing j , respectively.
EXERCISE 1 . Fin d th e primitive parabolic conjugac y classe s in r 0 (g), an d sho w
that the y correspon d t o th e se t o f cusp s a s describe d i n Sectio n 1 4.1 .
We com e bac k t o th e notatio n o f th e introductio n t o thi s section , wit h T a
discrete subgrou p suc h tha t th e quotien t r \ H ha s finite volume , an d ha s cusp s
(finitely many) . Th e lac k o f compactness create s som e analyti c difficulties , becaus e
 15. SPECTRA L THEOR Y O F AUTOMORPHI C FORM S 397

it turn s ou t that th e kerne l i s not integrable o n the diagonal , s o the trace , as

we hav e define d it , is infinite. Thi s proble m i s solved b y considering truncate d
fundamental domain s F(Y) whic h ten d t o F fo r Y—) > +00, i n the sense tha t
IF(Y) f^ z)^^z) " ^ IF /( 2 )^/ i ( 2 ; ) f° r a n y integrabl e / . Th e kerne l K restricte d to
F(Y) i s no w integrabl e o n th e diagonal , an d on e ca n comput e th e truncate d trac e

TrY K= [ K{z,z)dfji(z)
JF(Y)

using th e spectra l expansio n fo r th e restrictio n o n th e on e hand , an d th e decom -

position int o conjugac y classe s o n th e othe r hand . O n th e spectra l side , thi s give s
for Y — > +0 0 a n expressio n o f th e for m T r K — A logy - h T\ + o(l), an d o n th e
geometric sid e another expressio n T r y K — A l o gF + T2 + o(l), fo r th e som e A > 0.
Hence w e have th e tautolog y A = A, an d th e non-trivia l formul a T\ =T 2, which i s
the trac e formula .
A smal l analytica l miracl e occur s whe n doin g th e exac t computations : th e
complicated Harish-Chandra/Selber g transfor m disappear s fro m th e picture , an d
all tha t i s left i s the functio n h on on e side , an d it s Fourier transfor m h on the
other side .
We no w giv e indication s o n th e shap e o f th e variou s parts , an d stat e th e com -
plete formula i n the specia l case T = 5T(2, Z). O n the spectra l side , the cus p form s
and residua l spectru m Uj contribute, a s i n th e cas e o f a compact quotien t

(15.24) £ > ( * ; )
3

and th e Eisenstei n serie s ar e show n t o contribut e

(15.25) M ^ ^ i ^ i . f +c°h{ty{i + lt)dt

where $(s) is the scatterin g matri x forme d b y th e first Fourie r coefficient s o f the
Eisenstein serie s an d tp(s) — det<3>(s) i s it s determinan t (se e Sectio n 1 5.4) .
The geometric computations ar e done independently fo r each type of (primitive )
conjugacy classes . Th e clas s o f th e identit y motio n ha s a contribution give n b y

(15.26) Voir\ H J th^t^h^dt.

A hyperboli c conjugac y clas s P with associate d primitiv e clas s PQ contribute s

A primitiv e ellipti c clas s R of orde r m (as a n isometry ) contribute s

+
^/ TT£\- 1 /* °° e- 2nt£/m
(15.28) £ (2msin- ) / ^—-^h^dt .
0<£<m J ~°°
Finally, al l primitiv e paraboli c conjugac y classe s contribut e th e sam e amount ,
which is

(15.29) ^ - ^ - { f e ( 0 ) l o g 2 - / $(l+it)h(t)dt}
398 15. S P E C T R A L THEOR Y O F AUTOMORPHI C FORM S

s
where ijj(s) = jr( s) * ^n e logarithmic derivativ e o f the Gamma function .
For SL(2 , Z), afte r som e rearranging , an d taking int o accoun t th e descriptio n
of the cusps an d of the scattering matrix , w e obtain
T H E O R E M 1 5.9 . Let h be a function satisfying the conditions (1 5.1 9), h its
Fourier transform. Let (p(s) = 0(1 — s)6(s)~ 1 , where 0(s) = 7r~ sT(s)((2s), be the
scattering matrix o/SX(2,Z) . Then we have the formula

ft
(15.30) J2 (*j) + T " / — ( 2 + it)Wt)dt =

f h(t) (^ tanh(jrt ) - - ^ log 2 - i-tf>( l + it))dt

Z7r KZ7r J
P NP 2 - NP~2
hyperbolic

+E E (2 m sin -) ^^^KDdt.
elliptic

15.8. Hyperboli c lattic e poin t problems .

One o f the mos t direc t application s o f Theore m 1 5.7 , similar i n spiri t t o the
examples give n fo r th e Poisso n summatio n formul a an d th e Vorono i formul a i n
Chapter 4 , is to the study o f hyperbolic lattic e poin t problems .
Let D C I b e an open subset . Th e lattice poin t proble m fo r D i s the problem
of countin g th e number o f translate s o f a give n ZQ G HI by element s o f a discret e
subgroup T C G whic h ar e in D. Th e classical problem s o f Gaus s an d Dirichle t
already mentione d i n Chapte r 4 ar e exactl y analogue , takin g th e actio n o f the
lattice 1 ? on R2 b y translation an d D the points i n a circle or under a n hyperbola .
As i n these cases , analyti c numbe r theor y ca n have it s say only i f one considers a
sequence o f subsets gettin g larger , an d sufficiently regular .
We tak e a geodesi c bal l o f radiu s tendin g t o infinity , i n th e mor e convenien t
form
B(w, R) = {zeM | 4u(w, z)+2^R}
and loo k fo r good asymptotic s fo r the counting functio n
N(R) = \{7 eT\7w e B(w,R)}\.
As in the euclidean case , the expected valu e of N(R) i s the area of the geodesi c
ball, but whereas Gauss' s classica l argumen t (packin g with unit squares ) establishe s
the correspondin g resul t easil y wit h a fairly goo d erro r ter m R 1 /2 i n the euclidean
case, thi s simpl e argumen t i s very inefficien t i n the context o f hyperbolic geometry .
Indeed, th e error ter m i s proportiona l t o th e length o f the boundary o f the disc,
which i s small i n euclidean geometry , bu t is of the same orde r o f magnitude a s the
area i n the hyperbolic case . Thi s fac t follow s fro m th e isoperimetric inequality : i f
the boundar y dD o f D is smooth, A — Vo l (D) and L = length(cL4), the n w e have
4TTA + A 2 < L 2 , henc e L ^ A.

One ca n at leas t prov e elementar y uppe r bound s whic h ar e quite usefu l (i t is

used i n the proof o f Proposition 1 5. 8 and in Chapter 22).
 15. SPECTRA L THEOR Y O F AUTOMORPHI C FORM S 399

PROPOSITION 1 5.1 0 . Let T be a discrete subgroup o/SL(2,R ) such that the

volume Vo l (r\M) is finite and let a be a cusp ofT.
(1) For any z e H we have

|{ 7 er a \r | im^-V > Y}\ < 11 0

cay
(2) For any z, w e H , an d J > 0 , we have

| { 7 € a- xTaa | u( 7 2, w) < S}\ < y ^ + T ) ( I m («, ) + c" 1 ) + * +

/ + 1
c a l m (it; )
u>z£/i an absolute implied constant.
Here we let

ca — min ^ c > 0 | I , G cr~1rcra /o r som e a , 6 , d I.

See [1 5] , Lemma 2.1 1 fo r th e proof .

Applying Theore m 1 5. 7 for suitabl e kernels , one can get a n asymptotic formula .
THEOREM 1 5.1 1 . Let T be a discrete subgroup o/SX(2,R ) with Vo l ( r \ H) <
+oo. LetI D G I . We have

2<5,<1

with an implied constant depending on T and z only.

SKETCH O F PROOF . W e conside r onl y T = T 0(q) a s before . I f k(u) ^ 0 is an y
function wit h k(u) ^ 1 for u < (R — 2)/4 , an d i f th e correspondin g automorphi c
kernel K(z,w) satisfie s th e assumption s o f Theore m 1 5.7 , we hav e b y positivit y

(15.31) N{R) ^ Y^ k{u(w, ~fw)) = 2K(w, w).

By Theore m 1 5.7 , we hav e

(15.32) K(w,w) = Th(t J)\uJ(w)\2 + T±- f h(t)\E a{w,\+it)\

2
dt

with notatio n a s i n th e theorem . W e take k(u) = 1 for u < (R — 2)/4 an d k(u) = 0

for u ^ (- R + S — 2)/4 , an d / c linear between .
We estimat e h(t) usin g Lemm a 1 5.6 , fo r a suitabl e functio n / o n E I which i s
eigenvalue o f A wit h A = \ + 1 2 . Fo r example , f(z) — y s yield s
s
h(t) = / k(i,z)y d/j,(z).
Ju
One finds first tha t

r(s + i)
for \ < s ^ 1 , wher e th e implie d constan t depend s o n s. Th e constan t eigenvalu e
for j — 0 yield s th e mai n ter m 7ri?(Vo l ( r \ H ) )_ 1, an d th e exceptiona l eigenvalue s
(if any ) wit h 0 < Xj < \ yiel d th e secon d term , u p t o th e erro r ter m 0(R 1 ^2 - f S).
1400 5 . SPECTRA L THEOR Y O F AUTOMORPHI C FORM S

Let T — RS~ 1 . Fo r Re (s) = |, integratin g b y parts yield s

5/2
h(t) < \s\- R^2(mm(\slT) + \ogR),
with a n absolute implie d constant . Henc e w e estimate th e Eisenstein serie s contri -
bution t o (1 5.32 ) a t a given cus p a by

/ h(t)\E a(w, \ + it)\2dt « R 1 '2 [ ^ (( 1 + tf + log R)Wa{^l^t)? dt

JR J-VT U + ^J
R/
I _( T + log^)'
0 V , , ^5 d t
l\t\>Vr^ ' ~ ~'( * + *)
1 /2
< (#T)
by integratio n b y part s usin g Propositio n 1 5.8 . Th e same boun d hold s fo r the
contribution o f the non-exceptional discret e spectru m b y partial summatio n fro m
Proposition 1 5.8 . Usin g thes e estimate s wit h 5 = it! 2/3, (1 5.31 ) an d (15.32) giv e

2<Sj<l

Similarly wit h k(u) = 1 for u < (R - S - 2)/ 4 an d k(u) = 0 for u ^ (R - 2)/4 , we

get th e lower boun d an d Theorem 1 5.1 1 is proved. •

R E M A R K . Not e tha t fo r T 0(q) w e have A i ^ § s o Sj < | b y (5.88) , s o th e

contribution o f exceptional eigenvalue s i s smaller tha n th e error term .
COROLLARY 1 5.1 2 . (1 ) Let Ni(X) be the number of integer solutions to the
equation ad — be = 1 with a 2 + b 2 + c 2 + d 2 ^ X. We have
2 3
N1(X) = 6X + 0(X / ).
(2) Let r(n) be the number of representations of n as a sum of two squares. We
have
J2 r(n)r(n + 1 ) = SX + 0{X 2/3).

PROOF. Not e first tha t fo r w = i, we have 4u(ji, i) + 2 = a 2 + 6 2 + c 2 -f <i2 for

any 71
=( . So (1) follows immediatel y fro m Theore m 1 5.1 1 for T — SX(2, Z)
and w = 2 , since Vo l (SX(2, Z)\H) = STT" 1 an d Ai = 91 .1 4.. . > \ fo r the modular
group.
For (2) , consider th e group

!•={(• 5 ) 6 5 « a , z , i » j { ^ ) } . ( ; - I)"'r0,2)(; ->

We hav e a + d = 2k, a — d = 2l, b Jre = 2m , 6 — c = 2n , wit h ad — be —
k2 - i 2 - m 2 + n 2 = 1 and a2 + 6 2 + c 2 + d 2 = 2(/c 2 + £ 2 + m 2 + n 2 ), s o Theorem
15.11 give s
J2 r(m)r{m + 1 ) = N(AX + 2) - 8 X + 0 ( X 2 / 3 ) .

a
 15. SPECTRA L THEOR Y O F AUTOMORPHI C FORM S 401

15.9. Distributio n o f lengt h o f close d geodesie s an d clas s numbers .

The simples t applicatio n o f th e Selber g Trac e Formul a i s t o th e distributio n
of th e length s o f closed geodesie s o n the quotien t spac e X — r \ H, als o calle d th e
length spectrum. I n fact, close d geodesies are in one-to-one correspondenc e with the
hyperbolic conjugac y classe s of V. Indeed , i f g G T is an hyperbolic element , the n g
maps th e geodesic 7 (half-circle ) joinin g it s two fixed-points t o itself. I n particular ,
for an y 2 G 7 , th e geodesi c segmen t fro m z t o gz i s mappe d i n r \ H t o a close d
geodesic. Takin g th e model situatio n wher e g = a(p) wit h p > 1 as in (1 5.21 ) an d
z = i, observ e tha t th e length o f this close d geodesi c i s d(i,g(i)) — log p = log Ng.
Conjugates o f p, whic h hav e th e sam e fixed points , giv e ris e t o th e sam e close d
geodesic, an d conversely an y closed geodesi c i s obtained i n this way.
THEOREM 1 5.1 3 . Let T be a subgroup o/SL(2,R ) with Vol(r\H ) < +00 . We
have
8
Y^ logNP = X+ J2 s^X * +0(X 3'A)

where the sum is over primitive hyperbolic conjugacy classes and the implied con-
stant depends on T.
REMARKS. (1 ) For r wit h compac t quotient , thi s was proved b y Huber [Hub] .
(2) Th e exponent ca n be improved , se e e.g. Luo , Rudnick an d Sarna k [LRS] ,
where 3/ 4 i s replaced b y 7/1 0 + e fo r an y e > 0. Althoug h th e analog y wit h th e
Prime Numbe r Theore m suggest s tha t i t should b e X1/ / 2 + £ , an d the analogue of the
Riemann Hypothesi s holds , thi s i s yet unknown . Th e main difficult y i s that ther e
are man y mor e eigenvalue s tha n zero s o f the zeta function .
Also, a s i n Theore m 1 5.1 1 , i f T = r 0 (g), Selberg' s inequalit y A i ^ ^ mean s
that th e exceptional eigenvalue s hav e smalle r contributio n tha n th e error term .
SKETCH O F PROOF . Th e argument i s quite simila r i n principl e t o tha t o f the
previous theorem . W e choose h such tha t th e Fourier transfor m h localizes the sum
over hyperboli c conjugac y classe s t o thos e wit h NP ^ X + Y fo r som e paramete r
Y ^ 1 , say h(t) = 2 cosh(7rt)q(27rt) where q is a smoot h function , even , supporte d
on \t\ ^ log( X + Y) an d wit h 0 < q ^ 1 an d q = 1 on [-logX,logX] . Th e
contribution o f the discrete an d residual spectru m t o the spectral trac e is

Y,Htj) = X+ J2 s^X's+OiY + X^T)

1<s <
i2 3

with T = X 7 " 1 , an d th e sam e estimat e hold s fo r Eisenstei n serie s (t o se e this ,

estimate h b y Fourie r inversio n o n ^ < s ^ 1 an d o n R e (s) — 1 separately) .
The identity , ellipti c an d parabolic motion s hav e contribution s < C X1/2T b y simple
estimations, s o one gets
1
^9(log7VP)(log7VP)=X+ J2 ^7 X ^ + 0 ( y + X1/2T).
2<SJ<1

Subtracting thi s applie d a t X - f Y an d X, b y positivity, w e see that

J2 (log7VP)< F + X 1/2 T
X<NP<X+Y
402 15. S P E C T R A L THEOR Y O F AUTOMORPHI C FORM S

hence th e resul t b y choosin g Y = X 3 / / 4 . •

In [Sa2] , P . Sarna k deduce s th e followin g corollary .

T H E O R E M 1 5.1 4 . For any integer d > 1 such that d = 0, 1 (mod 4) , let h(d)
be the class number of Q(yd) and Ed be the fundamental unit. We have
2
] T M^)=Li(X ) + 0(X 3 / 2 (logX) 2 )

for X ^ 2, with an absolute implied constant.

P R O O F . Th e proo f depend s o n the followin g ma p betwee n indefinit e quadrati c
forms an d close d geodesies . Le t Q(x,y) — ax2 + bxy + cy2 b e a primitive integra l
indefinite quadrati c for m o f discriminant d > 0, i.e. (a , 6, c) = 1 and d = b 2 — 4ac >
0. B y linea r chang e o f variable , SL(2,Z ) act s o n suc h form s an d th e numbe r o f
equivalence classe s i s h(d).
The equation s Q(0, 1 ) = 0 has tw o rea l roots , 6\ = (—6 + Vd)/2 an d 62 =
(—6 — \/d)/2, t o which i s associated th e geodesi c i n HI joining # 1 an d #2 - The
stabilizer o f Q unde r th e actio n o f 5X(2,Z ) i s equa l t o th e stabilize r o f Q\ (or # 2 ),
and a generator i s give n explicitl y b y

n(n\ - f( to-bu0)/2 -cu 0

9W)
V ^ 0 (t 0 + bu 0)/2
where Ed = (to + uoVd)/2 (i n othe r words , (to,^o ) i s th e fundamenta l solutio n t o
the equatio n t^ — du^ = 4). Th e nor m o f thi s hyperboli c elemen t i s the n e\.
The ma p Q H > g send s primitiv e integra l quadrati c form s t o hyperboli c conju -
gacy classe s o f 5L(2,Z) . W e clai m i t induces a bijection betwee n classe s o f form s
and primitiv e hyperboli c conjugac y classes .
To se e this , le t g be an y primitiv e hyperboli c elemen t o f 5L(2,Z) . Le t 0\ an d
62 be th e fixed point s o f g an d K — Q(0i , #2)- Clearl y K is a real quadrati c field,
say K = Q(Vd) fo r a unique d > 0 , d = 0, 1 (mo d 4) . The n th e stabilize r o f 8\ is
generated b y g becaus e g i s primitive . Th e for m Q(x,y) = (x — 6\y){x — 62V) is a
primitive quadrati c for m wit h g(Q) = g. Thi s implie s surjectivit y o f th e ma p an d
injectivity i s easy t o check .
Hence th e norm s o f conjugac y classe s o f primitiv e hyperboli c element s ar e e 2d
with multiplicit y h(d). I n thi s cas e Theore m 1 5.1 3 become s
3 4
(15.33) Y. 2h(d)logs d = X + 0(X / ),
ed^Vx
hence th e resul t afte r summatio n b y parts . •

REMARK. Siege l [Sie2 ] ha s show n tha t

7
(15.34) Yl 1Hd) loge d =
8C(3 ^) 7^+0(XlogX)
ttx
for X ^ 2 . Notic e th e differenc e betwee n th e orderin g o f the discriminant s in
(15.33) an d (1 5.34) . A n asymptoti c formul a fo r h(d) ove r d ^ X i s no t known .

EXERCISE 2 . Prov e (1 5.34 ) usin g th e Dirichle t Clas s Numbe r Formul a (2.31 ) .

 http://dx.doi.org/10.1090/coll/053/17

CHAPTER 1 6

SUMS O F K L O O S T E R M A N SUM S

16.1. Introduction .
Individual Kloosterma n sum s S(m, n ; c) ar e quit e wel l examine d b y algebrai c
methods (se e Chapte r 1 1 ) . Th e resul t i s Weil's boun d

(16.1) |5(m,n;c) | < (m,n,c)^c^r(c)

which i s th e bes t possible . Thi s an d earlie r estimate s hav e bee n use d i n analyti c
number theor y indirectl y i n many ingeniou s ways (se e for instanc e [Li2] , [At] , [110],
[Wi3]). Yu . V . Linni k [Li3 ] pointe d ou t tha t fo r certai n application s on e actuall y
needs estimate s fo r sum s o f typ e

(16.2) S mn(X) = Yl c^Sim, n ; c).

c=0(mod q)

The uppe r boun d (1 6.1 ) implie s

(16.3) 5 mn (I)«r((m,n)g)(m,n,^lhogI

for X ^ 2 where th e implie d constan t i s absolute . However , Linni k suggeste d tha t

a smalle r boun d shoul d b e true , i f X i s sufficientl y large , becaus e o f cancellatio n
of Kloosterman sums , whic h i s due t o a regula r chang e o f sign . Linni k conjecture d
that

(16.4) S mn(X) « X e,

where th e implie d constan t depend s o n m , n an d £ , an d h e als o provide d som e

heuristics t o suppor t thi s bound .
In hi s investigation s Linni k wa s drive n b y a n analog y o f th e zet a functio n
c
(16.5) z mn(s) = J2 ~2'sKn;c)
c=0(mod q)

to the hypothesi s o f Hasse on the L-function s o f elliptic curve s (t o be precise Linni k

spoke onl y fo r q = 1 and use d a differen t notation) . Thi s analog y turn s ou t t o b e
close onl y i n th e analyti c aspect s o f th e serie s (1 6.5 ) (lik e analyti c continuation ,
but no t th e Eule r product) . B y independen t analysi s A . Selber g [S7 ] revealed tha t
Zran{s) admit s a spectral decompositio n wit h respect t o automorphi c form s (fo r th e
group To(q)). Selber g le d th e foundatio n fo r th e spectra l theor y o f Kloosterma n
sums yet h e did no t delive r th e necessar y technica l tool s to cultivat e thi s fertil e soil .
403
40 4 16. SUM S O F KLOOSTERMA N SUM S

In this chapter w e sketch proofs o f central results of this profound theory . Ther e
are soli d presentation s o f selecte d topic s i n th e article s [1 3] , [1 1 3] , [DI ] an d i n th e
books [1 5] , [Sa3] , [Motl] , however , a conclusiv e expositio n i s yet t o b e written .
16.2. Fourie r expansio n o f Poincar e series .
There ar e variou s method s t o develo p th e spectra l decomposition s o f sum s o f
Kloosterman sums . Arguabl y th e bes t structura l treatment s appea l directl y t o
the Gree n functio n (se e [1 5]) , o r th e resolven t operato r (se e [GS]) . Nevertheless ,
we follo w th e traditiona l metho d whic h begin s b y dua l treatment s o f th e Poincar e
series (fo r th e grou p T = T 0(q))
F
(16.6) P m(z)= Yl (m7^), m = l , 2 , 3 , . . . .

Here F(z) i s a functio n o n H I which i s periodi c i n x o f perio d on e an d i t satisfie s

F(z) < C ya wit h som e a > 1 so th e serie s (1 6.6 ) converge s absolutel y (a = 1 will
be attaine d b y continuity) . Ou r treatment s o f P m(z) ar e paralle l t o thos e alread y
given fo r th e holomorphi c form s i n Sectio n 1 4.2 , an d ou r goa l i s th e Kuznetso v
formula (1 6.34 ) whic h i s a n analo g o f th e Petersso n formul a (1 4.1 5) . Th e extr a
care i s neede d t o justif y infinit e summation s ove r th e spectru m o f th e Laplac e
operator. T o thi s en d on e need s som e crud e estimate s fo r th e Fourie r coefficient s
of cus p form s i n th e spectra l parameters . W e shal l prov e muc h stronge r estimate s
than necessar y i n Sectio n 1 6.5 , bu t onl y afte r th e final Kuznetso v formula s ar e
established. However , on e coul d deriv e th e require d auxiliar y estimate s fro m th e
early form s o f Kuznetsov' s formulas . W e encourag e th e reade r t o d o s o an d t o
organize our argument s int o a correct logica l order (w e compromised th e order , bu t
not th e completeness , fo r reducin g th e lengt h o f presentation) .
Applying th e doubl e cose t decompositio n (1 4.1 4 ) an d the n th e Poisso n sum -
mation formul a a s i n th e proo f o f Lemm a 1 4. 2 w e arriv e a t th e Fourie r serie s fo r
Pm(z),

nz
Prn(z) = F(mz) + Yl E Y.< )f F (m(ar * H)z)e(-nZ)dt.

Now assum e tha t th e generatin g functio n F{z) satisfie s th e rul e F(z-\-t) = e(t)f(z)
for t £ M so b y th e actio n o f T o n H I we hav e

F m =e
( (c d)0 (v) F ( - m / c ( £ * + d ) ) -
From thi s actio n on e get s th e Kloosterma n sum s i n th e Fourie r serie s

(16.7) P m{z) = F(mz) + ^2 E S

(m'n; c )Fc(m>n:' y) e nz
( )
n c=0(mo d q)

where
2
F c (m, n; y) = f F(-m/c Z)e(-nOdi
Jim £=y
by shiftin g £ to £ — d/c. W e pu t
(16.8) F(z) =p(47ry)e(z)
 16. SUM S O F K L O O S T E R M A N SUM S 405

and cal l p(y) th e tes t functio n fo r th e correspondin g P m (z), gettin g

(16.9) F c(m,n;y) = £ ^ . ( ^ f )e(-% " *)*.

In thi s generalit y w e cannot comput e th e integra l (1 6.9) . A n interestin g choic e

for th e tes t functio n i s

(16.10) p(y) = ys
where s i s a comple x paramete r a t ou r disposa l wit h R e s > 1 (ther e ar e othe r
eigenfunctions o f A whic h lea d als o to interesting results) . I n this cas e the Poincar e
series become s
s
(16.11) P rn(z,s) = ^ (47rml m ^z) e(myz)

and it s Fourie r expansio n become s

(16.12) P m (z, s) = {47rm) s{e(mz) + ] T Z mn(s; y)e(nz)}

n

where
c
Zmn{s\ y) = ^2 ~2sS(m, n ; c)Is(m, n ; y)
c = 0 ( m o d q)
and
2s
/ 8 (m, n;y)= [ |£|- e ( - < ~ m/Z<?)d£.

The genera l Poincar e serie s P m(z) ca n b e represente d b y P m(z, s) b y applyin g th e

Mellin transfor m t o p(y).
The particular Poincar e series P m(z, s) wa s introduced b y A. Selberg [S7] . Thi s
is nearl y a cus p form , bu t no t exactl y so , becaus e i t fail s t o b e a n eigenfunctio n o f
A. Precisel y w e hav e

(16.13) ( A + s(l - s))P m(z, s) = sP m(z, 5 + 1 ) ,

because ( A + s( l — s))F{z) = ^ixsyF{z) fo r F(z) = y se(z). Henc e w e ge t th e
recurrence formul a

(16.14) P m(z,s) = sRxP m(z,s + l)

where R\ = ( A + A ) - 1 i s the resolven t o f A a t A = s( l — s). B y th e spectra l theor y
R\ i s meromorphic , therefor e P m(z,s) ca n b e continue d meromorphicall y t o th e
whole comple x s-plane . Moreover , i n th e half-plan e R e (s) > \ th e pole s o f R\
are a t th e point s Xj = Sj(l — Sj) > 0 i n th e discret e spectru m o f A , an d the y ar e
simple, s o ar e th e pole s o f P m (z, s) a t s — Sj. I n fac t i t i s conjecture d b y Selber g
that Xj ^ | , s o Sj = \ + itj wit h tj rea l an d P m (^, 5) i s holomorphic i n R e (s) > | .
In what follow s we avoid the analytic continuation o f the Selberg-Poincar e serie s
and th e resolven t operato r altogether . Instead , w e rel y heavil y o n th e spectra l
decomposition o f automorphi c functions .
40 6 16. SUM S O F K L O O S T E R M A N SUM S

16.3. Th e projectio n o f Poincar e serie s o n Maas s forms .

We return t o the Poincar e serie s P m(z) fc> r a generating functio n o f type (1 6.8) .
Let f(z) b e a Maas s for m o f eigenvalu e A = | + r 2 give n b y th e Fourie r serie s

(16.15) f(z) = ay^ r + by^ r + 2 /^ a nKir{2*\n\y)e{nx)

(see Lemm a 1 5.1 ) . B y th e unfoldin g metho d w e deriv e

/•OO pi

(f,Pm)= / / f(z)F(mz)d liz.

Jo Jo

Inserting (1 6.1 5 ) an d (1 6.8 ) w e pic k u p th e ter m a m b y integratio n i n 0 < x < 1 .

Then, changin g th e variabl e y — > y/2irm w e obtai n th e formul a
/»oo

(16.16) </,P m) = (2^m)*a m / e-*K ir(y)p(2y)y-*dy.

Jo

Hence observ e tha t P m(z) i s orthogona l t o constan t functions . I n particula r fo r

p(y) — V s th e abov e integra l i s a produc t o f gamm a function s givin g

(16.17) (/ , P m) = 2nT(s - \ + r)T(s - \ - i r ) r ( s ) - W a m.

This i s clearl y a n analo g o f th e formul a fo r th e inne r produc t o f a holomorphi c

modular for m wit h a holomorphi c Poincar e serie s give n i n Lemm a 1 4.3 .

16.4. Kuznetsov' s formulas .

Now w e ar e read y t o deriv e explici t relation s betwee n Kloosterma n sum s an d
Fourier coefficient s o f automorphi c forms . T o thi s en d w e comput e i n tw o way s
the inne r produc t ( P m , Q n ) o f tw o Poincar e serie s Pm(z) an d Q n(z) generate d
by F(z) — (47ry)2p(4r:y)e(z) an d G(z) = (Airy)^ q(A7ry)e(z) respectivel y (not e w e
modified th e tes t function s p(y),q(y) b y th e facto r y*).
First b y th e spectra l decompositio n (se e Theore m 1 5.2 ) w e get
(16.18)
OO

J =l
OO

(P m ,#oo(*, \ + ir) > (£«>(•, \ + ir),Q n)dr.

/

where (UJ(Z)) i s a n orthonorma l basi s o f cus p form s an d (E 00(z, \ + ir); r G R) i s

the eigenpacke t o f Eisenstei n serie s associate d wit h a syste m o f inequivalen t cusps .
Suppose Uj(z) ha s th e Fourie r expansio n

(16.19) Uj(z) =Pj{0)y s> +y l

^Pj(n)Kltj{27T\n\y)e{nx)
n^O

where Sj = \ +itj wit h tj real , o r | < Sj < 1 because Xj = Sj(l — Sj) — \+i] > 0.
Then
(16.20) (uj,Q n) = {^n)^p 3{n)ujq{t3)
 16. SUM S O F K L O O S T E R M A N SUM S 40 7

where
/•OO
1
(16.21) u q(t) = / e-yK lt(y)q(2y)y~ dy
Jo
by th e formula (1 6.1 6) . Similarl y fo r the Eisenstein serie s

(16.22) E a(z, ±+ir) = 5 ay^r + y> fl(± + ir)yi~ ir

+ 2/ 5 ^2^a(n,r)K ir(27r\n\y)e(nx)
n^O

we hav e
(16.23) {Ea&\ +ir),Q n) = (4^n)*r a(n,rH(r).
Inserting (1 6.1 9 ) an d (1 6.23 ) int o (1 6.1 8 ) w e obtain
OO

(16.24) (P m,Qn) = 47ry / ^|y^pj(m)pj(n)cjp(tj)a; g (tj)

00
1 Z* " i

an
Next w e compute ( P m , Q n ) b y unfolding wit h respec t t o Q n d applyin g th e
Fourier expansio n (1 6.7 ) fo r P m. W e obtain
/•OO /» 1

(PrmQn}= / / P rn(z)G(nz)d/JJZ
Jo Jo

\6(m, n)F(imy) + ^ 5(m , n; c)Fc (ra, n; y)e(iny))G(iny) — .

q\c

Changing th e variables w e arrive a t

(16.25) (P m, Qn ) = 47rvm^(^(m , n)(p, g) + ^ ^ ^ ( m , n; c)Vm[ 7r ^mn^

q\c

where
/•OO
1
(16.26) (p,q)= / e-yp(y)g(y)y' dy
Jo
and

We simplify V pq(x) b y changing £ = 2z(( 2 + 1 ) x , where £ runs ove r th e semi-circle

|£| = 1 , Re(£) > 0 clockwise. Lettin g R e (£) — r] w e have £ = ^(C^) _1 , |£l = ^ _ 1
and ICI"" 1 ^ = z(C^) -1 ^C- Henc e

(16.28) V :+y)C^ dyd(

pq{x) = if J p{^J^q{xyrj)e-^v
/o rKy'')^*'"" y(rj'

Combining (1 6.24 ) an d (1 6.25 ) w e arrive a t th e followin g

408 6 . SUMS O F K L O O S T E R M A N SUM S

PROPOSITION 1 6.1 . Let m,n > 0 andp(y),q(y) be smooth, bounded functions

on R + . Then we have
(16.29)

Y^Pj{rn)p3{n)uv(t3)ujq(tj) + ^— / f a (ra,r)r a (n, r)uj p(r)cuq(r)dr

4 7 F
3=1 a ^-oo

e/ w N V ^ - I n / x-r rf ^ y^TOU
= d{m,n)(p,q) + 2_ ^c M m
> n ; c )*>7(
c = 0 ( m o d g)

Now let us choose p(y) = q(y) = yz+tl/ wit h v G R. Fo r these tes t function s we
get
ujp(t) = ^T(\ + i* + « ^ ) r (| - i * + iv)T(l + iz/)"1,
whence
(16.30) (jj p(t)ujp(t) = 7r2 sinh(7rz/)/^ cosh7r(^ — t) cosh7r(z/+ t).
On th e other hand , (p,q) = 1 and Vpp{x) = B2i v(x), wher e

1
_
(16.31) B s(x) = 2ix / X s (Cx)C ^C.

In thi s cas e (1 6.29 ) become s (a n auxiliary versio n o f Kuznetsov's formula )

COROLLARY 1 6.2 . For ra,n> 0 and v e R we have

Pj(m)pj(n)
(16.32) £
_^ cosh7r(i / — tj) cosh7r(z/ + tj)
, \ ^ 1p T a (m,r)r a (n,r) ^
i -
^- ' 47 r J-oo cosh7r(z/ — r) cosh7r(i/ + r)
7r__2zy
/r / N v ^ 1- 0 / x^ , /47rA/^m\ \
= —- —- d(ra,n ) + > c S(m,n;c)B 2iv[ ) ) .
smn(7ri/) V ^— 'V c / /
vJ
c=0(mo d q)
The integra l (1 6.31 ) ca n be expressed i n terms o f the Bessel functio n J s(y) of
real positiv e argument . Indeed , b y Cauchy's theore m

Bs(x) = 2ix([ + f) KsiCxX^dC

oo
l
(Ks(ixy) - K s{~ixy))y' dy.
/
For y > 0 we have (se e (5.14.4), (5.22.5) , (5.22.6 ) o f [Leb] )

^Hy)-^^(e^J.M-e-^Js(y)).
Hence
— 7T ?
K.(iy)-K.(-iy) = ^^jT)(My) + J. B(y)).
 16. SUM S O F K L O O S T E R M A N SUM S 409

Finally, thi s yield s

(16.33) B s{x) = / (J s(y) + J-sWjy^dy.

COS(7TS/2) J x

With th e real paramete r v one can generate quit e a lot of functions h(t) out o f
the particula r on e h(t) = (cosri7r( ^ — £)cosh7r(z / + t))~ 1 . N . V. Kuznetsov [Kuz]
produced fro m (1 6.32 ) th e following
T H E O R E M 1 6.3 . Suppose h(t) satisfies the conditions (1 5.20). Then for any
ra, n > 0 we have
(16.34)
^, , . N h{tj) v- ^ 1 f°° , , , .h(r)dr
> Pj{m)pAn) T^— + > — / T a(m,r)Ta(n,r)
£_jfj\ ;rj\ ^ Q g h ^ Z^ 4nJ a v
' a vy
cosher
w
j —1 ^ a
1x '4ir \/mn
5(ra,n)# 0 + ^2 c c ${m,n\c)g\
S{m,n;c)g(-
c = 0 ( m o d q)

where
oo
(16.35) 9o = n : rh(r) tanh.(nr)dr
/ -oo

and
2% rft(r)
(16.36) 9\x) = - J2irWT T ^ r
-
7T 3 cosh(7rr )

R E M A R K S . Th e formula (1 6.34 ) wa s also establishe d independentl y b y R. W.

Bruggeman [Bru] , bu t only fo r a certain typ e o f test function s h(t) an d with g(x)
in less refined for m tha n (1 6.36 ) o f Kuznetsov. Th e resul t als o holds (an d th e proof
is similar ) fo r ran < 0, except fo r the integral transfor m (1 6.36 ) whic h change s to
4 f°°
g (x) = — / K2ir{x)rh{r)smh{irr)dr.

The Kuznetso v formul a (1 6.34 ) ca n be derived fro m (1 6.32 ) b y means o f the

following
LEMMA 1 6.4 . Suppose h(t) satisfies the conditions (1 5.20). Then
oo cosh(7rr)dr —2h(t)
(16.37) (h(r+i)+h(r-i))
/ -OO
cosh7r(r — t) cosh7r(r + t) cosh(7rt )

PROOF ( S K E T C H ) . Mov e the integral o f h(r+ | ) dow n to the lin e I m (r) = — \

excluding smal l half-circle s centere d a t r = t — |, r = — t — |, an d mov e th e
integral o f h(r — | ) u p to the line I m (r) = \ excludin g smal l half-circle s centere d
a t r = £ + | , r = £ — | . Th e horizontal integral s cance l ou t because h(t) i s even.
The fou r integral s ove r th e small half-circle s yiel d i n the limit — 2h(t)/ cosh(7rt) by
computing residues . D
410 16. SUM S O F K L O O S T E R M A N SUM S

By integratin g (1 6.32 ) accordin g t o (1 6.37 ) on e obtain s immediatel y th e spec -

tral term s o n th e lef t sid e o f (1 6.34) . O n th e righ t sid e on e obtain s th e diagona l
term wit h

»-5?£<« r + * > + *<'-*"ds5S)

- 1 f°° rdr 1 f°°
7T2 J_ QO 2 tanh(Trr ) TT 2
J_00 2

which i s exactl y (1 6.35 ) becaus e h(r) i s even . Moreover , on e obtain s th e su m o f

Kloosterman sum s wit h

Inserting (1 6.34 ) an d interchangin g th e integratio n w e ge t

because h(r) i s even. Next , b y th e recurrenc e formul a

(s + 1 ) Js+1 (y) - (s - l)J s-i(y) = -2sy{J s(y)/y)\

we ge t
-2ix [<*> [<*> rh(r) 1( ,
g(x) = / / M2ir(y)/y) drdy
* Jx J-oo cosh(Trr )
which coincide s wit h (1 6.36) .
The formul a (1 6.34 ) finds it s basi c application s fo r estimatio n o f th e Fourie r
coefficients o f cusp forms i n spectral parameter s an d th e leve l (se e the next section) .
In th e leve l aspect alon e the preliminar y formul a (1 6.32 ) i s often sufficien t an d eas y
to use . However , th e extr a flexibilit y offere d b y th e larg e famil y o f function s h(t)
in th e Kuznetso v refine d versio n (1 6.34 ) i s helpful fo r tacklin g delicat e issues , suc h
as exceptiona l eigenvalue s o r th e distributio n o f lengt h o f close d geodesie s o n th e
Riemann surfac e r\M ; cf . [Ill] , [1 1 2] .
For problem s i n analyti c numbe r theor y w e nee d t o trea t th e sum s o f Kloost -
erman sum s rathe r tha n th e Fourie r coefficient s o f cus p forms . T o thi s en d w e
need a n equatio n i n whic h th e tes t functio n attache d t o th e Kloosterma n sum s i s
given a priori , an d th e one s attache d t o th e spectra l term s ar e obtaine d b y integra l
transforms wit h explici t kernel s (somethin g lik e th e Rieman n explici t formul a fo r
primes). Unfortunately , th e formul a (1 6.34 ) i s no t sufficien t becaus e on e canno t
represent ever y requeste d functio n f(x) b y th e integra l (1 6.36) . Moreover , eve n i f
g(x) admit s th e integra l representatio n (1 6.36) , i t i s no t eas y t o fin d it s origina l
function h(r).
Historically speaking , i t wa s E . C . Titchmars h [T3 ] wh o firs t trie d t o inver t
i » g give n b y (1 6.36 ) fo r g i n a dens e subspac e o f L 2 (R + , x~ ldx), bu t
the ma p h —
 16. SUM S O F KLOOSTERMA N SUM S 411

not i n th e contex t o f automorphi c forms . The n jointl y wit h D . B . Sear s [ST ] the y
realized tha t th e imag e o f this ma p i s no t dense . Indeed , writin g (1 6.36 ) a s

9(X) = - / (J2it(x) - J-2it)(x)) } .M

IT JQ cosn(7rr )
it i s clea r tha t th e Besse l function s Jk-i(x) o f positiv e integra l orde r k — 1 with k
even ar e missed , becaus e the y ar e orthogona l t o J2it{x) — J-2it(x). However , the y
showed tha t { Ji(x)\ £ = 1 , 3, 5 , . .. } together wit h {J2it{x) — J-2it(x)\ 0 < £ < oo }
form a complete , orthogona l syste m i n L 2 (M + , x~ ldx), th e latte r i n th e sens e o f
continuous spectra l measure .

Precisely, le t f(x) b e a functio n o f clas s C 2 o n [0 , oo) suc h tha t

(16.38) /(0 ) = 0 , /<">(* ) < (s + l)- a

for a = 0,1 , 2 with a > 2 . The n w e hav e

(16.39)
/(*) = r w M - j-2u(x))(r(j2u(h) -j-2*mf{v)-). 2 /f.
Jo V oy }
smh(27rt )
1
+ Y. W - l)Jk-i(x) J k-i{y)f{y)y' dy.
k>0,k eve n ^ °

Here the integral component o f f(x) i s in the image of the integral transform (1 6.36) .
Therefore, fo r completenes s on e needs , i n additio n t o (1 6.34) , a formul a fo r sum s
of Kloosterma n sum s S(m,n;c) twiste d b y th e Besse l function s J/ c _1 (47r v / mn/c).
These w e alread y kno w fo r eac h k a s th e Petersso n formul a (1 4.1 5) . Pu t
• /"OO

1
(16.40) M f(t) = , / (J 2it(x) - J^u(x))f(x)x- dx

and
A k
(16.41) AT f(k) = \ ~Sl r J 1
k-i(x)f(x)x- dx.

Let pj{m) an d r a (ra, r) b e the Fourier coefficient s o f a complete orthonormal syste m

of cus p form s an d th e Eisenstei n serie s respectively . I n additio n le t ipjk(m) fo r
1 ^ j ^ dimSfc(r ) b e th e Fourie r coefficient s o f a complet e orthonorma l syste m o f
holomorphic cus p form s
oo

(16.42) f jk{z) = Y2 ^jk{m)m^~e(mz)

of weight k = 2 , 4, 6 , . .. Addin g Kuznetsov' s formul a (1 6.34 ) t o Petersson's formul a

(14.15) i n accordanc e t o th e decompositio n (1 6.39 ) on e derive s
412 16. SUM S O F K L O O S T E R M A N SUM S

THEOREM 1 6.5 . Let f satisfy (1 6.38). For mn > 0 we have

/A / \ °°
(16.43) Y. c-'S(m,n;c)f(^^\ = £ M/ft^ftHftfn )
c = 0 ( m o d q) ^ ' j =l

1 f°°
+ Y1 A~ Mf{r)f a(m,r)Ta(n,r)dr

0<fc=0(mod 2 ) l^j^dimS k{T)

Observe tha t th e diagona l ter m o n th e spectra l sid e i s gone !

One als o needs a formula fo r sum s of Kloosterman sum s S(ra , n; c) with ra,nof
different sign . Thi s i s a similar case , actuall y i t i s simpler becaus e th e holomorphi c
cusp forms d o not appea r (the y have no Fourier coefficient s a t negativ e frequencies) .
Put
/>oo
1
(16.44) K f(t) =4 / K 2lt{x)f{x)x- dx.
Jo

THEOREM 1 6.6 . Let f satisfy (1 6.38). For mn < 0 we have

(16.45)
1
£ c- S(m,n]c)f(Any^\ =f^K f(tj)pj(m)pj{n)
c = 0 ( m o d q) ^ ' j =l

1 f°°
+ AX ^ 4 ^ / fcf(r)f a(m,r)Ta(n,r)dr.
a

PROOF. Thi s follow s fro m Theore m 1 6. 3 (wit h g(x) replace d b y g (x)) an d

the followin g Kontorovich-Lebede v inversio n (se e (5.27.1 4 ) o f [Leb] )
1
(16.46) f{x) =^ J K 2lt{x) (J K 2lt{y)f{y)y- dy\ tsinh(27rt)dt .

We clos e thi s sectio n wit h remark s o n th e Kloosterma n sum s zet a function .

Applying th e formul a (1 6.43 ) fo r f(x) = x 2 s _ 1 formall y on e derive s th e followin g
spectral decomposition :
k 2s-l
/.^ „„ x ~(2ixJmn) „ , x v ^ T( s — s 7 -)r(s — 1 + 57-) , x , x
(16 47) 2 z (s)=
- ^fer 2 1
-fe
g coihU ) ''ftMftfr o
+ ^ 7r- (47r) - (fc-l)!r(a-|)r(s-l + |) ^ ^(m)^ fe(n)
fc>0,fc eve n l ^ j ^ d i m SK(T )

V- 1 r r ( 3 - | - i r ) r ( 8 -|+ir)_
T
Z . I47r
T /j _ T^i^A
0 0 cosh(7rr)
« ( m ' r) T » ( n ' ^

Of course , th e functio n f(x) = x 2s~l doe s no t satisf y th e condition s (1 6.38) , nev -

ertheless on e ca n prov e rigorousl y tha t th e formul a (1 6.47 ) hold s true .
 16. SUM S O F K L O O S T E R M A N SUM S 413

The serie s
AnJmn
(16.48) L mn(s) = ^ c^Sim, n ; c)J t
c E 0 ( m o d q)

seems mor e natura l tha n Z rnn(s) becaus e i t satisfie s a suitabl e functiona l equatio n

(16.49) L mn(s)-Lmn(-s) + — ^ r a ( ^ m , — J ra ^ n , - — J = TT -1 sin( — -JS(m, n) ,

a

(see Theore m 9. 2 o f [1 5]) .

16.5. Estimate s fo r th e Fourie r coefficients .
In thi s sectio n w e deriv e fro m th e Petersso n an d th e Kuznetso v formula s im -
mediate estimate s fo r th e Fourie r coefficient s o f basi c modula r forms . T o thi s en d
we simply estimat e th e involve d sum s o f Kloosterma n sum s b y usin g Weil' s boun d
for th e individua l sum . Mor e advance d method s an d stronge r result s hav e bee n
established i n [1 3] , [1 1 3 ] an d [DI] .
Our problem s reduc e t o estimatin g th e serie s (1 6.48) . Firs t w e sho w tha t

(16.50) ] T c-^IS^njc) ! < 8(cr - ^)- 2 r((ra,n))(m,n, q)^r(q)q-i- a

c = 0 ( m o d q)

if | < a ^ 1 . Indeed , b y applyin g (1 6.1 ) th e lef t sid e o f (1 6.50 ) i s bounded b y

oo

c=l
and th e las t serie s i s bounded b y
cv+D E T ^d~°-
d\(m,n)
l
Hence (1 6.50 ) follows , becaus e r(d) ^ 2dS an d ((a + \) < 2(< r - \)~ .
For k ^ 2 we hav e
y.k-1

J f e _ 1 (*)« m in(l ) ? !--)«(f]

(fc-1)!
for an y 0 ^ a < 1 . Therefor e (1 6.50 ) yield s
_2tV™n\ r(q)^
(16.51) L mn (A: - 1 ) « (2 a - l)~ z ( \~ ) " y M ( m , n))(m, n , q)*

for an y | < a ^ 1 , where th e implie d constan t i s absolute .

Next fo r s = a + it wit h \ < a < 1 we have (se e (23.41 1 .4 ) o f [GR] )

Therefore (1 6.50 ) yield s

(16.52) L mn ( S ) « ( 2 a - l ) - 2 e ' r l s l / 2 ( / ^ ' ) lM T((m,n))(m,n,q^

\ \s\q J y/q
414 6 . SUM S O F KLOOSTERMA N SUM S

where th e implied constan t i s absolute.

Note th e right sid e o f the Petersson formul a (1 4.1 4 ) equal s
S(m, n) + 27ri kLrnn(k - 1 ) .
Hence, usin g (1 6.52 ) wit h 2< J — 1 = (log3ran) _ 1 , w e conclude
T H E O R E M 1 6.7 . Let k > 2 , k even. The Fourier coefficients of an orthonormal
system of holomorphic cusp forms of weight k and level q satisfy

(16.53) / 47rNfc_i Yl^3k{m)^jk{n) = 5(m, n)

3

-f O i ——T((m, n))(m, n , q) * (ran) * log 2 3ran J

where the implied constant is absolute.

Next notic e tha t th e right sid e o f the Kuznetsov formul a (1 6.34 ) equal s
1
f r 1 N h(is/2)s .
6(m,n)g0 + — / L m n (s) — —ds
2m J {(T) COS(7TS/2 )
for an y | < a < 1 . Insertin g (1 6.52 ) wit h 2< r — 1 ^ (log3mn) - 1 w e conclude
THEOREM 1 6.8 . Suppose h(r) satisfies (1 5.20). Then for m , n> 0
h(r)dr
(.6.54) £ ft(m)ft<B)_»|L+ £ ^ ra(m,r)ra(n,r)-
7=1 ^ a
cosh(7rr)

8(m,n)go + Of H r((ra , n))(ra, n, q)2 (ran)4 )

q
where

(16.55) H = rj~ 2 f \rih{r)\dr

_1
for any 0 <7 7 ^ (41 og3ran) , and the implied constant is absolute.
In particular , Theore m 1 6. 8 yield s

f e / T ) 2 lft(n)| 2
(16 56) V e - +Y— f°° e- (t/T)2^^-dt
4^* COSh(7T^ )^ 47 T J_00 COSh(7Tt )

= g 0 + O ^ J r ^ g - V ^ ) ^ , # n i log 2 3n)
where g o — TT~ 2T2 -\- 0(1) an d the implied constant s ar e absolute.
Another interestin g choic e o f the test functio n i s
(16.57) h(t) = (X u + X~ u)2(t2 + I ) "2
where X ^ 1 . Thi s i s non-negative o n the spectrum, i.e. , for t o r it real , an d large
for th e exceptiona l points , precisel y w e hav e h(t) ^> X2s~l i f s = \ + it > | .
Clearly g 0 < C 1 and the integral (1 6.55 ) satisfie s

7/ < rj- 2xi+2r] < C X2 (log3mnX) 2

 16. SUM S O F K1
L O O S T E R M A N SUM S 4 5

_1
by choosin g rj = (41 og3ranX) . Therefor e (1 6.54 ) fo r th e tes t functio n (1 6.57 )
yields

2
(16.58) Yl |P;H| ^ 2 S j1
_
« l + — r(n)(n,^(nX)*(lo g3nX)
2
.
i<sj<i q

According to Selberg' s eigenvalue conjecture fo r th e grou p To(q) th e sum on th e

left sid e o f (1 6.58 ) i s void . Thi s ha s no t ye t bee n proved . However , th e inequalit y
(16.58) show s tha t th e exceptiona l point s appea r infrequently . Specifically , takin g
n — 1 and X — q 2 w e ge t
2
(16.59) £ |p J(l)| g 21
^- )«r(g)(log2g)2.

Here one can choose a basis ( a Hecke-Pizer basis ) suc h that |pj(l) | ^ > q~^~ £, gettin g
2(251
j ) 1 +£
(16.60) Yl 4 " «4 .

Hence w e deriv e th e followin g densit y estimat e

3 4a+£
(16.61) \{j; Sj > a}\ « g "

for an y a ^ \ an d e > 0 , th e implie d constan t dependin g onl y o n e . I f a > ^ , thi s

bound i s quit e smal l relativ e t o th e volum e o f JT 0(g)\H.
16.6. Estimate s fo r sum s o f Kloosterma n sums .
Now w e hav e everythin g read y fo r estimatio n o f sum s o f Kloosterma n sums .
The origina l su m o f Linni k give n b y (1 6.2 ) fo r q = 1 wa s treate d i n 1 97 7 b y
Kuznetsov [Kuz] . H e applie d hi s formul a (1 6.43 ) fo r a suitabl e f(x) an d imple -
mented Weil' s boun d fo r th e Kloosterma n sum s nea r th e boundar y o f th e summa -
tion rang e gettin g

(16.62) S mn(X) « X*(log2X) ^

for X ^ 1 an d ra, n ^ 1 , wher e th e implie d constan t depend s o n ra, n . Th e

exponent ^ her e i s remarkabl e b y compariso n t o th e | i n (1 6.3) , thoug h i t i s no t
yet arbitraril y smal l a s conjectured b y Linnik. I n practice on e is satisfied wit h sum s
of 5(ra , n; c) ove r th e modulu s c restricted smoothl y i n a n interva l rathe r tha n b y
the shar p cut-off . Thes e smoothe d sums , whil e no t compromisin g rea l applications ,
can b e give n th e tru e estimat e (se e (1 6.72)) .
Fix ra, n,q ^ 1 an d a functio n / o f clas s C 3 supporte d o n [|,1 ] . W e shal l
estimate

(16.63) S(X) = V cT^m , n ; c)f(2ny/mnX/c)

c = 0 ( m o d q)

for al l X ^ 1 . B y Theore m 1 6. 5 th e proble m i s transferre d t o estimatio n o f th e

Fourier coefficient s o f basi c modula r form s an d th e integra l transform s (1 6.40) ,
(16.41) o f f(xX/2).
416 16. SUM S O F K L O O S T E R M A N SUM S

Introducing th e power serie s expansio n

tJ2l+v
(-l)V
(16.64) J u(2x) = ^2
£=0
£T(£ + v + l)

into
M f J 1
f( ) = — T ^T / ( 2,-i(2x) - J 1 ^2s(2x))f(xX)x- dx

where s = h2 -\- it, we get

w M _ _ z L _ f ( t W + 28-i) 2a ^.-21
sin(27rs) ^ ' £\Y{£ + 2s)
+ th e same expressio n wit h 5 changed int o 1 — 5 ,

where f(s) i s the Melli n transfor m o f fix). Al l we need t o kno w abou t / i s tha t
\f^\ ^ K for a = 0,1, 2, 3 and some K, > 0. Integratin g b y parts u p to thre e time s
we deduc e tha t
(16.65) f(s) < K(\S\ + l ) " 3 i f Re (5) > 0,
where th e implied constan t i s absolute. Moreover , w e use the Stirlin g formul a

5 S
(16.66) r ( S )=(f) Q ( l+ o ( ^ ) ) ifRe( S)>0.

Hence, estimatin g al l term s o f th e powe r serie s fo r Mf(t), excep t fo r £ = 0 , we

deduce tha t

(16.67) M f{t) = u>(s, X) + 0 ' KX~3 X l 0 g 2X

\s\ cosh(7rt )
where

W ( a , X
- ^ j ( l ^ ~T(2-2s) J'
Next b y the functional equatio n r ( $ ) r ( l — s) = 7r/sin(7rs) w e write

(16.68) UJ(S, X) = T(2 s - 1 )/( 1 - 2s)X 1 ~2s - T( l - 2s)f(2s - l)X 2s

~\
We als o estimat e th e leadin g ter m LJ(S,X), bu t onl y fo r s = ^ + it wit h £ real,
getting
*;log2X
(16.69) M 7 ( t )< C
s| 3 cosh(7rt )

Similarly fo r the integral transfor m (1 6.41 ) w e deduce tha t

3
(16.70) Af f(k) ^K,k- X-\

Finally introducin g (1 6.69 ) an d (1 6.70 ) int o (1 6.43) , excep t fo r the points \ <
Sj < 1 , the n applyin g Cauchy' s inequalit y an d th e estimate s (1 6.53) , (1 6.56 ) w e
arrive a t
 16. SUM S O F K1
L O O S T E R M A N SUM S 4 7

THEOREM 1 6.9 . Let f be of class C3 with compact support in [\, 1 ] and | / ^ | <
1 for a = 0,1 , 2, 3. Then for m n, q, X ^ 1 we have

(16.71) S(X) = £ u(s 3,X)p3(m)p3(n) + R(X)

\<Sj<l

where UJ(S,X) is given by (1 6.68) and R(X) satisfies

(16.72) R(X) < ( l + ( m , ^ ) 1 ( l + (n , g ) ^) "r(g)r(mn)(log2mnX) 3

u^'£/i £/& e implied constant being absolute.

Assuming Selberg' s eigenvalu e conjectur e th e su m ove r ~ < s 3<1 i s void ,
so S(X) — R(X) an d (1 6.72 ) become s a pur e boun d fo r (1 6.63) . I n particular ,
it i s know n tha t th e Selber g conjectur e hold s tru e fo r th e modula r grou p wher e
q = 1 (fo r a proof , se e e.g . [DI] ; Hejha l compute d tha t th e smalles t eigenvalu e i s
Ai = 91 .1 4...) . I n thi s cas e w e get th e unconditiona l resul t

(16.73) y ^ c~ 1 S(m, n ; c) f (2ir y/mnX / c) < r(mn)(mn) ^ (log 2mn)2 .

oo
Notice w e replaced (log2ranX) 3 b y (log2ran) 2 becaus e th e facto r log2 X i n (1 6.69 )
appears fo r smal l t whic h i s no t i n th e spectru m fo r th e modula r group .
Denote th e su m ove r th e exceptiona l spectru m (hypotheticall y empty ) b y

(16.74) E(X)= J2 u(s 3,X)p3(m)p3(n).

\<s 3<l

Suppose s\ = \ + OL i s the larges t point , i.e. , A i = s i ( l — s\) — \ — a 2 i s the lowes t

eigenvalue i n th e spac e o f cus p forms . Sinc e OJ(SI,X) ^ > X 2a fo r sufficientl y larg e
X w e deduc e b y comparin g (1 6.71 ) wit h (1 6.3 ) fo r m = n wit h pi(n) ^ 0 tha t
a ^ | whic h correspond s t o th e Selber g lowe r boun d A i ^ ^ . Th e bes t know n
estimate i s a < ^ du e t o Ki m an d Sarna k [KSa ] (se e Sectio n 5.1 1 ) . Th e estimat e
a
^ n * s °btained i n [1 1 4 ] b y simple r arguments . Applyin g Cauchy' s inequalit y
and LJ(S,X) < C X 2s~l log2 X uniforml y fo r \ ^ s ^ 1 we ge t

E(X)<&Em(X)*En(X)*log2X
where
2
Em{X) = Y J\Pj{m)\ X^-\
3
2a
For an y K r ^ I w e hav e E m(X) ^ (X/F) E m ( r ) , an d (1 6.58 ) yield s
1
Em(Y) < C l + T(q)T(m)q- (m,q)2(mY)^{log2mY)2.
Choosing Y = min(X , max(l, q 2 /m(m, q))) w e deriv e

Em(X) « [ 1 + ((m,q)m/q 2)iX2a + ((m , q) m X / q2)2a}r(q)r(m) (log 2m X ) 2 .

Combining th e abov e estimate s w e obtai n

418 16. SUM S O F K L O O S T E R M A N SUM S

PROPOSITION 1 6.1 0 . Suppose the lowest cuspidal eigenvalue for T$(q) is A i =

\ — o? with 0 < a ^ \. Then the exceptional sum (1 6.74) satisfies
(16.75)
E{X) < ( l + {{m,q)m/q 2)^Xa + ((m , g ) m / g 2 ) a X a )

( l + ( ( n , g ) n / g 2 ) ^ I a + ((n , g ) n / g 2 ) a X a ) r ( ^ ) r ( m n ) ( l o g 2 m n X ) 3

iu/iere the implied constant is absolute.

Because th e boun d (1 6.72 ) i s absorbe d b y (1 6.75) , th e latte r constitute s ou r
final estimat e fo r th e su m (1 6.63) . Althoug h th e Selber g eigenvalu e conjectur e i s
not ye t proved , ou r boun d fo r S(X) turn s ou t t o b e stron g enoug h i n th e leve l
aspect t o yiel d adequat e result s i n th e mos t importan t applications . Ver y ofte n i n
practice i t suffice s t o know that A i ^ ^ . Puttin g a = \ w e get th e followin g simpl e
unconditional estimat e

THEOREM 1 6.1 1 . For m,n,q,X ^ 1 we have

(16.76) S(X) < ( l + (m , q) ^ f) * (l + (n , q) ^) \{q)r(mn) (log 2mnXf

where the implied constant is absolute.
Similar argument s wor k an d th e sam e result s hol d tru e fo r th e su m (1 6.63 ) i f
S(ra, n; c) is replaced b y S(—m, n ; c) for an y m, n ^ 1 . T o this end replac e Theore m
16.5 by Theore m 1 6.6 .
 http://dx.doi.org/10.1090/coll/053/18

CHAPTER 1 7

PRIMES I N ARITHMETI C PROGRESSION S

17.1. Introduction .
The Prim e Numbe r Theore m assert s tha t ever y residu e clas s a (mod q) wit h
(a, q) — 1 contains th e sam e proportio n o f primes , i.e. , fo r fixed q we hav e

(17.1) TT(X; q, a) ~ — — TT(X),

(17.2) ^ ( x;g>a)„ J_^(s )

as x tend s t o infinity . On e eve n ha s reasonabl e estimate s fo r th e erro r term s i n th e

asymptotics (1 7.1 ) , (1 7.2) . However , mor e importan t tha n th e erro r ter m i s th e
range o f uniformit y fo r th e modulu s q in term s o f x. Th e Siegel-Walfis z Theore m
(Corollary 5.29 ) assert s tha t

A
(17.3) il>(x ]q,a) = - ^ - + 0(x(logx)- )

for an y q > 1 , (a , q) = 1 , x ^ 2 and A ^ 0 , the implie d constan t dependin g onl y o n

A. Notic e tha t thi s estimat e i s non-trivial onl y i f q <C (logx) A .
w
The Gran d Rieman n Hypothesi s fo r L(s,x) ^ n x( m
° d q) implie s

(17.4) ^(x ; g, a) = - £ - + 0(x ^ (lo g xf)

where th e implie d constan t i s absolute , an d thi s i s non-trivial fo r q <C x? (logx) - 2 .

It i s conjecture d b y H . Montgomer y tha t

e
(17.5) 4( x;q,a)=*+0(q-ixi+ )

where th e implie d constan t depend s onl y o n e. Thi s implie s tha t th e asymptoti c

formula (1 7.2 ) hold s tru e uniforml y i n q ^ x 1 ^6. I t wa s shown , however , b y Fried -
lander an d Granvill e ([FG] , [Gra]) , extendin g previou s wor k o f Maie r [Mai ] o n
primes i n shor t intervals , tha t (1 7.1 ) an d (1 7.2 ) canno t hol d fo r q ^ x(logx)~ A fo r
any A > 0.
The GR H ma y b e ou t o f reac h fo r severa l generation s o f researchers , bu t w e
have alread y a satisfactor y substitut e o f (1 7.4) , o r rathe r th e extensio n o f (1 7.3) ,
in th e followin g form .

419
42 0 17. P R I M E S I N A R I T H M E T I C P R O G R E S S I O N S

THEOREM 1 7. 1 (E . BOMBIERI , A . I . VINOGRADOV , 1 965) . We have

A
(17.6) 7 ma x ?/>(#;</ , a) -—\<^x(logx)~
^ ( a . ? ) = i l <p{q)\

for any A ^ 0 , where Q — x 2 ( l o g x ) _ B with B = B(A), the implied constant

depending on A.
Here ar e brie f comment s o n th e histor y o f th e problem . Al l treatment s us e i n
one wa y o r anothe r th e ide a o f larg e siev e du e t o Linni k [Lil] . First , A . Reny i
[Re] establishe d (1 7.6 ) wit h Q = x e~£ fo r som e 0 > 0, the n K . Rot h [Rot ] go t
this fo r 0 = | whil e M . B . Barba n [Bar ] succeede d wit h 9 = | (an d die d soo n
after). Finall y A . I . Vinogrado v [Vi ] reache d th e leve l 0 = | whil e E . Bombier i
[Bo5] workin g independentl y i n th e sam e tim e prove d a slightl y stronge r resul t a s
stated i n Theore m 1 7.1 . Bot h author s derive d thei r result s fro m ne w densit y the -
orems fo r zero s o f L-function s whic h the y establishe d first . Bu t Gallaghe r [Ga3 ]
and Motohash i [Mot2 ] simplifie d th e argumen t t o avoi d th e us e o f zeros , instea d
feeding th e mai n idea s directl y int o the estimatio n o f bilinear form s ove r arithmeti c
progressions. Thes e development s ar e relate d t o th e identitie s o f Vinogradov , Lin -
nik, Heath-Brow n o r Vaugha n typ e fo r sum s ove r prime s discusse d i n Chapte r 1 3 .
In thi s chapte r w e borro w fro m thes e innovation s a s muc h a s possibl e t o prov e
Theorem 1 7. 1 with B(A) = 2A + 6 by employin g minima l work .
EXERCISE 1 . Prov e th e asymptoti c formul a fo r th e Titchmars h diviso r prob -
lem:
V r(p - 1 ) - ex wit h c = ^ 7 ^ 3 ) = 1.943596 . . .

as x—)• +oo. Thi s formul a wa s prove d firs t b y Linni k [Li7] . [Hint : Us e Dirichlet' s
hyperbola metho d t o reduc e t o countin g prime s = 1 (mo d d) wit h d ^ y/x, the n
use Theore m 1 7. 1 an d th e Brun-Titchmars h inequalit y (6.95). ]
See [Kol ] fo r a fairl y natura l analogu e o f thi s proble m fo r ellipti c curves .

The Bombieri-Vinogrado v theore m i s expecte d t o b e tru e wit h Q = x x~£ (th e

Elliott-Halberstam conjecture ) whic h follow s fro m (1 7.5) , bu t no t fro m th e GRH .
There ar e som e unconditiona l result s o f typ e

(17.7) T\(q)U(x;q,a)-^-) « x(logz)-4

with Q — £ 7_ £ , wher e a ^ 0 i s fixe d an d X(q) i s an y well-factorabl e functio n i n

the sens e o f siev e theor y (se e [BFI]) . Her e e > 0 , A > 0 are arbitrar y an d th e
implied constan t depend s o n a, £, A. Thi s resul t contain s usefu l informatio n abou t
equidistribution o f prime s p ^ x in a fixed residu e clas s a (mod q) o f modulu s q
larger tha n x^ , i n whic h rang e th e GR H doe s no t apply .
One ca n handl e th e large r modul i q if th e averagin g ove r th e classe s a (mod q)
is permitted . Thus , P . Tura n [Tu2 ] derive d fro m th e GR H tha t
2
4
(17.8) Yl* ( ^ * ; g , a ) - - 7 - r ) <x(logx)
a ( m o d q)
 17. PRIME S I N ARITHMETI C PROGRESSION S 421

where th e implie d constan t i s absolute. Thi s estimat e i s non-trivial fo r q a s larg e

as x(log:r)~ 5 . Introducin g th e additiona l averagin g ove r th e modulu s M . B . Bar -
ban [Bar ] an d H . Davenpor t an d H . Halbersta m [DH2 ] establishe d unconditiona l
results o f comparabl e strength .
THEOREM 1 7. 2 (BARBAN , DAVENPOR T AN D HALBERSTAM , 1 966) . We have

( 17 - 9 ) E I T (^;4,a)--^) 2 «x(io gx)-A

for any A > 0, where Q = x(logx) - j B with B = B(A), the implied constant de-
pending on A.
There ar e analogou s result s fo r th e Mobiu s functio n i n arithmeti c progressions .

EXERCISE 2 . Prov e tha t

(17.10) y ^ ma x ^ fi(m) <x (logx)~A
m=a(mod q)

for an y A > 0 , where Q = ^ ( l o g x ) - ^ wit h B — B(A), the implie d constan t

depending o n A.
17.2. Bilinea r form s i n arithmeti c progressions .
Given a reasonable arithmeti c functio n / i t is natural t o expect it s value s to
be equidistribute d ove r primitiv e residu e classes , i.e. , on e shoul d hav e a non-trivial
bound fo r

(17.11) D f(x;q,a)= £ f(n) - -L£ /(n )

n^x ^ ^ n^x
n=a(mod q) (n,g ) =l

for al l (a,q) = 1, provide d g is not extremel y large . I n many case s on e ca n sho w

that

(17.12) D f(x;q,a) « ( £ I / H I 2 ) " * * (lo g = r ) ^.

n^x

This i s non-trivial onl y fo r q ^ (log;r) A . Havin g (1 7.1 2 ) on e ca n prov e b y th e larg e

sieve inequalit y (7.31 ) a bound fo r Df(x; q, a) fo r som e typ e o f / whic h i s good fo r
almost al l q $J x^ (logx) _ j B . I t is surprisin g ho w littl e i s require d o f /. Essentiall y
/ need s t o b e represente d a s a convolution o f tw o sequences , sa y / = a * (3, one of
which, say /?, is equidistributed ove r primitive residue classes uniformly wit h respec t
to th e modul i q ^ (logx) A , o r mor e generally , / shoul d b e wel l approximate d b y a
linear combinatio n o f suc h convolutions .
We begi n ou r investigation s b y examining a sequence (3 — (/3 n) o f complex
numbers supporte d o n 1 ^ n ^ N. W e assum e tha t fo r som e 0 < A ^ 1 it satisfies
9
(17.13) \D e(N;q,a)\^m\N^A

for al l (a,q) — 1, wher e

=c*>
42 2 17. P R I M E S I N A R I T H M E T I C P R O G R E S S I O N S

First w e show tha t (1 7.1 3 ) implie s th e following boun d fo r character sums .

LEMMA 1 7.3 . For a non-principal character \ (mo d r) and a positive integer
s we have
(17.14) | J2 Mn)\ < ||/3||iV^A3rT(S).
(n,s) = l

P R O O F . B y Mobius' inversio n w e chang e th e condition (n , s) = 1 and then we

split th e sum as follows:

(n,s) = l fc|s n=0(mo d fc)

= £Mfc)£/*w E ^x(n ) + EMfc) E &»*(") •

fe|s £|/ c (n,£ ) = l fc|s n=0(mo d fc)

Next w e estimate th e sum of ^nx{n) o v e r ( n> ^) — 1 by splitting int o classe s modul o

£r, an d for each clas s w e apply th e hypothesis (1 7.1 3) . Thi s give s u s (the leading
terms cance l ou t because \ ls non-principal )

||/3||A^A 9 £ £ | / i ( * M * r ) < \\(3\\N^ 9

Kv(r)T(s).
k\s £\k

ove rn
Then w e estimate th e sum o f finX( n) = 0(mod k) usin g Cauchy' s inequality .
This give s us
\\/3\\Ni Y, k ~"2 ^ MN'K-iT(s).
k\s
k>K
Adding bot h estimate s w e obtain (1 7.1 4 ) b y taking K = A- 6 . •

Now we proceed t o the main resul t

T H E O R E M 1 7.4 . Suppose ft = (/3 n) is a sequence supported on 1 ^ n ^ N
which satisfies (1 7.1 3). Let a — (am) be any sequence supported on 1 ^ m ^ M.
Then we have
(17.15)
Y ] ma x \D a*p(MN;q,a)\

« ||a| | ||/3||(AM5iV5 + Mi + Nl* +Q)(\ogQ) 2

where the implied constant is absolute.

PROOF. Usin g Dirichle t character s w e writ e

DMf}{MN;q,a) = -— £ x(a)(£amX(m))(£/3„x(n)) .
^ ' x ( m o d q) m n
x/Xo

Hence, reducin g t o primitive characters , th e left sid e o f (17.15) i s bounded by

£ i£ ^r) S * I £ «rnX(™)\\ J2 ^X(n)[

v v
s^Q ^ ' Kr^Q ^ ' X ( m o d r ) ( m , s ) = l (n,s )=l
 17. PRIME S I N ARITHMETI C PROGRESSION S 423

If r i s small , sa y r < R, w e appl y (1 7.1 4 ) gettin g

||/3||tfiA3 E 4 4 E r
« H/5||A^iA 3^2(logQ)2.
5^Q ^ ^ r<C H

For th e remainin g r > R w e spli t th e rang e int o dyadi c segment s P < r ^ 2P an d

for eac h o f such reduce d sum s w e apply th e larg e siev e inequalit y (7.31 ) gettin g
1
I N ||/3| | E ^ ( ^ 2 + M)^(P 2 + A0ilog Q
R<P<Q
1
< H | ||/3||( Q + M i + JV * + M ^ / r )(logQ)2.
Adding bot h result s fo r i ? = A - 1 w e get (1 7.1 5) . •

17.3. Proo f o f th e Bombieri-Vinogrado v Theorem .

Suppose x s < n ^ x. B y (1 3.39 ) wit h y = z = x$ w e writ e A(n ) = A#(n ) - f
b
A (n), wher e

A«(n) = 5 2 A(%(m) ,
I
m^x 5

Ab(n)= V J A(%(m) .
im—n
14
a: 5 <m^ix 5

Here A(£ ) i s th e incomplet e logarith m

A
(17.16) X(e) = logl- E (*0-
I

Accordingly w e spli t

(17.17) D A(x;q,a) = D At(x;q,a) + D Ai>(x;q1 a) + 0 ( x ^ lo g a;)

where th e erro r ter m come s fro m th e contributio n o f term s n < x$ fo r whic h th e

above identitie s d o no t apply .
If / i s continuou s an d monotoni c o n [l,y] , the n b y elementar y argument s w e
derive
\Df(y;q,a)\^2\f(l)\+2\f(y)\.
Applying this fo r f(£) = log £ an d f(£) = 1 we obtain D\(y; q, a) < C xs logx . Henc e
DA$ (x ; q,a) « x s lo g x an d summin g ove r g ^ Q w e ge t
ma x
(17.18) E |^A«(x;g,a) | < Qx ^ logx .

We shal l estimat e D Ab(x;q,a) b y a n appea l t o Theore m 1 7.4. T o thi s en d w e

need A b(n) t o b e i n a for m o f convolutio n o f tw o sequences , an d w e almos t hav e i t
except tha t th e restriction n = £m ^ x make s £ and m interconnected . T o relax thi s
and t o kee p hol d o f th e size s o f Im, w e ar e goin g t o divid e th e interva l 1 < n ^ x
1424 7 . PRIME S I N ARITHMETI C PROGRESSION S

into shor t subinterval s o f type y < n ^ ( 1 -f 6)y wit h x~~s < 6 ^ 1 . Ther e are
0(5~1) suc h subintervals . W e cove r A b(n) b y partial sum s o f type

(17.19) £ £ A(<)/i(m )
£rn=n
L<^(1+<5)L
Af<ra^(l+<5)Af

with L , M taking value s ( 1 + <S)J in the rang e x 5 < L,M < x*, LM < x, except for
n < xs an d (1 + J ) _ 1 x < n < ( 1 + <5)rr in which range s th e coverin g i s not exactl y
with multiplicit y one . Estimatin g th e contribution i n these exces s area s triviall y
we obtai n

(17.20) D A, (x; q,a) = ^2^2 D (LM'i ^ a

) + 0(6q~ lx lo g x)
M

where D(LM\q,a) stand s for

1
L L A(%(m)- — J2H Wtirn). W
L<£^(1+<S)L ^ L<l^{l+8)L
M<m<(l+(5)M M<m^(l+(5) M
£ra=a(mod g ) (^ra,<? ) =l

For eac h o f D(LM;q,a) w e ca n appl y Theore m 1 7. 4 wit h A = (logx)~ A becaus e

the sequenc e /i(m ) satisfie s th e hypothesis (1 7.1 2 ) b y the Siegel-Walfisz Theore m
(also th e sequenc e X(£) satisfies thi s hypothesis) . Thi s give s

(17.21) 2 ^ ma x |D(LM;g,a) | <(5Ax(logx )

, (^<?) = ! '

2
for Q — Ax 2. Summin g ove r L , M (ther e ar e 0(6 ) such segments ) w e obtain b y
(17.20) an d (17.21) tha t
1
(17.22) ^T m&x\D A>(x;q,a)\<z:(6- A+ 5)x(\ogx) 3.
^Aa,q) = l

We choos e 6 — A 2 gettin g th e boun d A2x(logx) 3 . Addin g (1 7.22 ) t o (17.18) we

obtain

(17.23) V ma x \i/>(x;q,a) - ^-fi~\ < A2x(logx) 3 .

(a,q) = l <p(q)

Here w e ca n replac e ip(x) b y x + O(Ax) b y the Prim e Numbe r Theorem . Havin g

done this , Theore m 1 7. 1 follows wit h

(17.24) B(A) = 2, 4 + 6.

17.4. Proo f o f the Barban-Davenport-Halbersta m Theorem .

We shal l establis h Theore m 1 7. 2 for arithmetic function s / mor e genera l tha n
A. I t does no t requir e / t o be a convolution typ e a s in Theorem 1 7.4 , but onl y tha t
/ i s well distribute d ove r residu e classe s t o quite smal l moduli .
 17. P R I M E S I N A R I T H M E T I C P R O G R E S S I O N S 425

THEOREM 1 7.5 . Suppose that for some 0 < A ^ 1 we have

9
(17.25) \D f{x-g,a)|< ||/MA
for all (a,q) = 1. Then
2
(17.26) Y, E * l^/(^9^)| « | | / | | 2 ( A x + Q)(logQ) 2
q^Q a ( m o d q)

where the implied constant is absolute.

PROOF. B y the orthogonality o f characters

E * l^/(*;<7,a) l = ^y E |E/(")X(») . m
o ( m o d (? ) X ( ° d <? ) n ^x
XT^XO

Averaging ove r q, then reducin g t o the primitive characters , w e see that th e lef t
side o f (17.26) i s bounded by

E ^ j E ^r) E* | E /(»)x(n)f .
VJ Vy
s^Q ^ l<r<C Q ^ x ( m o d r ) n<C x
(n,s) = l

At thi s poin t w e apply th e same argument s a s in the proof o f Theorem 1 7. 4 with

a — ^ — f gettin g th e bound
| 2 { x A 3 # 2 + Q + xH xR- l
}(\ogQ)2.
Choosing R = A * we arrive a t (1 7.26) . •
COROLLARY 1 7.6 . Let the conditions be as in Theorem 1 7.4- Let ab ^ 0 . Then
we have
(17.27)
aml3n a
E | E E
<7^Q a r a = 6 n ( m o d q )
-~rtn){ E -)( E)&>
(m,q) = l (n,g =l
(g,ab) = l ( m n , g )=l

« ||a| | ||/3||(M + Q)5(Ai V + Q)i(logQ) 2 .

PROOF. Th e difference o f the sum s i n (1 7.27 ) i s also give n by

Y* E * D a(M;q,u)Dp{N;q,v).
au=bv(mod q)

Hence th e resul t follow s b y Cauchy' s inequalit y fro m Theore m 1 7. 5 applied fo r

/ = a wit h A = 1 and for / = / ? with A . •
This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/19

CHAPTER 1 8

THE LEAS T PRIM E I N A N

ARITHMETIC PROGRESSIO N

18.1. Introduction .
After havin g unifor m distributio n o f prime s i n primitiv e residu e classe s t o a
given modulu s q the natura l questio n t o as k i s how bi g i s the firs t prim e i n a give n
class? Denot e
(18.1) p(q, a) = min{p : p = a(mo d q)}.
Clearly p(q, a) canno t b e smalle r tha n ( 1 + o(l))ip(q) logq fo r som e a(mo d q). I t i s
conjectured tha t
1 +e
(18.2) p(q,a)<£q
while th e Rieman n Hypothesi s fo r Dirichle t L- functions o f character s x(mo d q)
implies
(18.3) p{q,a)<^^{q)\ogq)\
Unconditionally, i t i s easy t o deriv e fro m th e Siegel-Walfis z Theore m (se e Corollar y
5.29) tha t
p(q,a) <exp(<f )
for an y e > 0 , th e implie d constan t dependin g o n e (ineffectively) . Moreover , i f
there i s no exceptional characte r modul o g , then i t follow s fro m th e Prim e Numbe r
Theorem tha t
p(<7,o)<«clo«
where c i s a positiv e constant . Th e celebrate d theore m o f Yu . V . Linni k [Li4] ,
[Li5], [Li6 ] o f 1 94 4 assert s
THEOREM 1 8. 1 (LINNIK) . There exist absolute constants c ^ 1 and L ^ 2
such that
L
(18.4) p(q,a)^cq .

This beautifu l theore m i s on e o f th e greates t achievement s i n analyti c numbe r

theory. Originall y Linni k di d no t giv e a numerica l valu e o f L , thoug h hi s metho d
was effective , i.e. , bot h constant s c an d L coul d hav e bee n compute d give n th e
time an d will . Her e i s a selecte d lis t o f th e Linni k constan t produce d b y variou s
researchers
L = 1 0,00 0 Pa n (1 957) , L = 2 0 Graha m (1 981 ) ,
L = 1 1 1 Che n (1 965) , L = 1 6 Wan g (1 986) ,
L = 8 0 Jutil a (1 977) , L = 5. 5 Heath-Brow n (1 992) .
427
42 8 18. TH E LEAST PRIM E I N A N ARITHMETIC PROGRESSIO N

In thi s chapte r w e prove Linnik' s theore m withou t determinin g th e constant L.

Our treatment s borro w substantiall y fro m th e magnificent wor k of S. Graham [Gl],
[G2]. Fo r readers intereste d i n learnin g othe r method s (th e Turan power-sums) ,
we recommen d Chapte r 6 of E. Bombieri [Bo2] . Th e sharpest tool s ar e developed
in Heath-Brow n [HB4] .
All proof s o f Linnik's theore m us e in one form o r anothe r th e following thre e
principles abou t th e zeros o f the Dirichlet L-functions . Throughou t w e denote

(18.5) L q(s)= f jL ( S >X)-

x(mod q)
w m
A zer o o f L(s , x) b e denote d b y p = (3 + ry , o r b y p x = /3 X + i"f x i f the
dependence o n th e characte r nee d b e displayed . Fo r \ < a ^ 1 an d T ^ 1
we denot e b y N(a, T , \) th e number o f zero s p x counte d wit h multiplicit y i n the
rectangle
(18.6) a < a ^ 1 , \t\^T.

Thus
N
(18.7) N q(a,T)= Y, (<*>T>X)
x(mod q)

is the number o f all zeros of Lq(s) counte d wit h multiplicit y i n the rectangle (1 8.6) .
PRINCIPLE 1 (TH E ZERO-FREE REGION) . There is a positive constant c\ (ef-
fectively computable) such that L q(s) has at most one zero in the region
(18.8) a^ 1 - c i / l o g g T , \t\^T.
The exceptional zero, if it exists, is real and simple and it is for a real, non-principal
character.
PRINCIPLE 2 (TH E LOG-FRE E ZERO-DENSIT Y ESTIMATE) . There are positive
constants c\, C2 (effectively computable) such that for any \ ^ a ^ 1 and T ^ 1 ,
C 21
(
(18.9) N q(a,T) ^c(gT) -a).

PRINCIPLE 3 (TH E EXCEPTIONA L ZER O REPULSION) . There is a positive con-

stant C 3 (effectively computable) such that, if the exceptional zero fii exists, say
L{P1,Xi) = 0with
(18.10) l - c i / l o g g T ^ / ? i < l ,
then the function L q(s) has no other zeros in the region

(18.11) v>l-c 3 j ^, \t\^T.

The firs t principl e i s classical (du e to Landau) , an d is proved i n Sectio n 5.9.

The secon d principl e i s due to Linnik [Li6 ] and the third principl e i s a quantitativ e
version o f the Deuring-Heilbronn phenomeno n (als o due to Linnik [Li5]) . Bombier i
[Bo2] establishe d a somewha t stronge r versio n o f Principl e 3 . H e showed tha t i f
there i s a n exceptiona l zero , the n Principl e 2 can be mad e stronge r a s well . W e
extract fro m hi s result th e followin g
 18. T H E L E A S T P R I M E I N A N A R I T H M E T I C P R O G R E S S I O N 429

PROPOSITION 1 8. 2 (BOMBIERI) . There is a positive constant c 2 (effectively

computable) such that, if the zero fi\ satisfying (1 8.1 0) exists, then
(18.12) N q(a,T) « ( 1 -/JxHloggTXgT)^ 1 -")
for any | ^ a ^ 1 , where the implied constant is absolute (and effectively com-
putable).

EXERCISE 1 . Sho w tha t Propositio n 1 8. 2 implie s Principl e 3 .

EXERCISE 2 . Deriv e fro m Propositio n 1 8. 2 th e Siege l boun d (5.73 ) fo r rea l
zeros o f Dirichle t L-functions .

The numerica l value s o f the constant s ci , c2, c% an d c yield th e Linni k constan t

L, bu t fro m ou r estimate s thi s wil l b e quit e large . Assumin g (a s w e can ) tha t q
is sufficientl y larg e an d T ^ log g th e abov e principle s ar e know n t o hol d fo r th e
constants c\ = y^ , c2 = 3 , c$ = \.
18.2. Th e log-fre e zero-densit y theorem .
Recall th e Huxle y densit y estimat e (se e Theore m 1 0.4 )
(18.13) N q(a,T) « (qT)^'^ (logqT)A
where A i s a n absolut e constant . Therefore , (1 8.9 ) i s ne w onl y fo r zero s nea r th e
line R e s = 1 , namely , fo r a wit h
1
-
(18.14) l - a < AC log£
where
(18.15) C = \ogqT.

In thi s sectio n w e ar e goin g t o prov e Principl e 2 wit h th e constan t c 2 — 47,

however, th e metho d i s capabl e o f givin g a muc h smalle r constan t b y direc t modi -
fications.
Note tha t fo r th e principa l characte r N(a, T , Xo) equals th e numbe r o f zero s of
the Rieman n zet a functio n i n th e rectangl e (1 8.6) . Thu s b y th e Vinogrado v zero -
free regio n (Corollar y 8.28 ) an d th e Huxle y densit y estimat e (se e Chapte r 1 0 ) i t
follows tha t
13
(18.16) i V ( a , T , x o ) « T ( -).
This coul d als o b e establishe d withou t appealin g t o th e zero-fre e regio n alon g th e
lines o f this section , bu t w e exclude th e principa l characte r fro m consideratio n her e
to simplif y ou r exposition .
The mai n ide a behin d th e proo f o f Principl e 2 i s th e sam e a s fo r th e densit y
estimates i n Chapte r 1 0 , i.e. , w e construc t a Dirichle t polynomia l whic h serve s
as a zero-detecto r becaus e i t assume s enormou s value s a t th e zero s o f L(,s,x) -
Then, usin g th e dualit y ide a followe d b y estimatio n o f resulting characte r sums , w e
obtain th e desire d boun d fo r N q(a,T). I f w e ha d applie d thi s directl y t o L(s,x )
then a los s o f logarithmi c factor s woul d occur , whic h i s no t acceptable . Therefor e
some refinement s ar e necessary . W e eliminat e th e logarithmi c factor s b y reducin g
the coefficient s o f L(s,x) wit h a devic e simila r t o tha t whic h gav e u s a positiv e
43 0 18. T H E L E A S T P R I M E I N A N A R I T H M E T I C P R O G R E S S I O N

proportion o f zero s o f ((s) o n th e critica l lin e (se e Chapte r 24) . Ther e ar e othe r
methods fo r establishin g Principl e 2 , suc h a s th e Powe r Su m Metho d o f P . Tura n
(used b y Bombieri , Jutila , Montgomery) .
We introduc e tw o kind s o f mollification s t o L(s , x). Pu t
OO

(i8.i7) ^ ) = E ( EA ^ ) ( E ^ ) n " s
1 d\n b\n

where A ^ i s a continuousl y cut-of f Mobiu s functio n an d 0^ is a kin d o f coefficien t

used i n siev e methods . Thoug h A ^ an d 0^ appear t o hav e simila r structure s the y
play her e distinc t roles . Th e firs t on e i s applied t o annihilat e th e earl y term s o f th e
series (excep t fo r n = 1 ) whil e th e secon d on e i s applie d t o sif t ou t (o r rathe r t o
diminish th e contribution ) o f th e term s whic h hav e smal l prim e factors . Precisel y
we choos e

(18.18) A d = /i(d ) minfl, i^SfZ l

for 1 < d ^ z wher e 1 < w < z, an d w e se t A ^ = 0 if d > z. The n w e choos e

(18.19) ^ =* E ^ t
(a,bq) = l

where G i s the normalizatio n facto r suc h tha t 0\ — 1, i.e. ,

/i2(a)
(18.20) G =Y . . .
V
a<:v ^ '
(a,9) = l

The coefficient s Ot, ar e comin g fro m th e A 2-sieve an d hav e th e followin g propertie s

(see Sectio n 6.5) :

(18.21) |0 6| <1

Ol,C>2

(18.23) G^ —log y
q
(in fac t on e coul d choos e Ob fro m essentiall y an y upper-boun d siev e o f leve l y) .
Since A ^ agrees wit h fi(d) fo r 1 < d ^ w i t follow s b y Mobiu s inversio n tha t

v—-v (I i fn = 1 ,
18.24 Y\ d=\
V}
4 ^I 0i f 1< n ^ w.
d\n
For an y n > w w e have th e trivia l boun d
Xd r(n)
(18.25) \52 \ ^ '
d\n
 18. T H E LEAS T P R I M E I N A N A R I T H M E T I C P R O G R E S S I O N 4 3 1

however, i t wa s show n b y S . Graha m [Gl ] tha t

(18.26) E ( E A d ) % ^ . (1 +0(^)).
b to
w<n^x d\n ' <

Since w e d o no t see k th e bes t constant s w e mak e th e comfortabl e choic e o f th e

levels y,w,z, namel y
2 7
(18.27) V=(qT) , w=(qT) ,z = (qT) s.

For an y x(mo d q) we hav e th e twiste d serie s

oo

(18.28) ^(*,x ) = E ( EA " ) ( E ^ ) x ( n ) n " i -

1 d\n b\n

This factor s int o

(18.29) K(s,x) = L(s,x)M(s, X)

where

(18.30) M ( « ) = E ( E EA A ) ^ m ) ™ " s .
m [b,d]=m

Actually w e shal l tak e onl y a partia l su m o f K(s, x) 5

(18.31) K x(s,X)= J2 ( E A ^ ) ( E ^ ) x ( n ) n " S

l^n^x d\n b\n

with
23
(18.32) x = (qT) .
Notice tha t m = [6 , d] ^ bd ^2/2 and fo r X 7^ X o
s
(18.33) \Y^x(n)n- \^2q\s\(~y.

Hence, fo r s = a + it i n th e rectangl e (1 8.6 ) th e tai l o f K(s, x) satisfie s

(18.34) \K(s, X) ~ K x(s,x)\ < 2q\s\yzx^ < AqTyzx~ a < 1.
For s = p, a zer o o f L(s , x), w e hav e K(p, y) — 0 by (1 8.29) , s o (1 8.34 ) become s
(18.35) \K x{p,X)\^\-

Extracting fro m K x(p, x) th e firs t term , thi s inequalit y become s th e zer o detecto r

(18.36) | ]T Q T A d)(5»X(n)n- >3-

w<n^x d\n b\n

For th e trivia l characte r X — Xo th e detecto r o f zero s i s somewha t differen t

from (1 8.36) , s o w e shal l coun t thes e zero s separately . Le t <S(x ) denot e th e se t o f
zeros o f L(s,x) (counte d wit h multiplicity ) i n th e rectangl e (1 8.6) . B y (1 8.36 ) w e
deduce tha t th e tota l numbe r
(18.37) R= "£, \ S(X)\= £ j V ( a , T , X )
XT^X o x / x o
432 18. TH E LEAS T PRIM E I N A N ARITHMETI C PROGRESSIO N

satisfies

*<2E E | E (E^)(E^KH ^
X s£S(x) w<n^.x d\n b\n
2
E |(E^)(E^)I| E E c*(*)x(n)r
w < n ^ x d\n b\n X s€<S(x)
for som e number s c x(s) wit h |c x (s)| = 1 . B y Cauchy' s inequalit y w e obtai n
R2 ^ AUV

where
, 2
(18.38) U= ) n ™ '~ -2a
Y, ( E \ *d)
w<n^x d\n

and

(18.39) ^ = E / (n ) ( E ^ ) 2 n 2 Q _ 1
|EE c
x(s)x{n)n
n b\n X seS(x)

Here / i s an y non-negativ e functio n suc h tha t f(n) ^ 1 for w < n ^ x.

Using (1 8.26 ) on e derive s b y partia l summatio n tha t fo r an y \ ^ a ^ 1 ,

(18.40)
log z/w I \ log z/w/ J

To estimat e V w e square ou t an d chang e th e orde r o f summatio n gettin g

(18.41) y ^ ^ ^ ^ | £ ( X l ^ 2 j S l+ s 2+ i_ 2a )|
Xl X 2S iS 2

where
B(X,s)=Y,f(n)(Y/^)2x(n)n-s
n b\n
for x — X1X2 and s = si - f 5 2 + 1 — 2a. Not e tha t R e (s) ^ 1 . Tak e an y / supporte d
on [w/v, xv] whic h i s continuous, bounde d an d piecewis e monotonic . The n w e have

E /(*,)(*)--sa +0(|.|i)
n=a(mod q )

where

(18.42) F(s) = J f(QC"dZ

and th e implie d constan t i s absolute . Henc e w e deriv e tha t fo r an y (d,q) — 1 ,

E f(n)x(n)n-° = 5(x)^F(s)+0(\s\q^)
n=0(mod d)
a nd
where S(xo) = 1 <Kx ) = 0 otherwise. Thi s yield s b y (1 8.28 ) an d (1 8.29 )

(18.43) S (XjS)= 5 ( x )r^__^ + 0 | s(gr_y.

qG \ w
 18. T H E L E A S T P R I M E I N AN A R I T H M E T IC P R O G R E S S I O N 43 3

Introducing (1 8.43 ) int o (1 8.41 ) w e ge t b y (1 8.30 )

(18.44) V^Yl J2J2 \F(si + s2 + l-2a)\(logyr1 + 0(R 2 1 2

Tqvw- y)

One ca n find / suc h tha t fo r R e (s) ^ 1,

(18.45) F(s) < ( l + |5-l|logv)- 2logx.
For example , w e ma y tak e

^}
logw/f1 _ logg/ .
/(()=min 1 - ^ ,1 , 1
logv ' ' lo g
for ly/i ; ^ £ $J xv and /(£ ) = 0 otherwise. Indeed , writin g

/ ( 0 lo g v = l o g + ( W 0 - l o g + ( ^ / 0 - \og+(w/0 + \og +
(w/v0
one verifie s b y th e formul a

Jo
that

Fd - s)=/(S) = r / ( o * - 1 * = K - z , a / + w
Jo 5 lo g ^
(t/s - l)(x s — w 2v~s) . 1 / \
=2 l <mi n logx , —y -z .
s^logi; V \s\ \ogvJ
This give s (1 8.45) .
For s = s i + s 2 + l - 2a w e have j s - l j = j/^ i + /32 - 2 a + i(7i -721 ) ^ |7 i ~1'2\-
By (1 8.44 ) an d (1 8.45 ) w e ge t

(18.46) y « ^ V VV ( l + |7 i - 7 2 | l o g ^ ) - 2 + ^ V ^ 1 ? / 2 -
log y ^ rrf v

LEMMA 1 8.3 . Let \ (mo d g ) 6 e a non-trivial character, ^ ^ a ^ l , i > ^ 2 an d

t a real number. Then
1
(18.47) Yl t +1 7 "* l log^)" 2 < ^ (l - a + ~) lo g A ^ ( | t| + 1)

where A is an absolute constant.

PROOF. W e firs t suppos e tha t x * s primitive. W e hav e

where th e summatio n range s ove r al l zeros o f L(s,x) wit h /3 > 0 and B(x) i s a
constant wit h
Re£(x) = - ^ R e -
p
43 4 18. T H E L E A S T P R I M E I N A N A R I T H M E T I C P R O G R E S S I O N

(see (5.29)) . Henc e i t follow s tha t fo r 1 < R e s ^ 2 ,

(18.49) V / R e— = \ log(gM ) + R e ^-(s, x ) + 0(1 ) .

-^i s — p2 L
p

For s — a + it wit h a > 1 we hav e

2
Rp _J _ = (^-/3 )> i / , , h-*h-
+
s-p (a - /3) 2 + ( 7 - ty ' (a - /3 ) V a - :

-( S ,x)| < EA ( n ) n_CT = ~\^ = v^i + ° ( 1 ) -

Hence (1 8.49 ) yield s (dro p th e zero s wit h f3 < a)
^| 7 -i|x-2
E ( 1 + 73r)"2 ^( CT - a )(5iog«(i*i + i) + ^T + o(i)).

Choosing c r = 1 + 1 /log v w e obtain (1 8.47) . I f x(mo d q) is induce d b y a primitiv e

character x*(mo d q*) with q* \ q, then th e lef t sid e o f (1 8.47 ) agree s wit h tha t fo r
X*, so th e resul t hold s fo r an y \ 7 ^ Xo -
Notice tha t 1 + ( 1 - a) lo g v ^ v 1 '*. Therefor e Lemm a 1 8. 3 an d (1 8.46 ) yiel d

(18.50) V « Rv ^^^pL + R2Tqly2 << c Rv l-a

log y lo g v w
by th e trivia l boun d R < C qT log gT an d b y choosin g
(18.51) v = qT.
Multiplying (1 8.40 ) an d (1 8.53 ) w e obtain b y (1 8.38 ) tha t
(18.52) R < (x 2vy-a = {qT) 47
^-a\
Hence w e ge t (1 8.9 ) b y addin g th e numbe r o f zero s o f L(s,xo ) whic h i s negligibl e
by virtu e o f (1 8.1 6) . •

18.3. Th e exceptiona l zer o repulsion .

In this section we are going to prove Principle 3 , however, ou r metho d i s capable
of givin g (1 8.1 3) . I t begin s b y a suitabl e modificatio n o f th e argument s whic h w e
used fo r th e proo f o f Principl e 2 in th e previou s section .
Suppose xi(mo d q) i s the exceptiona l characte r an d /3 i i s the exceptiona l zer o
of L(s , xi) whic h satisfie s
(18.53) 5 1 = 1 - ft ^ ci(loggT ) - 1
Throughout w e assume (a s w e can) tha t c\ i s a sufficientl y small , absolut e positiv e
constant. Ou r goa l i s to prov e ther e exist s a n absolut e constan t C o ^ 2c\ suc h tha t
the functio n L q(s) ha s n o zer o othe r tha n (3i i n th e regio n

(18.54) a ^ i _ ________ _ | t |<r.

Clearly thi s resul t implie s Principl e 3 .

 18. T H E L E A S T P R I M E I N A N A R I T H M E T I C P R O G R E S S I O N 435

To tak e ful l advantag e o f th e repulsio n propert y o f /3 i w e appl y th e metho d

of zero-detector s fro m th e previou s sectio n t o th e functio n £(s)L( s + 5i , Xi) rathe r
than t o C(s). W e hav e
oo
s
(18.55) C(s)L(s + 5i , xi) = £ p(n)n-
l

where p(n) i s a positive , multiplicativ e functio n give n b y

Sl
(18.56) p(n)=J2xi(a)a- .
a\n

REMARK. On e coul d wor k a s wel l wit h th e produc t

oo
T n
(18.57) a*)L(s,Xi)=J2 ( >^n~S
1

where

(18.58) r ( n , x i ) = 5 > i ( a )
a\n

but th e sligh t shif t i n L(s + <5i , Xi) turn s ou t t o simplif y som e technica l arguments .

As befor e w e introduce tw o kind s o f mollification s t o ((s)L(s -f - <Si , xi) gettin g

oo
f
(18.59) K( S) = Y, (ZXd){J2e<>)p(n>~S
1 d\n b\n

where A ^ are define d b y (1 8.1 8 ) an d 0^ are define d b y (1 8.1 9 ) wit h th e parameter s

y, w, z give n by (1 8.27) . The n fo r an y non-principal characte r x(mo d q) we consider
the twiste d serie s
oo

(18.60) K(s, x) = £ ( E Xd ) ( E 0b)p(n)x(n)n- s

.
1 d\n b\n

This factor s int o

(18.61) K(s, x ) = L(s, X )L(s + 5 uXXi)M(s', x )

where

(18.62) M(s, X) = Y,{J2Y, X^) T\(P(P)-X{p)p- s 2h

- Mrn)m-s.
m [b,d]=m p\m

a s
To se e thi s w e arrang e K(s,x) follows :

s
K(s> X) = J2{1 2Y, ^dO b)x(m)m- ^ p(mn)x{n)n-* .
m [6,d]= mn
436 18. TH E LEAS T PRIM E I N A N ARITHMETI C PROGRESSIO N

Opening p(mn) b y (1 8.56 ) w e se e tha t th e innermos t su m i s

a n=0(mo d a/(a,m)) a '

s 6
= L(s, X)J2xi(c)c-^ £ XiX(a)a- ->
c\m (a,c ) =l

= L(s , x)L(s + ^ xxi ) £ X i (c)c"* I T 1 - X1 X(P))P" 8-41

c]ra p| c

and thi s yield s (1 8.61 ) an d (1 8.62 ) becaus e m i s squarefree . B y (1 8.61 ) i t follow s

that K(s,x) i s holomorphi c i n th e whol e comple x plan e (i f x — Xi? the n th e pol e
of L(s + #i , Xo) a t s = 1 — Si = fii cancel s wit h th e zer o o f L(s, Xi)) -
As before w e take onl y a partia l su m o f K(s, x) ,

s
(18.63) K x(s,X)= £ ( £ > ) ( £ > ) P(")x(")«~
l^.n^x d\n b\n

with x give n b y (1 8.32) , an d w e show by contour integratio n tha t th e tai l o f K(s, x)

satisfies

(18.64) \K(s,x)-K x(s,x)\^%

for 5 = a + it wit h < J ^ | an d |£ | $ J T. The n fo r s = p , a zer o o f L(s , x) whic h i s

different fro m /3 i i f % = xi ? w e have K(p , x) — 0 by (1 8.61 ) , therefore (1 8.64 ) yield s

(18.65) | J2 (£^)(Efc)/>Mx(")«~i H
if p = / J + i 7 wit h / ? ^ | an d |7 J < T .
The inequalit y (1 8.65 ) i s viewe d a s th e zero-detector . Thi s wouldn' t b e usefu l
if no t fo r th e fac t tha t p(n) i s ver y smal l quit e frequently . On e canno t exploi t
a possibl e cancellatio n o f term s becaus e o f th e complexit y o f th e coefficient s A^ .
Remember tha t i t i s du e t o th e introductio n o f th e factor s

(18.66) cj{n) = Y, Xd
d\n

that th e earl y term s wit h 1 < n ^ w ar e annihilated . Now , ou r strateg y wil l b e t o

remove these factor s (b y applying Holder' s inequality ) t o create a sum o f v{n) 2p(n),
where
(18.67) is( n) = Y2°b
b\n

over n wit h w < n $ J z. Whe n applyin g Holder' s inequalit y on e coul d als o remov e
the facto r ^(n) 2 , bu t i t woul d resul t i n a los s o f log g whic h i s no t acceptable . Fo r
this reaso n w e retain v{n) 2p(n). Still , th e su m o f u(n) 2p(n) i s manageable becaus e
the suppor t o f #5 is relatively smal l b y comparison t o th e rang e o f n. I n thi s wa y b y
taking advantag e tha t p(n) nearl y vanishe s quit e ofte n (dependin g o n ho w clos e t o
one is the exceptiona l zero ) w e derive fro m th e inequalit y (1 8.65 ) tha t /3 i s not clos e
to one , precisel y thi s inequalit y prohibit s p — /3 + ij t o b e i n th e regio n (1 8.54) .
 18. T H E L E A S T P R I M E I N A N A R I T H M E T I C P R O G R E S S I O N 43 7

Having outline d th e strateg y w e proceed t o details . B y (1 8.65 ) w e ge t

£ \oj{n) V{n)\p{n)n-P > \.

w<n<^x

Hence applyin g Holder' s inequalit y

2
(18.68) 6U VW ^ 1
where

2
V= J2 v (n)p3(n)n-\
w<n^x
2 1
W= Y, v (n)p{n)n- .
w<n^.x

For U w e hav e b y (1 8.40 )

(18.69) U^x^-n.
In V w e estimat e p 3(n) b y r 3(n) an d appl y a siev e o f dimensio n 8 (se e th e Funda -
mental Lemm a 6.3 ) gettin g

(18.70) V « fe^) « 1 .

We shal l sho w tha t

(18.71) W < £ i l o g z .
Multiplying thes e estimate s w e derive fro m (1 8.68 ) tha t
1 4{
(18.72) S 1 {logx)x ~^ > 23c 0,
where C Q is a positiv e absolut e constant . Assumin g (1 8.53 ) hold s wit h 2c\ ^ C o we
conclude b y (1 8.72 ) wit h x = {qT) 23 tha t p = 3 + 2 7 is not i n th e regio n (1 8.54 ) a s
claimed.
We no w procee d t o th e proo f o f (1 8.71 ) , i t i s her e wher e th e exceptiona l zer o
plays it s role . Ou r treatmen t i s essentiall y elementary , ye t som e argument s ar e
quite subtl e (w e borro w thes e fro m th e wor k o f Bombier i [Bo2]) . W e begi n b y a n
asymptotic evaluatio n o f W. T o thi s en d w e examin e th e generatin g zet a functio n
00
2
(18.73) W(s) = J2 v {n)p{n)n-s.
1

This i s just (1 8.60 ) wit h \ d = 0 d an d \ = Xo, so b y (1 8.61 ) an d (1 8.62 )

W(s)=L(s,xo)L(s + 6 1 ,Xi)M(s)

where
M(s)= J2 °ml[( P(p) - P-s-2Sl)m~*
(m,g) = l p\m
1438 8 . TH E LEAS T PRIM E I N A N ARITHMETI C PROGRESSIO N

and

[b1,b2]=m

Hence W(s) i s holomorphic everywher e excep t fo r a simpl e pol e a t s = 1 with

res W(s) = ^ L ( l + 5 U Xi)M(l) .

s=i q

By contou r integratio n w e derive tha t

(18.74) W = ^ M ( l ) L ( l + <Ji , xi) lo g - + O (-) .

Now we need estimate s fo r M(l ) an d L( l + 5i, Xi)- Fro m siev e theory (se e Chapte r
6) w e ge t

1
(18.75) M ( l ) =£ ^n^P)-^" "1
2
' ) " 21
"
(ra,g) = l p|r a

«no-T)-*n('-T)-
w e nrs
To estimat e L( l + 5i,xi ) ^ examin e

1
(18.76) P(x,y)= Y, ( + Xi(p))p~ 1 -

LEMMA 1 8.4 . For x > y ^ q 2 and q sufficiently large we have

(18.77) P(x,y) ^4(Jilogx .

P R O O F . Conside r

where r(n, Xi ) i s given b y (1 8.58) , s o it i s a multiplicative functio n wit h r(p , xi) =

1 -f - Xiip)- Multiplyin g P{x,y) an d S(y) w e deduc e fro m th e non-negativit y o f
r(n, xi ) tha t
P(x,y)S(y)^S(xy)-S(y).
Then b y th e elementar y formul a (se e (22.1 1 ) )

S(x) = L(l , Xi)(logx + 7 ) + L'(l , xi) + 0(qix~* logs )

we ge t

S(xy)-S(x) = L(l,xi)logx + 0(qiy~hogy) ^ 2L(1 , Xi) logx

 18. TH E LEAS T PRIM E I N A N ARITHMETI C PROGRESSIO N 43 9

because L ( l , x i ) ^ > q~? an d y ^ q 2. O n th e othe r hand , w e deriv e a lowe r boun d

for S(y) a s follow s

S ( y ) > y - * £ T ( n , ; a ) ( i - - ) n~^

J(i)) s(
2iri ./(i s+ l)

-M+«»-'«*> >sr^.x.).
Combining thes e bound s w e complete th e proo f o f (1 8.77) . •

REMARKS. Lemm a 1 8. 4 show s tha t i f Si = 1 — fi\ — o(l/logg) , the n XI(P) —

— 1 for almos t al l primes i n segment s q 2 < p ^ q A. Fo r ver y smal l prime s th e abov e
arguments ar e rathe r crude , bu t se e Chapte r 22 , i n particula r (22.1 00) .
Now w e ar e read y t o estimat e L( l + 5i , Xi)- Fo r an y 1 < s ^ 2 we hav e
s
as)Hs,xi)~l[(l + T(p,xi)p- )
p

and b y Lemm a 1 8. 4 w e obtain b y partia l summatio n

J | ( l + r(p,xi)p~s)^exp(^r(p,xi)p"s) < e x p ( — ^ ) .
p>y p>y

Hence
s
L(slXi) « (s - l)ex p ( ^ ) H(l + r(p,xi)p- )-
p^y
In particular , w e ge t

(18.78) L( l + <5i,Xi)«<5i n ( ! + — ) •

Combining (1 8.74) , (1 8.75 ) an d (1 8.78 ) w e get W < <5 i logx-f-g - 1 whic h complete s
the proo f o f (1 8.71 ) , an d o f Principl e 3 .

18.4. Proo f o f Linnik' s Theorem .

Throughout ci , C2, C3 and c are th e absolut e constant s fro m Principle s 1 , 2 an d
3. W e assum e (a s w e can) tha t

(18.79) c,c 2 > 1 > ci,c 3 > 0 .

Let
4c2
(18.80) x ^ q

and pu t

(18.81) R = x 1 /2C2.
44 0 18. TH E LEAS T PRIM E I N A N ARITHMETI C PROGRESSIO N

We us e th e truncate d explici t formul a (se e (5.65) )

YWJ Y\H) x ( m Q d q) px H X

where J2 restrict s th e summatio n t o zero s p x = f5 x + ryx o f L(s , x) i n cr ^ | , \t\ ^

R. Firs t w e estimat e th e su m ove r zero s p x wit h \T < \*yx\ ^ T for 1 < T ^ R by
using Principl e 2 as follows :

x
|£ x(a )2 ^ x l< |E E ^ = "I / ' ^ ^ ( « , r )
/ 2
X(modg) l T < | 7xKT X | 7x|<T

a
= ^ Nq(lT) + ?-(\ogx) [ N q(a,T)x da
1l
Jl/2
2c 2
^ — x HqT)^ + %x(\ogx) I ((qT^/x^da
T 1
Jl/2
2c x log x ^ , _ x
^ T l o g ( ^ T ) ^ ) ^ 2 c ( C 2+ 1 )
T
provided x ^ (gT) C2+1 whic h i s our case . Notic e th e absenc e o f logx i n the las t
bound. Henc e th e explici t formul a reduce s to
0
(18.82) , ( , ; , , „ ) , _ - _ £ ««> £ 7 7 + ( ^ + « <^
X(mod q) P x

where Y restrict s th e summation t o zeros p x wit h |7 X | ^ T , and th e implied

constant i s absolute. W e choos e
(18.83) T = q,
so th e erro r ter m i n (18.82) i s O(x(logq)/q(p(q)).
Now w e procee d t o a more precis e treatmen t o f the su m ove r th e zero s wit h
|7 X | ^ T . I f there exist s a real characte r xi(nio d q) for which L(s,\i) n a s a r e a l
zero f3\ — 1 — 5\ wit h
(18.84) J i <ci/21 ogg ,
an
we exclud e i t fro m J2 d we estimat e th e remainin g su m Yl b y an appeal to
Principle 1 as follows:

Px
X m
x(mod q Px x( od q) P x
pl-rj pl-rj
a a
-2 / x dNq(a,T) = 2x*Nq(\,T) + 2(logx) / N q(a,T)x da,
Jl/2 J\
where
(18.85) rj = Cl /2logq
if th e zer o /3 i doe s no t exist , an d
(18.86)7 7 = c 3 |log(2(Jilogg)|/21 ogg
 18. TH E LEAS T PRIM E I N A N ARITHMETI C PROGRESSIO N 441

if fiiexists. B y Principl e 2 we hav e

2c
Nq(a,T) ^ cq ^-a\
so ou r su m ove r th e zero s p x othe r tha n /3 X is bounded b y
rl-ri
2cx2qC2+2cx(logx) I (q^/xy-^da

2«iogx / ^ y . . ^
log(xq~2c2"> \ x J
Therefore w e hav e establishe d th e followin g
PROPOSITION 1 8.5 . Let c,ci,C2,C 3 be the absolute constants from Principles
1, 2, 3. Then for x ^ q 4°2 we have

(18.87) ^ ( x;g,a) = ^ | i _X l ( a ) ^ + ^ x - ^ + 0 ^)|

where the term involving Pi does not exist if Pi does not, r\ is given by (1 8.85) or
(18.86) respectively, \9\ ^ 4 and the implied constant is absolute.
In th e first cas e o f Propositio n 1 8. 5 we deduc e
THEOREM 1 8.6 . Let ci , c2 and c be the absolute constants from Principle 1 and
Principle 2 respectively. Suppose for any real character \i (mo d q) there is no real
zero of L(s,xi) ^ n
(18.88) a> l - c i / 21 o g g .
4c2
Then for x ^ q we have

(18.89) ^ ; , , a )= 4 - (l + ^cexp(-^^)+ 0(^

where \0\ ^ A, and the implied constant is absolute (effectively computable).

In th e secon d cas e whe n fi\exists, w e assum e tha t x ^ q u wher e
/ 4 4 log: 8c
(18.90) z / = max(4c 2 ,—,
Ci C 3 | l o g C i |

and w e estimate a s follows :

r £ i— 1 r —Si
1 " X i ( a ) - r- ^ 1 - - 5 - > Pi ~ * " * >Pi-q
Pi P i
= i - g °' -6i^ — —— 5 i
1 + i/dilog g
^ i l o g g 4* i
l + z/ci/ 2 3c i
„/2 ^ - * „ / 2 {2di logq) »c3/4 ^ (25 i l ogq)c-3/4-l ^ »1 <*i_ l Q g g _
=
loggjc x - ^ -
4cci

Inserting thes e estimate s int o (1 8.87 ) w e deduc e

442 18. T H E L E A S T P R I M E I N A N A R I T H M E T I C P R O G R E S S I O N

T H E O R E M 1 8.7 . Let c, ci,C2,C3 6 e the absolute constants from Principles 1 , 2

and 3. Suppose there exists a real zero fii of L(s, \i) f or a rea ^ character \i (mo d q)
such that Si = 1 — /3i ^ c\/2\ogq. Then for x ^ q v with v given by (18.90) we
have

where the implied constant is absolute.

From Theorem s 1 8. 6 and 1 8. 7 it follow s tha t i n an y case w e have (becaus e
S1logq > q~i)
COROLLARY 1 8.8 . If q is sufficiently large and x ^ q L with
{4 4 los8 c 1
(18.92) L = max<Uc 2 ,— log8c , — — - —r \ ,
\c1 c 3 | log ci I J

then
(18.93) ^ ( x ; 4 , a ) >

where the implied constant is absolute.

This resul t i s a quantitative versio n o f Linnik's theorem .
 http://dx.doi.org/10.1090/coll/053/20

CHAPTER 1 9

T H E GOLDBAC H P R O B L E M

19.1. Introduction .
Christian Goldbac h (1 690-1 764) , a membe r o f th e Petersbur g Academy , i n a
letter o f 1 74 2 t o Leonar d Eule r pose d th e proble m o f provin g tha t ever y numbe r
TV ^ 5 ca n b e represente d a s th e su m o f thre e primes . Eule r answere d wit h th e
equivalent hypothesi s tha t ever y eve n numbe r N ^ 4 i s a su m o f tw o primes .
Inspired b y many failure s o n these problems , Edmun d Landa u challenge d (i n 1 91 2 )
to sho w tha t ever y numbe r N > 1 is a su m o f a t mos t k prime s wit h k sufficientl y
large bu t fixed. Landau' s proble m wa s solve d i n 1 93 0 by L . G . Shnirelma n (se e e.g .
[Kh]). Th e wor k o f Shnirelma n opene d a ne w direction i n genera l additiv e numbe r
theory.
In 1 93 7 I . M . Vinogrado v [V6 ] succeede d i n solvin g th e origina l Goldbac h
problem fo r al l od d i V which ar e sufficientl y large . Befor e him , G . H . Hard y an d
J. E. Littlewood [HL1 ] made a serious attack b y means of a then ne w circle method .
Indeed, the y succeede d conditionall y unde r th e Rieman n Hypothesi s fo r Dirichle t
L-functions. Vinogrado v followe d thei r lin e of attack (wit h his own beaut ification o f
the circl e method) , an d remove d th e us e of GR H i n the estimat e o f the exponentia l
sum ove r prime s

a
(19.1) £ < P^

which i s required b y th e circl e method . Hi s treatment o f this particula r su m (siev e

and doubl e sum s methods ) becam e a fundamenta l too l fo r estimatin g a larg e clas s
of sum s ove r prime s (se e Chapte r 1 3) .
In thi s chapte r w e deriv e Vinogradov' s Thre e Prime s Theore m b y a mixtur e
of hi s origina l an d somewha t moder n ideas . W e replac e th e exponentia l su m ove r
primes (1 9.1 ) b y th e exponentia l su m wit h th e Mobiu s functio n fo r whic h w e hav e
a unifor m boun d (1 3.54 ) du e t o H . Davenport . I n thi s wa y w e minimiz e th e rol e
played b y th e circl e method , an d w e do no t appea l t o th e distributio n o f prime s i n
residue classe s (th e us e o f thes e propertie s i s accumulate d i n th e proo f o f (1 3.54)) .
One ca n avoi d completel y th e ide a o f circl e metho d b y usin g onl y th e elementar y
dispersion metho d (se e th e Rutger s Lectur e Note s o f 1 994 , Chapte r 6) .
None o f th e curren t method s promise s an y chanc e t o solv e th e Goldbach-Eule r
problem fo r tw o primes. However , i t wa s established independentl y b y N . G . Tchu -
dakov [Tchu] , va n de r Corpu t [Cor3 ] an d T . Esterman n [Est ] tha t almos t al l (i n
the sense of density) eve n numbers ar e represent able as the sum of two primes. Thi s
follows by a suitable generalization o f the Vinogradov resul t (on e of the three prime s
443
444 19. T H E G O L D B A C H PROBLEM

can b e take n fro m an y sequenc e o f integer s whic h i s relativel y dense) . H.L . Mont -
gomery an d R.C . Vaugha n [MV3 ] showe d tha t th e numbe r o f exceptiona l eve n
integers (thos e whic h ar e no t representabl e a s th e su m o f tw o primes ) i s quit e
small.
For eve n number s N w e conside r

(19.2) G 2(N)= Y, A(m)A(n 2 ),

n1-\-n2=N

and fo r od d number s N w e conside r

(19.3) G 3(N) = J2 A(m)A(n 2 )A(n 3 ).

ni+n2+n3=iV

Therefore
A
(19.4) G 3{N) = J2 (n)G2{N - n).
n<N

Hence a goo d estimat e o f G 2(N') fo r sufficientl y man y N' < N wil l yiel d a corre -
spondingly goo d estimat e o f Gs(N).
The heuristi c Mobiu s Randomnes s La w describe d i n Sectio n 1 3. 1 justifies th e
following stronge r for m o f th e Goldbac h proble m (whic h implie s it s solutio n fo r
sufficiently larg e TV):
CONJECTURE. For even numbers N ^ 4 ,
A
(19.5) G 2{N) = <5 2(N)N + 0(N{\ogN)- )

where

(19.6) © 2 (jV)=C2TT^
p\NF
p>2

and

(19.7) C 2 = 2;[](1 -(P-1 )- 2)-

Here A is any positive number and the implied constant in the error term depends
only on A.
We shal l prov e th e conjectur e fo r almos t al l eve n numbers .
T H E O R E M 1 9.1 . Let A,B,X ^ 4 . The number of even integers N with 4 ^
N ^ X for which
A
(19.8) \G 2(N) - <S 2(N)N\ > BN(logN)-

does not exceed CX(logX)~A where C is a constant which depends only on A,B.
The celebrate d resul t o f Vinogrado v fo r thre e prime s follow s immediatel y b y
(19.4) an d Theore m 1 9.1 .
 19. T H E G O L D B A C H P R O B L E M 44 5

THEOREM 1 9. 2 ( V I N O G R A D O V ) . For odd numbers N ^ 7 we have

(19.9) G 3{N) = <5 3(N)N2 + 0(N 2 (lo g N)~ A)

where 6 3 (N) > 0 is given by

(19.10) ©aw= \ na - (p -1)-2) n(i+(? -1)-3) > 0,

p|iV pfT V

and A 25 any positive number, the implied constant depending on A.

19.2. Incomplet e A-functions .

Given z ^ 2 we split A(n ) a s follows:

(19.11) A(n ) = - ^ / / ( r a ) l o g m = A^r O + A^n )

m|n

where A^(n ) an d A b(n) ar e the partial sum s ove r th e divisors m ^ z an d m > z

respectively. The n w e write

(19.12) G 2(N) = Gf{N) + 2Gf (JV ) + C?2 b(W)

where
G«»(iV)= X ) A V ) A » M
ni+77 2 =AT
b bb
and G^ (N),G (N) define d similarly .
LEMMA 1 9.3 . Let N be an even number ^ 4 and z > 2 . T #e /za?je

(19.13) Gf(A 0 = e 2(N)N + 0(N (log z)~A + T(N)Nz'i + z3 )

/or an y A ^ 0, the implied constant depending only on A.

PROOF. W e hav e

G*J(N)= ] TM (m
1 )/i(m 2 )(logm
1 )(logm 2 ) ^ 1 .
r
mi, 771 2^2 : £i mi-\-l2'm2=N

The equatio n fim i + £ 2ra2 = A f i s equivalen t wit h th e congruenc e condition s

(rai,ra 2 )|Af an d

£mi(mi,m2)~ 1 = ^ ( ^ I , ^ ) " " 1 (mo d ra 1_

2(rai,ra2) )

with 1 < £ < Nm^ 1 . Therefor e th e number o f solutions i s Af[rai,ra 2 ] - 1 + 0(1)

giving
2
Gl\N) = N<5 2(N;z) + 0((zlogz) )
where

777.1,7712^2
(mi,m2)\N
446 19. T H E G O L D B A C H P R O B L E M

It remain s t o estimate 02(iV;2) . W e write

W;*)= £ 4f £ ^^(logdmOOogdm ^
*-^ d ^— '1
1
17717 2
1
1
1d\N 77 ,77 2^ 2
(mi,m2) = l
(mi77l2,d) = l

= E^ E ^( E ^x***)'-
d\N c^z/d m^z/cd
(m,cd) = l

For cd >Z2 w e estimate triviall y by

2
]T ^ c~ d~l(log z) 4
^2r(N) Z-^ {log z)4.
l
^ Cd>z 2

If cd ^ z? we derive fro m th e Prime Numbe r Theore m tha t

_£ ^2logcdm=^-{ l + O(r(cd)(l0gZ )- A )}.

m^.z/cd
(m,cd) = l

Hence

Now extending th e sum ove r cd ^ z^ t o the infinite serie s (a s we can wit h the error
term whic h i s already present ) w e obtain

d\n Y V }
(c,d ) = l ^ V }

This complete s th e proof o f Lemma 1 9.3 . •

Lemma 1 9. 3 reveal s tha t th e main contributio n t o the binary Goldbach-Eule r

equation come s fro m th e incomplete functio n A^(n) . I n other words , th e comple-
mentary functio n A b (n) give s considerabl y less . Thi s i s due to the sign chang e of
the Mobiu s functio n involve d i n Ab(n) whic h range s ove r a large segment. However ,
we do not have yet rigorous argument s to justify thi s heuristic. Bu t it is a relatively
simple matte r i n ternary additiv e problems .
19.3. A ternary additiv e proble m wit h A b.
In thi s sectio n w e give a bound fo r a general sum

Tb(N) = £ u evmA\n)
£+m+n=7V

where U£ an d vm ar e arbitrary comple x numbers . T o this en d we need a n estimate

for th e exponential sum

(19.15) S b
(a) = 5Z A b (n)e(an)
 19. TH E GOLDBAC H PROBLE M 44 7

which i s unifor m i n a. B y Theore m 1 3.1 0 w e derive b y partia l summatio n tha t fo r

any a G l an d x ^ 2 ,

(19.16) 2_, ^( m)0-ogm)e(am) ^ x(logx) "

with an y A ^ 0 , th e implie d constan t dependin g onl y o n A. Sinc e

|S b (cOK 5 ^ 1 X ! MO ) (lo g m)e(c^m)

^<CiV/z z<m^N/£

we get b y (1 9.1 6 ) tha t

(19.17) S b
(a) < C 7V(log7V)(log^)-A.

REMARK. Any sequenc e o f number s i n plac e o f A b(n) coul d b e use d provide d

the associate d exponentia l su m (1 9.1 5 ) satisfie s S b(a) < C N(logN)~A fo r an y A >
0.

Now w e ar e read y t o prov e th e followin g

LEMMA 1 9.4 . For any complex numbers u£ and v m we have
b A
(19.18) ] Tu iVrnA (n) « \\u\\ \\v\\ N(logN)(logz)-
l+m+n=N
with any A ^ 0 , the implied constant depending only on A.
PROOF. W e hav e

Tb(N) = (J2 u ee(a£)^j ( ^ v me(am)^S

b
(a)e(-aN)da.

Hence (1 9.1 8 ) follow s b y (1 9.1 7) , Cauchy-Schwar z inequalit y an d th e Parseva l for -

mula
pi , , 2
/ X^ ( ^ da=^2\ui\ 2.
e a

19.4. Proo f o f Vinogradov' s thre e prime s theorem .

Let CN be an y comple x number s fo r N ^ X,N odd . W e writ e
b b
Y, c NG2(N)= £ c N(Gl\N) + 2G^(N) + G 2 (N)).

Then w e us e Lemm a 1 9. 3 with z = X+ gettin g

A
£c NGiH.N)= £ c N<52(N)N + 0(\\c\\xH\ogX)- ).
N^X N^X

For th e sum s o f CNG^(N) an d CNG^(N) w e appl y Lemm a 1 9. 4 gettin g th e sam e

bounds a s th e erro r ter m above . Henc e
C A
(19.19) ^2 NG2(N) = ^2 c N®2(N)N + 0(\\c\\xi(logX)- ).
N^X N^X
448 19. THE G O L D B A C H P R O B L E M

For th e specia l coefficient s CN = C^AT ) — &2(N)N thi s yield s

PROPOSITION 1 9.5 . For any X ^ 3 we have
-1A
(19.20) Yl (°2( N
)~ ®2(N)N) 2 < X 3 (logX)
N eve n

where A is any positive number and the implied constant depends only on A.

Clearly (1 9.20 ) implie s Theore m 1 9.1 . Fo r th e proo f o f Theore m 1 9. 2 we appl y

(19.19) t o (1 9.4 ) gettin g
3
G3(N) = J2 A(n)C 2 (iV - n)(N - n) + 0(N {logN)-A).
n<N

By (1 9.1 4 ) th e mai n ter m i s equal t o

V V
(c,d) = l ^ ^ ^ n<N

n=N(mod d)

The inne r su m i s equa l t o

N2

if (d , A/") = 1 by th e Prim e Numbe r Theorem . Sinc e

(d,ciV) = l r \ JV \ J

this complete s th e proo f o f (1 9.9) .

 http://dx.doi.org/10.1090/coll/053/21

CHAPTER 2 0

T H E CIRCL E M E T H O D

20.1. Th e partitio n number .

The circl e metho d originate d i n th e pape r o f 1 91 8 b y G . H . Hard y an d S . Ra -
manujan [HR ] o n partitions . An y decompositio n n = n\ + ri2 + n 3 + • • • int o
non-negative integer s withou t regar d t o thei r orde r i s calle d a partitio n o f n. Le t
p(n) denot e th e numbe r o f suc h (unrestricted ) partitions . Thi s i s als o equa l t o th e
number o f ordere d solution s (xi,X2,... ) i n non-negativ e integer s t o th e equatio n
n — X\ + 2^ 2 + 3^ 3 H— • . Algebraically, i t i s also th e numbe r o f conjugacy classe s
in th e symmetri c grou p <5 n.
L. Eule r wa s fascinate d b y th e remarkabl e propertie s o f p(n). H e introduce d
the generatin g powe r serie s
oo oo
n 1
m
F(z) = J2p(n)z = U(l-z )-
0

and showe d tha t

F(z) =n(i-*m)=D-1)'* (3^-1)^/2

which is known a s Euler's Pentagona l Number s Theorem . Playin g with other powe r
series Eule r derive d man y amazin g formulas , suc h a s

P(n) = ~ ^2 cr(h)p(n - h)
0<h^n

where cr(h) is th e su m o f divisor s o f h.

Ramanujan noticed , an d late r gav e proof s o f th e congruenc e propertie s
p(n) = 0(mo d 5) , i fn = 4(mo d 5) ,
p(n) = 0(mo d 7) , i fn = 5(mo d 7) ,
p(n) = 0(mo d 1 1 ) , i fn = 6(mo d 1 1 ) .

He wa s als o fascinate d b y analyti c expression s fo r p(n) i n term s o f modula r forms .

Changing th e variabl e z int o e(z) on e transforms th e powe r serie s F(z) int o th e
Fourier serie s
oo

(20.1) f(z) = Y,p(n)e(nz).

o
449
450 20. T H E CIRCL E M E T H O D

We writ e f(z) = e(z/2A)/rj(z) wher e

OO

is th e Dedekin d et a function . Thi s i s a modula r for m o n th e uppe r half-plan e H I

with respec t t o th e modula r grou p SL2CZ) o f weight | an d a suitabl e multiplie r
system. Precisel y Dedekin d prove d tha t
(20.2) r](-fz) = 0(j)(cz + d)l2r](z)
for an y z G H an d 7 = ( a ) G SLiffi), wher e th e multiplie r #(7 ) i s define d b y th e
1
properties 0 ( - 7 ) = e(i)0( 7 ), fl^ J) ) = e ( £ ) an d

»((: s))-(^-i-5-<* 4
if c > 0 . Her e s(d, c) i s the Dedekin d su m

x(mod c ) V 7 V 7

(recall tha t ip(x) — x — [x] — | ) .

Note tha t th e cus p 0 0 i s non-singula r wit h respec t t o th e Dedekin d multiplie r
system (se e [1 4 ] fo r definition) , s o th e Fourie r expansio n o f rj(z) ha s n o constan t
term; i t i s a lacunar y serie s

^z) = ^2a ne(n

2
z/2A)
n

with a n = 1 i f n = + 1 (mo d 1 2) , a n = — 1i f n = ± 5 (mo d 1 2 ) an d a n = 0

otherwise. However , th e Fourie r coefficient s p(n) o f f(z) = e(z/24)/rj(z) ar e quit e
large. Not e th e trivia l boun d p(n) $ C 2 n whic h follow s fro m
00 0 0
n m
Y,P(n)2- = ~[[(l-2- )-\
10

One ca n deriv e a bette r boun d fro m th e modula r equatio n

'«=@^H)M4)
(see (20.2 ) fo r 7 = ( ~ )) . I n particular , fo r z = z/ y wit h y ^ 1 this yield s

p(n) < y ^ e x p ( 2 ^ ( | j + - ) ) .

Hence b y takin g y = 7Ty/n/6 on e get s

B
(20.3) p ( n ) « n ^ e ^
where i ? = 2TT/^/6. Hard y an d Ramanuja n prove d th e asymptoti c formul a
p{n) ~ (4v/3n)~ 1 e B v ^.
They als o gav e a n asymptoti c expansio n wit h a n estimat e fo r th e erro r term .
 20. T H E C I R C L E M E T H O D 451

Refining th e metho d o f Hardy-Ramanujan , H . Rademache r [R ] establishe d i n

1937 the exac t formul a
THEOREM 20.1 . Forn ^ 1 we have

(20.4) P („) = - l 1 f ; c U ( n ) A i - 8 i n h ( | A n )
V
C= l

where B — 2TT/^6, X n = (n — 7^)2 and

(20.5) A c(n)= ^ * e{\s(a,c)-an/c).

a(mod c )

Notice tha t th e Rademache r serie s (20.4 ) converge s absolutel y b y virtu e o f th e

trivial estimate s |A c (n)| ^ c an d

—-( — sin h — An ) < c" 3 exp(— A n ).

dn\Xn c J V c/
Taking onl y th e firs t ter m i n (20.4 ) on e alread y get s a ver y goo d approximatio n

(20.6) KB )= J^(1_^. + 0 ( «-t«.)).

The startin g poin t o f the Hardy-Ramanuja n metho d i s the Cauch y integra l fo r

the coefficient s o f a powe r series . I n ou r cas e
1n
(20.7) p(n) = — I F{z)z- ~ dz
27T« J\z\=r
(this integratio n ove r a circl e gav e th e metho d it s name) . Changin g z int o e(z) w e
get b y periodicit y

(20.8) p(n) = / f(z)e(-nz)dz

Jw
where w i s an y poin t i n th e uppe r half-plan e E I and th e pat h o f integratio n i s an y
continuous curv e i n H I from w t o w + 1 . W e shal l choose th e pat h o f integratio n
as a continuou s chai n o f certai n circula r arc s o n whic h f(z) ca n b e analyze d quit e
precisely b y exploitin g it s modularity .
To construc t th e arc s i n questio n w e nee d som e well-know n propertie s o f th e
Farey sequenc e or series (se e e.g. [HW]) . Le t C b e a positive integer. Th e collectio n
of al l reduce d fraction s ^ wit h 1 ^ c ^ C an d (a , c) = 1 arranged i n th e increasin g
sequence
a' a a"
cc c"
is calle d th e Fare y sequenc e o f orde r C. Give n a poin t ^ i n tha t sequenc e le t ^ 7
and ^7 7 denote th e adjacen t points . Th e denominator s c r, c" ar e determine d b y th e
conditions
{C-c<c'^C, ad = l(modc) ,
v(20.9)
J < ^
\C-c<d'K:C, ad' = -l(mo d c )
45 2 20. T H E C I R C L E M E T H O D

and th e numerator s ar e a! — (acf — l) c \ a' (ac" + l) c * , so the point s ar e

a 1 a" a 1
(20.10) -c + —
c cc cc
Between a/c an d a'/c' ', a"/c " ther e are the mediants (a'+a)/(c'+c) , (a+a") / (c+c")
which ar e als o reduce d fraction s bu t d o no t belon g t o th e Fare y sequenc e o f orde r
C. The y hav e denominator s i n th e segmen t (C , c -f C] and satisf y

a' + a 1a a 1
cf + c c c(c -h d) +
c c^c-f-c' )
a" + a1 a 1 a"
c" + c c c( c + c") " c" c"{c + c")'

To eac h reduce d fractio n ^ wit h c ^ 1 , (a , c) = 1 we associat e th e circl e

2 1
(20.11)
2? 2^

which i s denote d b y C(a/c) an d i s calle d a For d circle ; i t i s a circl e i n H tangen t

to th e rea l lin e a t z = ^ . Tw o differen t circle s d o no t intersect ; the y ar e tangen t
if an d onl y i f the y belon g t o consecutiv e fraction s i n th e sam e Fare y sequence . Al l
the circle s C(a/c) ar e obtaine d a s th e image s o f th e horizonta l lin e I m z = 1 under
the actio n o f SL 2 (Z) . Indee d writin g

az + b 1
(20.12) 72: =
cz + d c(cz + d )

for 7 = (1 ac )d wit h c ^ 1 it follow s tha t

a% 2 2 + d/ c
Z —
^
c„ ++ 2c
o ^ 92
2c r» 22 "z + d/ c

if I m (z) = 1 , whence i t i s clea r tha t 72 : satisfies (20.1 1 ) .

Now le t C ^ 1 be given . Suppos e a!/c' < a/c < a"/c" ar e thre e consecutiv e
points fro m th e Fare y sequenc e o f orde r C. The n C(a/c) touche s C(a'/c') an d
 20. T H E C I R C L E M E T H O D 453

C(a"/c") a t th e point s

-1
(20.13) + C' ac Wit h &
c(d + ic) '
// 1
(20.14) + C wit h c ac c(c" — ic)

Indeed, w e see that th e point (20.1 3 ) i s an appropriat e mea n of the centers of C(a/c)
and C(a'/c f)

-1
I /a' i \ I /a i \ l 11
2 ? W + 2d*) + 2d* V c + 2c^)\ 2^ + 2^2 c c(c ' -f ic )

by (20.1 0) , so this i s the tangenc y poin t o f the circle s C(a/c) an d C(a'/c f). Similarl y
one verifie s th e poin t (20.1 4) .
On eac h circl e C(a/c) wit h 1 ^ c ^ C w e choos e th e uppe r ar c 7 a c whic h
connects th e tangenc y point s (20.1 3 ) an d (20.1 4 ) wit h th e circle s o f th e adjacen t
fractions. Th e chai n o f suc h arc s j a c give s u s a n infinit e continuou s curv e whic h i s
periodic o f period one . W e choose a s th e pat h o f integratio n i n (20.8 ) th e fragmen t
of thi s curv e whic h i s containe d i n a vertica l stri p o f widt h on e beginnin g i n a
tangency point . W e ge t

(20.15) P(n) = YJ 1 2 H
^n)
c^.C a ( m o d c )

where

Hac(n) = / f(z)e(-nz)dz

by virtu e o f th e periodicit y o f f(z) an d e(—nz).

We shall evaluate each arc integral H ac(n) separatel y b y applying specific trans -
formation 7 e SL2{ rl). First , changin g th e variabl e z int o a/c + iz/c 2 (translatio n
- magnificatio n - rotation ) w e transform (20.1 1 ) ont o th e circl e \z — | | = \ gettin g

i ( an J a iz\ ( inz\
dz
cz \ c

where z i s o n th e righ t ar c o f th e circl e betwee n th e point s

(20.16) -ifC'ac C + IC
22/-H —ic
(20.17) z" = -ic Cac
c" — ic '
45 4 20. T H E CIRCL E M E T H O D

Given c ^ 1 and (a , c) = 1 the modula r relatio n (20.2 ) wit h z replace d b y —d/c

1/cz translate s int o

\c cJ \% J\ «ii.(*=)+if'+«)V-'-''
24cC\z J '
J \ c czs
where ad = l(mo d c) . Changin g z int o iz/c w e write thi s equatio n a s

where E c(z) = z? exp(j^( ^ — Jj) . Applyin g thi s t o th e integra l H ac(n) w e ge t

Hac(n) = ic~*e{\s{a,c) - an/c)I ac(n)

where
2
Iac(n) = I ^ / ( - - + -^E c{z)e ™*/Sdz.
Z
ac
(More precis e bu t cumbersom e notatio n woul d b e I a^c(n)).

To evaluat e 7 ac (n) w e replac e / b y it s constan t ter m i n th e Fourie r expansio n

(20.1) whic h i s p(0) = 1 , an d the n w e exten d th e ar c o f integratio n t o th e whol e
circle. Thi s wil l giv e u s th e mai n ter m

hip) = Ec(z)e2nnz/c2dz.

Changing z int o \jz w e transfor m th e circl e \z — || = \ int o th e vertica l lin e

Re (z) — 1 getting

Jc(n)=
L exp U+^vn-2i)J^2^
Therefore w e hav e

(20.18)
'•<»>-*W'.(3W
where /^(y ) i s th e Besse l functio n (se e [GR] , 8.4-8.5) . (Thi s unfortunat e coinci -
dence i n notatio n wil l en d shortly! )
 20. T H E C I R C L E M E T H O D 455

It remain s t o estimat e th e differenc e I ac{n) — / C W = Iac( n) ~ ^ac( n) ~ Kc{ n)->

say, wher e

/:» = j'" (f(-- c + j) - 1) ^ ) ^ 4,

C
I'ac(n)= r Ec(z)e2™^2dz,
JO

2
C(n) = f E c(z)e ™^dz.
1
JZ '

For estimatio n o f I^ c(n) w e move the pat h o f integratio n fro m th e ar c t o th e chor d

of th e ar c becaus e o n thi s chor d i/z ha s larg e heigh t s o f(—d/c + i/z) — 1 i s
small. Indee d o n thi s chor d w e hav e \z\ ^ min( |
z'nX WnA) ^ 2cC~ l an d Re z ^
max(Re < c , R e < c ) < c 2C~2 b y (20.1 6) , (20.1 9 ) an d (20.9) . Th e lengt h o f th e
chord i s bounde d b y \z' ac\ + \z'^ c\ ^ AcC~ 1 . Moreover , o n th e chor d w e hav e

ra=l x 7

/ c V (2irn\
*{c) exp V^J
because R e (z' 1 ) = 1 , R e {z) ^ c 2C~2 an d p(m ) < 2 m . Therefor e
3

(20.19) / a*c(n) « (£)* ex p ( ^ ) .

Next w e estimat e I' ac(n). Th e lengt h o f th e ar c fro m 0 t o z' ac o n th e circl e

\z — 1 1 = \ i s bounded b y f | ^ c | ^ 7rcC _1 an d fo r z o n that ar c we have \z\ ^ cC~ l.
Moreover, w e derive (a s w e di d whe n estimatin g Ia C{n)) tha t

Therefore I' ac{n) satisfie s th e boun d (20.1 9) . Similarl y w e show that J^ c (n) satisfie s
the boun d (20.1 9) . Collectin g thes e bound s an d th e formul a (20.1 8 ) w e ge t
2
Hac(n) = ^ e ( | 5 ( a , c ) - ™)(1 2A n)-ija ( T ^ ^ ) + O ^ C ^ exp(27mC- )).

Inserting thi s int o (20.1 5 ) w e ge t

p(n) = J2 — A c(n){\2\n)-iU2 ( - A n) + 0(C"* exp^nC" 2)).

C C
c^C

Letting C ten d t o infinit y w e obtai n

(20.20) p(n) = -^-\-3>J2c-lAc{n)I± (-A„) .

C=l
456 20. TH E CIRCL E METHO D

Using th e formula s

I±(y)=(—-J sinhy , h(y) = y^—(y-*Ii(y))

we conclude (20.4) .
REMARKS. Sinc e rj(z) i s essentiall y a thet a function , th e multiplie r syste m
#(7) ca n b e expresse d i n term s o f Gaus s sum s rathe r tha n th e Dedekin d sums .
Consequently th e coefficient s A c(n) i n th e Rademache r formul a (20.4 ) ca n b e als o
expressed b y simpler functions . Indee d A . Selber g showe d (usin g directl y th e Eule r
Pentagonal Number s Theorem ) tha t

(
(20.21) A c(n)=(lY J2 -^COS(H^
^(mod 2c )
^(3^-l) = -2n(mod 21 )

Hence \A c(n)\ ^ 2r(c)c^ .

20.2. Diophantin e equations .

After hi s historica l pape r wit h Ramanuja n o n th e partitio n functio n Hard y
continued workin g o n th e circl e metho d jointl y wit h J . E . Littlewood . Fro m 1 92 0
to 1 92 8 the y publishe d a serie s o f eigh t paper s unde r th e commo n title , "Som e
problems o f Tartiti o Numerorum'" . The y realize d tha t th e modularit y o f th e
generating Fourie r serie s i s no t a n essentia l propert y fo r th e method , bu t rathe r
good estimate s fo r th e relevan t exponentia l sum s ar e sufficient . Thi s observatio n
opened numerou s ne w applications , th e mos t popula r on e bein g fo r th e Warin g
problem.
This concern s th e solvabilit y o f th e equatio n

(20.22) x\+xl + '-- + xl = N

in positiv e integer s # i , . . . , xn, wher e k ^ 1 and N ^ 1 are give n numbers . I n 1 77 0

Edward Warin g asserte d tha t (20.22 ) doe s have a solution provide d n i s sufficientl y
large i n term s o f k alone . Thi s wa s prove d i n 1 90 9 by D . Hilber t [H] . By th e circl e
method Hard y an d Littlewoo d obtaine d a n asymptoti c formul a fo r th e numbe r
of representation s (20.22) . I n thi s sectio n w e illustrat e th e metho d b y showin g a
somewhat late r resul t du e t o L-K . Hu a [H2] .
THEOREM 20.2 . Let n > 2 k. Then the number u(N) of representations of N
as the sum of n positive k-th powers satisfies

(20.23) u(N) = <5(N) r(\ + ]{® n N*- 1 ^ + 0(JV~*)} .

1 [njk)
Here &(N) is given as the sum of infinite series of the multiplicative function

(20.24) *.*) = E '( IE .(£))".( -

^q * —' v q / / v q J
a(mod q) i(mo d q)
 20. T H E C I R C L E M E T H O D 457

The series

(20.25) e(N) = Y,c(q,N)

9=1

converges absolutely and satisfies c\ ^ <S(N) ^ C2, where c\, C2 are positive numbers
depending on k and n, but not on N; also 5 > 0 and the implied constant depend
at most on k,n.

We begi n b y applyin g th e metho d t o a genera l equatio n

(20.26) / ( z i , . . . ,z n) = 0
where / i s a polynomia l o f degre e k ^ 1 with intege r coefficients . W e see k integra l
solutions x = ( x i , . . . , xn) restricte d t o a bo x B C [—X, X]n wit h X ^ 1 . Thu s

(20.27 ) f(x) <^X k ifxeB.

The method work s if the numbe r o f variables i s much larger tha n th e degre e and th e
values of f(x) var y considerably (s o f(x) canno t b e constant i n many variable s afte r
any unimodula r transformation , se e ou r analyti c conditio n (20.35)) . Therefore , b y
a statistica l reasonin g w e anticipat e tha t th e numbe r Vj(JS) of solution s t o (20.26 )
in x e B fl Z n shoul d satisf y
uf(B)<^Xn-k.
Exactly, th e numbe r o f suc h solution s i s given b y th e integra l (Vinogradov' s varia -
tion o n th e Hardy-Littlewoo d powe r serie s setting )

(20.28 ) M®)= I ( E e(af(x)j)da.

xeBnz
As in the Hardy-Ramanuja n wor k o n the partitio n numbe r w e expect her e tha t th e
main contributio n come s fro m th e integratio n i n neighborhood s o f rationa l point s
with smal l denominators .
According t o th e Dirichle t approximatio n theore m ever y a satisfie s

(20.29) a < —, withl^g^P , M )= l ,

where P i s a fixed positiv e number . Le t P ^ 2Q ^ 2 . I f a satisfie s (20.29 ) wit h

q ^ Q , w e sa y a belong s t o th e "majo r arc "

a_
(20.30 )
m |a; J ~ | ^ ~ ^ ' withg ^ Q,(a,g ) = l j .

Notice tha t th e interval s o f Wl centere d a t differen t point s a/q d o no t overla p

because
\ - - ^ \ >— > — > —
\q qi\ qqi qQ qP
if a/q ^ a\/q\ wit h q\ ^ Q. Th e remainin g par t o f th e segmen t m = [0,1 ] \ Tl i s
called th e "mino r arc" .
458 20. T H E C I R C L E M E T H O D

REMARK. I n th e Hardy - Ramanujan wor k th e circl e a (mod 1 ) wa s divide d ex -

actly b y th e mediant s o f th e Fare y sequenc e o f orde r C — P, s o ther e wa s nothin g
in th e mino r arc .
For aGOJIw e evaluat e th e exponentia l su m

Sf(a)= £ e(af(x))

by splittin g int o residu e classe s

e e x
Sf(<*) = Yl (^ /(M) ) £ <n
^V
u ( m o d q) x(zBnZ
x = u(mod q)

where 6 = a — ajq. Sinc e |0 | i s small , namel y |0 | ^ (qP)~ l b y (20.29) , w e ca n

replace th e su m ove r residu e classe s u(mo d q) b y a correspondin g integra l wit h a
reasonable erro r term . Suppos e tha t P ^ > X k~x wit h th e implie d constan t larg e
enough s o tha t

£«*>
P
$J — fo r 1 ^ v ^ n , x G B.
Hence \0df /dx v\ < l/2g , s o Lemma 8. 8 applie d successivel y fo r eac h variabl e x v
i^(mod q) yield s
X\^-U
VJ e(0f(x)) = q-nBf{6) + o((l + ^ " " ' j
Q
a;Eu(mod q)

where Bf(0) i s th e integra l i n questio n

(20.31) B f(0) = [ e(6f(x))dx.

JB

A trivia l erro r ter m woul d b e ( 1 + X/q) n, s o w e onl y go t savin g o f th e siz e o f

one variabl e o f summation . Not e tha t Bf(0) doe s no t depen d o n th e residu e clas s
u(mod q) (i n mor e sophisticate d version s o f th e circl e metho d th e residu e clas s
does appea r i n th e mai n ter m wit h a profoun d effect) . Therefor e summin g ove r th e
classes w(mo d q) w e ge t
1
(20.32) S f{a) = B f{e)Cf(a/q) + 0{q{q + X)^ )
where Cf(a/q) i s the normalized , complet e exponentia l su m

(20.33) C f(a/q) = ±; £ e(^/(«)) .

ii(mod q)

Integrating (20.32 ) ove r th e majo r arc s w e ge t

/ S f(a)da = ] T c f(q) f B f(0)d0 + 0{P~ lQ2{Q + X)"' 1

)
Jm q ^Q J\e\^i/ qp

where
c
(20-34) <*(« ) = £ J2* £ (^/(«))-
a ( m o d q) w(mo d q)

Note tha t Cf(q) i s rea l an d \cf(q)\ ^ q.

 20. T H E C I R C L E M E T H O D 459

First w e evaluat e asymptoticall y th e integra l o f 23/(0 ) ove r th e segmen t \0\ ^

1/qP. Suppos e tha t fo r an y 6 > 0 ,
k
(20.35) B f(6) < {6X )-l^Xn
where 7 > 0 . Hav e i n min d tha t X n i s th e trivia l boun d an d 6X k i s abou t 0f(x),
so we postulate a saving which i s only slightly large r tha n th e siz e of the amplitude .
This conditio n i s quit e realisti c fo r a smoot h functio n f(x) whic h change s fas t i n
several variables . Observ e tha t th e boun d (20.35 ) remain s th e sam e i f a constan t
is adde d t o / . Fo r example , i f
(20.36) f{x) = N - x\ xk
for 0 < x v < X = Ni, the n

k n/k
\Bf(0)\ = I / e(0x )dx <C 1

so (20.35 ) hold s wit h 7 = | - 1 > 0 , i f n > k. Usin g (20.35 ) w e ca n exten d th e

integration ove r \0\ ^ 1 /qP t o th e whol e rea l lin e gettin g

(20.37) / 8 f(0)d0 = V f(B) + 0{(qPX~ ky X n k

~)

where
oo
Bf(0)d0.
/ -OO

The conditio n (20.35 ) als o implie s tha t

n k
(20.39) V f{B)<^X -

(use (20.35 ) i f |0 | > X~ k an d estimat e triviall y i n |0 | < X~ k

).
Next w e evaluat e asymptoticall y th e su m o f Cf(q) ove r th e modul i q ^ Q.
Suppose tha t fo r an y q ^ 1 , (a, q) — 1 ,
2_r7
(20.40) C / ( a / g ) < < T
wThere r\ > 0 . Observ e tha t thi s boun d remain s th e sam e i f a constan t i s adde d t o
/ . Henc e
1
(20.41) c f(q)<^q- -ri.
Using (20.41 ) w e ca n exten d th e summatio n ove r q $C Q t o al l modul i gettin g
c
(20.42) £ / t e )= <5
/ + °( £rT?)

where
00

(20.43) 0/=5>/O?) -
1

Now collectin g th e abov e result s w e obtai n

n k
(20.44) f Sf(a)da = (5 fVf(B) + 0(AX -)
46 0 20. TH E CIRCL E METHO D

where

(20.45) A = Q- 1 1 + Q 2
{QPX~1
k
).

Observe tha t w e can take A ^ 2X~ 67r] wit h 5 a small, positiv e number b y choosin g
Q = X6^ an d P = xk-6^2^^\ bu t som e othe r choice s of Q an d P can b e better ,
see th e lin e abov e (20.62) .
The infinit e integra l (20.38 ) an d th e infinit e serie s (20.43) , ar e calle d th e "sin -
gular integral " an d th e "singular series " fo r the equation (20.26) . Thes e ca n be
brought t o expression s whic h allo w meaningfu l interpretation s i n terms o f loca l
densities.
First w e work ou t th e singula r integral . W e hav e

sin
Vf(B) = li m I* ( I e(9f(x))dx)d9 = li m / (27r{(*)T)dx,

Integrating ove r th e leve l se t {x € B; f(x) = t) an d usin g

r
°° sin27r t
-dt = 1
/ DO Kt

one can show that Vf(B) i s the ( n — l)-dimensional volum e of the se t {x E B; f(x) =
0} with respec t t o a suitable measur e whic h depend s o n /. I t is also approximatel y
equal t o the n-dimensional Lebesgu e measur e o f the set {x e B;\f(x)\ ^ \).
Precisely, w e deriv e b y th e conditio n (20.35 ) tha t fo r any V > 0 ,

k
(20.46) V f(B) = ^\{xe B; \f(x)\ < V}\ + 0{{VX' yXn-k)

where |^4 | denote s th e n-dimensiona l Lebesgu e measur e o f A. Thi s i s non-trivial

for V <C Xk, an d i t yields

(20.47) Vf{B) = Hi n ^\{x G B; \f(x)\ < V}\.

For th e proo f o f (20.46) w e use th e formul a

7-oo *0 I 0,i f \y\ > V.

and th e estimate s sin27r6» V = 2ir9V + 0(92V2), si n 2 ?^ < C 1. Henc e b y (20.35)

we obtai n

l
+ oU 9V\B f{9)\d9+ j {eV)- \Bf(9)\de)

l
-^\{x e B; \f(x)\ ^V}\ + 0((TV + {TV)- ){TXk)^Xn-k).

Choosing TV = 1 we ge t (20.46) .
 20. T H E C I R C L E M E T H O D 461

For example , i f f{x) i s given b y (20.36) , the n (20.47 ) yield s

k k
V(N) = l J"-J (N-x 2 x S*-ldx2--.dxn.

x%-\ \-x„<N

EXERCISE 1 . Sho w tha t th e abov e integra l i s

(20.48) v(N) = r(i + irr^)"1^*- 1.

Now w e wor k ou t th e singula r series . Th e su m ove r a(mo d q) wit h (a , q) — 1

in (20.34 ) i s th e Ramanuja n su m

E* .(f/<«> ) - E »$y
a ( m o d q) d\q
d\f(u)

therefore

°f^ = i&(TH |Mmod q);f{u)=o(mod d)}|=E^W^

where d n~1 ujf(d) i s th e numbe r o f solution s t o th e congruenc e
(20.49) f(x) = (mo d d).
Since cjf(d) i s multiplicative , w e hav e

£/(<?)=n( w /(p a )- w /(p a _ i ))-

pa\\q

Hence th e singula r serie s i s als o give n b y th e infinit e produc t

(20.50) <S / = II*/(P )
V

where
oo
a 1
(20.51) 5 f{p) = 1 + $ > / ( p ) - ^ ( p " " ) ) .

Note tha t th e serie s (20.51 ) an d th e produc t (20.50 ) converg e absolutel y b y

virtue o f (20.41 ) . Th e number s 5f(p) ar e called "loca l densities" fo r th e reaso n tha t
a
(20.52) 5 f(p) = li m Uf(p ),
a—>oo

so Sf(p) i s a densit y fo r th e p-adic solution s o f f(x) = 0 . B y th e analogou s formul a

(20.47) on e can interpret Vf(B) a s a density measur e fo r th e real solutions of f(x) =
0.
If ujf{p oc) stabilizes , i.e. ,
1a+
(20.53) ^ / ( P ) = " / ( P Q ) , for a ^ a,
46 2 20. T H E C I R C L E M E T H O D

then 5f(p) = ojf(p a) i s the densit y o f solution s o f th e congruenc e

a
(20.54) f(x) = 0(modp ).
For som e polynomial s / (such a s (20.36) ) th e stabilit y propert y (20.53 ) hold s fo r
all p wit h a commo n a which i s sufficientl y larg e dependin g o n / . I n thi s cas e
(20.55) 6 /= rjW/(pa).

We now procee d t o th e estimatio n o f the integra l o f the exponentia l su m Sf(a)

over th e mino r ar c m = [0,1 ] \dJl. Firs t Hard y an d Littlewoo d estimate d 5/(a ) fo r
each a i n m and the n integrate d trivially . Fo r a polynomia l i n one variable w e have
LEMMA 20. 3 (WEYL) . Let g(x) = cx k + c i x ^ - 1 + • • • with c and k positive
integers. Suppose

(20.56) L - - U c — z, with (a,q) = 1 .

I q\ q
Then
a 1
(20.57) Y. < d(n)) « X ^+ | + ^) 21 K

for any £ > 0 and X ^ 1 , the implied constant depending only on e.

REMARK. Th e rang e o f summatio n i n (20.57 ) ca n b e shifte d t o an y segmen t
of lengt h X withou t affectin g th e bound .
PROOF. B y Propositio n 8. 2 th e exponentia l su m i s bounde d b y

1
2X1 X~ k ] T •••V J minpf,||adb!Tii---nfe_i||- )}
— X < 7 l i , . .. , 7 l f c _ i < X

<X{X"1+X-'C V JT 1
k^l(m)mm(X,\\acklm\\- )}

where th e firs t ter m comes fro m n\ . . . n^_i = 0 an d Tfc_i(m ) appear s a s a boun d

for th e numbe r o f representations o f m — ri\ • • • n&_i wit h l^n 1 l/<X sol^m<
M = X k~l. Afte r applyin g Tk-i(m) < C 22 m £ w e exten d th e summatio n t o

(20.58) Yl min(X , H^ir 1 )

where L = ck\M (i.e. , we ignore th e conditio n £ = 0(mo d cfc!)) . I f £ ranges ove r a n

interval o f lengt h q/2 th e point s a£(mod 1 ) ar e space d b y q~ x b y virtu e o f (20.56) ,
therefore th e su m (20.58 ) i s bounde d b y

(! + f ) E mi n ( x ) | ^ |1
" ) « ( l + ^ ) ( X + ,log ( ? ).
£(mod q)

This estimat e lead s t o (20.57) . •

From Weyl' s lemm a fo r a = a/q an d X = q we obtai n

 20. T H E C I R C L E M E T H O D 46 3

COROLLARY 20.4 . Let g(x) be a polynomial of degree k ^ 2 with integer coef-

ficients. Let (a,q) = 1 . Then

l 2l
(20.59) £ e(^g{x))«{c, q)H - ~k+s
x ( m o d q)

where c is the leading coefficient of g, and the implied constant depends only on s.
R E M A R K . B y the Riemann Hypothesi s fo r curves ove r finit e field s (prove d by
A. Wei l [Wei] ) on e can obtain th e bound q^ +e, but wit h th e implie d constan t
depending o n g. However , ou r exponent 1 — 2l~k is more tha n sufficien t fo r appli-
cations i n this chapter .

For a in the minor ar c m we have (20.56 ) wit h Q < q < P, so the saving facto r
in the bound (20.57 ) is
0 i-fc

I_ L _ ^
Q + X + Y k/
which i s rather small . However , thi s ca n be multiplied i f / ( # i , . .. , x n ) has many
summands i n distinct singl e variables . Fo r example le t us consider

(20.60) f(x u ...,x n) = /i(:ri) + . .. + f n{xn)

where fi(xi) — C£Xk + • • • wit h Q > 0 for all 1 ^ £ < n. Then , b y Lemma 20.3

1
/I P \n2 l-k
n+
(20.61) Sf(a) = Sfl (a) • • • S fn(a) « X ' (_ + _ + _)

for an y a G m. Th e same boun d hold s fo r the integral o f Sf(a) on m. W e require

this to be of smaller orde r tha n X n~k whic h w e accomplish b y taking n sufficientl y
large i n terms o f k. Choosin g Q = X* an d P = Xk~$ w e ge t

(20.62) / S f(a)da<^X n—k — 5

with som e S > 0, if n > /c2/c+1. Not e tha t th e condition (20.35 ) fo r the polynomial
(20.60) hold s with 7 = n/k—1. S o the abov e choice of Q and P also gives A <C X~6,
see (20.45) . Moreover , w e verify (20.40 ) usin g Corollar y 20.4

Cf(a/q)<^q~n2l~k-^q-2~^

if n > 2k. Therefore (20.44 ) become s

n k 6
(20.63) J S f(a)da = (5fVf(B) + 0(X --)

where 0 / and Vf(B) ar e the singular serie s an d the singular integra l fo r the poly-
nomial (20.60) . Addin g (20.62 ) t o (20.63) w e conclude
46 4 20. T H E C I R C L E M E T H O D

T H E O R E M 20.5 . Let / i ( # i ) , . . . , f n(xn) be polynomials of degree k ^ 2 with

integer coefficients. Suppose n > /c2 /c+1 . Then the number of integral solutions to
the equation
(20.64) /l(Zl ) + " - + /n(*n)= 0
n
in a box B C [— X, X] with X ^ 1 satisfies
(20.65) u f(B) = <S fVf(B) + 0 ( X n " f e - 5 )
/or som e 5 > 0 ; where 0j an d V}(S) ar e t/i e singular series and the singular integral
for the polynomial (20.60).
The implie d constan t i n th e erro r ter m o f (20.65 ) ma y depen d o n th e poly -
nomials / i , . . . , fn bu t rathe r mildly ; i t doe s no t depen d o n th e constan t term s
of / i , . . . , / n . I n particular , th e asymptoti c formul a (20.65 ) applie s nicel y t o th e
equation (b y addin g a constan t term )
(20.66) fi(xi) + -- + f n(xn)=N.

Let c i , . . . , cn b e th e leadin g coefficients , an d suppos e al l ar e positive . Le t ff(N)

be the numbe r o f solutions to (20.66 ) i n positive integers . The y ar e contained i n th e
box B = [0 , X]n wit h X — cN* for som e constan t c ^ 1 depending o n / i , . . . , f n.
Therefore Theore m 20. 5 yield s
1 6
(20.67) i/ f(N) = & f(N)Vf(N) + OiN^- ')
for som e 5 > 0 . I n thi s cas e w e ca n comput e th e singula r integra l Vf(N) asymp -
totically. Firs t w e know fro m (20.46 ) tha t
n k
|{0 ^x u...,xn^X; |/i(a;i ) + • • • + f n(xn) -N\^ U}\ « UX ~
for an y U > 0 . Henc e an d b y (20.46 ) w e deriv e usin g th e approximation s fi(xi) =
QX£ + O p ^ " 1 ) tha t

Vf(N) = ^ | { 0 < X!,.. . ,x n < X ; | C l x{ + • • • + c*x $ - JV | < t / } |

+ 0{U- 1 Xn~1 + (C/X-* 5)7^-*).

In th e erro r ter m w e choos e U = x k~l^lJr^ gettin g 0{X n-k~^^l+^). I n th e
main ter m w e ca n reduc e U t o an y smalle r numbe r V withou t changin g th e abov e
error ter m b y virtu e o f (20.46) . Therefor e
1
(20.68) V f{N) = (c i • • • c nyiv{N) + O^- "*)
where V(N) i s th e sam e on e a s i n (20.48) . Insertin g thi s t o (20.67 ) w e conclud e
that
S
(20.69) u f(N) = ( Cl • • - c n ) - * r ( l + l^T^y'N^H&fiN) + 0(N~ )).

In particular , w e complet e th e proo f o f Theore m 20. 2 i f n > /c2 /c+1 , excep t fo r

the analysi s o f th e singula r serie s <5(N) about whic h w e shal l spea k later .
In orde r t o ge t result s fo r smalle r n w e look fo r improvement s i n the treatmen t
of th e mino r arc . Suppos e / ha s th e first variabl e separated ,
(20.70) f(x u ... ,x n) = g{xi) + h(x 2,... ,x n).
 20. T H E C I R C L E M E T H O D 465

Then Sf(a) = S g(a)Sh(&), s o we ca n writ e

(20.71) I / S f(a)da\ < ( m a x | ^ ( a ) | ) / \S h(a)\da.

In th e mino r ar c w e estimat e S g(a) b y Weyl' s lemma . Recal l tha t fo r a G m w e

have (20.44 ) wit h Q < q ^ P. W e choos e Q = X* 5/3 an d P = X k~l+5, wher e J i s
a smal l positiv e number . Assumin g tha t g i s a non-constan t polynomia l w e ge t a
non-trivial boun d

(20.72) S g(a) < X 1 " 2 ^, i f a € m.

Since rj is small , w e cannot wast e muc h i n th e estimatio n o f Sh (a). Ou r goa l i s

n1
(20.73) / |S /l(a)lda<X - -fc+T'
JO

which i s sufficien t t o yiel d

n k
(20.74) / S f{a)da<^X - -^.

Moreover, th e abov e choice of Q and P make s the error term i n (20.44 ) o f admissibl e
order X 71 -^'. Addin g (20.44 ) t o (20.74 ) w e ge t
n k
(20.75) v f{B) - e fVf(B) + 0(X - ^).

It remain s t o prov e (20.73) . Not e tha t thi s boun d i s clos e t o th e tru e orde r
of magnitude . Indeed , b y droppin g th e absolut e valu e o f Sh (a) w e estimat e fro m
below b y th e integra l whic h i s exactly equa l t o th e numbe r o f solution s t o

(20.76) /*(x 2 ,... ,x n) = 0

in th e relevan t box , an d thi s numbe r i s expecte d t o b e o f orde r X n~1 ~k. I n vie w
of thi s remar k th e tas k o f showin g (20.73 ) ma y see m t o b e almos t a s difficul t a s
the origina l problem , o r eve n mor e difficul t becaus e w e hav e on e variabl e les s i n
(20.76) tha n i n th e origina l equatio n (20.26) . However , thi s situatio n i s not a s ba d
in practice, th e reaso n bein g that i t i s easier t o give an upper boun d fo r th e numbe r
of point s o n a variet y tha n t o coun t thes e wit h asymptoti c precision .
In 1 93 8 L-K . Hu a [H2 ] succeede d i n provin g (20.73 ) wit h an y rj > 0 fo r th e
polynomial whic h i s th e su m o f sufficientl y man y fc-th powers .
LEMMA 20. 6 ( H U A ) . Let k ^ 1 and

fe
(20.77) S(a)= J2 e(an ).

Then for any 1 ^ £ ^ k we have

2
(20.78) / |S(a)| 'da<A:2'-'+e
Jo
the implied constant depending on £ and k.
46 6 20. T H E C I R C L E M E T H O D

P R O O F . Fo r £ =1 w e get

I \S{a)\2da = X.

Suppose tha t (20.78 ) hold s fo r £Jowith 1 < £ < k. W e shall prov e i t hold s fo r £ + 1.
By repeated applicatio n o f the Weyl "differencin g process " £ times (se e Propositio n
8.2) w e arrive a t the inequality
e1
\S(a)f ^(2Xf- ~ £ " • £ ^(a/u.-./i^n;/n,.. . A) )
— X<hi,... ,h £<X nel

where / = I {hi,... , hi) is a subinterval of [1 , X] andp(x; / i i , . .. , hi) is a polynomial

of degre e k — £ with intege r coefficients . W e simplify thi s b y writing
1t
\S(a)f ^ (2Xf- - J2ame(am)

where a m i s the number o f solutions t o the equation

hi •••/i£p(n;fti,.. . ,ht)=m
in th e abov e range . Therefor e a^ < C X 1 an d a m < C X s i f m / 0 , becaus e
p(x; / i i , . .. , /i^) i s not a constant polynomial . W e can also writ e

\s{a)\2 = y]bme{am)

where 6 m is the number o f solutions to

" ™ks ~ nk s+i n\ 8 = m

£
with s = 2 ~ \ 1 ^ n i , . . . , n2s < X. Clearl y
2£
2 6 m = |5(0)| =X 2 £

and
1
60= [ \S(a)\2'da^:X2t-e+e
JO
by th e induction hypothesis . No w we derive b y the Parseval identit y

f \S(a)f +1
da < (2I)2 '-'- 1Va m () m « X 2 ^ " 1 * 2 ^ = X2
Jo m

completing th e proof. •

From Hua' s lemm a wit h £ — k on e gets (20.73 ) fo r any polynomial h in which

at leas t 2 k variable s appea r exactl y a s k-ih powers . Thi s hold s fo r the polynomial
(20.36) provide d n — 1 ^ 2 k (w e used on e variable fo r th e non-trivial estimatio n
(20.72) i n the minor arc) . I n particular, thi s complete s th e proof o f Theorem 20.2
except fo r the analysis o f the singular series .
In genera l thi s analysi s o f 0/ i s not simple. I n the special cas e of the equation
(20.23) on e can estimate th e series (20.43 ) b y elementary ye t long arguments , see
the comment s below .
 20. T H E C I R C L E M E T H O D 46 7

The asymptoti c formul a

n k 6
(20.79) v f(B) = <5fV f{B) + 0(X ~-)
for th e equatio n f(x) = 0 in x £ B n Z n wil l b e meaningfu l onl y i f on e know s tha t
the singula r serie s <5f i s bounde d belo w b y a positiv e quantit y whic h doe s no t o r
only barel y depend s o n / , an d tha t th e bo x B i s larg e enoug h s o tha t Vf{B) ha s
order o f magnitude X n~k. Th e latte r i s essentially tru e i f the equation f(x) = 0 has
solutions i n rea l variable s layin g deepl y i n th e interio r o f B. However , a goo d lowe r
bound fo r <5f i s a more subtle questio n whic h w e shall addres s o n special occasions .
The positivit y o f 0/ follow s i f the equatio n f(x) = 0 is solvable in every p-adic field
by virtue o f the produc t formul a (20.50) . Actuall y th e loca l densities 5f(p) ar e ver y
close t o on e uniforml y fo r al l p sufficientl y large , s o th e proble m i s t o giv e goo d
lower bound s fo r 5f(p) a t finite numbe r o f place s whic h depen d o n / . I f f(x) = 0
has loca l solution s an d th e numbe r o f variable s n i s considerabl y large r tha n th e
degree /c , then (subjec t t o some extra incidental conditions ) ther e ar e plenty o f local
solutions s o tha t (5f ^ c(fc , n) > 0. Fo r example , i f f(x) — N — x\ — • • • — x\ (th e
case o f th e Warin g problem ) i t i s known tha t (se e H . Davenpor t [Da2] )

(20.80) <S(N) ^ c(fc , n) > 0 i f n > 4fc .

20.3. T h e circl e m e t h o d afte r K l o o s t e r m a n .

The circl e metho d i s a too l whic h i s suitabl e fo r detectin g th e equatio n m — n
by means of the additive characters e(ara) , wher e m runs over a given set of integers
and n i s fixed. I n a mor e genera l settin g on e ha s a sequenc e o f comple x number s
A — ( am ) wit h
^2 \ am\ < O O
mez
and th e proble m i s to evaluat e asymptoticall y a selecte d ter m o f A. W e hav e

Oin — I Sj,{a)e{—an)da
Jo
where
SA(OL) = ^ a m e(am).
m
Assuming w e hav e a n adequat e knowledg e abou t th e distributio n o f A i n residu e
classes t o smal l modul i w e ca n evaluat e SA (a) a t th e point s a whic h ar e clos e t o
rationals wit h smal l denominators . Th e se t o f suc h point s i s calle d th e majo r ar c
because th e mai n ter m i n th e asymptoti c fo r a n come s fro m thi s area . However , a
harder wor k i s neede d fo r estimatio n o f Sj,(a) fo r a i n th e remainin g are a whic h
is calle d th e mino r arc . Her e th e treatmen t depend s ver y muc h o n th e structur e o f
the sequenc e A. W e ar e particularl y please d wit h A bein g a n additiv e convolutio n
of many independen t sequences . I n thi s cas e SA(&) factor s int o several exponentia l
sums, s o eve n a smal l savin g i n estimatio n o f eac h o f thes e ca n lea d t o a success -
ful conclusion . I t i s intrinsi c o f th e Hardy-Littlewoo d typ e argument s tha t on e
needs a t leas t thre e summand s t o resolv e th e additiv e equatio n i n question . Th e
arguments certainl y fai l fo r binar y additiv e problem s (sometime s on e sumrnand ca n
be partitione d int o tw o s o th e additiv e proble m unde r consideratio n seem s t o b e
46 8 20. T H E C I R C L E M E T H O D

of ternar y type , bu t th e rang e o f resultin g variable s i s reduce d drasticall y s o thi s

rearrangement doe s no t reall y chang e th e problem) .
Another characteristi c o f th e equation s m = n fo r whic h th e circl e metho d a s
developed by Hardy-Littlewood work s is that th e expected number of solutions mus t
be larg e relativ e t o n , an d i t shouldn' t depen d essentiall y o n n . Fo r example , th e
method ma y wor k fo r a diophantin e equatio n f(x) = n i f th e numbe r o f variable s
is muc h large r tha n th e degree .
A rea l challeng e fo r th e circl e metho d i s t o detec t th e inciden t m = n whe n i t
is expecte d t o happe n irregularly . Moreover , a ne w ide a i s neede d t o trea t binar y
additive problem s

(20.81) "(n) = Y, J2 a™^

m-\-£=n

where A = (a m) an d B — (/^) ar e give n sequence s o f som e arithmetica l nature . I n

1926 H . D . Kloosterma n [Klo ] succeede d wit h th e equatio n
(20.82) <nx\ + a 2x\ - f a^x\ + a±x\ = n.
Actually hi s metho d work s fo r th e equatio n
(20.83) Q(x) = n
where Q(x) i s an y positiv e definit e quadrati c for m i n r ^ 4 variable s an d integra l
coefficients. (I n particular , usin g i t on e ca n recove r th e fou r square s theore m o f
Lagrange, whic h th e Hardy-Littlewoo d versio n coul d not) . I n th e nex t sectio n w e
shall trea t th e equatio n (20.83 ) b y extendin g th e origina l wor k o f Kloosterman .
To explai n th e novelt y o f th e Kloosterma n metho d le t u s speculat e o n th e
general additiv e binar y proble m (20.81 ) (th e quadri c equatio n (20.82 ) i s o f thi s
type, fo r on e ca n tak e m — a\x\ + a2x\ an d £ = a^x^ + #4X4 as the summand s wit h
a m , fy bein g th e representatio n number s b y th e correspondin g binar y forms) . W e
have
u(n) = I 5^(a)5e(a)e(—an)d a
Jo
where SU(a ) an d S& (a) ar e th e correspondin g exponentia l sums . Th e bes t bound s
for thes e exponentia l sum s whic h on e hope s t o prov e fo r almos t al l a ar e th e £2-
norms
MII = (£K*I 2 )*
77 1

and ||S| | respectively . Eve n i f th e bound s S^a) < C ||*4| | an d S 13(a)<C \\B\\ were
accomplished fo r a i n th e mino r ar c (whic h i s rarel y th e case) , i t i s no t sufficien t
because th e expecte d mai n ter m doe s no t excee d ||^4| | ||B|| . Therefor e on e canno t
rely onl y o n estimates . Fo r som e sequence s on e ca n transfor m th e correspondin g
exponential su m int o anothe r sum , sa y

(20.84) S A(a) = ^7 m ( « ) + E A(a),

77 1

where th e ne w term s 7 m (a) hav e th e statu s o f bein g "dual " t o a me(—am) an d

the erro r ter m E^(a) i s significantl y smalle r tha n \\A\\. O f course , th e dua l su m
is essentiall y a s larg e a s th e origina l one , bu t i t ma y hav e fewe r terms . Now ,
 20. TH E CIRCL E METHO D 469

instead o f estimating , on e integrate s th e produc t o f dua l sum s i n a gettin g a n

extra savin g fro m th e cancellatio n whic h i s du e t o th e variatio n i n th e argumen t
of eac h individua l ter m 7 m (a). I n thi s wa y on e break s th e barrie r ||^4| | \\B\\ which
couldn't b e passed b y absolute valu e integration . Th e mai n ter m emerge s a s befor e
from th e integratio n i n smal l neighborhood s o f rational s wit h smal l denominators .
The diophantin e natur e o f a play s a rol e i n th e transformatio n (20.84) , s o i t i s
appropriate t o divid e th e circl e a (mo d 1 ) b y th e Fare y point s ^ wit h 1 ^ c <
C, (a,c) = 1 . However , i n contras t t o th e previou s practice , neithe r gap s no r
overlapping o f thes e segment s i s allowe d (becaus e o f th e prohibitio n o f employin g
estimates fo r th e exponentia l sum s SA(&) an d S& (a)). T o cove r th e circl e exactl y
we choos e naturall y th e divisio n point s t o b e th e mediant s o f th e Fare y sequenc e
of orde r C. Th e resultin g segment s
a' + a a + a" 1 1
m (D - (7 + c ' c + c" c c(c + c'Y c
+ •
c{c + c")

have lengt h whic h change s i n arithmeti c fashio n (se e (20.9) ) an d thi s propert y ha s
to b e considere d whe n integratin g differen t piece s o f th e dua l exponentia l sum .
Kloosterman control s th e lengt h o f th e segment s 9#(f ) b y Fourie r analysis , an d
consequently create s exponential s o f typ e e ( ^ ) fo r variou s u G Z , wher e d i s th e
multiplicative invers e o f a( mo d c) . B y thi s analysi s th e integratio n i n a (mod 1 )
amounts t o the summatio n ove r th e Fare y point s ~ with factor s e(^f). A t th e sam e
time a nic e coincidenc e ma y occu r tha t th e dua l term s 7 m (a) fo r a clos e t o - als o
have a phase of type ^-(~) fo r variou s v E Z (thi s happens fo r th e sequenc e of values
of a quadrati c form) . Multiplyin g thes e factor s an d summin g ove r a(mo d c ) on e
arrives a t th e Kloosterma n su m
n/ x v ^ fdm + an\
S{m,n;c)= 2 ^e [~ )
ad=l(mod c )

with m — —u — v. No w an y non-trivia l boun d fo r S(m,n;c) correspond s t o a

cancellation i n th e integra l o f th e dua l sum s givin g a critica l improvemen t o f th e
direct boun d ||^4| | ||S|| . Kloosterma n succeede d i n showin g tha t

S(m,n;c) < C {m,n,c)^c^r(c),

while th e Wei l boun d (Corollar y 1 1 .1 2 ) assert s tha t

(20.85) |5(m,n;c) | < (m,n,c)h^r(c).

Let u s emphasize tha t i f Q(x) wa s not a quadratic polynomial , the n th e dua l terms ,
if they exist , ma y hav e a phase othe r tha n e ( ^ ) , i n which cas e another exponentia l
sum t o modulu s c would appea r i n place o f the Kloosterma n sums , an d th e metho d
still work s b y a n appea l t o appropriat e result s fo r th e relevan t sums .
We no w procee d t o a detaile d presentatio n o f Kloosterman' s idea s i n a some -
what abstrac t form . W e begi n b y establishin g a Fourier-Kloosterma n expansio n o f
the delt a symbo l
if n = 0 ,
(20.86 ) 5(n) =
if n + 0 .
47 0 20. TH E CIRCL E METHO D

PROPOSITION 20.7 . Let C be a real number ^ 1 . We have

(20.87) 5(n) = 2R e / ^ ^ * (cd)"^ ^- ^)d x

w/iere • restricts the summation by (c , d) = 1 an d J z s the multiplicative inverse to

d modulo c.
REMARK. Fo r n = 0 the formul a (20.86 ) give s a cut e identit y
1 =
(20.88) J2Y, M" i
c^C<d^c+C
(c,d) = l

EXERCISE 2 . Prov e (20.88 ) directly .

For th e proo f o f (20.87 ) w e ma y assum e tha t C i s a positiv e integer . Le t

/ : M.— > C b e a periodi c functio n o f perio d 1 . W e shal l evaluat e th e constan t ter m
of th e Fourie r serie s o f / whic h i s equa l t o th e mean-valu e

/!(/) = / f(x)dx.
Jo
Splitting int o Fare y segment s 97t(^ ) w e obtain b y th e periodicit y o f / ,

M/) = £ £ / , /(*)<* *
^^ fl/c(c+c")

= £ £/ /(?+* >
(o,c) = l

the las t lin e followin g b y (20.9) . Suppos e tha t / ha s th e symmetr y f(—x) = f(x)
(i.e., / ha s rea l Fourie r coefficients) , the n
ri/cc * / /7 \
(20.89) M / ) = 2R e £ / £ /( * - -)dx .
c^C**0 C<d^c+C
cdx<l

Applying thi s t o f(x) — e(nx) an d changin g x int o x/cd on e get s (20.87) .

In th e remainin g par t o f thi s sectio n w e elaborat e (20.89 ) further , no t fo r im -
mediate application s bu t rathe r t o uncover mor e structure o f the formula. W e write
(20.89) a s

(20.90) fi(f) = 2R e f°° ( ^ c^Kffa c ) x 2 dx

 20. T H E C I R C L E M E T H O D 471

where Kf(x; c) is a kind o f incomplete Kloosterma n su m associated wit h /

(20.91) K f(x;c)= £ * f(~-^]

C <d^m'm(x ,c+C) c ex

This ca n be expressed a s a complete su m in d(mod c) by applying th e formul a

~x — d' 1i f d ^ x,
)-{ 0 i
•C-d-
mm (•• . c . - c. fd> x
for C < d < c -f C and x ^ C. Henc e w e ge t
[x — d C-d
(20.92) K f(x;c)= ^ min M
J/ Vc ex J
d(mod c)

Moreover, th e "min " facto r i n (20.92 ) equal s 1 if x > c + C, and it is equal to

x-C
n—)+n—)
"x — d' •C-d' d
-
.c . c.
if C < x < c + C, therefor e

(20.93) +r W £^)_^))/(l.i)
d(mod c)

where th e last su m is present onl y i f C < x < c + C. Insertin g thi s int o (20.90 ) we
obtain
P R O P O S I T I O N 20.8 . Let C be a positive integer and f a periodic function on
R of period 1 . Then the constant term of f in its Fourier series is given by
(20.94)
,</)-2R. E c-/;»,!„(», i ^ )( £ * / ( £ - i ) ) . - *
l
dx.
c<C d(mod c)

R E M A R K S . Quit e ofte n i n applications th e first lin e o f (20.94) yield s th e main

term whil e th e second lin e ca n be successfully estimate d b y means o f exponentia l
sums ove r finite fields. Expandin g \p and / int o Fourier serie s the sum Kf(x\ c) can
be expressed i n terms o f complete Kloosterma n sum s (20.85) , but this optio n i s not
always practical . Fo r f(x) = e(—nx) on e get s by (20.94) anothe r expressio n fo r the
delta-symbol

5(n) = 2Re ^2c' lR{n\c) I e(-—) minfl , X ~ )x~ 2

dx
c^C

c" 1 / (R c(n;c) - R x(n;c))e( )x~ 2

dx
J c V CXJ
c<C
472 20. T H E C I R C L E M E T H O D

where R(n; c) = S(0 , n; c) i s the Ramanuja n su m an d

x — d\ fnd
Rx(n;c)= 2 ^ if; I-
d(mod c)

Note tha t
i? x (n;c)<C ( ^T d"2^c^r(c ) log2 c
<i|(c,n)
by applyin g th e Fourie r expansio n

<^)=E<^-M^)+o((1+c||^|)-')
V7
0<|^|< c

and Weil' s boun d fo r th e Kloosterma n sum s S(£,n;c).

20.4. Representation s b y quadrati c forms .

We procee d t o trea t th e equatio n (20.83 ) b y th e Kloosterma n circl e method .
Our goa l i s
THEOREM 20.9 . Let Q(x) be a positive definite quadratic form in r ^ 4 vari-
ables and integer coefficients. Then for any n > 0 the number r(n, Q) of integral
solutions m = ( m i , . . . , mr) to Q(m) = n satisfies

[ +£
(20.95) r(n,Q)= ' " &(n,Q) + 0(n^~i )
T(k)y/\A\
where k = | , \A\ is the determinant of Q and <5(n , Q) is the singular series
CO

(20.96) 0 ( n , Q ) = X)c- r0c(n,Q)

c=l

(20.97) 9c(n,Q)= ]T *£ e(^(Q(/i ) - n)) ,

<i(mod c ) h ( m o d c )

£/ie implied constant depends on e and on the form Q.

The quadrati c for m Q(x) i n x = ( x i , . . . ,x r ) i s associate d wit h a symmetri c
matrix A = (a^ ) o f ran k r wit h a ^ G Z an d a ^ G 2Z whic h i s positiv e definite . I n
SiegePs notatio n

\cg\X) = =
^./i[3? J — 2 X.ri X — ~F: / Q,ii%i \ / / OijXiXj .

The discriminan t o f Q i s define d b y

( D — (—1)5 |A| i f r i s even ,
\ D = \{-l) r-¥\A\ i f r i s od d
where |^4 | is th e determinan t (on e show s tha t D = 0 , l(mod 4 ) i f r i s even). Pu t
CO

(20.98) 0{z) = ] T e (Q(m)z) = ]TV(n , <2)e(nx )

 20. TH E CIRCL E METHO D 47 3

where
r
r(n,Q) = \{meZ ;Q(m)=n}\.
This thet a functio n i s a modula r for m o f weigh t k = | , leve l 27V , where N i s suc h
that NA~ l i s integral, an d i t transform s b y a suitabl e multiplie r syste m define d i n
terms o f the Jacob i symbo l (se e Theorem 1 0. 8 of [1 4]) , but w e do no t us e thi s fact .
All we nee d i s the followin g Jacob i inversio n formul a
LEMMA 20.1 0 . Let z e H and x e R r. Then

(20.99) J ; e ( | ^ [ m + x] ) = | ^ | - i Q ) f c £ e {^A^[m]^mx\

SKETCH O F PROOF . Thi s follow s b y th e Poisso n summatio n formul a

J2 f(m + x)= ] T f{rn)e^mx)

mezr mez r

for f(x) = e(|A[x]) . Th e computatio n o f the Fourie r transfor m f(m) i s carried ou t

by a linea r chang e o f variable s whic h diagonalize s A , s o i t reduce s t o th e familia r
integral i n on e variabl e

JRily2--y)dy={^2e(-^).
D

Let ad = l(mo d c) . W e shal l us e Lemm a 20.1 0 t o transfor m 0(z — 7 ) int o a

theta functio n fo r th e adjoin t quadrati c for m

(20.100) Q*(x) = ^A-^x].

REMARK. Th e adjoin t for m Q*(x) doe s not hav e integra l coefficients . I fT V is a

positive intege r suc h tha t NA' 1 i s integral, the n NQ*(x) ha s integra l coefficients .
For example , N = \A\ satisfie s thi s condition , ye t a smalle r numbe r suffice s quit e
often. Not e tha t N r = 0(mo d \A\) s o every prim e facto r o f \A\ i s in N.

Splitting th e summatio n i n (20.98 ) int o residu e classe s h = (hi,... ,h r) o f

modulus c we ge t
0 Z
( ~l)= £ e
(-? 0 ( / l ) ) E e(Q(m)z).
h(mod c ) m=h(mod c)

Then b y th e Jacob i inversio n formul a (20.99 ) fo r x = hc~ l w e ge t

m=h(mod c) rn

Inserting thi s int o th e forme r su m an d changin g h int o dh modul o c we arriv e a t

47 4 20. TH E CIRCL E METHO D

LEMMA 20.1 1 . Let z e H I and ad = 1 (mo d c) . Then

,,0,0, , (l.£).|4-^(J)'EOm(^W_^0)

where Q* is the adjoint quadratic form and G m (d/c) is the Gauss sum

Gm
(20.102) (T)= H ^ W W + ' H ).
/i(mod c )

To pick u p th e coefficien t r(n,Q) i n (20.98 ) w e appl y (20.89 ) fo r f(x)

e(—nz)0(z), wher e z = x + iy wit h 2 / > 0 to b e chose n later . W e obtai n
,1/cC
(20.103) r(n , Q) = 2R e 2 ^ / T(c,n;a:)e(-n^)<i x

where
d
(20.104) T(c,n;x)= J^ e(n-)#(z--) .
c
cdx<l
Then b y Lemm a 20.1 1 we obtai n
.k
(20.105) T(c,n;x) = \A\-h- rl-\ ] TT m (c,n; x)e(-Q*(ra)/c 2 £)
Zj
m£l
where

(20.106) T m(c,n;x)= ^ * e(n-\G,

C«L^c+C
cdx<l

Now w e ar e goin g t o estimat e T m(c, n ; x). Firs t w e estimate th e Gaus s su m

LEMMA 20.1 2 . Let (c,d) = 1 and m eZ r. We have

<<c i
(20.107) Grn (~c)
where the implied constant depends only on the quadratic form Q.
PROOF. W e hav e

Gm
(l)\ = E * e(^(Q(x)-Q(j/)+*(a:-j/)m) )
^C^ ' - -(mo d c )
x,y(mod

J2* e[- cCyAz + Q(z) tzm

y,z(mod c)
< c r |{z(mod c):Az = 0(mo d c)} |
which yield s (20.1 07) . •

Next w e evaluate th e Gaus s su m precisely , bu t onl y i f th e modulu s i s od d an d

coprime wit h th e discriminant .
 20. T H E C I R C L E M E T H O D 475

LEMMA 20.1 3 . Let (c, 2\A\d) = 1 and rn e Zr. We have

W
<*«*> ^-(T)HT)*R^ )-
P R O O F . I t reduces t o the one-dimensional cas e b y a local diagonalizatio n and
completing th e square,

(20.109 ) £ e(^-)=eJfjV~c
V
y(mod c) ' ^ '

if (c , 2a) = 1 (see (3.38)). Recal l tha t a nonsingular symmetri c matri x ove r a field
of characteristic ^ 2 is equivalent t o a diagonal matrix . Henc e one shows tha t ther e
exist a n integral matri x V and an integral diagona l matri x B = diag (&i,.. . , br)
such tha t (c , |V|) = 1 and tVAV = B(mod c) . Changing x to Vy modulo c , we ge t

A[x] = B[y\ = Y,b,yl.

Moreover, lettin g V tm = [di,.. . , dr ], we get txm = t yVtm = ^u d vyv. Henc e

rr r

Q(x) + txm=\ ]T(&„}/ 2 + 2dvy„) = \ ^ b v ( y v + bvdv)2 - \ ^ M ^-

Here th e last su m is congruent modul o c to

Moreover, w e have bi--b u = \B\ = \A\ |T/|2(mod c) . Therefore , applyin g (20.1 09 )

in eac h variabl e y v we conclude (20.1 08) . •

To estimat e th e sum (20.106) w e first complet e it as follow s

(20.110) T m(c, n;x)= Yl 7 WW rn, n\ c)

£(mod c)

where

e
'£d + nd\^G(m d
(20.111) K(£,m,n;c)= £* ( ^ T ^ Kc)
<i(mod c)

and

C<6^min(c+C,l/cx)

If |^ | ^ | w e have 7(£ ) < C (1 -f Kl)- 1- Therefor e th e problem reduce s t o estimatio n

of th e complet e sum s K(£, m, n; c).
The complet e su m K{t,,m,n',c) i s multiplicative i n c. Precisely , i f c — c§c\
with (co,ci ) = 1 , then K(l,m,n;c) — K^c°\i, ra, n ; ci)if ( Cl^(£, m, n; Co) where the
superscripts (co),(ci ) indicat e tha t th e additive character s e( — x) an d e( — x) ar e
used i n place o f e( —) an d e( —) respectively . Le t Co denote th e largest facto r o f c
having all its prime divisor s in 2\A\ and c\ be the remaining factor , s o (ci, 2|^4|) = 1 .
476 20 . T HE C I R C L E M E T H O D

We estimat e th e sum s (20.1 1 1 ) t o moduli CQ an d c\ separately. B y (20.1 07 ) an d

trivial summatio n ove r d o (mod Co) , we ge t
1
Cl)
(20.112) K( {e,m,n;c0)<Z:4+
where th e implie d constan t depend s onl y o n th e for m Q. The n b y (20.1 08 ) w e ge t

(20.113) K^(£ 1 1 )= 0^\ fei(f)v^)

m,n;c S r{c0£, c0(n + Q*(m)))

where

d(mod c i)
is eithe r a Kloosterman su m i f r is even , o r a Salie su m i f r is odd . I n eac h cas e
this su m satisfie s th e boun d (20.85 ) givin g
(
(20.114) i f ^ ) ( ^ m , n ; c i ) < ( ^ n + Q*(m),ci)ic^T(ci) .
Inserting (20.1 1 3 ) an d (20.1 1 4 ) int o (20.1 1 0 ) an d summin g ove r \£\ ^ | w e deriv e
LEMMA 20.1 4 . We have
(20.115) T m{c, n ; x) < ( n + Q*(m), cjitfc^ r(c) log2 c
where the implied constant depends only on the form Q.
We shall appl y th e boun d (20.1 1 5 ) t o al l terms i n (20.1 05 ) excep t fo r m = 0 in
the rang e 0 < x < l/c( c + C) i n whic h To(c , n; x) i s equa l t o

e
d\ v ^ / d
(20.116) T(c,n)= J2* (nc) ^ e {--Q(h))-
cJ ^— ' \ c
d(mod c) h(mod c )
We obtai n

T{c,n;x) = \A\-h-r(-) T(c,n)

k
^i / 2ixyQ' {m)"
+ o((c 0c)Mc)(log2c)(^
r
mez
where ^ mean s that m = 0 is excluded fro m th e summatio n i f 0 < x < l/c( c + C).
We estimat e thi s su m b y

Recall tha tT V i s a positive intege r suc h tha t NQ* i s integral an d (ci , 27V) = 1 . W e
take
(20.117) C = n^, an d y = C'2 = n " 1 .
This choic e implie s tha t fo r z = x + iy wit h 0 < x < (cC) _ 1 ,

and if (c(c + C ) )_ 1 < x < (cC) _ 1 , the n w e als o hav e th e lowe r boun d
c\z\y~i > (2C)" 1 i/-^ = |.
 20. T H E C I R C L E M E T H O D 47 7

Moreover, w e hav e Q*(m) ^ > \m\ 2. Hence , fo r an y 0 < x < (cC) -1 w e deduc e

E V ^ W e x p ( - ^ g M ) <(c|*|y-i) «
for an y K ^ 0 . Applyin g thi s wit h K — k w e ge t

T( C ,n;x) = | ^ | - i c - r ( ^ ) f e T ( c , n ) + O(C(c 1 )( C0 c)5T( C )(log2c)n*)

where
2
aci) = Y /(ciJ + nN)i(e+l)- .

Inserting thi s int o (20.1 03 ) w e ge t

fl/cic+C)
r(n,Q) = \A\-* Y^ C-rT(c,n) / (i/z) k
e{-nz)dx
c^C J-l/c{c+C)

+ O(n^- iyj^(ci)(co/c)5r(c)log2c).

Here th e erro r ter m i s bounde d b y

---+£
n2 4 .

On th e othe r hand , th e integra l i s equal t o

1
(i/z)ke(-nz)dx = (270^)-V"
/ oo

up t o a n erro r ter m 0((cC) ) . Summin g thi s erro r ter m ove r c ^ C w e find tha t
its total contributio n i s absorbed b y th e erro r ter m alread y presen t (us e th e boun d
(20.115) fo r T{c,n)), therefor e

T(k)y/\A\^C
Finally, extendin g th e summatio n i n th e mai n ter m t o al l c w e obtai n (20.95 )
(notice tha t T(c,n) = g c{n,Q) b y changin g th e variable s h int o dh an d d int o — d
in (20.1 1 6)) .
Theorem 20. 9 i s als o vali d fo r ternar y form s ( r = 2k = 3) , bu t th e erro r
term i n (20.95 ) i s just to o larg e t o yiel d a n asymptoti c formula . However , a sligh t
improvement i s possibl e b y exploitin g additiona l cancellatio n i n sum s o f th e sum s
(20.113) fo r variou s modul i o f the relevan t Fare y sequence . I n th e cas e o f r = 3 one
obtains th e Sali e sum s
e
(20.118) K{a,b;c)= ^ {^) {—~
d(mod c )

These ca n b e compute d explicitl y i n term s o f quadratic root s modul o c. Fo r exam -

ple, i f (c , 2a) = 1 , then (se e Lemm a 1 2.4 )

(20.119) K(a,b;c)=e c{~)d ] T e(^) ,

x2=ab(mod c)
478 20 . TH E CIRCL E METHO D

which easil y yield s \K(a,b;c)\ ^ c^r(c). Clearly , thi s boun d i s bes t possible , ye t
one can deriv e a better estimat e o n averag e ove r th e modulu s c due t o th e variatio n
of e(^f) . Suc h a resul t belong s t o th e spectra l theor y o f automorphi c form s o f
weight k = § , an d i t lead s t o th e followin g formul a (se e [Du2 ] an d [1 6] )
T H E O R E M 20.1 5 . Let Q(x) be a ternary, positive definite quadratic form with
integral coefficients. Then for any n > 0 the number of integral representations
Q(m) = n satisfies

(20.120) r(n,Q ) = 4 7 r ^ V ( 5 ( n , Q ) + 0(ni-5iT )

where &(n,Q) is the singular series given by (20.96)-(20.91 ), and the implied con-
stant depends on Q.
The formula s (20.95 ) an d (20.1 20 ) ar e tru e asymptotic s onl y i f th e singula r
series <5(n , Q) doe s no t vanish . W e complet e thi s sectio n b y highlightin g a fe w
features o f (S(n , Q). W e kno w fro m th e genera l consideration s i n th e previou s
section tha t th e singula r serie s (20.96 ) i s th e produc t o f loca l densitie s (o f p-adi c
solutions t o Q(x) = n)

(20.121) 0(ra,Q )= IppM)-

p

Also th e leadin g facto r

1
(20.122) < W n , Q ) - '
T(kW\A\
embodies th e densit y o f rea l solution s t o Q(x) = n . W e shal l comput e th e loca l
densities S p(n, Q) fo r p\ 2\A\ usin g th e formul a (20.1 08 ) fo r Gaus s sums ; i t gives

/ 2
(20.123) &(«,<?) = ( ^ W v ]T * (^Ve(-r^ )i f (c,2|^| ) = l .

Suppose r i s even. The n th e abov e su m i s the Ramanuja n su m givin g

gc{n,Q) = XD(c)c k ] T V>(p)q

q\{c,n)

where D = (—l) r/2|^4| i s the discriminan t o f the quadrati c for m Q(x) an d XD(C) —
( ^ ) i s the Jacob i symbol . Henc e fo r an y P wit h (P , 2D) = 1

c\P q\{n,P) p| f

In particular, i f every prime powe r factor o f P i s larger tha n tha t o f n, thi s simplifie s
to
^c-^^^^^x^^y-^na-XD^p-^).
c\P q\n p\P
Taking P — p " + 1 , where v — ord p n, w e ge t th e loca l densit y

(20.124) 5 p(n,Q) = (1 + X D ^ P " * ) - ^! - XD^P 1 -")'1 ^ ~ XD(p")p" (1 ~fe))

 20. T H E C I R C L E M E T H O D 479

for an y p \ 2D. Henc e i t i s clear tha t

(20.125) S p(n,Q)^0 iip\2D.
Moreover, takin g P t o be the product o f arbitrary larg e numbe r o f primes p \ 2D
we deduce tha t

(20.126) <5(n , Q) = <Ti_ fc(n, X*D)L% X 4 D )1_ f[ 6 p(n, Q)

p\2D

where L(S,X4D) denote s th e Dirichlet L-function , an d

(20.127) < 7 s (n,xH$>(dK.

d\n

The question when S p(n, Q) > 0 for p\2D is quite delicate. I f r > 4 one can show
that th e non-vanishing o f all the local densitie s 5 p(n, Q) wit h p\2D is equivalent t o
the existenc e o f m e 7U such tha t
(20.128) Q(m) = n (mo d 27 |A| 3 ).
Assuming thi s condition , on e can infer b y (20.96)-(20.97 ) tha t

(20.129) <5(n , Q) x JJ(1 + XD^P 1 ^)

p\n

if r i s even, r = 2k > 4. Fo r r = 4 a similar analysi s applies , excep t fo r 5 2 (n, Q).

If r i s odd, r ^ 5 , then th e exact computatio n o f the local densitie s o f S p(n, Q)
are mor e involved ; nevertheless , on e can derive directl y b y the trivial estimatio n
1 +
(20.130) | £ c(n,Q)|^cS if( c ,2|A|) = l
that th e singular serie s <S(n , Q) is bounded fro m belo w an d above b y positive con -
stants depending only on Q, provided the congruence (20.1 28 ) has integral solutions.
REMARK. I f r = 3 , we know onl y tha t 0(n , Q) ^ > n~ £ fo r an y e > 0 wher e
the implie d constan t depend s o n s, Q, still assumin g th e solvability o f (20.1 28 ) and
that S2(n.Q) > 0 . A t presen t th e implie d constan t i s not effectivel y computable ,
because th e result use s Siegel' s boun d (5.76) .

There i s mor e t o sa y abou t th e formul a (20.95 ) i n term s o f modula r forms .

Indeed, th e representatio n numbe r r(n,Q) i s th e n-t h Fourie r coefficien t o f th e
theta functio n (20.98 ) whic h belong s t o the space Mk(2N,$) o f modular form s of
weight fe, level 2A^ , and a suitabl e multiplie r # . Suppos e r — 2k ^ 4 , so the whole
space Mk(2N,d) i s spanne d b y th e Eisenstei n serie s (on e for eac h singula r cus p
with respec t t o $) an d a finite numbe r o f cusp forms . Therefor e
0(z,Q) = E(z,Q) + F(z,Q)
where E(z, Q) is a unique combinatio n o f the standard Eisenstei n serie s (calle d the
Eisenstein serie s of the form Q) and F(z, Q) is a cusp for m uniquel y determine d by
Q. Fro m thi s decompositio n
r{n,Q)=p{n,Q) + T(n,Q)
48 0 20. T H E C I R C L E M E T H O D

where p(n, Q) , r(n, Q) ar e th e correspondin g Fourie r coefficients . I t turn s ou t tha t

p(n< Q) coincide s wit h th e mai n ter m i n th e formul a (20.95 )
/o \ kk 1
[
p(n,Q)=800(n,Q)Y[6p(n,q) = ^n<5(n,Q).
(kW\A\
I (k)\/\A\
In vie w o f thes e observation s (du e t o Siegel ) th e erro r ter m i n (20.95 ) i s jus t a
bound fo r th e Fourie r coefficien t r{n, Q) o f the cus p for m F(z, Q). Havin g reveale d
this fac t w e coul d appl y th e Ramanujan-Petersso n conjectur e

r(n,Q) < r(n)n^~

(the implie d constan t dependin g o n Q) whic h wa s prove d b y P . Delign e i f k i s a n
integer, k ^ 2 , i.e. , i f r = 2k i s even , r ^ 4 . Ou r resul t (20.95 ) correspond s t o th e
Selberg boun d fo r th e Fourie r coefficien t r(n, Q). I f r i s odd, th e theor y o f Delign e
is not applicable .
Siegel [Sie3 ] als o observe d tha t th e Eisenstei n serie s E(z, Q) fo r a for m Q i s
genus invariant . Actuall y h e showe d tha t
w
E(z,Q)= J2 (QvM*M
Qve ge nQ

where Q v run s ove r inequivalen t form s i n th e genu s o f Q , an d w(Q u) ar e th e genu s

masses. T o defin e thes e number s w e first recal l som e definition s an d basi c fact s
from th e theor y o f quadrati c forms , i n particular , genu s theory .
Two quadrati c form s Qi,Q 2 (i t i s assume d throughou t tha t al l form s unde r
consideration hav e the sam e ran k r ^ 2 ) ar e equivalen t i f one ca n b e obtaine d fro m
the othe r b y a unimodula r (invertibl e ove r Z ) chang e o f variables, i.e . i f A\ — A 2U
for som e U £ M r (Z) wit h \U\ — ± 1 . Th e equivalen t form s for m a class . Th e
determinant i s a clas s invariant .
A quadrati c for m Q(x) = ^A[x] i s equivalen t t o itsel f i n a numbe r o f ways .
Denote b y
0(Q) = {U e MriZy^UAU = A}
the grou p o f automorphs o f Q. Thi s i s finite an d th e sam e on e for equivalen t forms ;
we denot e it s orde r b y \0(Q)\. I f r = 2 th e numbe r \0(Q)\ depend s onl y o n th e
determinant \A\, bu t i f r > 2 it varie s fro m clas s t o class .
Two forms Qi , Q2 are in the sam e genu s if they ar e equivalent ove r every p-adi c
field and ove r the rea l numbers (th e latte r conditio n i s redundant sinc e both Qi , Q2
are assume d t o b e positiv e definite) . Th e determinan t i s a genu s invariant . On e
can sho w tha t tw o form s Q\,Q2 o f th e sam e determinan t \A\\ = |^4_2 1 are i n th e
same genu s i f an d onl y i f the y ar e equivalen t ove r Z t o form s congruen t modul o
8|Ai||A2|. Th e numbe r o f classe s i n a genu s o f a fixed determinan t i s finite.
Clearly, th e numbe r o f representation s r(n,Q) i s a clas s invariant , bu t i t ma y
vary within a given genus, because the number o f automorphs ma y do so. Therefore ,
it i s natura l t o ad d th e followin g weigh t w(Q) t o th e representations :

^-i^L^isib)" 1 -
 20. T H E C I R C L E M E T H O D 481

Note tha t th e tota l mas s o f th e form s i n a genu s i s

Que ge n Q

Accordingly, th e averag e numbe r o f representatio n o f n b y form s o f a give n genu s

is define d a s th e weighte d su m

r(n, ge n Q) = ^ w(Q u)r(n, Q v).

Q„e ge n Q

Collecting th e abov e notatio n an d result s w e arriv e a t th e Siege l mas s formula ,

V
(27r)knk~1e n
(20.131) r(n , gen Q) = p(n , Q) = / /rn ( > Q)-
1 {k)yJ\A\

20.5. A n o t h e r decompositio n o f t h e delta-symbol .

Dissecting th e uni t circl e b y th e Fare y point s ~ c of orde r C w e develope d a
Fourier typ e expansio n fo r S(n) i n term s o f th e additiv e character s e(n^) , wher e
dd = l(mo d c) . W e cal l ^ th e Kloosterma n fractions . Thes e appea r i n th e expres -
sion (20.87 ) wit h d ranging ove r th e interva l C<d^c + Co£ lengt h exactl y c , bu t
not alone ; the y ar e escorte d b y th e othe r factor s involvin g d. Thi s featur e seem s t o
be insignifican t fro m a n analyti c viewpoint . I n orde r t o rela x thes e "minor " factor s
one uses Fourier technique , an d inevitabl y on e is led to complete Kloosterma n sum s
S(m,n;c) i n additio n t o th e Ramanuja n sum s 5(0 , n;c). On e canno t avoi d the m
in Kloosterman' s treatment , eve n fo r smal l c.
In thi s sectio n w e giv e a differen t expressio n fo r o(n) entirel y i n term s o f th e
Ramanujan sum s

s ( 0 ,„ ; c ) = £ * „(2!)_ £„(')„.
"nd^
VC /
d(mod c ) d|(c,ra )

We begi n b y a simple-minde d treatment . A n intege r n i s equa l t o zer o i f an d onl y

if i t ha s a diviso r large r tha n \n\. Therefore , fo r an y q > \n\ we hav e

'<»> = ?a (£m o d q)•(=) •

Writing - i n it s lowes t term s w e ge t

(20.132) S(n) = -^5(0,n;c) .

c\q

Next w e average ove r q to gai n a n extr a variable . Le t w(u) b e a function supporte d

in a segmen t C < u < 2 C an d normalize d b y
CX D

(20.133) ^Tw(q) = l.
9=1
482 20. T H E C I R C L E M E T H O D

Summing (20.1 32 ) ove r q we ge t

oc o o
1
(20.134) 5(n) = ^ c f ^ O , n ; c) ]T r " ^^).
c=l r = l

This formul a i s vali d onl y fo r \n\ < C ; consequently , i t employ s th e character s

e(nd/c) t o moduli a s large as |n| , whic h is not a good featur e i n applications. I n thi s
respect a Kloosterman typ e expansion ha s the advantag e o f employing character s t o
much smalle r modul i relativ e to |n| , precisel y one can make it usefu l wit h c ^ 2|n | 2.
To refin e th e abov e ide a w e list her e th e followin g observations :
— ever y positiv e intege r divide s zero ,
— a non-zer o intege r ha s fe w divisors ,
— of th e tw o complementar y divisor s o f n > 0 one doe s no t excee d n 2.
Let w{u) b e a functio n o f u ^ 0 which vanishe s a t u = 0 and decay s rapidl y t o
zero a s u —> oc . W e normaliz e w{u) b y th e conditio n (20.1 33) . The n b y th e abov e
observations w e deduc e th e followin g identity :

6(n) = Yl{"(9)-»(!j))-
q\n

Detecting th e conditio n q\n b y orthogonalit y o f additiv e character s e(an/q), an d

then reducin g th e fraction s a/q w e obtai n
PROPOSITION 20.1 6 . Let w(u) satisfy the above conditions. Then for any in-
teger n we have
00
5
(20.135) S(n) = ^2 ( ° ' n ; c )Ac(n)
c=l

where A c(u) is a function on R given by

00

(20.136) A c(u) = Y^^r^iwicr) - w(\u\/cr)).

r=l

Since A c(u) i s not compactl y supporte d (eve n if the test functio n w(u) i s chosen
with compac t support) , w e make a practical alteratio n b y multiplying (20.1 35 ) wit h
a nic e functio n f(u) supporte d i n \u\ < 2N an d normalize d b y /(0 ) = 1 . W e ge t
00

(20.137) S(n) = ] T 5(0 , n; c)A c(n)f(n).

c=l

Now suppos e tha t w(u) i s als o compactl y supported , sa y i n th e dyadi c interva l

C ^ u ^ 2C (thin k o f w(u) a s a bum p function) . The n th e serie s (20.1 37 ) reduce s
to c ^ 2max(C , N/C) = X , say . Clearl y th e optima l choic e i s

(20.138) C = N l*
giving X = 2 C = 27 V i an d

(20.139) S(n)= J2 5(0,n;c)A c (n)/(n).

 20. T H E C I R C L E M E T H O D 483

Therefore i t take s character s o f modul i c ^ 2N 2 to detec t th e even t n = 0 withi n

the integer s \n\ ^ 2N.
Of course , th e optima l choic e fo r th e magnitud e o f modul i i s no t necessary ,
so w e continu e ou r investigatio n withou t th e conditio n (20.1 38) , becaus e ther e ar e
some technica l benefit s fro m havin g C independen t o f N. Yes , our ne w expression s
(20.135) an d (20.1 37 ) stil l contai n onl y th e Ramanuja n sum s (ther e ar e nowher e
hidden Kloosterma n fractions!) , bu t wha t abou t th e facto r A c (n)? Assumin g tha t
the tes t functio n w(u) i s smooth on e can regard A c(u) a s a continuous analo g of th e
Ramanujan su m 5(0 , n; c). T o be practical w e must b e abl e to contro l th e variatio n
of A c(u) i n bot h variable s c , n, an d t o separat e c from u a t a lo w cost . T o thi s en d
we ar e goin g t o prov e
LEMMA 20.1 7 . Suppose w(u) is smooth, supported in the segment C ^ u ^ 2C
with C ^ 1 , normalized by (20.1 33), and has derivatives satisfying
1a
(20.140) w^(u)<^C- -

for 0 ^ a ^ A. Then for any c ^ 1 and u E R we have

(20.141) A c(u) < - — —+

{c + C)C \u\ + cC
and for 1 ^ a ^ A,
fl
(20.142) A[ '(u) < (ccyWui + ccy a.
PROOF. W e spli t

(20.143) cA c{u) = V(c) - wO^)

where

(20.144) V{y) = y^r~ lw{yr) - I r~ l

w(r)dr,
r=l J °

(20.145) W% ) = y^r- 1 w{y/r) - I r~ l

w(r)dr.

By th e Euler-Maclauri n formul a

J2F(r)= f(F(r) + {r}F'(r))dr,

rJ

where F i s of clas s C 1 o n R an d {r } = r — [r] i s th e fractiona l par t o f r , w e ge t

(20.146) V(y) = yJ°°{-}(^p-)'dr.

Since {x} < min(l,x ) ^ 2x( l + x ) - 1 , w e deriv e b y (20.1 40 ) tha t
1
(20.147) V ( y ) « C - ( l + C/y)- 1 .
 48 4 20. T H E C I R C L E M E T H O D

In th e sam e way , b y applyin g (20.1 47 ) fo r w(u) change d int o w(l/u) an d y int o

1/?/, w e deriv e tha t
(20.148) W(y) < C~ l{\ + y/C)' 1 .
Inserting thes e estimates int o (20.1 43 ) w e obtain (20.1 41 ) . Fo r the proof o f (20.1 42 )
observe tha t Ac(u) vanishes , unles s \u\ > cC i n whic h cas e

Aia\u) = ^ ( - c r ) - a - V a > ( ^ ) < (cC)~ l

{\u\+cC)-a.
. cr.
•
Next w e shal l sho w tha t A c(u) approximate s t o th e Dira c distributio n quit e
strongly o n suitabl e tes t functions . T o thi s en d w e firs t establis h th e followin g
identity
LEMMA 20.1 8 . For any compactly supported function f of class C a on R we
have
(20.149)
J°° A c(u)f(u)du = /(0)t&(0 ) + / ( O K " 1 1 ° ° f3 a Q ( ^ ) {<l) dr

f°° / r \ f°°
-ca~l / p a{-) / w(u)u a w
f (ru)dudr,

where i3 a{x) = -^B a({x}) and B a(X) is the Bernoulli polynomial, so

a
0a(x) = ^2{-2iTim)- e(mx).

P R O O F . W e split th e integra l o n the lef t sid e of (20.1 49 ) int o two parts accord -
ing t o (20.1 43 ) an d evaluat e th e sum s ove r r b y th e Euler-Maclauri n formul a (se e
Theorem 4.2 )
F = {a
(20.150) Y
1 ^ l(F(r) + !3 a(r)F \r))dr.
rJ
First w e ge t

which lead s t o th e secon d ter m o n th e right-han d sid e o f (20.1 49) . W e are lef t wit h

r nu)w(^= r
J-00 \ C / C J-00
w{ \u\)±f{cru)duJM
r = 1
c
JO r
r ^d r
= r w{u) Vc f{cru)du - f(0)w(0) - M r ^ d r
Jo ^ h r
= J w(u) J {r}p a(r) — f(cru)drdu - f(Q)w{0)

by (20.1 50) . Changin g r int o r/c w e obtai n th e remainin g term s o f th e right-han d

side o f (20.1 49) . •
 20. T H E C I R C L E M E T H O D 485

COROLLARY 20.1 9 . Let the conditions be as in Lemma 20.1 7. For any f com-
pactly supported and of class C a on R, and for c > 1 we have
oo
Ac(u)f(u)du = /(0 ) + E c(f)
/ -OO

where
oo
(C-2a\f(u)\ + \f a
\u)\)du.
/ -oo

P R O O F . Thi s follow s b y \f3 a(x)\ ^ 1 an d trivia l estimatio n o f th e integral s

in (20.1 49) , bu t wit h th e mai n ter m f(Q)w(0) i n plac e o f /(0) . However , b y th e
normalization conditio n (20.1 33 ) an d th e Euler-Maclaurin formul a (20.1 50 ) w e have
1
a
w(0) = l + O(C- - ).
Here th e erro r ter m i s absorbe d b y (20.1 52 ) becaus e
oo
(\f(u)\ + \f^(u)\)du.
/ -oo

•
Suppose f(u) i s supporte d i n \u\ ^ 2N an d ha s derivative s
{a a
(20.153) f \u) <^N~
for 1 ^ a ^ A. The n (20.1 52 ) give s
2 a
TV / c \ °/ C \
E 1 +
(20.154) ^«7c{c) ( lv) -

In particular, i f C — ViV, then thi s simplifie s t o E c(f) < C (c/C)a (chang e a into a +
1). Recal l tha t th e serie s (20.1 37 ) terminate s a t X = X(C , N) = 2max(C , N/C) =
2\fN. Th e boun d (20.1 54 ) i s ver y smal l i f c ^ X 1 _ £ b y lettin g a b e sufficientl y
large.
When applyin g th e formul a (20.1 37 ) t o solv e a n additiv e proble m i t i s conve -
nient t o have A c (n)/(n) expresse d i n additive character s e(nv) wit h \v\ quite small .
To thi s en d w e conside r th e Fourie r transfor m
oo
Ac(u)f{u)e(-uv)du.
/ -oo

By Fourie r inversio n
oo
gc(u)e(uv)du.
/ -oo

Inserting thi s t o (20.1 37 ) w e ge t

oo
gc(v)e(nv)dv.
/
486 20. T H E C I R C L E M E T H O D

REMARKS. Her e w e kee p th e restrictio n c ^ X , becaus e afte r applyin g th e

Fourier integra l (20.1 56 ) fo r A c(n)f(n) on e loses sight o f the suppor t o f the origina l
test functions .

Now w e nee d estimate s fo r g c(v)- B y Corollar y 20.1 9 fo r f(u)e(—uv) i n plac e

of f(u) w e find tha t

(20.158) g c(v) = 1 + o(± ( £+ ^ + | V|cC)°).

Moreover, w e hav e b y partia l integratio n

oo
(Ac{u)f{u))^e(-uv)du.
/ -OO

Hence w e deriv e b y (20.1 41 ) , (20.1 42) , (20.1 53 ) anothe r estimat e

a 2
(20.159) g c(v) < (\v\cC)~ + ( 1 + NC~ )(\v\N)-a.
The formul a (20.1 58 ) i s fine fo r smal l \v\ whil e th e estimat e (20.1 59 ) i s fo r large r
|v|, th e \v\ — 1/cC bein g th e transitio n value .
The tes t function s f(u) an d w(u) d o no t nee d t o b e compactl y supported , bu t
only decayin g rapidl y enoug h outsid e th e crucia l ranges . W e propose th e followin g
EXERCISE 3 . Giv e explici t computation s fo r th e specia l tes t function s

(20.160) f(u) = e x p ( - £ ) , w(u) = ^exp(-^- £) ,

where K is a normalizatio n constant .

It i s understandable tha t th e Fourie r developmen t o f S(n) o f Kloosterman typ e

(20.87) an d ou r formul a (20.1 35 ) hav e essentiall y th e sam e powe r i n applications ,
because they employ the additive characters to moduli of comparable size . However ,
the absenc e o f Kloosterman fraction s i n (20.1 35 ) ca n simplif y som e treatments. Fo r
example, whe n solvin g th e diophantin e equatio n / ( # ! , . . . , x n ) = 0 b y mean s o f
(20.135) on e i s le d t o a n exponentia l su m fo r a polynomia l i n n + 1 variables ove r
a finite field, wherea s th e Kloosterma n metho d bring s u s a rationa l function .
 http://dx.doi.org/10.1090/coll/053/22

CHAPTER 2 1

EQUIDISTRIBUTION

21.1. Weyl' s criterion .

We begi n b y reviewin g th e genera l principle s o f equidistributio n theory . Le t
(X, 23 , djj) b e a probability space , where X i s a topological space , B is the a-algebr a
of Bore l sets , an d d/i i s a measur e normalize d b y

L dfjb(x) = 1 .

A sequenc e {x n} C X i s sai d t o b e equidistribute d (wit h respec t t o th e measur e

d/i) i f
lim h{n^N;x neB}\=fi{B)
N-+oc T V
for an y ope n se t B e B. Unde r som e mil d assumption s thi s propert y i s essentiall y
that
(21.1) li m1 J2 f( xn) = f f(x)drix)

for an y compactl y supporte d continuou s functio n / : X - > C .

Clearly th e equidistributio n implie s that th e sequenc e {x n} i s dense in X, ye t i t
is a more refine d property . Fo r showin g th e equidistributio n i t i s sufficient t o verif y
(21.1) fo r a syste m o f tes t function s / whic h generate s a dens e subse t o f CQ(X).
The syste m o f tes t function s ca n b e simpl y th e element s o f a n orthonorma l basi s
of L 2(X,dfi). I n particular , i f X = G i s a compac t group , th e matri x coefficient s
of irreducibl e unitar y representation s o f G for m a n orthonorma l basi s o f L 2(G)
with respec t t o Haa r measur e (th e Peter-Wey l theorem) . I f X = G* is the spac e of
conjugacy classe s of a compact group , then the characters of G form a n orthonorma l
basis o f L 2(G$,dg), wit h respec t t o th e measur e induce d b y Haa r measure .
Since the formul a (21 .1 ) alway s holds fo r f{x) = 1 which i s the characte r o f th e
trivial representation , specializin g th e abov e discussio n t o thi s cas e give s th e
W E Y L CRITERION . Let G be a compact group, G$ the set of conjugacy classes
in G. Then a sequence of elements x n G G^ is equidistributed with respect to
Haar measure if and only if for any non-trivial irreducible unitary representation
p : G - > GL(V), we have

(21.2) Y^ ^(P&n)) = o(N), as N - + oo .

A classica l exampl e i n numbe r theor y (initiate d b y H . Wey l [Wl ] i n 1 91 6 ) i s

the circl e X = R/ Z wit h th e Lebesgu e measur e dx. I n thi s cas e i t i s customar y
487
488 2 . EQUIDISTRIBUTIO N

to loo k a t a sequenc e o f rea l number s x n , an d fo r th e purpos e o f equidistributio n

to tak e th e fractiona l part s {x n} a s representative s o n th e circle . I f th e fractiona l
parts ar e equidistribute d o n th e circle , the n th e origina l sequenc e i s sai d t o b e
equidistributed modul o one . Th e character s e(hx) fo r h G Z, h ^ 0 , ar e the n th e
representations involved , an d th e Wey l criterio n fo r equidistributio n modul o 1 is
2_] e(hx n) = o(N), a s N — > oo , fo r an y h ^ 0 .

21.2. Selecte d equidistributio n results .

Originally Wey l introduce d hi s criterio n t o stud y th e distributio n modul o on e
of th e sequence s x n = / ( n ) , wher e / i s a polynomia l wit h rea l coefficients .
PROPOSITION 21 .1 . Let f e R[X] be a real polynomial of degree d^\. Assume
that f(x) = ax d + • • • with a £ Q. Then x n = f(n) is equidistributed modulo one.
PROOF. W e conside r f(x) — axd fo r simplicity . Le t h ^ 0 an d / ? = ha. Fo r
Q = X d~6, wit h 0 < 5 < 1 , appl y Lemm a 20. 3 (comin g fro m Propositio n 8.2 ) t o
an approximatio n \/3 — a/q\ ^ q~ 2 wit h (a,q) = 1 and 0 < q ^ Q. Thi s give s
] T e(/3n d) < X l+£(q~l + X " 1 + qX~ d d
y

with 7 d = 2 l~d. I f g ^ X dQ~1 = X 6, thi s implie s

d
^ e(pn ) < X 1 '1 fo r som e 7 > 0 .

If q < X dQ~1 1 denotin g g{t) — (a — -)t d, w e hav e b y partia l summatio n

d a
S(X) = £ e(Pn ) = £ e( ^ ) e ( 5 ( X ) ) - 2 « f fee(^)) 9'(t)e(g(t))dt.
q Jl
n^X n^X n^t ^

By splittin g int o residu e classe s modul o q we ge t

c( T )=- r + 0 ( g ) w i t h 6 =£ c ( T ) .
n^y ^ ^ x ( m o dq jH

Hence b y partia l integratio n agai n w e ge t

S(X) = ^j\(9mt + 0( q(l + ^ ) ) < < ^+ 0( q).

To th e complet e su m 6 w e ca n appl y eithe r th e bound s fo r Ramanuja n sum s (i f

d = 1 ) , Gauss sums (i f d = 2 ) or Weil's bound fo r higher degre e (als o Weyl's Lemm a
(20.59) would be sufficient), gettin g 6 < q 1 '2^ an d hence S(X) < Xq- l2 £
' + +X8.
Since g -> +0 0 becaus e / 3 ^ Q , w e hav e i n an y cas e
d
^ e(/3n ) - o(X ) a s l - i +00 .

The equidistributio n o f arithmetica l sequence s buil t ou t o f prim e number s ar e

fascinating an d mos t challenging . Fo r prime s themselves , i f on e take s a s spac e
 21. EQUIDISTRIBUTIO N 489

G = (Z/gZ) x wit h th e countin g measure , the n equidistributio n i s equivalen t wit h

the Prim e Numbe r Theore m i n Arithmetic Progression s (se e Chapter 5) . Similarly ,
if K/Q i s a finite Galoi s extension , the n on e wishes t o stud y th e distributio n o f th e
Frobenius conjugac y classe s a p i n th e Galoi s grou p o f K/Q. Ther e th e Chebotare v
Density Theore m give s th e result :
THEOREM 21 .2 . Let K/Q be a finite Galois extension with Galois group G.
For any set C C G which is stable by conjugacy, we have

A o o l|{p^| f f j , e c}| = {g}.

However, i n thi s cas e a s fo r arithmeti c progressions , question s o f th e siz e o f
the erro r ter m an d uniformit y i n term s o f parameter s i s mos t importan t an d quit e
different i n perspectiv e tha n th e simpl e equidistributio n statement .
One is naturally le d to consider sequences x p = /(p) , especially when x n = f(n)
has alread y bee n prove d t o b e equidistribute d modul o one . Ther e th e method s o f
Chapter 1 3 ar e relevant . Indeed , Theore m 1 3. 6 implie s th e followin g theore m o f
Vinogradov
T H E O R E M 21 .3 . Let a 0 Q be an irrational number. Then the sequence ap is
equidistributed modulo one as p— > +oo .

PROOF. Le t (3 = ah wit h h ^ 0 . W e mus t prov e

e
(21.3) J2 (#P) = o(X{logX)~ l) a s X - > +oo .
p^X

Let Q = x ( l o g x ) - B fo r som e B > 0 , an d le t a/q b e a rationa l approximatio n t o /?

such tha t |/ J — a/q\ ^ l/(qQ) ^ q~ 2. I f q ^ xQ~ x = (logx) 5 , applyin g Theore m
13.6 yield s
^e(/3p)<x(logx)3-B/2.

If q < x Q _ 1 , the n b y partia l summation , splittin g int o arithmeti c progression s

modulo q and th e Siegel-Walfis z Theorem , w e hav e

yq)
P^x ^

for an y A > 0. I f w e take B > 6 and A > B + 1 , then (21 .3 ) follow s sinc e q —> +o c
as X — > +o o fo r ft irrational . •

Similarly, (1 3.55 ) implie s that a^fp i s equidistributed fo r an y fixed real numbe r

a^O.
The technique s o f exponential sum s ove r finite fields, an d particularl y th e pow -
erful result s base d o n the Rieman n Hypothesi s prove d b y Delign e (se e Chapte r 1 1 ) ,
give very goo d tool s t o stud y man y interestin g equidistributio n problems . W e giv e
an exampl e o f Fouvr y an d Kat z [FK ] usin g Theore m 1 1 .43 .
490 2 . EQUIDISTRIBUTIO N

THEOREM 21 .4 . Let P 1 (X),..., P r(X) be polynomials in Z [ X i , . . . ,X n] such

that the total degree of any non-trivial integral linear combination a\P\-\ \-a rPr
is ^ 2 . Let <j){x) —> • -fo o as x —> +co . Then, for p — > +oo , £/& e sequence

{(^-••^)}
w/iere 0 < r r i , . . . , x n ^ ( / ^ ^ / p l o g p , * s equidistributed modulo one.
SKETCH O F PROOF . Le t w(p) = 0(p)y^log p an d fo r a = ( a i , . . . , a r ) ^ 0 , le t
r
a\P\{x) + • • • + a rPr(x)\
* = £ - £ <-
O^Zi^u* (p)

By th e Wey l Criterio n o f (R/Z) r , w e mus t sho w tha t

(21.4) S < C w(p)n a s p -> • +oc.
First assum e w(p ) < p . Usin g Fourie r inversio n o n Z/pZ , w e hav e

^h

where

P
a • P(x ) — h • x^

(here / i • x i s th e usua l scala r produc t h\X\ + • • • + h nxx). W e hav e

1
T(M)« [ ] minfoH/K/pir )

where ||t| | i s th e distanc e t o th e neares t intege r a s usual . Fo r th e complet e sum s

5i(a, P, /i), we hav e tw o bounds : first , a resul t o f Wei l show s tha t

(21.5) S ' 1 (a,P,/i)<pn-^

(the implie d constan t dependin g onl y o n th e Pi) becaus e a • P(x) — h • x i s neve r
identically zer o fo r a ^ 0 b y assumption . Thi s boun d i s insufficien t becaus e ther e
are to o man y h. Bu t b y Theore m 1 1 .4 3 wit h / = a - P , g = 1 and V — Ag, ther e
are varietie s Xj o f dimensio n ^ n — j suc h tha t Si(a,P,h) <^ C C pn / 2 + ^ - 1^ 2 fo r
hi Xj. Weil' s boun d (21 .5 ) show s tha t X n = 0 . Denotin g X 0 = A n w e ge t

S « p " n E p^'- 1
)/ 2 £ |T(fc,ti;(p))| .

Lemma 9. 5 o f [FK ] give s

E |T(ft,n;(p)) | < p n - « "1 ) i i ; ( p y -1 ( l o g p ) n - « -1 )

 21. E Q U I D I S T R I B U T I O N 491

from whic h
S<p1/2w(p)n_1(logp)
hence (21 .4 ) follows . I f w(p) > p, one dissects th e se t o f summation int o cube s wit h
sides < p an d appl y th e abov e boun d fo r eac h o f them . •

A differen t se t o f problems , stil l ver y natural , involve s th e "angles " o f expo -

nential sum s t o prim e modulus , o r th e loca l root s o f L- functions fo r varietie s ove r
finite fields. Quit e simila r i s th e questio n o f distributio n o f eigenvalue s o f Heck e
operators T p.
Maybe th e simples t non-trivia l exponentia l sum s ar e Gaus s sums . Le t x b e a
primitive characte r modul o p an d r(x ) th e associate d Gaus s su m (3.1 0) . B y (3.1 4 )
one ca n writ e
r(x) = e(a p(x))y/p,

where a p(x) £ R/Z i s calle d th e "angle " o f th e Gaus s sum . Th e questio n is , ho w

are thos e distributed ?
This questio n admit s a t leas t tw o variants . On e ca n conside r Gaus s sum s fo r
characters o f a fixed orde r d modul o prime s p = 1 (mo d d) , a s p —> +oc , o r on e
can loo k a t al l character s modul o p. I n th e forme r case , whe n d = 2 , Theore m
3.3 an d th e Prim e Numbe r Theore m sho w tha t e(a p(x)) i s equidistribute d amon g
the tw o element s {l,i } (not e tha t i n thi s cas e e(a p(x)) G { ± l , ± z } follow s fro m
r
( x ) 2 — r ( x ) r ( x ) = x(~l)P > s o th e angl e i s no t equidistribute d i n th e grou p
{±i,±i».
Kummer studie d th e argument s o f cubi c Gaus s sums . Fo r p = 1 (mo d 3) ,
there exist s a uniqu e prim e n e K = Q(e(l/3) ) C C suc h tha t TT = 1 (mo d 3 ) an d
NTT — p. Th e cubi c residu e symbo l XTT( X) = ( f ) 3 i s define d a s th e cub e roo t o f
unity suc h tha t
X(P-I)/S = Xn ( mod ^ fo rx e 0 / n 0

where O is the ring of integers in K. I t i s a character o f order 3 on the multiplicativ e

group o f the residue field O/TTO = Z/pZ , i.e . a (primitive ) cubi c Dirichlet characte r
modulo p. Usin g Jacob i sums , on e show s tha t th e Gaus s su m satisfie s

(^)'=«<*»<*•»--?•
see [IR] . It i s well-know n tha t TT/TT i s equidistribute d o n R/ Z (prove d usin g Heck e
characters o f K, se e Sectio n 3.8) , bu t t o ge t e(a p(x-K)) a certai n cub e roo t o f
unity "interferes" . Kumme r an d late r Hass e conjecture d a certai n non-unifor m
distribution base d o n som e numerica l evidence . Howeve r R . Heath-Brow n an d
S. Patterso n [HBP ] proved :

T H E O R E M 21 .5 . The angles e(a p{xx)) of cubic Gauss sums are equidistributed

in R/Z as p - > +oc ; p = 1 (mo d 3) .
The proo f i s base d o n technique s fo r sum s ove r primes , a s i n Chapte r 13, an d
properties o f a (metaplectic ) Eisenstei n serie s whic h ha s th e Gaus s sum s i n it s
Fourier coefficients .
In the case of all characters modul o p, we have the following theore m o f Deligne:
492 2 . EQUIDISTRIBUTIO N

THEOREM 21 .6 . Thep—2 angles (a p(x)) of Gauss sums of primitive characters

modulo p become equidistributed in R/ Z as p tends to oo , i.e. we have

-++00 p - 2 f-^ J 0
X (mo d p)

as p — > o c /o r an? / continuous function f : [0,1 ]— » C

PROOF. B y th e Wey l Criterion , i t i s enough t o sho w tha t
rT n
(2i-6) ii m - J - y- ( (xh
-m = 0
X (mo d p ) v

— see
for n ^ O . On e ca n reduc e t o n > 1 using th e formul a T(X)T(X ) = x ( l)p (
(3.14)).
We hav e

r(x)"= E ^ - x n ) e f1 +
; +
^ )

hence summin g ove r al l x (includin g \ — 1) w e get b y orthogonalit y o f character s

5 > ( x ) n = (p-l)tfn(l,P )
X

where if n (l,p) i s th e multipl e Kloosterma n su m define d i n (1 1 .55) . Subtractin g

the contributio n o f \ — 1 ? w e ge t

n
E* (^f) =p~ n/2 ((p - VW'P) + (-1)"*1)-
X (mo d p ) ^

As p— > +oc , w e hav e K n(l,p) < C p ( n _ 1 ) / 2 b y (1 1 .58) , wher e th e implie d constan t

depends onl y o n n , henc e (21 .6 ) follows . •

Of cours e fo r analyti c numbe r theor y a particula r interes t attache s t o th e clas -

sical Kloosterma n sum s (1 .56) . B y Weil' s bound , on e ca n writ e
S(a,b;p) = 2y/pcos(27r6 p(a,b))
for som e uniqu e 9 p(a,b) G [0,7r]. Th e proble m o f th e distributio n o f th e angle s i s
fascinating. Usin g al l a ^ 0 modulo p , N . Kat z [K4 ] solve d th e problem .
T H E O R E M 21 .7 . The p — 1 angles 9 p(a,a) for a ^ 0 are equidistributed with
respect to the Sato-Tate measure on [0,7r ] given by

djisT = 2TT _1 (sin 0)d0

as p —> +oo ; i.e we have

-IT E * M>(°)) ~ r WW-

a(mod p )

as p —> o o /or an? / continuous function f : [0 , n] —> C .

Moreover, h e state s th e followin g conjectur e wher e a i s fixed:
 21. E Q U I D I S T R I B U T I O N 493

T H E SATO-TAT E CONJECTUR E FO R KLOOSTERMA N SUMS . Let a, b be fixed

non-zero integers. For p prime let 0 p G [0, IT] be such that
S(a,b;p) = 2yfp(cos 0 P).
Then as X —> • +oc 7 the angles 0 P, p < X, become equidistributed with respect
to the Sato-Tate measure d/usr-
It i s not eve n known whethe r S(a, b\p) change s sig n infinitely often , o r that th e
statement
(2-g)^S(fl^;p)=2cosflp<2

(for al l p larg e enough) i s false, fo r an y e G]0 , 2[. Se e [FoM ] fo r th e bes t progres s i n

this direction , combinin g siev e method s an d cohomologica l studie s o f exponentia l
sums a s i n Chapte r 1 1 .
In Corollar y 21 .9 , o n th e othe r hand , w e prov e tha t th e angle s o f Sali e sum s
T(a, b;p) (fo r fixed a , b with ab not a square ) ar e equidistribute d a s p—> > +oo, bu t
with respec t t o Lebesgu e measure .
The sam e Sato-Tat e distributio n i s conjecture d t o hol d fo r th e distributio n o f
coefficients a p associate d t o a n ellipti c curv e E/Q withou t CM , o r mor e generall y
for th e eigenvalue s Xf(p) o f a non-dihedra l primitiv e holomorphi c cus p for m / o f
weight ^ 2 . I f p is a prime not dividin g the conducto r o f /, the n b y Deligne's boun d
we hav e
\f(p) = 2 cos 0 p(f)
for som e 0 p(f) G [0,7r]. Fo r E, thi s read s
aE(p) = 2y/p cos 0 P(E).
Then w e hav e
T H E SATO-TAT E CONJECTUR E FO R MODULA R F O R M S . Let f be a primitive
holomorphic cusp form of weight ^ 2 which is not of dihedral type. With the above
notation, the angles 0 p(f), p ^ X, become equidistributed as X— > -hoc with respect
to the Sato-Tate measure dfisT-
Using th e modularit y o f / an d th e analyti c propertie s o f symmetri c powe r L-
functions, mor e progres s ha s bee n mad e concernin g thi s conjectur e tha n abou t th e
analogue fo r Kloosterma n sum s (se e Serre' s lette r a t th e en d o f [Shah]) . I n fact ,
there i s a convincin g argumen t i n favo r o f thi s conjectur e base d o n th e expecte d
functoriality o f symmetric power L- functions an d th e analytic properties that woul d
follow b y th e result s o f Chapte r 5 . W e ca n sketc h thi s argument , takin g th e cas e
of th e Ramanuja n A functio n o f weigh t 1 2 a s a n exampl e - th e leve l i s 1 whic h
eliminates complication s arisin g fro m ramification . W e hav e th e Eule r produc t
factorizing a s
1s
L(A,s) = i[(i-a pP- r (i-PPp-1sr
v
with p 1 1 ^2(ap + f3 p) = r(p) an d a p/3p = 1 . Her e r(p) i s the Ramanuja n r-function ,
not th e diviso r function . Fo r n ^ 1 , th e n-t h symmetri c powe r o f A i s th e Eule r
product
494 21. EQUIDISTRIBUTIO N

The functorialit y principl e o f Langlands woul d impl y tha t L(Sym n A, s) is an auto-

morphic L- function o f degree n + 1 in the sens e of Chapter 5 , with conducto r q — 1
and wit h n o pole . Sinc e a p(3p = 1, we hav e

sin(n + 1)0 P
Asym-Atf) = r- ^" = Pn(COsO p)
S i l l (7r >

where 0 P = 0 P(A) an d P n is the Tchebyshe v polynomial . Applyin g Theore m 5.1 0

we find tha t fo r n ^ 1,

J2Pn(cos6p) = o(7r{X))

as X— > +oo , henc e

lim -— — Y] Pn(cos6 p) = 0 = / P n (cost)d/x 5T (f).

It i s easy t o sho w tha t th e functions P n (cost) spa n C([0,7r]) , S O the Sato-Tat e

conjecture follow s b y the Wey l Criterio n fro m thes e conjecture d propertie s o f the
symmetric powers .
It i s currently know n tha t L(Sym n A, s) is automorphic an d fo r n ^ 4 : n = 2 is
due t o Gelbart an d Jacque t [GeJ ] an d n = 3 , 4 to Kim an d Shahid i (se e [KSh]) ,
and tha t i t is an L-function i n the sens e o f Chapter 5 , possibly wit h extr a poles ,
for n < 9.
If / i s dihedral, whic h i s equivalent t o E having C M in the ellipti c curv e case ,
the distributio n i s known (b y wor k o f Deuring) an d i s quite different . Notic e tha t
one ca n determin e thi s distributio n b y th e abov e argumen t i f one know s th e orde r
of th e pol e of L(Symn /, s) a t s — 1. Similarly fo r weight 1 forms wher e th e Heck e
eigenvalues ca n onl y tak e finitely man y differen t values .
Finally w e shoul d mentio n tha t equidistributio n problem s i n hyperbolic plan e
are also very interesting . Ther e on e has to us e not onl y cusp forms a s test function s
in th e Wey l Criterion , bu t also Eisenstei n series . A s an example, W . Duke [Du2 ]
proved tha t Heegne r point s becom e equidistribute d wit h respec t t o the Poincare
measure a s the discriminan t goe s to infinity. Als o w e refe r t o [Sa4] fo r discussion
of th e problem s o f "quantu m chaos " an d the links betwee n th e quantum uniqu e
ergodicity conjectur e an d subconvexit y bound s fo r certai n high-degre e L-functions .
21.3. Root s o f quadratic congruences .
Let f(X) = aX2 + bX + c be a quadratic polynomia l wit h intege r coefficients .
For a given prim e p \ 2a the congruenc e

(21.7) /(i/ ) = 0(mod p)

has a t most tw o solutions ; precisel y th e numbe r o f solutions is

P(P) = i + (£
where D — b 2 — 4ac is the discriminan t o f / an d (— ) i s the Legendr e symbol . Henc e
50% of primes yiel d tw o roots an d th e othe r 50 % yields none . Therefor e p(p) i s one
 21. EQUIDISTRIBUTIO N 495

on averag e

(21.8) Y^p(p ) ~ 7r(x) , a s x —> • oo .

It i s a n interestin g questio n ho w th e tw o root s i/(mo d p) ar e distribute d a s p run s

over th e prime s wit h p \ 2a , p ) = 1 . Obviously , i f / ( X ) factor s ove r Q , the n th e
fractions {v/p} ar e no t dens e i n [0,1 ] ; i n fac t th e limi t point s ca n onl y b e ^ wit h
0 ^ u ^ a . Excludin g thi s cas e W . Duke , J . Priedlander , H . Iwanie c [DFI5 ] an d
A. Tot h [Tot ] showe d

T H E O R E M 21 .8 . Suppose f(X) is a quadratic polynomial with integer coeffi-

cients, irreducible over Q . Then for any 0 ^ a < (3 ^ 1 we have

(21.9) \{u (modp);p^xj(iy)=0 ( m o d p ) , ^ { - } < (3} ~ (/ 3 - a)n(x)

as x —> oo .
Equivalently, fo r an y continuous , periodi c functio n F(t) o f perio d on e

F F d t
(21.10) li m -±- £ £ (^ = / ^
V7 U
P ^ z /(i/)=0(modp ) ^

In othe r words , th e sequenc e o f number s i// p i s equidistribute d modul o one . Fur -

thermore, b y Weyl' s criterio n togethe r wit h (21 .8 ) thi s propert y i s equivalent t o

(21.11) l > f c ( p ) = O(T(*)) ,

p^x

as x —> oo , for an y positiv e intege r /i , where

(21.12) Ph (n)= £ e(^) .

i/(mod n )
/(i/)=0(mod n )

A consequence o f Theorem 21 . 8 is the uniform distributio n o f angles of the Sali e

sums
mx \~ ^ fd\/dm + dn\
T(m,n;P)= £ ( - ) ( ^ — ) •
d(mod p)
By Lemm a 1 2. 4 w e hav e

T(m,n;p) = 2co s ( — j T ( 0 , n ; p )

for p \ 2mn , wher e T(0,n;p ) — sPy/p(-) i s th e Gaus s su m an d i / i s a roo t o f

v1 = 4mn(mo d p) . W e thin k o f T(0 , n;p) a s th e normalizatio n facto r an d cal l
2-Kvjp the angl e o f th e Sali e sum . Therefor e Theore m 21 . 8 implie s
496 21. EQUIDISTRIBUTIO N

COROLLARY 21 .9 . If mn is not a square, then the angles of the Salie sums

T(m,n;p) are uniformly distributed modulo 2TT asp runs over primes.
This contrast s wit h th e conjecture d Sato-Tat e distributio n fo r Kloosterma n
sums.
In th e simples t cas e of the polynomia l f{X) = X 2 •+- 1 the root s ^(mo d p) wit h
p = 1 (mod 4 ) correspon d t o th e representation s o f p a s the su m o f two squares p =
r 2 + s 2 . B y choosin g — 5 < r < s wit h s odd , w e see that ever y suc h representatio n
gives th e uniqu e roo t v = rs(mo d p). Henc e

-1
= (mo d 1).
ps sp
In vie w o f thi s constructio n th e assertio n (21 .1 1 ) become s

J2* e(h^)=o(n(x))
r2-\-s2=p^.x

after clearin g th e almos t constan t perturbatio n e(h/sp) = 1 + 0(h/sp). Thi s resul t

is reminiscent t o th e followin g one :

r2+s2=p^x

which wa s establishe d i n [FI] . Othe r surprisin g application s o f Theore m 21 . 8 ar e

given i n [VdP ] an d [Ko2] .
In thi s chapte r w e present a proo f o f Theore m 21 . 8 fo r th e polynomia l

(21.13) f{X) = X 2 -D

with D < 0 (s o f(X) i s irreducible withou t furthe r conditions) . Th e genera l cas e of

/ wit h negativ e discriminan t ca n b e deal t wit h b y completin g th e squar e an d mod -
ifying slightl y th e arguments . However , th e cas e o f positiv e discriminan t require s
substantial change s (becaus e o f th e infinit e grou p o f unit s i n Q(y/D) i f D > 0) .
This cas e wa s settle d b y A . Tot h [Tot ] usin g a n alternativ e pat h passin g directl y
through Kloosterma n sums . Ou r approac h als o depends o n Kloosterman sums , bu t
indirectly vi a Sectio n 1 6.5 .
21.4. Linea r an d bilinea r form s i n quadrati c roots .
By th e Wey l Criterio n w e must sho w th e inequalit y (21 .1 1 ) fo r th e polynomia l
(21.13) wit h D < 0 an d h ^ 0 . W e ca n assum e tha t D i s squarefre e becaus e th e
Weyl su m (21 .1 1 ) wit h frequenc y h fo r D — m2E i s equa l t o th e Wey l su m wit h
frequency mh fo r E (u p t o bounde d amoun t fo r prime s dividin g m) .
According t o th e equidistributio n criteri a i t suffice s t o prov e tha t

=
(21-14) E^)s(D °("logx
as x — > oo , wher e h i s a fixed positiv e intege r an d g(y) i s a fixed smoot h functio n
supported o n 1 ^ y ^ 2 . T o thi s en d w e appl y Theore m 1 3.1 2 fo r th e sequenc e
 21. EQUIDISTRIBUTIO N 49 7

*4 = (a n) wit h a n — ph(n)g(n/x). W e have t o verif y th e condition s (1 3.67) , (1 3.68 )

for A = x e^ wit h e(x) — > 0 . Th e firs t conditio n i s weaker tha n th e estimat e

n
(2i.i5) Y, | E ^( )fl®I«*( lo g*r A

for an y A > 2 , wher e £ ? = 5(A ) an d th e implie d constan t depend s o n A, h an d g.

For th e secon d conditio n i t suffice s t o prov e th e followin g estimat e

A
(21.16) Y^p nPh{mn)g(^\\^x{\ogx)-

for an y comple x number s f3 n supported i n th e interva l

B
(21.17) (logx) ^ n < x^(logx )
with \/3 n\ ^ 1 . W e ca n als o restric t th e suppor t of/3 n t o prime s (se e (1 3.68) ) whic h
helps a t technica l points .
In thi s sectio n w e deriv e bot h (21 .1 5 ) an d (21 .1 6 ) fro m on e estimat e fo r th e
linear form s

(21.18) C d(x)= Yl PkWafe)-

n=0(mod d) ^ '

However, w e nee d a ver y stron g boun d fo r Cd(x) whic h i s unifor m i n larg e range s
of d an d h.
PROPOSITION 21 .1 0 . Letl^h^x. Then
2
(21.19) C d(x) < (x$ + d-*(d,h)ix*)T (dh)log2x
where the implied constant depends on the polynomial f and the test function g.
For d — 1 the spectra l theore m fo r th e modula r grou p woul d giv e a stronge r
(best possible ) resul t

n
where the implie d constan t depend s o n h. Withou t smoothing , th e spectral metho d
yields
Yphin) < X 3 log x

as show n b y V . A . Bykovsk y [Byk] . Ou r approac h t o Cd(N) wa s inspire d b y tha t

of Bykovsky . W e shal l wor k wit h th e grou p To(d) s o w e hav e t o dea l wit h th e
exceptional eigenvalue s tha t migh t b e there . All thes e wil l b e take n car e o f i n
estimation fo r a certai n Poincar e serie s i n th e las t section .
That (21 .1 9 ) implie s (21 .1 5) i s clear . No w w e ar e goin g t o deriv e (21 .1 6) fro m
(21.19). Le t B(x) denot e th e doubl e su m ove r ra,no n th e lef t sid e of (21 .1 6) . Firs t
we install th e conditio n (m,n) = 1 and estimat e th e missin g par t b y

X^IZ|^( m n ^(~)| <*]TKI n~2 <x(logx)- s

498 2 . EQUIDISTRIBUTIO N

because f3 n ar e supporte d o n prime s i n th e interva l (21 .1 7) . Le t B*(x) denot e th e

double su m reduce d b y (m , n) — 1, so B(x) — B*(x) + 0(x(\ogx)~ B). W e arrang e
B*(x) a s follow s

**« < E E I E ^»(v ) E «(£ )

m /((5)=0(m ) (n,m ) = l V 7
f(v)=0(mn) V 7

Then b y Cauchy' s inequalit y B*{x) 2 < C C*(x) logx, wher e

<™-5> E | E «•<= ) E •(£ )

m f(6)=0(m) (n,m ) = l X 7
f(^)=0(mn) X 7

z/=J(m)

Squaring ou t an d changin g th e orde r o f summatio n w e obtai n

C*(x) = Ty2!3 nJn2V*(n1 ,n2)

Til 77- 2

where
-iw \ V ^ (mn l\_(mn2\ sr^* (h v 2 v 2,\

V 7 V 7
(m,n 1 n 2 ) = l /(^)=0(mnj) *
^i=i/2(m)

For ni = n 2 w e use the trivial bound D*(n , n) < C n~2x2. I f ni ^ ri2 , then (ni , ri2) =
1 sinc e w e assume d the y ar e primes , therefor e

z
^-^ \ x J\ x —' V mnino J
V 7
(m,mn 2 ) = l \ / \ / /( v)=0(mmn2)

Note tha t | ^ ^ ^ 2 , or els e th e su m V*{ni,n 2) i s void. Her e w e remove th e

condition (ra , nin2) — 1 and estimat e th e exces s part b y 0((nin 2)~3^2x2) agai n by
taking advantage of ni, ri2 being primes. Wit h the condition (m , nin2) = 1 removed
the resultin g complet e su m T>(ni,n 2) i s just th e linea r for m (21 .1 8 ) fo r d = nin 2l
h replace d b y h(n 2 - n ±) an d th e tes t functio n g(y) replace d b y ^ ^ g ( ^ ) g : ( ^ - ) .
After rescaling , Propositio n 21 .1 0 yield s
V(n1,n2) < C ( ( n i n2 ) - 4 £ 2 + ( m n 2 ) ~ ^ 4 ) x l 0 g 2 x
where th e implie d constan t depend s onl y o n f^g,h. Puttin g togethe r th e abov e
estimates on e ca n easil y complet e th e proo f o f (21 .1 6) .
21.5. A Poincar e serie s fo r quadrati c roots .
In this section w e interpret th e linea r for m Cd(x) give n by (21 .1 8 ) a s a Poincar e
series Ph{z) f° r th e grou p To(d) o f th e kin d considere d i n Section s 1 6. 2 an d 1 6.3 .
Throughout z = x + iy denote s a poin t i n th e uppe r half-plan e H , s o x i s no longe r
the variabl e fro m th e previou s sections .
We begin b y recallin g variou s connection s o f positive definit e binar y quadrati c
forms wit h modula r forms . A fe w relevan t fact s ar e als o reviewe d i n Sectio n 22.1 .
Fix a negativ e intege r D. Conside r al l form s (wit h eve n middl e coefficient )
<p(X, Y) = aX 2 + 2bXY + cY 2
 21. EQUIDISTRIBUTIO N 49 9

of discriminan t b 2 — ac = D. Multiplyin g b y — 1 (if necessary ) w e can assum e tha t

a,c > 0 , s o p i s positiv e definite . Ther e ar e tw o zero s o f p whic h ar e comple x
conjugate. Choos e th e on e i n H

a
The modula r grou p T = SL 2(rL) act s o n th e quadrati c form s o f discriminan t D b y
linear chang e o f variable s

ip°{X,Y) = <p{aX + yY,pX + 6Y)

if a = f r. 1 G T. Thi s actio n i s compatibl e wit h takin g th e zeros , i.e. , az^

z^a. Le t F b e th e standar d fundamenta l domai n fo r T (se e (22.1 0 ) an d (22.1 1 )) .
Every for m o f discriminan t D i s equivalent (wit h respec t t o th e actio n o f T) t o th e
exactly on e for m wit h z^ G F. Denot e th e se t o f root s o f representin g classe s b y

A = {z (fie F] discr(<p ) = D}.

Therefore, ther e exist s a one-to-on e correspondenc e betwee n th e solution s o f b 2 —

ac — D wit h a, 6, c G Z, a , c > 0 an d th e point s o f th e orbit s {az;a G T/T z} fo r
z G A , wher e T z i s th e stabilit y grou p o f z. Henc e w e deriv e

e
ph(n)= J2 ( ^ ) = £ l r *1
l" £ e(hReaz).
2
b =D(mod n) ^ ' zeA aeroo \r
Im crz~yJ\D\/n

Next w e split th e inne r su m int o classe s with respec t t o th e subgrou p F 0 (d) gettin g

(21.20) P/. W = 5 ^ | r z | - 1 Yl E e(/iRear*) .

^EA rer 0 (d)\r aGr oo \r 0 (d)
Im arz—yJ\D\ln

We shall us e this formul a fo r number s n = 0(mo d d). Thi s congruenc e ca n b e buil t

into th e summatio n condition s fo r z an d r (bu t no t a) a s th e followin g congruenc e

(21.21) V\ D
\/lm TZ = ° ( m o d d).

Note tha t thi s conditio n doe s no t depen d o n th e choic e o f r i n th e cose t To(d)r.

Now w e ar e read y t o evaluat e th e linea r form s

(21.22) C d = ] T Ph(n)F(2nhy/\D\/n)
n=0(mod d)

where F(u) i s a smoot h functio n compactl y supporte d o n R + . Introducin g (21 .20 )

we arrang e thi s int o

(21.23) £ d = Elr *l - 1 Y! P
^rz)
z€A r e r 0(d)\r
500 21. EQUIDISTRIBUTIO N

where ^ restrict s r b y the congruenc e (21 .21 ) an d Ph(z) i s the Poincar e serie s fo r
the grou p To(d)

(21.24) P h{z) = ] T FQnhlm az)e(hRe az).

The formul a (21 .23 ) trul y represent s th e linea r for m Cd by th e Poincar e serie s
Ph{rz) becaus e ther e ar e relativel y fe w point s rz satisfyin g (21 .21 ) . Precisel y th e
number o f point s z G A is the clas s numbe r h(D) an d th e numbe r o f coset s To(d)r
with r G r i s th e inde x
[T:ro(d)]=dn(l + i ).

This alon e i s quit e large , however , no t ever y r satisfie s (21 .21 ) . Fo r ever y z G A,
let v(d,z) denot e th e numbe r o f coset s To(d)r fo r whic h (21 .21 ) holds . Le t (p =
aX2 + 2bXY + cY 2 b e th e quadrati c for m correspondin g t o z. The n notic e tha t
y/\D\/lm (z) = c = <p(0,1 ) . Henc e th e conditio n (21 .21 ) fo r

ap
75

is tha t

(21.25) <p((0 , l ) r) = ^(7,5 ) = a7 2 + 2b~/S + c6 2 = 0 (mo d d) .

Moreover, countin g th e coset s To(d)r mean s tha t w e coun t th e root s (7 , £) o f (f

modulo d u p t o a n invertibl e scala r modul o d (i.e . th e projectiv e solutions) .
We clai m tha t

(21.26) v(d,z) ^r(d).

To prove this, recal l we assumed tha t D i s squarefree, s o by the Chines e Remainde r

Theorem w e mus t prov e tha t i/(p , z) < 2 for p \ D prime . Sinc e p 2 doe s no t divid e
D, th e quadrati c for m aX 2 - f 2bXY + cY 2 i s non-zer o modul o p. Th e projectiv e
solutions t o (21 .25 ) fo r d — p wit h 5^0 (mo d p) correspon d t o root s o f th e non -
zero quadrati c polynomia l aX 2 + 2bX + c , henc e ther e ar e a t mos t 2 o f them . I f
p\ 5, the onl y possibl e projectiv e roo t i s the "poin t a t infinity " (l,p) , whic h occur s
if and onl y i f p | a. I n that cas e there was a t mos t on e solution i n the first situation ,
so th e tota l numbe r o f solution s i f alway s ^ 2 .
Note tha t (21 .1 8 ) become s (21 .22 ) fo r th e tes t functio n

(21.27) F(u) = g(2nhy/\D\/ux).

This particula r functio n i s supporte d i n Y" -i ^ u ^ 2 F -1

wit h

(21.28) 7rh^/\D\Y = x.

Therefore, b y th e abov e arrangement s an d observations , ou r proble m o f estimatin g

the linea r for m Cd(x) reduce s t o tha t o f th e Poincar e serie s Ph{z) fo r th e tes t
function F(u) give n b y (21 .27) . I n th e nex t sectio n w e trea t Ph(z) b y spectra l
methods.
 21. E Q U I D I S T R I B U T I O N 50 1

21.6. Estimatio n o f th e Poincar e series .

We shal l estimat e Ph(z) fo r al l z G HI uniformly i n h an d th e leve l d. Ou r goa l
is
LEMMA 21 .1 1 . Let Ph(z) be given by (21 .24) with F(u) supported in Y" -1 ^
u ^ 2Y' forY ^ 2 such that \F^\ < Y j for 0 ^ j < 4 . 77ie n /or any z G M an d
1

r G T w; e /mi; e

(21.29) P ^ ) « (2 / + y- 1 )1 - ((fcy) i + < H (d , h)i {hY)^)r{dh){\ogY) 2

where y = I m z and the implied constant is absolute.

It i s eas y t o se e tha t th e boun d (21 .29 ) fo r Y given b y (21 .28 ) togethe r wit h
(21.26) yiel d (21 .1 9) . Henc e on e complete s th e proo f o f Theore m 21 . 8 fo r th e
polynomial f(X) = X 2 - D.
We still have to prov e Lemm a 21 .1 1 . W e start fro m th e spectra l decompositio n
(see Theore m 1 5.2 )

Ph(z) = Y^(Ph, Uj)uj(z) + Y,-tJ{S' Ea{ '> * + {t ))E*(z> 5 + U)dt

where (UJ(Z)) i s a n orthonorma l basi s o f Maas s cus p forms , togethe r wit h th e
constant functio n fo r th e zer o eigenvalue , an d a runs ove r th e cusp s o f r 0 ( d ) .
By th e Cauchy-Schwar z inequalit y
2
(21.30) \P h{z)\ ^ K d{z)Rd{h)

where

(21.31) K d{z) = £>(*i)K(*)|2 + J2^jRHt)\Ea(z, \ + it)\2dt

and

(21.32) W = E ^ I ^^ + EyE^l(ft.^^«)>P*-
Here h(t) > 0 i s a functio n decreasin g fas t enoug h a t infinity , introduce d fo r con -
vergence. W e choos e
(21.33) h(t) = (1 + t 2)-1 - ( 4 + t 2)-1 = 3( 1 + t 2)-\A + t 2)"1 .

First w e estimat e Kd(z). T o thi s en d observ e tha t (21 .31 ) i s th e spectra l de -

composition o f th e automorphi c kerne l
k
Kd(z) = Yl «z^z))
~rer0(d)

where k(u) i s th e Harish-Chandra/Selber g transfor m o f h(t) (se e Theore m 1 5.7) .

Moreover, notic e tha t k(u) i s non-negativ e becaus e th e Fourie r transfor m o f h(t)
is positiv e an d decreasin g o n 1 R +. Thes e observation s sho w tha t K(z) will onl y
increase i f one replace s To(d) b y th e modula r grou p T = r 0 ( l ) . Havin g don e
this th e dependenc e o n d i s gon e an d th e enlarge d kerne l i s T- invariant, s o w e ge t
502 2 . EQUIDISTRIBUTIO N

Kd{rz) ^ K\{TZ) = K\(z) fo r an y r e T an d 2 G HI. Usin g crud e estimate s fo r

cusp form s an d Eisenstei n serie s fo r th e modula r grou p w e deriv e
1
(21.34) K d{rz)<z:y + y-
for an y r ET an d 2 E I , wher e th e implie d constan t i s absolute . Actuall y fo r ou r
application w e coul d sto p a t th e inequalit y K^(rz) < Ki(z) becaus e z run s ove r a
fixed se t o f h(D) point s an d w e allo w th e implie d constan t t o depen d o n D.
Now i t remain s t o estimat e Rd(h). B y (1 6.1 6 ) w e hav e
(Ph,uj) = (2irh)hj(h)F(t j)

where pj (h) i s th e Fourie r coefficien t o f Uj (z) an d

/•O O

F(t)= / F(y)K it(y)y-Uy.

Jo
First w e explai n ho w t o estimat e F(t). Sinc e F(y) i s supporte d o n th e dyadi c
segment Y~ x < y < 2Y~ l wit h Y ^ 2 , i t i s usefu l t o appl y th e powe r serie s
expansion fo r Ka(y) gettin g a rapidly convergen t serie s of Mellin transform o f F(y)
at th e point s \ ± it + 2£, for £ = 0,1 , 2, The n integratin g th e Melli n transfor m
by part s fou r time s w e gai n th e facto r h(t) (se e (21 .33)) . Nex t w e estimat e th e
resulting term s b y Stirling' s formul a (5.1 1 2) , and su m the m u p t o deduc e tha t F(t)
on th e spectru m satisfie s
F(t) < /i(t)(cosh7rt)-^y^ {Y u + Y~ lt + 3 l o g y ) .
Actually thi s resul t i s a synthesis o f three differen t bound s derive d alon g th e abov e
lines separatel y i n th e range s 0 < it < | , 0 < £ < | , | ^ £ < 00 . Henc e w e obtai n
2
hit^KP^Uj)? < hYH( tj)\Pj(h)\ ,
where
H(t) = h(
f) AY 2it
+ Y~2it + 9log 2 Y).
COSh{7Tt)
A simila r estimat e i s deduce d fo r th e inne r produc t o f th e Poincar e serie s Ph an d
the Eisenstei n serie s E a. Fro m thes e estimate s w e ge t

Rd(h) « hYfeHMlpjih)? + £ ^ J^H(t)\ra(h,t)\2dt),

where r a (/i, t) i s the Fourie r coefficien t o f E a(z, \ +it) (se e (1 6.22)) . The n applyin g
(16.56) an d (1 6.58 ) w e arriv e a t
(21.35) R d(h)<^hY {I+ d- 1 (d,h)^(hY)1 *}T2{dh) log 4
Y.

Finally insertin g (21 .34 ) an d (21 .35 ) int o (21 .30 ) w e finish th e proo f o f Lemm a
21.11.
 http://dx.doi.org/10.1090/coll/053/23

CHAPTER 2 2

IMAGINARY QUADRATI C FIELD S

22.1. Binar y quadrati c forms .

We are primarily intereste d i n the imaginary quadrati c field K ~ Q(v^D) , but
to pu t our treatment int o historica l perspectiv e w e begin b y reviewing th e theory
of binar y quadrati c form s

(22.1) <p(X, Y) = aX2 + bXY + cY 2

of discriminan t

(22.2) D = b2 - 4ac < 0.

We d o not assum e tha t D i s fundamental , however , w e only conside r primitiv e

forms, i.e.,

(22.3) (a,6,c ) = l.

This i s not a serious restrictio n becaus e an y form i s a multiple o f a primitive form .

Changing th e sign of all coefficients (i f necessary) w e can assume tha t

(22.4) a > 0, an d c> 0

so (f is positive definite . W e call a, b, c the first, the middle an d the last coefficient s
of (p, respectively. Th e quadratic for m <p(X, Y) factor s ove r comple x number s int o
linear form s

(22.5) <p{X, Y) = a(X + z aY)(X + z aY) = c{zcX + Y)(z cX + Y)

with

(22.6) z a = — an dz c= — .
2a 2 c
Actually th e roots z a,zc depen d o n b but we do not displa y it . Throughou t w e
choose the root s o that z a an d zc ar e in the upper half-plan e H . Not e that z azc = 1 .
The modula r grou p T — SL 2 (Z) act s on the primitive form s o f a given discrim -
inant D by unimodular transformation s

(22.7) ^(X,Y) = <p(aX + 1 Y,pX + 5Y), * * = ( " j ) e T

= a aX2 + baXY + c aY2,

503
504 22 . IMAGINAR Y QUADRATI C FIELD S

say, wit h
aa = aa 2 + baf3 + c/3 2
(22.8) 6* = 2acr y + 6(a< 5 + £7 ) + 2c/3<5
c a = a7 2 + 67< S + c<5 2.
This actio n i s compatible wit h actio n b y linea r fractiona l transformatio n o n H, an d
the root s ar e preserved .
Two form s p,i\) ar e sai d t o b e equivalen t i f the y belon g t o th e sam e T-orbit ,
i.e., ip = (p a for som e a £ T ; in this case we write cp ~ ip. We denote the equivalenc e
class whic h contain s (p b y
(22.9) [<p] = {ip° : a e T}.
Obviously, ever y equivalence class is represented b y a unique form <p(X, Y) = aX 2 +
bXY + cY 2 whos e roo t z a i s i n th e standar d fundamenta l domai n o f th e modula r
group F = F~ U F+ , wher e

(22.10) F' = {zeU: \z\ > 1 , - - < x < 0},

+
(22.11) F = { ^ e M : \Z\ ^ 1 , O ^ x ^ i } .

such a for m <p i s called reduced ; i t i s characterized i n term s o f coefficient s b y

(22.12) -a<b^a^c wit h b^O i fa = c.
Since I m (z a) = v /TDJ/2a ^ \ / 3 / 2 , th e first coefficien t o f a reduce d for m satisfie s

(22.13) a < V|L>|/3 .

Hence i t follow s tha t th e numbe r o f reduced form s i s finite. Thi s numbe r h — h(D)
is called the class number o f discriminant D becaus e it equal s the number o f distinct
equivalence classes .
The transformation s r eT whic h fix a form (p ar e calle d automorph s o f (/?; they
form a finite cycli c grou p Aut(<p ) = { r G T : (pT = (p} of orde r

(22.14) u> =
f6 if
if
D = -3,
D = -4,
4
2 if D < -A.
Equivalent form s hav e conjugat e group s o f automorphs , namel y
1-
A\lt((pa) = c r A u t ( < / ? ) c r .

For reduce d form s Aut(c^ ) i s generated b y

2
# = ( \ "/ ) if(^(x,y)-x + xr + F2,
1 -
s
=(i ) if^(x,r ) = x 2 + F2 ,
l
T=\ ~ J i f D<-4.
 22. IMAGINAR Y Q U A D R A T I C FIELD S 505

Note tha t fo r all integers k the forms p(X + /cY , Y) ar e equivalent; suc h form s are
said t o be parallel. Th e parallel form s hav e th e same firs t coefficien t a while the
middle coefficient s ar e in the same residu e clas s modul o 2a.
In "Disquisitione s Arithmeticae " Gaus s introduce d compositio n o f form s o f
a give n discriminant . Hi s construction wa s subsequently simplifie d b y Dirichle t
but onl y fo r pair s o f form s p 1 (X,Y) = a-^X 2 + b\XY + c\Y 2 an d <p 2(X,Y) =
2 2
a2X + b 2XY + c 2Y satisfyin g

(22.15) ( 6Ji +_ 62 2 )
fll} > a 2 1.

Note that 61 , b2 have the same parit y (namel y 6 1 = b 2 = D(mo d 2) ) so | ( 6i +62) is

an integer . I f the Dirichle t conditio n (22.1 5 ) holds , the n th e forms p>\,p 2 ar e said
to b e united. Clearly , thi s relationshi p extend s t o the classes of parallel forms . For
the unite d form s ther e exist s a unique 6(mo d 2aia 2 ) suc h tha t
2
(22.16) b = 61 (mod 2a x ), b = 6 2(mod 2a 2 ), b = D(mod 4aia 2 ).
Putting b 2 — D = Aaia 2c w e derive th e form
(22.17) p(X, Y) = aia2X2 + bXY + cY 2
which i s primitive o f discriminant D; it satisfie s th e identity

(22.18) <f(x,y) = ^1 (^1 ,2/1 )^2(^2,2/2 )

with x = X1 X 2 — q/1 2/ 2 an d y = a\X\y 2 + 62/1 2/ 2 + ^2^22/1 - Thi s i s the Dirichle t
formula fo r composition o f united classe s o f parallel forms .
Now we use the Dirichlet formul a t o define compositio n o f any tw o equivalenc e
classes. W e sa y that a positive intege r m i s properly represente d b y p if
(22.19) m = <p(a , 7) wit h (a , 7) = 1 .
Since ip is positive definite th e number of proper representations , denote d by r* (ra),
is finite . Equivalen t form s represen t properl y th e same numbers ; thes e ar e exactly
the number s whic h appea r a s the first coefficient s o f forms i n the equivalent class .
Every primitiv e for m represent s properl y a number m whic h i s co-prime wit h any
fixed positiv e integer . Therefor e give n tw o classe s A\,A 2 w e can choose represen -
tatives (/?i,(/? 2 wit h (ai,a 2) = 1 ; such form s ar e united. The n (22.1 7 ) determine s
the thir d clas s A = A\A 2 = [<p\. On e can show tha t A doe s no t depend o n the
chosen representative s pi,(p 2, s o the Dirichlet compositio n induce s a well-define d
multiplication o f classes . Th e law of compositio n i s commutativ e an d also asso -
ciative (th e latter propert y i s not obvious). Thi s make s th e set H o f all distinc t
equivalence classe s a finit e abelia n grou p o f order h. Th e identity elemen t o f 'H ,
which w e denote b y 1, is the class whic h contain s th e principal for m

p = X 2 - jY 2 i f D = 0(mod 4),
(22.20 )
ip = X 2 + XY + ^—T^-Y 2
HD = l(mod 4).

The invers e clas s o f <p(X, Y) — aX2 + bXY + cY 2 i s th e clas s whic h contain s

Tp(X,Y) = aX 2 — bXY + cY 2; thes e ar e calle d th e opposit e forms . Not e th e
Jp(X,Y) i s equivalent t o (p(Y,X) (th e for m whic h i s obtained b y interchanging the
506 22. IMAGINAR Y QUADRATI C FIELD S

first an d th e las t coefficients) . Therefor e th e set s of numbers properl y represente d

by a class A an d it s invers e A~ 1 coincide.
The question of which numbers ar e represented properl y by a given form, ha s no
simple answer . However , ther e is a simple, sufficien t an d necessar y conditio n fo r a
positive integer m to be properly represented b y some primitive form o f discriminant
D, namel y i t is the solvabilit y o f th e congruenc e
2
(22.21) b = D(mod 4m) .
Let Rf)(m) denot e th e numbe r o f al l representations o f m by a complete syste m o f
inequivalent form s o f discriminan t £> , thus

iMm) = £ £ ^(m/d 2)
2
Vd \m

where ip runs ove r inequivalen t forms . I f (m, D) = 1 we hav e

(22.22) R D{m)=wY,XD(d)
d\m

where XD i s the Kronecker symbol . W e shal l se e tha t fo r special discriminant s

(idoneal numbers ) th e abov e formul a yield s th e numbe r o f representations fo r a n
individual for m (becaus e exactl y on e clas s o f forms represent s m) . Thi s "conve -
nient" situatio n i s explained i n th e genu s theor y du e to Gauss, whic h w e ar e goin g
to presen t i n more moder n terms .
First w e classif y th e binar y quadrati c form s o f a given discriminan t accordin g
to rationa l unimodula r transformation s rathe r tha n th e integra l ones . Tw o form s
if, ip are sai d t o b e i n th e sam e genu s if ijj(X, Y) = ip(aX + 7Y, fiX + SY) fo r som e
a = I r. 1 G 5Z'2(Q)- I n fact i t suffice s t o apply a wit h al l entries havin g
the denominato r SD. Obviously , equivalen t form s ar e i n the sam e genu s bu t no t
conversely. Indeed , th e invers e class A~l i s in the genu s of A, though A' 1 doe s no t
always coincid e wit h A. Ever y genu s ha s th e sam e numbe r o f classes. Th e genu s
which contain s th e principa l clas s is called th e principa l genus ; w e denot e i t by Q
and pu t
(22.23) h x = \G\.
A beautifu l theore m o f Gauss assert s tha t Q consists o f square s o f classes ,
(22.24) Q = {A2 : A G H}.
Thus Q is a subgroup o f W. Th e facto r grou p T — Ti/Q i s called th e genu s group .
We sa y tha t a for m <p i s ambiguous i f it is equivalent t o Tp; i n other words ,
the clas s A — [tp] ha s exponen t tw o i n the clas s grou p %. Suc h classe s for m th e
subgroup
(22.25) £ = {AEH:A2 = l}.
Thus £ is isomorphic t o T becaus e it is th e kerne l o f th e homomorphis m A —> A 2.
Putting
(22.26 ) ft 0 = \S\
 22. IMAGINAR Y Q U A D R A T I C F I E L D S 50 7

we hav e h — h§h\.
The grou p o f ambiguou s classe s £ i s th e easie r par t o f %. A n ambiguou s for m
(p whic h i s reduced i s characterized b y havin g it s roo t z a o n th e boundar y o f F +,

(22.27 )
2a
Hence a reduce d for m ip(X, Y) = aX 2 + bXY + cY 2 i s ambiguou s i f an d onl y i f
a = 6 , a = c or b = 0 in which cases the discriminant factor s int o D = a(a — 2c), D =
(6 — 2a ) (6 + 2a) o r D = —4ac , respectively. Usin g thes e characterization s on e ca n
compute ho; we obtai n

(22.28 ) h0 •yr+S — l

where r i s the numbe r o f distinct, od d prim e divisors of D an d s = 0,1 , 2. Precisely ,

if D = l(mo d 4) , the n s = 0 , an d i f D = -AN, the n

s= 0 i f N E E 3 , 7(mod 8) ,
(22.29 ) s= 1 i f J V = l,2,4,5,6(mod8) ,
5=2 i fT V = 0(mo d 8) .

If a positiv e intege r m prim e t o D i s represente d properl y b y a for m o f dis -

criminant D (i.e. , (22.21 ) i s solvable) , the n i t i s represente d b y form s onl y o f on e
genus, actuall y th e residu e clas s m(mo d D) alon e determine s thi s genus . I n thi s
connection Gaus s assigne d t o a discriminan t D a syste m o f r + s genu s characters .
These ar e th e Legendr e symbol s \p f° r ever y p \ D,p > 2 , an d i f D = —AN, w e
have extr a s character s modul o 8 listed below

none if i V = 3,7(mod8) ,
X4 if i V = l,4,5(mod 8
X8 if A ^ = 6(mod8) ,
X4X8 if 7 V - 2 ( m o d 8 ) ,
X4 and x s if i V = 0(mod8) .

Let S denot e th e syste m o f genu s characters . W e hav e

(22.30 ) n X=X D
xes
Hence i f m i s represente d b y a for m o f discriminan t D an d (m , D) — 1 , the n b y
(22.30) an d (22.21 ) w e obtai n Yl xesX(m) = Xr>( m) = 1 - Tw o suc h number s m,n
are represented b y forms o f the sam e genus if and onl y if x(m) = x( n) f° r al l % ^ <5.
Therefore a genu s o f form s i s characterize d b y a collectio n o f sign s e x = ± 1 suc h
that

(22.31)
ii'
xes
and w e hav e exactl y 21 5' x possibilities .
508 22. IMAGINAR Y QUADRATI C FIELD S

Now we return t o th e formul a (22.22 ) which , i n vie w o f the genu s theory , give s
the number o f representations o f m b y forms representing the genus whose invariant s
are
(22.32) e x = x(m) fo r al l x G 5.
If ther e i s onl y on e clas s i n eac h genus , the n Rr>{m) reduce s t o th e numbe r o f
representations b y a singl e form . Henc e a question : whic h discriminant s D satisf y
hi = \G\ = 1 , o r equivalentl y
2
(22.33) A = 1 fo r ever y AeH.
After Eule r w e call suc h a discriminan t "numeru s idoneus" , (i n Englis h i t i s calle d
"idoneal number" , o r "convenien t number" ; i n Frenc h "nombr e convenable") . A s
a matte r o f fac t Eule r considere d discriminant s o f typ e D = —AN an d calle d N a
numerus idoneu s rathe r tha n D. Gaus s gav e a lis t o f 6 5 such numbers :
( N = 1,2,3,4,7 with h(D) = = 1 ,
N = 5,6,8,9,10 (plu s te n more ) with h(D) = = 2 ,
I N = 21,24,30,33 (plu s twent y more ) with h(D) = - 4 ,
N = 105,120 (plu s fiftee n more ) with h(D) == 8 ,
{ N = 840,1320,1365,1848 with h(D) = = 1 6 .
The las t numbe r i n Gauss ' lis t N = 1 84 8 i s th e larges t know n numeru s idoneus .
W. E . Brigg s an d S . Chowl a [BC ] showe d tha t ther e i s a t mos t on e suc h numbe r
beyond 1 0 65 .
Idoneal numbers N hav e the property tha t th e form x 2 + Ny 2 represent s prime s
essentially i n n o mor e tha n on e way , wherea s a composit e numbe r ha s severa l rep -
resentations i f an y (fo r detail s se e [Cox]) . Thi s propert y offer s a n algorith m fo r
primality testin g whic h wa s a n inspiratio n fo r Eule r t o stud y th e idonea l number s
in th e first place .
22.2. Th e clas s group .
Prom no w on we assume tha t D i s a negative fundamenta l discriminant . A fun -
damental discriminan t i s characterized b y th e propert y tha t ever y for m tp(X, Y) =
aX2 + bXY — cY 2 wit h b 2 — 4ac — D i s primitive , i.e. , (a , 6, c) = 1 . Suc h negativ e
discriminants ar e o f typ e D = l(mo d 4) , D squarefree , o r D = — 4N, A f squarefre e
with
(22.34) N = 1 ,2,5 , 6 (mo d 8) .
The associate d Kronecke r symbo l X£>( m) = ( ^ ) ( s e e (3-43) ) i s a real , primitiv e
character o f conducto r —D. I t turn s ou t tha t th e fundamenta l discriminant s ar e
exactly th e discriminant s o f quadrati c fields .
Throughout w e us e th e notatio n fro m th e theor y o f th e imaginar y quadrati c
field K = Q(y/D) a s describe d i n Sectio n 3.8 . Th e quadrati c form s (£>(X , Y) (wit h
a > 0 , c > 0 , b 2 — 4ac = D < 0 ) correspon d i n a one-to-on e wa y t o th e primitiv e
ideals
r b+y/D-)
(22.35) a = a, = a[l,z a]cO.
 22. IMAGINAR Y Q U A D R A T I C F I E L D S 50 9

Hence th e terminolog y an d result s fo r quadrati c form s whic h wer e use d i n th e

previous sectio n translat e naturall y fo r ideals . I n thi s sectio n w e go through thi s
transition.
Recall th e following :
/ - the group o f non-zero fractiona l ideals ,
P - the subgroup o f principal ideals ,
H = I/P - the class group ,
h=[H\- th e class number ,
D = (VD) - the different .
Every clas s A G H ha s a unique, primitiv e idea l (22.35 ) wit h z a i n the funda -
mental domai n F — F~ U F+ o f the modular grou p Y = SL 2(Z); suc h idea l i s said
to b e reduced. Thu s th e number o f reduced ideal s i s h. Th e subgroup

G = {A 2 : A e H}
is calle d th e principa l genu s an d the facto r grou p T = %/Q i s calle d th e genu s
group. Henc e tw o non-zer o ideal s a , b belon g t o th e sam e genu s i f an d onl y i f
a — be2 for som e eel. On e can also sho w tha t a , b are in the same genu s i f and
only if

(22.36) Na = NbNj fo r some 7 e K*.

A clas s A G % is said t o be ambiguou s i f A = A - 1 , i.e. , A 2 = 1 . Th e group of

ambiguous classe s
£ = {AeH:A 2 = l},
as the kernel o f the homomorphism A —) • A 2, i s isomorphic t o the genus grou p T.
The correspondenc e betwee n quadrati c form s an d primitive ideal s cast s ambiguou s
forms int o ambiguou s ideal s ( a is ambiguous i f a2 is principal); i t proves tha t ever y
ambiguous clas s A ha s its reduce d primitiv e idea l a wit h z a o n the boundar y o f
F + . Recal l tha t a reduce d ambiguou s for m ip(X,Y) = aX 2 + bXY = cY 2 ha s
a — b, a = c or 6 = 0 , hence i t yield s a factorization o f the discriminant int o

-D = d ±d2 = a(4 c - a) , (2 a - b) (2a + 6) , 4ac ,

respectively, wit h d 2 > d\ > 0 and ( d i , ^ ) = 1 (the coprimality o f these factor s

results fro m D bein g fundamental) . Conversely , a factorizatio n — D = d\d 2 wit h
d2 > d\ > 0 , (di,d 2) = 1 yields a reduce d ambiguou s form . Fo r example , i f
D = l(mo d 4 ) we obtain

d l 2 2
f <p(X, Y) = dxX2 + d xXY + ^ r ,i f d 2 > 3a7!,

[ y,(X,y ) = ^ X2 + d l_^XY^dl±AY2, i f * < 3d!.

Correspondingly, a reduce d primitiv e ambiguou s idea l yield s a factorizatio n

of th e differen t D = (y/D) int o D = D1 U 2 with (^1 ,^2 ) = 1 ; not e tha t Cl(d\) =
C£(D2). Distinc t suc h factorization s (u p to the order o f factors di,d 2) yiel d distinc t
ambiguous classes , therefor e

(22.37 ) ft 0 = \S\ =2t~1

510 22. IMAGINAR Y QUADRATI C FIELD S

where t = u(\D\) is th e numbe r o f distinc t prim e divisor s o f D (not e tha t (22.37 )

agrees wit h (22.28 ) i n cas e o f fundamenta l discriminants) .
The genu s o f a n idea l ca n b e determine d b y value s o f rea l character s o f th e
class group . Thes e character s ar e assigne d t o ever y factorizatio n D = D1 D2 int o
two fundamenta l discriminant s D\,D2- Not e tha t Di,i^ 2 hav e differen t sign s an d
are coprime , s o w e hav e exactl y 2 t~l distinc t factorization s u p t o th e order . W e
define XD 1 ,D2 n r s ^ o n prim e ideal s b y

lXD2(Np) i f p f D 2

(this i s wel l define d becaus e XD(NQ) = 1 if (a , D) = 1 ) , an d w e exten d XD 1 ,D2 t o

all non-zer o fractiona l ideal s b y multiplicativity . W e obtai n a characte r XD UD2 :
I— > {±1} suc h tha t XD 1 ,D2(a) — 1 f° r a ^ d E P. Indeed , fo r a = (a ) wit h
a = ±(m + ny/D), (a,D) = 1 we have XD UD2{O) = XDA*) = XDi(^(m 2 - n 2D)) =
XD1(n^) — 1 (similarl y on e ca n chec k th e othe r cases) . Therefor e XD 1 ,D2 € W;
these ar e exactl y al l rea l character s o f th e clas s grou p calle d th e genu s characters .
Two ideal s a , b G / ar e i n th e sam e genu s i f an d onl y i f

x(a) = x(b ) fo r al l genu s characters .

The genu s theor y o f Gaus s wit h it s explici t characterizatio n o f ambiguou s

classes prove d t o b e ver y usefu l fo r primalit y testin g an d factorizatio n techniques .
A metho d fo r factorin g — D whic h exploit s th e clas s grou p structure , a s describe d
by D . Shank s [Sha] , require s onl y 0(|D|4 ) operations .
If — D i s prime, then th e whole class group i s the principal genu s Ti = Q, I n thi s
case H tend s t o b e cycli c mor e ofte n tha n not . Afte r Gauss , a discriminan t D (no t
necessarily negativ e prime ) i s said t o b e regula r i f the principa l genu s i s cyclic (thi s
is not th e sam e a s th e regula r prime s occurrin g i n th e theor y o f cyclotomi c fields).
We stil l d o no t kno w i f ther e ar e infinitel y man y regula r discriminants . However ,
the numerica l evidenc e support s th e existenc e o f a large proportio n o f these. I t ha s
been conjecture d i n [Ge ] tha t th e proportio n o f regula r t o al l discriminant s (bot h
negative) i s
00 _ . .

(C(6)IIC(n)) =0.8469... .
n=4

Gauss genu s theor y gives u s a ful l descriptio n o f th e 2-Sylo w subgroup s o f th e

class group . Ther e ar e som e interesting , however , incomplet e development s fo r
other p-Sylo w subgroup s o f H, i n particular, fo r p = 3 by Davenpor t an d Heilbron n
[DH]. Accordin g to the Cohen-Lenstr a heuristic s th e probabilit y tha t a n imaginar y
quadratic field ha s i n it s idea l clas s grou p a n elemen t o f orde r p i s
00

l-n^-p - ")-
71=1
 22. IMAGINAR Y 1
QUADRATI C FIELD S 5 1

22.3. Th e clas s grou p L-Functions .

Let x : H—) > C * b e a characte r o f th e clas s group . Defin e
(22.39) L K(s,X) = J2x(*)(Nay s
a
=
for R e (s) > 1 wher e th e serie s converge s absolutely . Notic e tha t LK(S,X)
LK{S, X)- Fo r an y primitiv e idea l a = [a , b+ 2 ] we hav e

J2X(a)LK(s,X) = hJ2(Nb)- s = ^a- s £ \a\~ 2s

by writin g b = (a)a . Sinc e a - 1 = Z + z a Z, w e deriv e

E i«r 3'= E E i™+^r 2s

O^aGa- 1 (m,n)^(0,0 )

m+n 2s = 2 2s W\\-s
c(2«)E E i ^i" ^ ) ( ^ ) " ^«> s)
(m,n) = l
where
E(Z, S )= 5 3 (im 7«r = ^EEj'*i m+n;j: i" 28
Ter^vr (m,n)= i
is th e Eisenstei n serie s fo r th e modula r group . Thi s give s u s
1h/J\D
(22.40) ^ x(a)L K(s, X ) = — ( ^ - [ ) C ( 2 * ) £(*„ , *)•

We pu t
s
(22.41) 6>(s)=7r- r(s)C(2s),
(22.42) E*(z,s)=0(s)E(z,s),
s
(22.43) A K(S,X) = (2n)- r(s)\D\iLK(s,X).
In thi s notatio n (22.40 ) become s

(22.44) £ x(a)A/ f (*, x) = ^E*(z a, s).

By Fourie r inversio n (th e orthogonalit y o f characters ) w e obtai n

(22.45) A K(s,X) = -J2x(a)E*(z a,s)

a
where a runs ove r a set o f primitive, inequivalen t ideal s ( a good choic e being the se t
of reduced ideals) . B y virtu e o f this formul a th e Heck e L- function inherit s analyti c
properties o f th e Eisenstei n series .
It i s well know n (se e (1 5.1 3) ) tha t E*(z, s) ha s th e Fourie r expansio n
(22.46) £?*(* , s) = 6{s)y s + 0(1 - s)y 1
~s
oo

+ 4 v ^^r s _ 1 / 2 (n)K s _ 1 / 2(27rny)cos(27rnx)

512 22. IMAGINAR Y QUADRATI C FIELD S

where
(22.47) ,(«)-E(sr- ad=n

The Fourie r expansio n yield s analyti c continuatio n o f E(z, s) to the whol e comple x
5-plane, i t shows tha t E(z,s) i n Re (s) ^ \ ha s onl y a simple pol e a t s = 1 with
constant residu e ~ , and i t satisfies th e functiona l equatio n
(22.48 ) E*(z,s) = E*(z,l-s).

Inserting (22.46 ) int o (22.45 ) w e obtain th e Fourie r expansio n

,
(22.49) A Jf (, )X )=ij;xW{'W(f) +»(l-«)(f)1 '
4
\D\\W D\
+ ( ^ ) £T._ Va ( » ) * .1
_ /2 (*n—)cos(™-)}

where a runs ove r primitiv e representative s o f ideal classes . Not e tha t th e Fourie r
series converge s rapidl y sinc e th e Besse l functio n K J/(y) ha s exponentia l deca y

1 / 2
^(y)=(^) e - ( i+ ^ ) .
2yJ
for y > 0, v e C where \0\ ^ \v 2 - \\ (se e (23.451 .6 ) o f [GR]).
We deduce b y (22.45 ) or (22.49) tha t LK(S, X ) n a s analytic continuatio n t o th e
whole comple x s-plane , i t is entire excep t fo r a simple pol e a t 5 = 1 if \ i s trivial
with
27T/ i
r
(22.50) esCx(s)
5=1
w\/\D\
and i t satisfies th e functiona l equatio n

(22.51) A / f (s,x) = A jK:(l-s,x).

All th e above propertie s o f LK(S,X) c a n ^e also deduce d fro m th e integral

representation ( a la Riemann, se e (4.77) ) du e to Hecke

<22-52> «''* > - s ^ ij + /"(»- + . -U (-4fe,)*

where 5(% ) = 1 if x is trivial, o r else <5(x ) — 0 , an d

(22.53)/ XW = E# W
a

where a runs ove r al l non-zer o integra l ideals . Integratin g termwis e w e ge t

 22. IMAGINAR Y QUADRATI
1 C FIELD S 5 3

where a =T V a an d T(s,x) i s the incomplete gamm a functio n

s l
T(s,x)= / e-yy - dy.
JX

Hecke derive d th e formula (22.52 ) b y splitting LK{S, X) m ^o a sum of Epstein zet a

functions (apparentl y introduce d first b y Dirichlet)

s
(22.55) Z a(s)= J2 1 2 ^a(m,n)-
(m,n)#(0,0)

where (p a(X,Y) = aX 2 + bXY + cY 2 i s the quadratic for m correspondin g t o the

ideal a . Thi s i n turn i s the Mellin transfor m o f the theta functio n

mn

Both 6 a(z) an d f x(z) ar e modular form s o f weight on e for the group ro(|D| ) an d
character XD- I f X i s n ° t real , the n f x(z) i s a primitiv e cus p for m wit h Heck e
eigenvalues

(22.56) A x (n) =£ X (a).

Na=n

If x i s real , sa y x — XD 1 ,D2 wit h D1 D2 — D, the n th e Heck e L-functio n factor s

into Dirichle t L- functions. Precisely , i t ca n be verified b y comparing loca l factor s
of Eule r product s an d using (22.38 ) togethe r wit h th e factorization la w of ideals in
K - Q(y/D) tha t
KRONECKER FACTORIZATIO N FORMULA .

L
(22.57) L K(s,XDltD2) = (*,XD
1 )L(S,XD2).

For th e trivial characte r w e have th e Dedekind zet a functio n

00

(K(S) = Y,( Na)~S = C ( s ) ^ XD) = £ r(n, XD )n-°.

1a

In thi s case , (22.45 ) become s

(22.58) A K(s) = -J2E*(z a,s).

a

Comparing th e residues a t s = 1 on both side s w e obtain th e celebrated

DIRICHLET CLAS S NUMBE R FORMULA .

(22.59) L( ,XD) = ^m.

W\/\D\
514 22 . IMAGINAR Y Q U A D R A T I C F I E L D S

EXERCISE 1 . Prov e tha t fo r D < - 4,

(2-XD(2))h(D)= J2 XD(n) .
0<n<^

Another interestin g formul a come s fro m (22.45 ) a t s — \. B y th e Fourie r

expansion (22.46 ) w e see that E(z, \) = 0 and
oo

E'(z,\) = ,/ylogy + 4y/y^T(n)K0(2nny) cos(2>Knx).

l

Hence th e central valu e o f LK(S, X) * S give n by

L
*(§'X) = ^ r 1/ 4 I > ( a ) £ ' ( M ) -
a

Inserting th e Fourier expansio n (22.61 ) w e arrive a t

1av

Using Ko(y) < C y~*e~y on e derives b y trivial estimatio n tha t

av

This is interesting fo r the trivial clas s character xo- Assumin g L(^,XD) ^ 0 (whic h
follows fro m th e Rieman n Hypothesi s fo r L(S,XD), bu t ma y conceivabl y b e es-
tablished sometim e withou t recours e t o th e GRH ) th e lef t sid e i s LK{\,XO) —
C,{\)L\,XD) ^ 0 so we derive
D
D\\K V\ \
MD)ȣ(^p)'iog
where a runs ove r th e complete syste m o f primitive, reduce d ideals . Not e tha t w e
dropped th e facto r 2 in th e logarith m (fin d ou t why this i s permissible!) . Recal l
that a = Na < y/\D\/3 s o al l term s abov e ar e positive . Thi s implie s tha t h(D)
cannot b e very small , namel y

h(D)^>\D\*log\D\,

where th e implie d constan t i s effectivel y computable . Moreover , thi s implie s tha t

a positiv e proportio n o f the primitive , reduce d ideal s hav e th e nor m Na x \/\D\.
Duke [Du2 ] has show n unconditionall y tha t th e point s z a ar e equidistribute d i n
the fundamenta l domai n F wit h respec t t o the hyperbolic measur e d/j, = y~ 2dxdy.
His resul t i s ineffective becaus e i t use s Siegel' s boun d h(D) » \D\2~ S (se e (5.76)).
The Clas s Numbe r Formul a (22.59 ) reveal s th e algebraic natur e o f the specia l
value L(l , XD)- On e can derive infinitel y man y formula s o f that kin d b y comparin g
 22. IMAGINAR Y Q U1
ADRATI C FIELD S 5 5

coefficients i n the Taylo r expansio n a t s — 1 on bot h side s of (22.45) . O f particula r

interest ar e th e constan t terms . O n th e lef t sid e w e hav e

(22.61) A K(s,x)~^-LK(l,x)

if x is non-trivial , an d

(22.62) A K(s,X) ~ - ( ^- + l o g ^ l +
1 ^( , XD))

if x is trivia l b y th e followin g expansions :

r ( s ) = i - ( * - i )7 + "-

C(«) = —7 + 7 + ---
s—1
L(s, X D) = L(l , XD) + (S- 1)1/(1, XD) + • • •
and th e Clas s Numbe r Formul a (22.59) . O n th e righ t sid e o f (22.45) w e nee d a n
expansion o f E*(z,s) whic h w e shal l deriv e fro m th e Fourie r serie s (22.46) . Firs t
notice tha t 20(s)y s ~ — s- 1 + 7 — log47ry a s s—» 0 by th e followin g expansions :
(y-Y = i +siogy+...
\7T/7 T

r(s) = - -7 + • • •
S

C(25) = - - -5 l o g 2 7 r + .--

using th e value s C(0 ) - - § and C'(0 ) = -| log27r. Moreover , 20( 1 - s ) y 1 - 2 - §

by usin g th e valu e £(2 ) = ^-. Henc e

(22.63) 2£*(z , s) - - 1 - + 7 - l o g 4 ^ + \y + 2D(s)

5—1 O

where D(z) i s the tai l i n th e Fourie r expansio n (22.46 ) a t s = 1 , precisel y

00

D(z) = 4 v /^^r 1 /2(n)K 1 / 2 (27rny)cos(27rnx).

1
y
Since Ki(y) — {^I2y)^e~ an d Ti/ 2 (n) = <j_i(n)n2, thi s become s

(22.64) £>(z ) = 2Re ^ a _ i ( n ) e ( n ^ ) .

1

If w e execute th e summatio n a s follow s

00 0 0 0 0
_ 1
D(z) = 2Re ^ ^ m e ( m ^ ) = - 2 ^ l og | 1 - e{nz)\
11 1

we find tha t
(22.65) D(z) = ~y-2\og\ V(z)\,
516 22 . IMAGINAR Y QUADRATI C FIELD S

where rj(z) i s th e Dedekin d et a functio n

oo

(22.66) V(z) = e(j i)'[[(l-e(nz)).

The et a functio n i s a modula r for m o f weigh t \ o n th e grou p T = SL 2 (Z) wit h a

suitable multiplie r syste m involvin g Dedekin d sum s (se e e.g . [1 4]) . The n 77(z) 24 =
A(z) i s th e famou s Ramanuja n A function , th e primitiv e cus p for m o f weigh t 1 2 .
Therefore th e functio n
2
(22.67) F(z)=(-%L^\r,(z)\

is T-invarian t (her e th e facto r (2/y^DJ) 1 / 2 i s introduced fo r normalizatio n s o tha t

F(za) = a~ 2 \rj(z a)\2). I n thi s notatio n (22.65 ) become s

(22.68) D(z) = ~y - \ lo g ^ - logF(z) .

Inserting (22.68 ) int o (22.63 ) w e obtai n

l
(22.69) 2E*(z,s) —- + 7 - \og2ir^J\D\ - 2\ogF(z).
5—1
Next insertin g (22.69 ) int o (22.45 ) w e obtai n

(22.70) A x (s,x)^^(x)^(^-+7-log27rV^l)--Vx(a)logF(za).
w \s — 1 / w ^—^
0
ls
Now matchin g (22.70 ) wit h (22.61 ) , o r (22.62 ) i n cas e x trivial , w e deriv e a n
exact equatio n whic h i s calle d th e
KRONECKER LIMI T FORMULA . If x is a non-trivial class group character, then

(22.71) L * ( l , x) = -^£ x (a)logF(zn).

W
V \ D\ a
Moreover,

(22.72) ~(l, XD)=log\D\-'y+^logF{za).

a

Consider (22.71 ) fo r a non-trivia l genu s characte r \ — XB,C which correspond s

to th e factorizatio n D = BC int o non-trivia l fundamenta l discriminant s B, C. On e
of thes e i s negativ e an d th e othe r i s positive , sa y B < 0 < C. B y th e Kronecke r
formula (22.57 ) w e have 1/^(1 , XB,C) = L(l, x#)L(l , \c) a n o - b y the Dirichle t Clas s
Number Formul a (22.59 ) w e hav e L ( 1 , X B ) = 27rh(B)/w(B)y/\B\. Dirichle t als o
established th e clas s number formul a fo r positiv e discriminant s (se e (2.31 )) , namel y

(22.73) L ( l , xc) = 2h(C)\C\~i loge(C )

where h(C) i s th e clas s numbe r o f th e rea l quadrati c fiel d Q(VC) an d e(C) i s th e
fundamental unit . Combinin g thes e formula s w e infe r fro m (22.71 ) tha t
(22.74) £ic) 2h(B)HC)/w(B) = -Q F(a) - x(»)
 22. IMAGINAR Y Q U A D R A T I C F I E L D S 517

Besides thei r significanc e fo r algebrai c number s th e Kronecke r limi t formula s

can b e use d t o estimate derivative s o f L- functions. Fo r thi s purpos e i t is more
convenient t o operat e wit h D(z) rathe r tha n F(z), s o we retur n t o (22.68 ) gettin g
by (22.72 ) an d (22.59 )

(22.75) I U x D , . | j : ( i+ _ ^I ( 7+ l0 ,1 £t+1 D(^))

where a runs ove r inequivalen t primitiv e ideal s an d a = Na. I t is no t easil y see n

that th e right sid e of (22.75 ) doe s not depen d o n the choic e of representatives o f the
ideal classe s whil e thi s i s clea r i n (22.72) . A goo d choic e i n practic e i s th e syste m
of primitive , reduce d ideal s becaus e th e point s z a have th e larges t heigh t makin g
D(za) eas y t o estimate . Estimatin g triviall y D(z a) < C 1 we obtai n

where th e implie d constan t i s absolute .

wa s
REMARKS. A differen t formul a fo r 1 /(1 , XD) derive d b y S . Chowl a an d
A. Selber g [CS] , namel y
2
\D\iL'(l,XD) = -* E XD(n)logr(^ j) + ^( 7 + log27r).
W
0<n<\D\ ' '

22 A. Th e clas s numbe r problems .

Estimation o f th e orde r o f the clas s grou p o f a numbe r field i s one o f th e mos t
intricate problem s i n arithmetic. I n particular, on e wants t o know which fields have
class numbe r on e sinc e i t means tha t th e rin g o f integer s o f suc h a field has th e
unique factorizatio n property . A smal l clas s numbe r (relativ e t o th e discriminant )
is no t a common feature . A s wit h man y storie s i n arithmeti c thi s on e start s wit h
Gauss. I n hi s Disquisitione s Arithmeticae , Gaus s i s seriousl y concerne d abou t th e
class numbe r h(D) o f th e imaginar y quadrati c field K = Q ( A / D ) .
GAUSS CONJECTURE . We have h{D) -» oo as -D -> oo .
Even i f thi s conjectur e i s prove d th e rea l issu e i n practic e i s
GAUSS CLAS S NUMBE R PROBLEM . Find an effective algorithm for determining
all imaginary quadratic fields K = Q(\/D) with a given class number h = h(D).

Gauss kne w tha t h(D) = 1 for th e followin g nin e discriminant s

(22.77) -D = 3,4 , 7,8,1 1 ,1 9,43,67,1 63 .
Here i s a remarkable characterizatio n o f thes e discriminant s
PROPOSITION 22. 1 (Rabinovitch , 1 91 3) . For D = l(mod 4 ) we have h(D) = 1
if and only if the polynomial
f(X) = x*-x + i^

represents primes for all natural numbers x < ^ ^ .

The questio n becomes :
518 22. IMAGINAR Y QUADRATI C FIELD S

GAUSS CLAS S NUMBE R O N E PROBLEM . Prove that the above list of nine neg-
ative discriminants D with h(D) — 1 is complete.

Recall tha t h(D) i s linke d t o L ( 1 , X D ) b y th e Dirichle t formul a (22.59 ) bu t

it doe s no t hel p much . I n fac t thi s formul a serve d Dirichle t t o estimat e L ( 1 , X D )
rather tha n h(D) givin g
2-7T
(22.78) L( , XD)>
wJ\D\
since h(D) ^ 1 . Th e clas s numbe r on e proble m boil s dow n t o improvin g th e lowe r
bound (22.78 ) fo r \D\ > 163.
Next, recal l tha t th e effectiv e lowe r boun d

(22.79) ft(D)»|£>|ilog|D|,

follows fro m th e formul a (22.60 ) subjec t t o th e hypothesi s tha t L ( | , X D ) ^ 0 - Thi s

can solv e th e CNO P assumin g GR H (i f th e constan t ar e mad e explicit) .
There ar e muc h stronge r connection s betwee n h(D) an d zero s o f L ( S , X D ) > fo r
example
PROPOSITION 22. 2 (Hecke-Landau , 1 91 8) . If L{S,\D) does not vanish in the
region s > 1 — a/ lo g \D\, then

(22.80) h(D) > b ^ .

where a, b are some positive constants, effectively computable.

In moder n time s th e abov e resul t i s routine , an d on e ca n pus h thi s implica -
tion muc h deepe r whe n powere d b y ne w technologie s (Linnik' s densit y theore m
and Turan' s power-sum s method) . However , th e followin g implicatio n wa s quit e
surprising i n it s time .
PROPOSITION 22. 3 (Deuring , 1 933) . If the Riemann Hypothesis for the zeta
function ("(5 ) is false, then h(D) > 1 for all sufficiently large \D\.
A year late r Mordel l improve d Deuring' s resul t t o th e effec t tha t h(D) - > o o a s
—D—» 0 0 unde r th e sam e assumption . I n th e sam e tim e Heilbron n expande d th e
assumption t o th e falsit y o f th e Rieman n hypothesi s fo r an y Dirichle t L-series .
PROPOSITION 22. 4 (Heilbronn , 1 934) . / / GRH is false, then h(D) -> • 0 0 as
-D - > 00 .

Combining bot h Propositio n 22. 2 an d 22. 4 on e complete s a n unconditiona l

solution o f th e Gaus s Conjectur e tha t h(D) — > 0 0 a s — D— > 00 . O f course , th e
result i s not effectiv e becaus e w e do no t kno w whic h on e o f th e tw o proposition s i s
needed. Th e proble m i s that i f som e L(s , x) ha s a zer o of f th e lin e R e (s) = | th e
resulting lowe r boun d fo r h(D) depend s o n thi s hypothetica l poin t whic h ha s n o
numerical value . Therefor e th e CNO P stil l canno t b e reduce d t o a finite numbe r o f
computations. Som e degree o f effectivization wa s achieve d i n the followin g fashion .
 22. IMAGINAR Y Q U A D R A T I C F I E L D S 519

PROPOSITION 22. 5 (Heilbronn an d Linfoot, 1 934) . There is at most one imag-

inary quadratic field K = Q(V~D) with h = h(D) = 1 which is not in the list of
Gauss.
For th e above achievement s som e attentio n shoul d b e given t o E. Landa u fo r
his earl y idea s i n [La2] . Promptl y afte r Heilbron n an d Linfoot , h e wa s able t o
prove th e following remarkabl e estimate .
PROPOSITION 22. 6 (Landau , 1 935) . There exists an absolute, effective con-
stant c > 0 such that for every h ^ 1 all but one of the discriminants D < 0 with
h(D) = h satisfy
(22.81) \D\ ^c/i 8 (log3/i) 6 .

Next C . L. Siegel [Siel ] use d simila r argument s t o establish hi s famou s

THEOREM 22. 7 (Siegel, 1 935) . For any e > 0 there exists a constant c(e) > 0
(not computable) such that
€
(22.82) L( , XD)>C(S)\D\- .

Hence
2 £
(22.83) h(D) > c(e)\D\^ -.

Both paper s o f Landau an d Siegel appeare d i n the first volum e o f Acta Arith -
metica, 1 935 . A simple proo f o f Siegel's theore m wa s given b y D. Goldfeld [Gol ]
(see Chapte r 5) . The next interestin g resul t i s
THEOREM 22. 8 (Tatuzawa , 1 951 ) . SiegeVs bound is effective except for one
possible D for each e > 0. Precisely, if 0 < e ^ ^ and slog |D | ^ 1 , then with at
most one exception
£
(22.84) L( , XD)>IS\D\- .

In 1 98 0 J. Hoffstein [Hof ] established furthe r numerica l improvement s o f Tatu-

zawa's estimat e usin g Goldfeld' s technique .
The CNO P wa s solved fo r the first tim e b y Kurt Heegne r [Hee ] in 1 952 , but
his arguments (modula r forms , comple x multiplication) wer e understood an d recog-
nized onl y afte r hi s death i n 1968. A n additional not e of tragedy wa s that thi s firs t
recognition cam e tw o years afte r anothe r solutio n wa s found b y A. Baker. Baker' s
method [B ] is very different : i t use s a n effectiv e lowe r boun d fo r linea r form s of
three logarithm s o f algebrai c numbers . I n th e meantime , H . Star k [Stl ] gav e a
third solutio n whic h turne d ou t to be similar t o Heegner's, The n Star k looke d int o
Baker's proof an d realize d tha t i t sufficed t o use linear forms onl y in two logarithms,
so that th e problem coul d hav e bee n solve d alread y i n 1949 by Gelfand an d Linnik.
Retrospectively, i t seem s tha t H . Webe r wa s capable o f solvin g th e CNO P lon g
before. Se e [Cox] fo r a presentation o f the Heegner-Stark arguments .
Next wa s the problem o f class number two . Thi s has been solve d independentl y
in 1 97 1 by A. Baker [B ] and H. Stark [St2] . The y establishe d tha t ther e are exactly
eighteen negativ e discriminant s D with h(D) = 2 , namely
-D = 1 5,20,24,35,40,51 ,52,88,91 ,1 1 5,1 23,1 48,1 87,232,235,267,403,427 .
52 0 22. IMAGINAR Y QUADRATI C FIELD S

Our stor y culminate s a t tw o remarkabl e achievement s b y Goldfel d (1 976 ) an d

Gross an d Zagie r (1 983) . Goldfel d [Go2 ] gav e a n effectiv e lowe r boun d fo r h(D),
the qualit y o f whic h depend s o n th e ran k o f a n auxiliar y ellipti c curve , an d it s
satisfying th e Birc h an d Swinnerton-Dye r Conjecture . Gros s an d Zagie r ([GZ1 ] ,
[GZ2]) produce d a curv e wit h th e require d properties . Th e combine d result s o f
Goldfeld, Gros s an d Zagie r solve s th e Clas s Numbe r Proble m fo r an y fixed /i , u p
to a finite amoun t o f computation . J . Oesterl e [Oe ] reduce d th e implie d constant s
in Goldfeld' s wor k s o muc h a s t o mak e th e resul t practica l fo r computers . H e
succeeded i n showin g th e followin g nea t estimate :

(22.85) /l(jD)> _L(log|I?|)ri(l_M).

p\D

We shal l presen t Goldfeld' s approac h i n Sectio n 22. 7 wit h som e variation s o f ou r

own, an d w e survey th e constructio n o f Gross an d Zagier i n the Appendi x t o Chap -
ter 23 .
The bes t effectiv e lowe r bound s fo r h(D) whic h curren t technolog y allow s u s
to hop e fo r ar e o f typ e h(D) > c g(log\D\)9 fo r an y g > 0 an d som e computabl e
constant c g > 0, ye t i t i s too shor t t o solv e som e othe r popula r problem s suc h a s
E U L E R IDONEA L NUMBE R PROBLEM . Find all discriminants D for which the
class group ofQ(y/D) has one class in each genus.
If D i s an idonea l discriminant , the n h(D) = 2 t _ 1 wher e t = w(\D\) i s the num -
ber of distinct prim e divisors of D. Sinc e ^(|D|) ca n be as large as log \D\/ lo g log \D\
the proble m require s a n effectiv e lowe r boun d
(22.86) h(D) > |£>|c/iogiog|£> | wit hc > log2 .

We kno w b y Landau' s estimat e (22.81 ) tha t ther e ar e onl y finitely man y idonea l
discriminants bu t w e canno t lis t the m al l becaus e o f ineffectiv e constants .
22.5. Splittin g prime s i n Q(y/D).
If the clas s number h = h(D) i s small, the n ther e ar e only fe w prim e ideal s p of
degree one with smal l norm. Indeed , i f p = p p with (p,p ) = 1 , then p h i s a principa l
ideal generate d b y ^(m + ny/D) wit h n =fi 0 , whence p h — \{m2 — n2D) ^ A wher e

(22.87) A = M .

Therefore th e leas t prim e p\ — pi(D) wit h XD(J>I) — 1 satisfie s

(22.88) p i ^ VA.

Hence X£>( n) agree s with /x(n ) o n al l squarefree number s n ^ \/ A wit h (n , A) = 1 .

This propert y i s no t likel y t o hol d i n lon g segment s (becaus e XD i s periodi c whil e
\i i s not) , therefor e (22.88 ) suggest s tha t h i s rather large .
In thi s sectio n w e establis h severa l estimate s whic h hol d fo r an y D < — 4 bu t
are interesting onl y when h is relatively small . Le t po (p) be the number o f solution s
to
2
(22.89) b = D(mo d 4a )
 22. IMAGINAR Y Q U A D R A T I C F I E L D S 52 1

in fr(mod 2a). This i s the multiplicative functio n wit h po(p a) = 1 + XD(P) ifp\D,
PD{P) — 1 if P I D an d PD{p a) = 0 if p \ D,a > 1. Thu s th e generating Dirichle t
series fo r PD{°) n a s th e Euler produc t

a(-)=E^)°-=n(i+i) n (^iX 1 -^)"

Since p # (&) i s the number o f primitive ideal s o f norm a we have

CK(S) = C(2sr 1 CK(s) = C(2S)- 1 C(S)L(S,XD)

whence
(22.90) p D(a) = ^ XD(6)/X(C) .
6c21 a

First w e show th e following basi c inequalitie s

(22.91) £ ) / M a ) ^ ^] C M<0 -
a^>/A a ^ i / | D | / 3

This follows because the class number ft is equal to the number o f primitive, reduce d
ideals a = [ a , ^ ~ ] wit h b satisfying (22.89) . Fo r a ^ y/K th e points

<**) *->-*£•
have heigh t Im2; a ^ 1 , s o choosin g b with —a<b^a, thes e point s li e i n the
standard fundamenta l domai n of V = 51 ^(2 ) provin g the first inequalit y o f (22.91).
The secon d inequalit y follow s alon g th e same line s sinc e ever y reduce d idea l a has
a = Na ^ y/\D\/3.
Now suppos e p i < • • • < p r ar e the firs t r prime s whic h spli t completel y i n
K = Q(>/D). B y (22.91 ) w e have

Er ' < h,
where a i , . . . , ar ru n over non-negativ e integer s an d r' i s the number o f positiv e
exponents. Henc e i/ r(log A/2 log pr ) < ft where
27
i/P(a) = E "'-

Since u r(a) ^ ^ f , thi s implie s lo g A < (2ftr!) 1 / r (logp r ) « rh 1 ^ \ogp r i f r ^ 2),

i.e.,

(22.93) log* . > H *.

1
2
In particular , th e secon d prim e whic h split s completel y satisfie s p 2 ^ A ^.I f
ft ^ (logA)s , the n
(22.94) \{p ^ exp( x /loiA) : XD(P) = 1 } | < 2<?
522 22 . IMAGINAR Y QUADRATI C FIELD S

provided A is sufficiently larg e i n term s o f g, namel y lo g A > (4g) 4fir. Moreover ,

(22.91) yield s

X(P)=I
Hence fo r an y positiv e intege r r we obtai n
\{p^ A * :XD(P) = 1 }\ ^ r h i .

Choosing r = [| log 8ft] thi s give s

(22.95) \{p < A^ los8h : X(P) = 1 } | < log8ft .

Next w e exten d th e basi c inequalit y (22.91 ) t o large r ranges . T o thi s en d w e

use th e followin g elementar y estimat e (se e Propositio n 1 5.1 0) :

|{ 7 er 0 O \r:im 7 ^>y}|<i + y
which hold s fo r an y Y > 0 and z G H. Applyin g thi s fo r th e point s z a in the
standard fundamenta l domai n o f T — SL2 (Z) wit h Y — yfK/A w e deduce tha t
1(L4
(22.96) ] T p D(a)<fc(l +
a<A v A

for an y ^ 4 > 0. A s befor e thi s implie s

2
H(E ) <H 1 + 77x)-
X(P)=I
Hence fo r an y positiv e intege r r we obtai n

|{p < A : X D(p) = 1}| < rh^ + 5 r ^ ( —) r .

Choosing r = [| log 2ft] -f 1 we conclud e tha t

(22.97) |{ p < A : X D(p) = 1}| « ( 1 + A\D\~*'lo*2h) log2/ i

for an y A > 0, wher e th e implie d constan t i s absolute .
In Chapte r 2 3 we shall nee d th e abov e estimate s i n variou s forms , therefor e w e
end thi s sectio n b y derivin g th e require d results . Firs t restrictin g th e summatio n
in (22.96 ) t o squarefre e number s w e obtai n

(22.98) Y?T(a, XD)^h(l + ^ ).

Hence fo r an y positiv e A, B w e derive b y partia l summatio n

l 2
(22.99) Y! r(a, XD)a- / «h^ + ^y /2
B<a^A

1
(22.100) Y? T&XD)^ « h(± + ^ ).
 22. IMAGINAR Y QUADRATI C FIELD S 52 3

Next w e ar e goin g t o estimat e th e su m

b
(22.101) S(A,B)= 1
Y? 2
r{a)r{a,XD)a- 1

Ka^A
(a,C)=l

where C stand s fo r th e produc t o f prime s i n D an d al l th e prime s p ^ B wit h

XD(P) = 1 .
We hav e
T
S(A,B) < Y^Y? ^XD)T(.a2,XD){aia2rll2
l<a\a-2^A
(aia2,C) = l

1 /2
< 2 Y? ^Xz))a- + 2^ b^T(a1 ,XD)r(a2,xz))(aia2)-1 /2
.
B<a^A aia 2^A
a1,a2>B
Applying (22.99 ) an d (22.1 00 ) w e ge t

B<a^A v
2
, / l i4\V 2h / \ A\W , 2 / l logi4\/A\i/ 2
<</I
U +
A) +
7 I ( B + A ) +h (B + ^K)W •
Rearranging term s thi s boun d simplifie s t o

(22.102) S(A, B) < h(l + - - = ) ( ^ + ^ ) V 2 + ^ A ^ v ^ l o g A

IB*
where th e implie d constan t i s absolute , an d effective .

22.6. Estimation s fo r derivative s L^(1 ,XD) «

If th e clas s numbe r o f th e imaginar y fiel d K — Q(y/D) i s unusuall y small , sa y

(22.103) ^ ^ ( ^ H ) ,

then s o i s the valu e o f th e Dirichle t serie s L(S,XD) a t 5 = 1 , namel y

(22.104) L (
1 I X D ) = 0 ( _ J _ )

by virtu e o f th e Clas s Numbe r Formul a (22.59) . Thoug h ther e i s no direc t relatio n

between th e clas s numbe r an d derivative s o f L(S,XD) a t s = 1 all o f thes e value s
are als o affecte d b y th e improbabl e hypothesi s (22.1 03) .
Truncating th e Dirichle t serie s fo r L^(S^XD) w e ge t
!
(22.105) L ^ ( l , X D ) - ^ ^ ^ ( - l o g n ) / c + 0(|D|^- 1 aogx)fe+1 )
n
n^.x

by th e Pol y a-Vinogradov inequalit y (1 2.50 )

(22.106) Yl XD(n)<£\D\hog\D\.
 52 4 22. IMAGINAR Y QUADRATI C FIELD S

Hence b y absolut e summatio n

fc 1 fc+
(22.107) L< >(l,xz>)«(k>g|Z>|) .
It i s amazing tha t thi s trivia l boun d i s the bes t know n s o far (apart fro m som e
improvements o f th e implie d constant) . B y th e Rieman n Hypothesi s fo r L(S,\D)
one ca n sho w tha t
fc+1
(22.108) L^( ,
1 XD) « (loglog|Z?|) .
A sligh t improvemen t o n (22.1 07 ) ca n b e derive d fro m th e rathe r subtl e result s o f
Graham an d Ringros e [GRi ] fo r special D having onl y smal l prim e divisor s (se e
Chapter 1 2) . I n thi s sectio n w e give estimates whic h ar e bette r tha n (22.1 07 ) i f th e
class numbe r h(D) i s relatively small .
First w e examine 1 /(1 , XD)- W e know b y Landau's analyti c metho d (se e Chap -
ter 5 ) tha t L(s, XD) canno t hav e two real zero s (counte d wit h multiplicity ) clos e to
5 = 1 , therefore L(l , XD) being smal l shoul d impl y 1 /(1 , XD) i s not small . Thi s ca n
be show n quickl y b y elementar y arguments . T o thi s en d w e evaluat e th e su m

r - U_ V - XD[m)(^ x _ , „(m\\ , „(y/\D\

n
Mn,X»)=£^(l0g*
^—' m V m \x +7 + 0(™)) 0(^V,)
J JV + y J
m^.y

2
= L(l, XD)(logx + 7) + L'(l,XD) +
o (- + ^ log a:)
\x v /
for an y 2 < y ^ x , where the error terms above came from estimatio n o f the relevan t
character sum s wit h y < m ^ x b y means o f the Pol y a-Vinogradov inequality .
Hence w e deduc e th e approximat e formul a
1
(22.109) ^ n - r ( n , X D ) = L(l,XD)(logx + 7) + ^ ( l , X D ) + 0 ( | J D | ^ - h o g x ) .

Notice that o n the left sid e all terms are non-negative an d fo r n = dm2 wit h d \ D we
have T(TI,XD) ^ 1 - Summin g ove r thes e number s (wit h d squarefree t o ensur e th e
uniqueness o f th e representatio n n = dm 2) w e infer tha t (b y choosin g x = e~ 7D)

L(l,XD)\og\D\ + L\l,XD) > (^ +0( ]^-l)y(D),

where v(D) i s almost constant , precisel y

i+
(22.uo) ^)=n( ^)-
p\D V

Assuming (22.1 04 ) thi s yield s 1 /(1 , XD) > (^ + o(l))v(D).

Now we derive a precise asymptotic fo r 1 /(1 , XD) by appealin g to the Kronecke r
Limit Formul a (22.75 ) whic h allow s u s to take ful l advantag e o f the smal l clas s
number condition . On e coul d deriv e th e sam e thing s fro m (22.1 09 ) bu t w e choos e
the alternativ e pat h sinc e it works for an y class character an d reveal s other features .
First, estimatin g trivially , w e ge t

(22.111) L'(l , XD) = y * ( D) + 0 ( ^ == lo g \D\)

 22. IMAGINAR Y QUADRATI C FIELD S 52 5

where £(D) i s the su m ove r th e norm s o f primitive reduce d ideal s

(22.112) (D) = ^ V1.

a

We shal l approximat e 1 (D) b y the produc t

(22,13) i W - n H )n ( »+ ;)('-;)"•
p\D F p^\D\ F F

XD(P)=1

Clearly w e have th e uppe r boun d

1(D) < J2 PD(a)a~ l ^ P(D).

a^\D\

For a lower boun d w e apply Rankin' s tric k (recal l tha t A = |.D|/4) :

a a l s
e(D) > Yl Pv( )~ > CK(S) - £ PD(a)a~

where CK( S) ~ C(2s) _1 Cfi:(5) i s the zeta functio n o f K = Q(\/D ) (reduce d t o th e

primitive ideals )
-8
CK(s) = J2PD(a)a- s = H(l+p- s) [ J (I+P-^I-P ) -1 ,
a p\D XD(P)= 1

and s is any real numbe r > 1 to be chosen a t the end to optimize th e obtained
results. Applyin g (22.88 ) b y partial summatio n w e ge t
,_ Ush
E s

Here w e can replace l i s by 33 because i f s ^ 3 we have als o th e trivial boun d

3/s\/A derive d b y using PD(O) ^ a and h ^ 1 . The n w e estimate CK( 5 ) from
below b y the partial Eule r produc t restricte d t o primes p ^ \D\, say PS(D), an d
replace P S(D) by P(D) with admissibl e correction . Precisely , usin g th e inequalit y
wit n
f](l - x p) ^ 1 - J2 XP

^-'-('^)('-?)('^)"('-F)"
_2 1 -p 1 g 0/ „ ^plog p
J
" p + 1 l-ps ^ [
P21
-

we ge t P S(D)/P(D) > 1 - 2(s - l)rj{D) wher e r/(L> ) is given by

(22H4) , < C ) =£ £*? •

X£?(P)#-1

Hence £(D) > P(D){1 - 2(s - l)rj(D)} - 33h(D)/(s - 1 )>/A . Choosin g s close to 1 ,
namely 5 = 1 + (33h(D)/ri(D)P(D)y/\D\)^, w e derive th e unconditiona l formul a
2
(22.115) £(D) = P(D) - 0(rj(D)P(D)h(D)/^/\D\)^
where 0 < 6 < S V ^. Not e tha t rj(D) < C log \D\. Henc e w e obtai n
 526 22 . IMAGINAR Y QUADRATI C FIELD S

PROPOSITION 22.9 . Put e{D) = h(D)\D\~i lo g \D\. We have

(22.116) £(D) = P(D){1 + 0(e(D)i)}

where the implied constant is absolute and effective.

Inserting (22.1 1 6 ) int o (22.1 1 1 ) w e come u p wit h
PROPOSITION 22.1 0 . Suppose e(D) = h(D)\D\~i log\D\ ^ 1 . Then we have

(22.117) L'( ,
1 XD) = (^+0(e(D)i))p(D).

The produc t ove r splittin g prime s i n P(D) ca n b e reduce d considerabl y t o tha t

over a ver y fe w smal l prime s b y mean s o f (22.97) . Indee d (22.97 ) give s
4 /1
(22.118) Yl P' 1
< ( B~X + l ^ r °g2/llog|^|)log2/i.
B<p^\D\
XD(P) = 1

Applying thi s boun d wit h B — zlog2h w e deduc e fro m (22.1 1 7 ) th e followin g

PROPOSITION 22.1 1 . Suppose
(22.119) h^ l^i/iogiogiz ^
(22.120) 2^z^log\D\.

Then we have

1
(22.121) 1
L'( , XD) = ±«D) n ( ^ ) 1( +
J)" 1 ( 1 + ° @ )-
p^z log 2h
XD(P) = 1

where v{D) is defined by (22.1 1 0).

Now w e estimat e highe r derivatives . W e star t fro m th e su m

Sk(x) = J ^ X D M n - ^ l o g n ) *
n$Cx

which approximate s t o (— l) k L^ k \l,XD) wit h a n erro r ter m give n b y (22.1 05) . O n

the othe r hand , w e hav e

Sk(x) = £ ] > > ( n , XD )^(logmn)fc.

mn^x

By th e convergenc e o f ^ / i ( m ) m _ 1 ( l o g m ) ^ fo r 0 ^ £ ^ & , an d i n cas e o f £ = 0

using
^/i(m)m-1<(log22:)-1,

we obtai n
5 f c ( x ) « ^ ^ ^ l ( l o g 2 n ) ^ r i ^ 1 .
^-^v n log(2x/n) /
n^x '
 22. IMAGINAR Y Q U A D R A T I C F I E L D S 52 7

Estimating th e term s wit h y < n ^ x b y mean s o f (22.1 00 ) w e ge t

1
Sk(x) « J2 ^ ^ W n ) * - + -(logy)*" 1 + ^ = (log*)* .

Here w e estimat e log n b y logy , the n exten d th e summatio n t o n ^ D 2 an d appl y

(22.109) gettin g

Sk[x) « L / (l, X D)(logy) f c -1 + - ( l o g y ) ^ 1 + -^=(log*) fc .

yV A
We choos e x — D 2 an d y = 2h 2 t o arriv e a t
k
(22.122) L( \l,XD1 1
)<:\L'(l,XD)\(log2h) k
- +h\D\- H\og\D\)k.
Combining (22.1 22 ) an d (22.1 1 7 ) w e obtai n
PROPOSITION 22.1 2 . Suppose h(D)log\D\ < \f\D\ , then for any k > 1 ,
1k
(22.123) L^( ,
1 XD) « P(D)(\og2h(D)) -
where P(D) is the product given by (22.1 1 3) and the implied constant depends only
on k.
REMARKS. Notic e tha t (22.1 22 ) implie s th e trivia l boun d (22.1 07) , an d i t be -
comes stronge r whe n th e clas s numbe r i s relativel y small . I n th e derivatio n o f
(22.122) (whic h i s unconditional ) w e mad e a n appea l t o th e Prim e Numbe r Theo -
rem i n th e for m
(22.124) \ r ^ M ( i o g m ) *«i.
m

One ca n avoi d thes e no t s o simpl e result s altogethe r b y examinin g a specia l com -

bination o f th e truncate d derivative s Si(x) wit h 0 ^ £ < k.
EXERCISE 2 . Sho w tha t fo r x ^ 2 and a primitiv e characte r %(mo d |D|) ,
2
2 V r ( n ) x ) ^ ^ = (log x-7l)L(l,x)-27L,(l,x)-i"(l,x)+0(|JD|^-hog2x).
^—' n

Derive fro m thi s tha t

d\D

provided L(1 ,XD) < (lo g |Z)|) —2 an d \D\ i s sufficientl y large .

This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/24

CHAPTER 2 3

EFFECTIVE BOUND S FO R
THE CLAS S N U M B E R

We have already seen how the assertion of h(D) bein g small relatively to the dis-
criminant D lead s to simple, yet fake, approximations t o the special values L(^,XD)
and L ^ ^ ( 1 , X D ) - I n thi s sectio n w e conside r centra l value s o f derivative s o f certai n
L-functions associate d wit h automorphi c form s twiste d b y th e characte r XD- I n
this cas e th e leve l o f th e twiste d for m i s muc h large r tha n th e conducto r o f th e
character (i t i s abou t Z) 2), s o w e shal l b e abl e t o observ e positiv e effect s onl y fo r
derivatives o f orde r g > 2 and onl y whe n th e clas s numbe r i s extremel y smal l (es -
sentially h(D) < C (lo g |D|)2 _ 1 ). A s i n th e previou s case s ou r result s ar e effectiv e
but w e d o no t dwel l o n computin g al l th e implie d constants . Ou r motivatio n i s
to illustrat e th e masterwor k o f D . Goldfel d [Go2 ] whic h afte r Gros s an d Zagie r
[GZ1] ha s application s t o giv e effectiv e lowe r bound s fo r h(D). Althoug h fo r thi s
application Goldfel d (an d J . Oesterl e [Oe] ) considere d onl y th e L- function o f a n
elliptic curve , w e tak e an y primitiv e cus p for m o f a fixed leve l an d a centra l char -
acter (possibl y th e trivia l centra l character) . A t th e en d w e derive the lowe r boun d
for h(D) i n question , an d w e make suggestions fo r furthe r searc h o f the L- functions
which ca n b e employe d fo r estimatin g th e clas s number .
23.1. Landau' s plo t o f automorphi c L- functions.
Throughout, / i s a primitive cusp form o f weight k > 1 and characte r £(mo d N)
on th e grou p T 0(N) suc h tha t e(—l) = (— l) k . Therefor e (se e Chapte r 1 4 ) / ha s
the Fourie r expansio n
oo

f(z) = ^2\(n)n—e(nz)
l
with coefficient s X(n) whic h ar e th e eigenvalue s o f Heck e operator s T n fo r al l n .
Notice w e normalize d thes e s o that th e associate d L-functio n
oo

£(*,/) =£A(n)n- s
1

has a n Eule r produc t o f typ e

(23.1) L(s, /) = JJ(1 - \(p)p-° + e(p)p- 2s)-1

v
and th e complet e L-functio n

A(sJ)=^yT(s+^)L(s,f)
529
530 23. E F F E C T I V E B O U N D S F O R T HE CLASS N U M B E R

(which i s entire) satisfie s th e functional equatio n

(23.2) A(s,f) = w(f)A(l-sJ).
Here w(f) i s a complex numbe r dependin g o n / wit h \w(f)\ — 1 and / i s the cusp
form whos e Fourie r coefficient s ar e A(n) .
Let x b e the character associate d wit h th e field K = Q(y/D), tha t i s th e
Kronecker symbo l x(n ) = ( —); i t is a real, primitiv e characte r o f conductor \D\.
We creat e th e for m / 0 x b y twistin g th e coefficients wit h x,
oo

The twiste d for m / ® \ i s a cus p for m o f weight fc,character e and leve l ND 2, but
it i s not always primitive . However , ther e exist s a unique primitiv e cus p for m
oo
Z
fx( ) = J2 A x W n ^ e ( n 2 )
1

of leve l M\ND 2 an d character e x suc h tha t e x(n) — e(n) and Ax (n) = A(n)%(n )
for al l (n,ND) = 1. Therefore th e associated L-functio n

oo

1
has the Eule r produc t
23 1
Hs,fx)= I I (.l-KWp-'+extop- )-
p\ND

(23.3) [ ] ( 1 - X( P)x(p)p-s + E W P "2 * ) - 1.

The complet e L-functio n

A(s,/x) = ( ^r ) r ( s + ^i)L( S)/x)

is entir e an d it satisfies th e functional equatio n
(23.4) A(s,f x) = w(f x)A(l-s,Jx)
with \w(f x)\ = 1 .
Given / an d \ w e conside r th e produc t L-functio n (i n Landau's style )
oo

(23.5) L(s) = L(sJ)L(sJx) = Y,W S'

l

Note tha t L(s) has an Euler produc t o f degree four .

Now w e hav e al l the information abou t L(s, f) an d L(s, fx) tha t i s needed to
work wit h L(s). Th e complet e produc t L-functio n define d by

A(s) = QsT2(s + ^ ) L (5 , /)L(s , /x ) with Q = y/MN/Air 2

 23. E F F E C T I V E B O U N D S F O R T H E CLAS S N U M B E R 531

satisfies th e functiona l equatio n

(23.6) A(s ) = wA(l - s)
with th e roo t numbe r w = w(f)w(f x). Ou r first ai m i s t o evaluat e asymptoticall y
the derivativ e A^ 9\s) a t s — \ o f order g ^ 0 . W e make n o further hypothesi s unti l
the tim e come s t o examin e th e mai n term s i n ou r formula .
23.2. A partitio n o f A ^ ) ( \ ) .
We begi n b y expressin g A ^ ( ^ ) a s a su m o f tw o rapidl y convergin g conjugat e
series i n coefficient s o f L(s), simila r t o th e formul a (5.1 2) . T o this en d w e comput e
the contou r integra l
1=-^-. [ Ais+^s-^ds
2W(i)2
in tw o ways . Firs t movin g t o th e lin e R e (s) = — 1 we pas s a pol e o f orde r g + 1
at s — \ wit h residu e A ^ ) ( | ) . The n w e observ e tha t th e ne w integra l o n th e
line R e (s) = — 1 is equal t o — (—l)9wl b y th e functiona l equatio n (23.6) , therefor e
we obtai n A ^ ( | ) — I + (—l) 9wL Her e / denote s th e comple x conjugat e o f / .
Next w e comput e / b y integratin g termwis e th e Dirichle t serie s (23.5) . W e obtai n
/ = Q 1 /2S, s o
1 /2
(23.7) Q - A^(|) = 5+(-l)^ 5
where

,2,s, '-t^v®
and V(y) i s the invers e Melli n transfor m o f g\T 2(s + f ) s ~ ^ - 1 ,

s 2 9 l
(23.9) V(y) = ^~J y~ r (s + ±)s- - ds.

Note tha t (mov e th e integratio n t o R e (s) = | )

(23.10) V(y)= Yl c,(log^'+0(^log§ )

if 0 < y ^ 1 with th e leadin g coefficien t c g = T 2 ( | ) . Moreover , V(y) < y k 2

e~ ^
if y ^ 1 (mov e th e integratio n t o R e (s) = y/y an d estimat e b y Stirling' s formul a
(5.113)). Combinin g bot h case s w e deduce th e boun d fo r th e cut-of f functio n
1
(23.11) F ( 2 / ) < e - v ^ i o g ^ ( l - f 2/- )
which hold s fo r al l y > 0.
By virtue o f (23.1 1 ) th e serie s (23.8 ) begin s t o deca y exponentiall y wit h n ^ > Q
so w e ar e lef t essentiall y wit h Q terms , whic h i s stil l a lot . However , i f th e clas s
number h i s ver y small , the n man y coefficient s a(n) vanish . Unde r thi s fictitious
condition a mai n ter m fo r S wil l emerg e fro m a lacunar y subsequenc e o f th e coef -
ficients (essentiall y fro m a(m 2)). Thi s mai n ter m i s proportiona l t o th e valu e a t
s = 1 o f th e symmetri c squar e L-functio n attache d t o / (whic h doe s no t vanish ,
see Theore m 5.4 4 an d (5.1 01 )) .
53 2 23. EFFECTIV E BOUND S FO R TH E CLAS S NUMBER

Before estimatin g S we pul l ou t al l th e ramifie d place s an d a few smal l one s

which spli t i n the fiel d K = Q(y/D). Precisely , le t C be th e produc t o f primes in
ND an d al l prime s p ^ B with x(p) = 1 - W e shal l choos e B later, i t will b e quit e
small bu t fo r no w w e assum e onl y B ^ |D| . W e writ e

c|C°° V (m,C ) = lV ^

By th e Eule r produc t (23.1 ) th e coefficient s \{n) satisf y

2
(23.13) A(m)A(n ) = ] T s(d)X(mnd~ ).
d\(m,n)

Hence fo r (m , ND) = 1 we hav e

a(rn)= ^ ^(d)r(n,x)A(n )
d 2 n=ra

where ^ = £%. Accordingl y S is partitioned int o

a c 2
( ) V ^ ^(d) \~ ^ / .X(n)/cd n^
S=
c|C°° V (d,C ) = l (n,C ) = lV ^

We expec t tha t th e mai n contributio n t o 5 comes fro m n = m2 and wit h r(ra 2 , \)

deleted, whic h is
a c 2 2
() V ^ ^(d) ^ \(m ) /cd r"2y
Sl =
fe^.Xf .-*- X. —n-Q->
cjC 00

The remainin g term s yiel d

.= Sf.E.^E«^-i)M^) I
c|C~ v ~ (d,C)= l (m,C)= l

V'(d) v ^ ,A(n ) /cd 2 n

v v
„ Poo v
c|C°° A r*\ — -\
" (d,C)= l" ^ r*<\ — l'\
(n,C)= ^

Thus w e hav e 5 = Si + S2 -f 53. W e shal l trea t eac h o f these thre e sum s sepa -
rately. Th e firs t su m S i will b e evaluate d asymptoticall y b y a n appea l t o analytic
properties o f the symmetri c squar e L- function attache d t o the cus p for m / , an d
the othe r tw o sum s will be estimate d b y mean s of (22.86) an d (22.88 ) respectively .
We estimate S 3 essentially b y th e clas s numbe r whil e S 2 turns ou t t o b e negligible .
For th e coefficient s o f th e cus p for m / w e use Deligne' s boun d

(23.14) |A(n) | ^r(n).

This rathe r advance d resul t i s not essential , however , it is useful t o simplify expo -
sition.
 23. EFFECTIV E BOUND S FO R TH E CLAS S NUMBE R 53 3

23.3. Estimatio n o f £3 and 52-

We hav e

r(n,X)= J ] a 1( + x(jp)) I I ( l + a (l + x(p)))-

zot l a
p - \\n p' \\n

Hence, writin g n = am , where a is squarefree w e see that r(n, \) ^ T (a, x ) r ( m 2 ) -

2

Moreover, b y Deligne's boun d |A(n) | < r{n) ^ r{a)r(m 2). Therefor e

cIC00 vV
c|C°° dd m
ma>
a>
v
ll ^
(a,C) = l

Applying (22.88 ) an d (23.11) w e deduce tha t

1 0
S 3 « h(l + - ^ ) { ^ c ( l ) + i/c(!)B-*(logA)'+ }

where

c\C°°
fe fc
We hav e a ( / ) = £* A(p )A x (/- ), whenc e fo r any 5 > 0 by (23.14)
00 2

Ms) < n(E (p')p"' ) = ii(i -^ s)-4=fc(*),

r a
P|C 0 p| C

say. Fo r VD{1) we may have a more precis e estimat e

(23.15) MI ) ~ r K1+ — ) « ^ )
where v{D) is defined b y (22.110).
To estimate $ 2 first notic e tha t r(ra 2 , x) = 1 unless m has a prime diviso r wit h
x{p) = 1 , indeed i t is clear fro m th e formul a

r(m2,X)= IJ( l + a(l+X(p))).

p Q ||m

Therefore b y (23.11) an d Deligne's boun d

S2«M§)(logA)*+1 ] T ^^(r(m 2
,X)-l)e-™/^
(m,C) = l

mp/y/Q

m p\C
X(P)=I

Applying (22.86 ) w e infer tha t

l w
(23.16) S 2 « h(l + -^=y c(k)B- (logAy+

We simplif y bot h result s b y taking B in the following rang e

2
(23.17) /i (logA) 2 s + 2 ° < B < A 1 / l o s 2 h .
534 23 . E F F E C T I V E B O U N D S F O R T H E CLAS S N U M B E R

For s — |, 1 we have UQ{S) ^ £,c( s) ^ ^D(S) < C hr(\D\) ^ h 2, whenc e we conclude

that

(23.18) S 2 + S 3 <^c(l).

23.4. Evaluatio n o f 5 i .
To thi s en d w e us e comple x integratio n an d analyti c propertie s o f th e serie s
which com e out . Insertin g (23.9 ) fo r V(cd 2m2/Q) int o (23.1 2 ) w e ge t

(23.19) s 1 = 4 L- f Q s 2
^ {sJr!l)Z{2s + l)s- 1
9
- ds
27TI J (1 )

where Z(s) i s the correspondin g Dirichle t serie s

Z(S)=(£a(c)c-*/2)(£ ^d)d~')( ^ A(m 2

)m-*).
c|C°° (d,C ) = l (m,C )=l

Recall tha t ip = e^ . W e write Z(s ) = L(s,i/>)M(s, f)P(s, f) wher e

P(s,/) = (£ a (c)c~S/2)(E M)d- ay1 ( E A(m 2)m-S)"\

c|C°° d|C° ° m\C°°

L(s,xp) i s the Dirichle t L-functio n fo r th e characte r if) — e\ an d

oo
A
(23.20) M(s, f) = J2 (m2)m_S-
l

Every Dirichle t serie s abov e ha s a n Eule r product . T o compute the m w e factor th e

Hecke polynomia l fo r L(s , /) int o
(23.21)
1 - X(p)p- S + e(p)p- 2s = ( 1 - a(p)p- s)(l - ^P")
with a(p) + fi(p) — X(p) an d a(p)j3(p) = e(p). Similarly , thi s hold s fo r th e loca l
factors o f L(s , fx) wit h a x (p) = a(p)x(p), P x(p) = f 3(p)x{p) i n plac e o f a(p), /3{p).
This yield s

(23.22) A(p' .)= ;

a— p
Here an d hereafte r w e drop th e argumen t p i n th e root s a, ft, a x, (3 X for simplicity .
By (23.22 ) w e deriv e

£ Wfr-* = ~d - oVT1 - ^d - ft-)" 1

= (i - avr'fi+«/3p-s)(i - frr 1
- 1
=(1 - aV) ( l - aPp- syl{\ - p 2p~s)-\l - e 2
p-2s).
Hence th e C-par t i s equal t o

p(s, f)=n(i - ap-s'2ri{\ - pP-'2)-\i - axP-i2)-\i - n xP-s/2ri

p\C

rj(i - ^p-*)(i - ftoa+ep-T'a - ft-s)-

 23. E F F E C T I V E BOUND S F O R T H E CLAS S N U M B E R 53 5

After factorin g 1 — a 2p~s an d 1 — fi 2p~s w e arrange thi s produc t a s

(23.23)
^,/)=n( i+ ^" fl )" i ( i -^" fl )
p\C

H ( l + ap-*/2)(l + Pp~s^2)(l - a xp-l2)-\\ - P xp-

s/2 l
y.
P\c
2
Moreover, w e ge t M(s,f) = L(2s,e 2)~lL(s, Sym /) wher e
2
L(s, Sym /) = [ T ( l - a V 8 ) _ 1 ( l - a / J p ' T ^ l - J ^ P -8 -1
)
P

is the symmetri c squar e L-functio n attache d t o / ; se e Section 5.1 2 . Thi s L-functio n

appears a s a facto r i n th e Rankin-Selber g L-functio n
oo

(23.24) L ( s , / ® / ) = ]TA 2(n)n-s.

l

Indeed, b y (23.1 2 ) fo r m = n w e deduce tha t

2
(23.25) L(a, f ® /) = L(s, e)M(s, f) = J ^ LL ( 8 , Sym / ).

The //-functio n L(s,Sy m / ) admit s meromorphi c continuatio n t o th e whol e com -

plex s-plan e an d ha s a functiona l equation . A s prove d b y Shimur a [Sh2] , i t i s
holomorphic excep t possibl y fo r simpl e pole s a t s = 0 an d 8 = 1 .
The cases for whic h L(s , Sym2 /) ha s a pole at s = 1 are completely understoo d
(the dihedral forms ) an d ver y rare. W e shall assum e tha t ou r functio n L(s , Sym2 / )
has n o pol e a t s = 1 . B y (5.1 01 ) o r Theore m 5.44 , w e hav e
2
(23.26) L ( l , S y m /)^0.

Besides th e abov e arithmeti c propertie s o f th e zet a function s w e nee d crud e

estimates o n th e lin e R e (s) — ~. B y (23.23 ) w e ge t

(23.27) | P ( S ,/)|^c(|)«/>.
For the L-functio n o f the character I/J = e\ w e use the convexity boun d (se e Sectio n
5.9)

(23.28) L ( 5 , ^ ) < H | 2 3 | i .
Hence w e arriv e a t
4
(23.29) Z(s)<^ \s\ \D\*h
where th e implie d constan t depend s o n th e cus p for m / .
Having gathere d th e arithmeti c an d analyti c propertie s o f th e relevan t zet a
functions w e ar e no w read y t o evaluat e th e serie s Si . Movin g i n (23.1 9 ) t o th e lin e
Res = —1 / 4 w e obtai n
(23.30) S i =S 0 + O(h)
536 23 . E F F E C T I V E B O U N D S F O R T H E CLAS S N U M B E R

where So i s th e residu e a t s = 0, o r simpl y

d9
s 2
(23.31) S 0 = -—Q T (s + \)Z(2s + 1), a t s = 0.

REMARK. Th e error term i n (23.30 ) could be improved by factor | D | - 1 / / 1 6 usin g

Burgess' bound i n place of (23.28 ) bu t w e do not nee d anythin g better tha n (23.30) .

The dependenc e o f So o n th e characte r ip — e\ need s t o b e exposed . T o thi s

end w e se t

(23.32) R(s) = Q sT2{s + §)M(2s + 1, f)P(2s + 1, /)

and differentiat e L(2s + l,ip)R(s) b y Leibni z rul e gettin g

u 9
(23.33) S 0= Yl 2 ( u)L{u)(l^)Riv){0).
u-\-v=g

Next, whe n computin g th e derivative s R^(0) w e think o f Q s a s a dominant facto r

in R(s). Recallin g tha t al l prime s i n C are bounde d b y Q , w e deriv e b y repeate d
application o f Leibni z rul e an d th e produc t representatio n (23.23 ) tha t

(23.34) fl (*)(0) = J R(o)(logQr(l + o ( i ^ ) )

where

(23.35) t(C )= ^p-hogp.

P\c

This approximatio n woul d b e useles s i f the erro r ter m wa s no t small , bu t thank s

to (23.1 7 ) w e ge t b y (22.81 ) tha t t(C) = t(D) + <3(log2ft). W e hav e als o t(D) < C
u(\D\) < lo g 2ft, therefor e t(C) < log2ft .
Finally b y (23.30) , (23.34 ) an d (23.36 ) w e obtai n

(23.36) s 1 =;?(o) Y, 2 "(«)L(u)(i^)aogi^ir(i+o(^^))+o(/

u-\-v=g
log 2ft
l ).

23.5. A n asymptoti c formul a fo r A ^ ( | ) .

Adding (23.1 8 ) t o (23.37 ) w e obtai n

2 log 2ft
(23.37) S = R(0) J2 "(^) i(M) (l^)(log| J D|r(l + o ( ^ g f ) ) + 0 ( ^ c ( l ) )
u+v=g

where R(0) = r 2 ( f ) M ( l , / ) P ( l , / ) . Recal l tha t

(23.38) M (1, / ) = L(2, e2 ) " 1^ ! , Sym 2 / )

 23. E F F E C T I V E B O U N D S F O R T H E CLAS S N U M B E R 537

and P ( l , /) i s given by (23.23) . Hav e in mind tha t P ( l , / ) depend s on D bu t rathe r

mildly, precisel y

(23.39) J . ( l , / ) . n ( l+ ^)-,(l-^)(l + ^ )

\ y/i> *{ VP J ^ VP >
where th e produc t i s over prime s p \ D and p ^ B with \{p) — 1- Insertin g (23.37 )
to (23.7 ) w e obtain th e desire d expressio n fo r A ^ ( | ) i n term s o f derivative s o f th e
L-function fo r th e characte r ip = ex a t th e poin t 3 = 1 .
Prom now on we assume that e is a real character, th e symmetric squar e Sym 2 /
has rea l coefficient s (thoug h / ma y hav e bot h rea l an d imaginar y coefficients ) an d
P ( l , / ) i s also real. The n th e approximat e formul a (23.37 ) fo r S and S is the same .
Recall th e formul a (23.7) . Sinc e w e d o no t wis h th e mai n term s cancele d w e mus t
match th e parit y o f g with the sig n of the functiona l equatio n (23.6 ) fo r th e produc t
L-function L(s) = L(s,f)L(s,f x), i.e. , w e requir e
(23.40) w = w(f)w(fx) = (-If.
Now insertin g (23.37 ) int o (23.7 ) w e arriv e a t
(23.41)
Q-U<»>(!) = 2 r 2 ( f ) M ( i , / ) P ( i , / )

52 2«(^)L(«) (l^)(l0gW(l + 0 ( ^ ) ) + 0 ( ^ c ( l ) ) .
u-\-v=g ' '

We shal l refin e thi s rudimentar y formul a i n specia l cases .

Suppose e = 1 , so ty—XD- W e can use the estimates fo r L ^ ( l , XD) establishe d
in Sectio n 22. 6 unde r th e mino r assumptio n h(D)\og\D\ ^ \f\D~\ (se e (22.1 22) )
reducing (23.41 ) to
(23.42)
Q-*A^(i) = 2^r 2(|)M(l,/)P(l,/)

L'(l, XD)(\og \D\y~l (l + o ( j j ^ )) + 0(huc(l)).

Also we could replac e 1 /(1 , XD) b y ^-P(D) usin g (22.1 1 7) . Thoug h w e made s o fa r
rather mino r restrictions for the class number (exactl y such that th e segment (23.1 7 )
is not void) , our asymptoti c formul a (23.42 ) i s interesting onl y if h(D) < C (log \D\) 9,
or els e th e erro r ter m exceed s th e mai n term . Havin g assume d th e clas s numbe r
is tha t small , notic e tha t ther e ar e a t most 2g prime s p < exp(y/\og \D\) whic h
split i n K = Q(VD) (se e (22.95)) . Choosin g B = (log |Z>|)4^+20 (whic h i s withi n
(23.17)) w e find tha t i/ c (l) x u D(l) < v{Df b y (23.1 5) , the n 1 /(1 , XD) ~ v(p)
by (22.1 1 9) , an d

,2,43, P ( u )p\D, n H r V ^ p\D

v
+ihn(i+^f)
by (23.39) .
538 23. E F F E C T I V E B O U N D S F O R T HE CLAS S N U M B E R

23.6. A lowe r boun d fo r the class number .

As before x = XD is the character of the imaginary quadrati c field K — Q(\/D).
If the lef t sid e of (23.42) vanishes , we are left wit h a lower bound fo r the error ter m
which contain s th e class numbe r h — h(D). W e put this observatio n as
PROPOSITION 23.1 . Suppose f is a primitive cusp form of weight k ^ 1 , level
N and trivial central character. Suppose the product L-function vanishes at s = \
to order

(23.44) m = ordL(sJ)L(sJ x)^3.

Let g — m — 1 or g = m — 2 be such that (-1 ) 9 = w — w(f)w(f x). Then

1
(23.45) h(D) > 0(£>)(lo g \D\y-

where

(23.4S ) W = n(^n ' + ^ff-

p\D

The implied constant depends only on f and m and it is effectively computable.

The theor y o f ellipti c curve s provide s plent y o f example s o f cus p form s / G
S2(ro(iV)) for whic h th e associated L-functio n LE(S -f \) — L(s,f) shoul d hav e
large order of vanishing at s = \ , precisel y equal to the rank of the group of rational
points o n E, b y the Birch-Swinnerton-Dyer Conjecture . Beside s th e rank w e nee d
to kno w th e root number s w(f), w(f x) t o control th e parity o f g. I n general, fo r a
primitive for m / G Sk(To(N)) o f squarefree leve l N w e have (se e e.g . Propositio n
14.16 or [1 4] )
k
(23.47) w(f) = i /2(N)\(N).

There i s no simple formul a fo r w(f) i f N i s not squarefree. I f the greates t commo n

divisor (N,D 2) i s squarefree, the n th e twisted for m i s primitive, i.e. , fx = / ' ® % ,
has leve l M = [N, D2} an d the root numbe r is

(23.48) w(f x) = x(-Ni)n(N 2)\(N2)w(f)

where N = N XN2 wit h (Ni,D) = 1 an d 7V 2|£>. Thi s simplifie s t o w(f x) =

X(-N)w(f) i f (N,D) = 1 . Exploitin g (23.47 ) an d (23.48 ) w e shall deriv e fro m
Proposition 23. 1 our final result
T H E O R E M 23.2 . There exists an absolute, effectively computable constant c > 0
such that for any fundamental discriminant D < 0 we have

(23.49) h{D) ^c6{D)\og\D\

where

(23.50) 6 | (D) = Jl( i + I ) - 3 ( i + ^ ) - 1 .

p\D
 23. E F F E C T I V E B O U N D S F O R T H E CLAS S N U M B E R 53 9

PROOF. Firs t w e handl e th e cas e whe n p = 3 7 split s i n Q(y / D); i n thi s case
(22.72) implie s h(D) log37 ^ log|jD/4 | whic h i s bette r tha n (23.49) . Therefore ,
from no w o n w e assume tha t
(23.51) XD(37 )+ 1 .

We begi n wit h th e ellipti c curv e

(23.52) E 0 : y2 = x 3 + 1 0x 2 - 20 x + 8
which wa s considere d b y Gros s an d Zagier [GZ1 ] ; thi s i s a modula r curv e o f con -
ductor NQ = 3 7 and ran k VQ = 0 . Th e correspondin g primitiv e cus p for m i s
oo

f0(z) = £ X 0(n)nh(nz) e S 2(T0(N0))

l

and th e associate d Hasse-Wei l zet a functio n i s LE 0{S) = L(s — | , /o ) wher e

oo

L{$Jo) = ^ A 0 ( n ) n _ s .
l

This satisfie s th e appropriat e functiona l equatio n wit h sig n w(fo) = 1 . Fo r th e

proof o f Theore m 23. 2 w e tak e th e twiste d curv e
(23.53) E : -139y2 = x 3 + lOrr 2 - 20 x + 8 .
This i s a modula r curv e o f conducto rT V = 3 7 • 1392 an d ran k r = 3 . Th e corre -
sponding primitiv e cus p for m i s
oo

f(z) = J2 Kn)nh(nz) e S 2(T0(N))

i

with A(n ) = Ao(n)x~i39(n) , i.e. , f = f 0 0 x-1 39 , an d th e associate d Hasse-Wei l

zeta functio n i s LE(S) = L(s — |, / ) wher e
00

L(s,/)=]TA(n)n- s .
1

The sig n o f th e functiona l equatio n fo r / i s

w(/) = X - i 3 9 ( - l M / o ) = X-i39(-l ) = - 1
according t o (23.48) . Henc e L ( ^ , / ) = 0 , bu t Gros s an d Zagie r showe d tha t als o
L'{\if) — 0, therefor e L(s,f) vanishe s a t s = \ t o th e orde r a t leas t three . (I n
Appendix 23.A , w e sketc h th e proo f o f th e resul t o f Gros s an d Zagier) .
Now we examine the twisted for m f®XD £ S 2(YQ{ND2)) an d it s correspondin g
primitive for m f XD G S2{T0(M)) o f leve l M\ND 2.
C A S E 1 3 9 | D. The n M = ND 2, f XD=f®XD = fo® X-1 39D, an d b y virtu e o f
(23.48)
M / X D ) = X-1 39D(-A r i)M^2)A 0 (^ 2 )
where N XN2 = 3 7 wit h (N UD) = 1 and N 2\D. I f N x = 3 7 an d N 2 = 1 , w e ge t
wtfxo) = X-i39D(-37 ) = X-i39(-37 ) = -n;(/o ) = - 1 . If ^1 = 1 and N 2 = 37 , we
get w(f XD) = - X - I 3 9 D ( - 1 ) A 0 ( 3 7 ) = -A 0 (37) = - w ( / 0 ) = - 1 b y virtue o f (23.47) .
540 23. E F F E C T I V E B O U N D S F O R T H E CLAS S N U M B E R

CASE 1 39|D . W e writ e D — — 139C wher e C i s a positiv e fundamenta l discrimi -

nant coprim e wit h 1 39 . Accordingl y th e field characte r factor s int o \D — X-139XC
and / <g ) XD = / o ® X-139 ® Xc- Her e th e for m / o 0 X-1 3 9 * s n° t primitiv e becaus e
the trivial characte r X-1 3 9 simply kill s the place p — 139 in its L-function , however ,
it i s induced b y th e primitiv e for m /o - Therefor e th e primitiv e for m whic h induce s
/ ® XD i s f XD = / o ® Xc, i t ha s th e leve l M = [37 , C2 ] = 37C 2 /(37, C) an d th e
root numbe r
w(fXD) = xc(-N 1 )fi(N2)X0(N2)
by virtu e o f (23.48) , where N XN2 = 3 7 with (N UC) = 1 and N 2\C. I f N x = 3 7 and
N2 = 1 , we get w(f XD) = Xc(—37 ) = X-i39£>( —37) = —1 , as i n th e forme r case . I f
N± = 1 and N 2 = 37 , we get w(f XD) = -A 0 (37) = - 1.
Now we conclude tha t i n all cases w(f) — w(f XD) = —1 , whence the L-functio n
L(s) = L(s, f)L(s, f x) ha s roo t numbe r w — w(f)w(f XD) = 1 . Sinc e L(s) vanishe s
at s = ^ t o order m ^ 3 (actually m ^ 4 ) we can appl y Propositio n 23. 1 with g = 2
proving (23.49) . •

EXERCISE. Deriv e fro m (23.49 ) th e followin g effectiv e boun d

(23.54) h(D) >exp(-(loglog|D|)i)(log|£>|) .
[Hint: Us e genu s theor y i f D ha s man y smal l prim e factors. ]

23.7. Concludin g notes .

Our formul a (23.41 ) i s not restricte d t o the L-function s associate d wit h ellipti c
curves, an d on e hope s t o us e i t mor e efficientl y b y employin g othe r L-function s
which vanis h t o hig h order a t th e centra l point . A n interestin g propositio n i s to tr y
the L-function o f a cusp form with non-trivial centra l character, th e point bein g tha t
the leadin g ter m i n (23.41 ) ha s th e orde r o f magnitud e L(l,exr>)(lo g \D\) 9 whic h
for e ^ 1 is better tha n th e bound L'(l , xz))(log \D\) g~l deduce d fro m (23.42 ) whe n
e = l.
There ar e severa l example s o f cuspida l L-function s othe r tha n th e Hasse-Wei l
zeta function s o f ellipti c curve s whic h vanis h a t th e centra l poin t s — \ t o th e
order a t leas t tw o (bu t stil l with th e trivia l centra l character , unfortunately ) foun d
by F . Rodrigue z Villega s [RV] . Thes e ar e th e L-function s fo r Heck e character s o f
an imaginar y quadrati c field whic h ar e describe d i n Chapte r 3 by formula s (3.92) -
(3.95). Suppos e £ is such a character o f the field Q(y/—q) wit h q > 3 , q = 3(mo d 4 )
and o f weigh t £ ^ 1 ,£ = l(mo d 2) , i.e. , £ i s a characte r whic h o n principa l ideal s
takes

«(->»-(T)(R)'
for a = \{m + riy/^q). Ther e ar e exactl y h(—q) character s o f thi s type , eac h o f
which yield s a primitiv e cus p for m
/«(*) = ^2tta)(Na)h(zNa) G S e+
1 (T0(q))
a

(see (3.89)) , an d th e associate d L-functio n

a
 23. E F F E C T I V E B O U N D S F O R T H E CLAS S N U M B E R 541

satisfies th e following equatio n (self-dual )

(g)-r(.+ ^,0_.( g)'--r(,-. + <) i(1 -,o

(see Theore m 3.8 ) with th e root numbe r w = (—1 )^~( 2 ). I n the Main Theore m
of Rodriguez-Villega s [RV ] the centra l valu e L ( | , £ ) i s expresse d a s a squar e of
a linea r combinatio n o f certai n thet a functions , showin g tha t L ( | , £ ) ^ 0 . I t is
also prove d tha t L ( | , £ ) i s positive i f £ = 1 (because i t is true o n average ove r all
characters £ and these character s ar e Galois conjugate s o f each other) . Fo r £ = 3
the non-vanishin g o f L ( | , £ ) i s guaranteed onl y i f the class numbe r h(—q) is not
a multipl e o f 3 (thi s conditio n i s required t o show ther e i s only on e Galois orbit) .
Consistently, fo r q = 5 9 we have h(— 59) = 3 and L ( | , £) = 0 . Sinc e th e roo t
number is w = —(-^) = 1 we get three distinc t function s L(s,£) eac h o f whic h
vanishes at s = \ t o order a t least 2 , and the corresponding cus p forms hav e weigh t

23.A Appendix : Th e Gross-Zagier L-functio n vanishe s t o orde r 3 .

We will sketch the proof b y Gross and Zagier tha t th e Hasse-Weil zet a functio n
of the curve (23.53 )
-139y 2 = x 3 + 1 0x 2 - 20 x + 8
vanishes t o order ^ 3 at the central critica l point . Precisely , w e explain th e proof
of:
THEOREM 23.3 . For the above L-function, we have ! / ( / , \) = 0.
Since th e sign o f the function equatio n fo r L(f,s) i s —1 , thi s implie s tha t th e
order o f vanishin g i s a t leas t 3 a s desired . Th e proof i s base d o n a formul a o f
Gross and Zagier for the special valu e of the derivative of the L-function o f E (ove r
a suitabl e auxiliar y imaginar y quadrati c field) i n terms o f the height o f a specia l
point. T o state this , w e recall som e notation .
Let fcbea numbe r field an d M^ th e set of (normalized ) absolut e value s o f &; ,
denoted | • \v fo r v G M^; those includ e archimedea n a s well a s non-archimedea n
absolute value s an d we have the product formul a
TT \x\ v = 1 for any x G k*.
veMk
For an y x = [xo : x\ : x^\ i n the projective plan e P 2(fc), th e absolute logarithmi c
height o f x is

(23.55) h(x) =^ lo g Y[ max{|x 0 | v ,|xi| v ,|x 2 | v }

I- * ^ veM k

(this is well-defined becaus e of the product formula) . I f E c P 2 is an elliptic over /c,

then th e function x H- » h(x) i s almost a quadratic for m (wit h respec t t o the group
structure). Denotin g
hE(x)= h m,
n—>+oo 2
Tate an d Neron showe d tha t HE is a quadrati c for m wit h HE(X) = h(x) + O(l),
called th e canonical heigh t o n E. W e only nee d th e obvious fac t tha t 1 IE(0) — 0 ,
which follow s becaus e h(oo) = 1 ; see [Sil] for more details .
542 23. E F F E C T I V E B O U N D S F O R T H E CLAS S N U M B E R

Let E/Q b e an ellipti c curv e wit h conducto r N. I t i s modula r an d a (non-

obvious) consequenc e i s tha t ther e exist s non-constan t map s TT : XQ(N) — > E.
Among these , ther e i s a unique on e such tha t 7r(oo ) = 0 and 7r*(dz) = cf(z)dz fo r
some constan t c > 0, where dz i s the translation-invariant differentia l for m o n E
and / i s the primitive weight 2 cusp form associate d t o E. Th e map TT is then calle d
the modula r parameterizatio n o f E an d in fac t i t i s induced b y the holomorphi c
function H -> C defined by

f
TT(Z) = 2nic / f(w)dw.

Let K/Q b e any imaginar y quadrati c field with discriminan t D < 0 such that ever y
prime p dividin gT V splits i n K (i n particular, i s not ramified) . Not e tha t ther e
are infinitel y man y suc h fields becaus e th e conditions ar e XK{P) = 1 for p | iV,
which ar e congruence condition s fo r D. Ther e exis t 2 ^ ^ ) ideal s n c O suc h tha t
O/n ~ Z/iV Z (wher e O C K i s the ring o f integers) , thos e bein g i n one-to-on e
correspondence wit h solution s ft (mo d 27V ) to the congruence f3 2 = D (mo d 4iV) .
To each ideal class a of K i s associated the corresponding comple x point z f l G l
(see Sectio n 22.1 ) . On e easily show s tha t ther e i s an SL(2, Z)-equivalent poin t z a
such tha t
_ - E + V ^P
Za
~ 2A
with N | A an d B = / ? (mo d 2N). Thi s poin t i s uniqu e modul o T 0(N). Th e
Heegner poin t o n E associate d t o K an d n is defined t o be
y* = X>(*a) e £(C),
a

where the sum runs ove r al l ideal classe s o f K.

THEOREM 23. 4 (TH E GROSS-ZAGIE R FORMULA) . Let E/Q be an elliptic curve
with associated cusp form f, K an imaginary quadratic field as above with discrim-
inant D, x the Kronecker character associated to K. Let
LK(f,s) = L(f,s)L(f® X,s).

Then L K{f,\) = 0 and

where \\f\\ is the Petersson norm of f G S2{N), u G {1,2,3} is half the number of
units of K, deg(-7r ) ^ 1 is the degree of the map TT : XQ(N) —> E.
REMARKS. (1 ) In the applicatio n below , th e modularit y o f E an d th e existenc e
of TT wil l be obvious.
(2) Wha t i s implicit (an d already quit e deep ) i s that yx i s an algebraic poin t
in f£(C) , s o that it s canonical heigh t i s defined. Thi s come s fro m th e modular in-
terpretation o f X${N)\ point s r G ro(7V)\HI ar e in one-to-one correspondenc e wit h
pairs (E,H), wher e E is an elliptic curve over C and H c E(C) is a cyclic subgrou p
of orde r N. Thi s bijectio n i s induced b y the map r f- » (C/(Z 0 rZ) , ( N - 1 ) ) , and
the pai r correspondin g t o z a i s (C/a, q _ 1 a / a ) , wher e N K/qq = N. Thi s curv e has
complex multiplication , henc e finitely man y Galoi s conjugates , whic h implie s tha t
za i s defined ove r O.
 23. E F F E C T I V E B O U N D S F O R T H E CLAS S N U M B E R 54 3

More precisely, on e shows that yx G E{K). Th e z a, wher e a runs ove r the idea l
classes i n K, ar e i n fac t a singl e Galoi s orbi t unde r th e Galoi s grou p o f K.
SKETCH O F PROO F O F THEORE M 23.3 . Th e principl e i s tha t i f E/Q i s a n
elliptic curve , wit h associate d cus p for m / , suc h tha t th e roo t numbe r o f E i s + 1,
and i f D < 0 i s a fundamenta l discriminan t o f a n imaginar y quadrati c fiel d K/Q
such that th e twisted curve E^/Q ha s root number — 1, then one has L(f<g>Xi \) ~ 0 -
If i n additio n th e Heegne r poin t yx G E(C) turn s ou t t o hav e h(yx) = 0 , the n
£'*(/, §) = £(/,§)£'(/®x,|) = o.
If L(/ , \) 7 ^ 0 (whic h ca n b e rigorousl y checke d numerically) , the n i t follow s tha t
£'(/®X,5) = 0 .
We no w appl y thi s followin g [GZ2 ] t o th e curv e
E : y 2 = x 3 + 1 0x 2 - 20 x + 8
with th e imaginar y quadrati c fiel d K = Q(\/—1 39) , s o tha t £?£ > i s indee d th e
curve (23.53) . Thi s curve s satisfie s th e assumption s o f Theorem 23. 4 since w e hav e
X-i39(-37) = - 1 . We choos e th e idea l n = ( 1 + u> ) (where u = ( 1 + >/-1 39)/2) .
The discriminan t o f £ i s A = 2 1 2 -37 and th e j-invariant i s j = 2 ^ - = 409 ™00.
Thus th e give n Weierstras s equatio n i s no t minimal . A correspondin g minima l
model i s the curv e
E' : y 2 + y = x 3 + x 2 - 3x + 1 ,
the coordinat e chang e bringin g E' t o E bein g (x,y) \-> (4 x — 2, Sy + 4) . Th e curv e
E^ ha s discriminan t 37 , an d sinc e i t i s squarefree , th e conducto r o f E i s 37 . W e
denote b y / th e weight 2 primitive cus p for m o f level 37 corresponding t o E. Writ e
E(C) = C / ( Z 0 rZ ) fo r som e r G H , a s give n b y th e theor y o f ellipti c functions .
Since th e conducto r N = 3 7 i s squarefree , th e sig n o f th e functiona l equatio n
is give n b y SE = a#/(37 ) whic h give s SE = 1 , an d numerica l checkin g give s
£,(/, I ) = L ( £ , 1 ) « 0.725681 061 9361 5278.. . ^ 0
as required. I n fact, on e can check that E(Q) c ^ Z/3Z i s generated b y (rr , y) = (2,4) .
The idea l clas s grou p o f K i s o f orde r three , generate d b y a = [5 , 1 + v 2 ~ 1 3 9 ] .
The thre e Heegne r point s fo r th e chose n n ar e give n b y
„ _ 1 5 1 + x/ :r 1 39 _ _ - 7 1 + \/ z T39 . _ 3 + y /z: 1 39
Za Z
~ 37 0' ~a~ 37 0' Z
°~ 7 4'
Denoting b y TT th e modula r parameterizatio n X Q ( 3 7 ) —> E, w e nee d t o prov e tha t
7r(za) -f 7r(z- a) + TT(ZO) = 0 € E(C), o r equivalently , tha t
(23.57) ir(z a) + 7r(5_ a) + 7r(z a) G Z 0 rZ .
Notice tha t thi s i s a proble m state d i n analyti c terms . T o prov e it , on e use s th e
following lemm a fro m th e theor y o f ellipti c functions :
LEMMA 23.5 . Let A C C be a lattice, and let f ^ 0 be a A-periodic meromor-
phic function onC. Let si,... , Sd be a set of representatives of A-equivalence classes
of poles and zeros of f, with multiplicity. Then we have si + • • • + Sd G A.
This i s th e complex-analyti c analogu e o f par t o f Propositio n 1 1 .33 , whic h w e
proved ther e fo r ellipti c curve s ove r finite fields. I t i s i n fac t onl y a specia l cas e o f
the Abel-Jacob i Theore m (se e [Rey ] fo r instance) .
54 4 23. E F F E C T I V E B O U N D S F O R T H E CLAS S N U M B E R

P R O O F . Choos e a fundamenta l parallelogra m P C C fo r C/ A an d represen -

tatives fo r th e Si suc h tha t Si i s i n th e interio r o f P fo r al l i. The n conside r th e
integral

2*iJ9pZf(z)dZ Y l

by th e residu e theorem . I f 7 an d 7 + A o are paralle l side s o f P , wit h A o G A, the n

(taking int o accoun t orientation ) th e substitutio n z H * Z — zo o n th e secon d on e
yields b y A-periodicit y

™ y7 * -mdz+*n / 7+Ao
z
W)dz=2^ y 7 W)dz e Ao Z
since th e integra l i s th e windin g numbe r aroun d 0 o f th e close d loo p t \-t f(^{t)).
Hence th e resul t follows . •

We wil l exhibi t a n ellipti c functio n o n C/( Z 0 rZ ) wit h a pol e o f orde r 3 at 0

and simpl e zero s at th e thre e points n(zb) 1 thereb y provin g (23.57 ) b y Lemma 23.5.
Start wit h th e meromorphi c functio n u o n Xo(37 ) give n b y

u(z) = J * £ -
where7 7 is th e Dedekin d et a functio n give n b y

V(z)=e[^)l[(l-e(kz))

(see (22.66)) .
Letting q = e(z) , w e hav e a t infinit y th e expansio n

so u ha s a pol e o f orde r 3 a t 00 . Sinc e rj doesn't vanis h excep t a t infinity , thi s i s

the onl y pol e o f u. Moreover , becaus e7 7 is o f leve l 1 , th e substitutio n z »- > —1 /z
shows tha t u ha s a zer o of orde r 3 at th e cus p 0 , an d n o othe r zeros . No w conside r
v(z) = u(z) — U(ZQ). I t ha s stil l a pole of order 3 at infinity , an d a zero at ZQ. Thi s
is a simpl e zero , an d v a s tw o othe r simpl e zero s a t z a an d Z- a. Indeed , becaus e
there ar e a s many zero s as poles, it suffice s t o prove v(z± a) = 0 , or u(z± a) = U(ZQ),
to sho w this .
Notice tha t u(z) 1 2 = A(^)A(372:) -1 wher e A i s th e Ramanuja n function . Fo r
a poin t z G X Q ( 3 7 ) correspondin g t o a pai r (E,H) o f a n ellipti c curv e an d a
cyclic subgrou p o f orde rT V = 37 , a s i n th e remar k above , on e ha s A(z) = A(E)
and A(372: ) = A(E/H). Fo r a Heegne r poin t i b , thi s pai r i s (C/b,n _ 1 b/b). I t
follows tha t U(ZQ) 1 2 = u(z± a)1 2. The n on e ca n chec k (eve n numerically ) tha t
U(ZQ) = u(z± a) or , mor e precisely , prov e tha t U(ZQ) = 1 + w G K, wher e 3 7 =
N(l + CJ) , an d us e th e fac t fro m Comple x Multiplicatio n theory , tha t i f a i s th e
generator o f th e Galoi s grou p o f th e Hilber t Clas s Fiel d o f K correspondin g t o a ,
then u(z± a) — u(zo) a±a = U{ZQ).
To conclude , w e need t o pus h th e functio n v t o E(C). T o d o this , pu t

W(z)= n "t o
n(z')=z
 23. E F F E C T I V E B O U N D S F O R T H E CLAS S N U M B E R 545

for z £ E(C). I t i s easy t o check tha t thi s i s a well-defined meromorphi c functio n

on i£(C) , an d it has a pole of order 3 at 7r(oo) = 0 and simple zero s a t 7r(z b). Thi s
finishes th e proof . •

R E M A R K . Usin g program s suc h a s PARI/GP, on e can nowadays find explici t

equations an d formulas fo r all objects i n the above constructions . Som e ar e stated
in [MSD] , Sec . 5 . We thank C . Delaunay fo r computing th e Heegner points . I t is
easier t o deal wit h th e curve
F : y 2 + y = x3 + x2 - 23 x - 5 0
instead o f E. B y [MSD] , w e have E ~ F/fis, wher e fis i s the group o f 3-root s
of unit y (on e ha s F[3] ~ // 3 x Z/3Z as Galois-module), s o F i s isogenous t o E, i n
particular, E an d F (an d thei r twists ) hav e identica l L-functions .
The curv e Xo(37 ) i s the nonsingular mode l o f the plan e curv e wit h equation s
X 0 (37) : y 2 = -x6 - 9x 4 - Ux + 37,
and th e modular parameterizatio n o f F i s the ma p
TT(X, y) = (\(37x- 2 - 5) , |(37yx" 3 - 4)) .
The Hilber t Clas s Fiel d H o f K = Q(V~1 39 ) i s th e splittin g field ove r K o f
P = X 3 +_4X 2 + 6X + 1 . Denot e b y 6La the real roo t o f P, by 0o th e root i n H
and 6 a = 6Q. The n 7r(zb ) £ F(H) i s given by
7r(zb) = (Ai(0 b ) + A 2(0h)V=139,B1 (0h) + B 2 (fl b )>/=l39)
with
1_
Ax = -— 2 (1 1 52X 2 + 6398X + ^ | § 9)
412
1
4 2 = 1 Q n ^ l 2 ( 3 2 5 7 9 X 2 + 53474 X + ±2|8i )
ioy • 4i
2
B x = J_(_39658X + 225512X + 84810)
413
B2 = - * (2662280X2 + 6561700X + 1270577).
l o c / • 41
One ca n then chec k "b y hand" tha t 7r(z a) + 7r(z_ a) + 7r(5o ) = 0 in F(C). (Thi s
means tha t th e three point s 7r(zb ) a r e on a line i n P2 ). Note , however , tha t TT(ZO)
is of infinite orde r (i n F(H)) becaus e th e class numbe r o f Q(y/—139) i s larger tha n
the degre e ( = 2) of the modula r parameterizatio n n (thi s is a theorem o f Nakazato
[Nak]).
This page intentionally left blank
 http://dx.doi.org/10.1090/coll/053/25

CHAPTER 24

THE CRITICA L ZERO S O F TH E

RIEMANN ZET A FUNCTIO N

The zero s p = \ + ry o f £(s) o n th e critica l lin e ar e calle d th e critica l zeros .

Let N 0(T) denot e th e numbe r o f critica l zero s wit h 0 < 7 ^ T . Recal l tha t N(T)
denotes th e numbe r o f al l zero s p = /3 + 27 wit h 0 < /3 < 1 and 0 < 7 < T, an d
that w e hav e (se e Theore m 5.24 )

(24.1) N(T) = Ll o g l _ + O(iogT) .

The Rieman n Hypothesi s assert s tha t iVo(T ) = N(T). B y th e densit y theore m

(Theorem 1 0.1 ) w e kno w tha t almos t al l zero s ar e nea r th e critica l line . Rieman n
showed by hand computation s tha t ther e ar e zeros exactly on the line Re s = \ , th e
first critica l zer o being p\ = \ + i^\ wit h 7 1 = 14.1 3 . .. . Using moder n computer s
it ha s bee n checke d tha t million s o f th e firs t zero s o f £(s ) ar e o n th e lin e R e s = \
and the y al l ar e simpl e (se e e.g. [LRW]) .
24.1. A lower boun d fo r N 0(T).

In 1 91 4 , G . H . Hard y prove d tha t £(s ) ha s infinitel y man y zero s o n th e lin e

Re s = | , and seve n years later h e showed jointly wit h J . E . Littlewoo d [HL2 ] tha t
(24.2) 7 V o ( T ) > T
for al l larg e T . I n this sectio n w e giv e a proof o f this result . The n i n the nex t
section w e improve thi s boun d t o th e followin g estimat e o f A . Selberg ,
(24.3) 7 V 0(r)>TlogT,

which say s tha t a positive proportio n o f zero s lie s o n th e critica l line .

We se t

where g(s) = 7r~" S//2r(s/2). B y th e functiona l equatio n g(s)((s) = g(l — s)((l — s)

it follow s tha t f(u) is real an d even . W e conside r th e tw o integral s
rt+A
(24.5) I(t) = j f(u)du,
pt+A
J = du
(24.6) ^ J \fM\
where A is a fixed positive number . Clearl y \I(t)\ ^ J{t). I f
(24.7) |/(t) | < J{t),
547
548 24. T H E C R I T I C A L ZERO S O F T H E R I E M A N N ZET A F U N C T I O N

then f(u) mus t chang e sig n i n the interval (t,t + A), henc e f(u) mus t hav e a zero
in (t,t + A) , s o has C (| + ^ u) becaus e g(s) doe s no t vanish anywhere . Therefor e
our proble m reduce s t o showing tha t (24.7 ) occur s quit e often . T o this en d we ar e
going to prove a lower boun d fo r I(t) an d an upper boun d fo r J(t) o n average ove r
the segmen t [T,2T] .
LEMMA 24.1 . Let A ^ 1 . There exists a function K(t), which also depends on
A, such that

(24.8) J(t) ^A-K(t),

2
(24.9) / \K(t)\ dt<^T,

ifT^A2, where the implied constant is absolute.

LEMMA 24.2 . Let A ^ 1 . We have

2
(24.10) / \I(t)\ dt<^AT

ifT^A6, where the implied constant is absolute.

By Lemmas 24. 1 and 24.2 the estimate of Hardy-Littlewood ca n be deduced as
follows. Le t T b e the subset o f [T,2T] wher e \I(t)\ = J(£), thu s

/ \I(t)\dt= f J(t)dt.
IT JT

We get by Cauchy's inequalit y an d by Lemma 24. 2

/ \I(t)\dt^ f |J(t)|d*<A*T .

On th e other hand , w e get by Lemma 24. 1 and Cauchy's inequalit y

/ J(t)dt > A\T\ - / K(t)dt = A\T\ + 0 ( | T | * T * ) .

Combining th e abov e bound s w e deduce tha t th e measur e o f T satisfie s \T\ <C

A~2 T wher e th e implied constan t i s absolute. Fo r A sufficientl y larg e thi s give s
\T\ ^ \T. I n other words , the set S = [T, 2T)\Ton whic h (24.7 ) hold s has measur e
\S\ ^ \T. Clearl y S contain s a sequence {£i,.. . , tn} o f A-spaced point s o f length
R > A _1 |<S| ^ T/2A . Fo r each t r, 1 < r < R, ther e i s a sig n chang e o f f(u) i n
tr < u < tr + A, hence ther e i s a critical zer o p r = \ + i^ r wit h t r < 7 r < t r + A.
This prove s tha t N 0(T) > T / A which i s (24.2).
It remain s t o prove th e two lemmas .
P R O O F O F LEMM A 24.1 . W e hav e
pA PA
J(t) = / |CC i + it + iu)\du ^ / C ( | + ^ + iu)du
JO ' J O
i rA i
^A - / (C( | + it + iw) - l)d u = A - #(*) ,
'Jo '
 24. TH E CRITICA L ZERO S O F TH E RIEMAN N ZET A FUNCTIO N 549

say. T o estimate K(t) w e us e th e approximatio n

(24.11) <(*) = J2 n-' + 0(T-*]

l^n^T

which i s valid fo r s = \ + it wit h T < t < 3T (se e (8.3)) . Henc e

iA
1 - n~
K(t) = \ J2 *-*" " logn + 0(AT"2) ,
Kn^T
and b y Theore m 9. 1 w e conclude tha t
r2T
2 2 2
/ |lf(*)| d t < T J2 n-\logn)- + A <z:T.
Kn^T
D

P R O O F O F LEMMA 24.2 . Usin g th e convexit y boun d ((s) < C \s\* w e arrang e

the integra l o f (24.1 0 ) a s follow s
r2T r 2T,r A
2
/ = / \I(t)\ dt= / / f(t + u)du\
JT JT UO '
pA pA p2T
= f(t + u 1 )f(t + u 2)dtdu
1 du2
Jo Jo JT
rA pA p2T
= f(t)f(t + u 12-u )dtdu
1 du2^0(A3T2)
Jo Jo JT
A n2T
/ 3
( A - M ) / f(t)f(t + u)dtdu + 0(A Ti).
-A JT
Now w e approximat e f(t)f(t + u) b y a Dirichle t polynomial . Firs t b y Stirling' s
formula
T(a + it) = (27r) i (it)*"* ( t / ef e - 5 ' ( l + O ^ " 1 ) )

(-°m>
for < J > 0 , t > 0 we find tha t
g{\ + tf)£(± + i t + m ) /2TT \"
|fl(| + *t)s(5+ ** + *«)! V *
Then b y (24.1 1 ) an d th e convexit y boun d £(s ) < C |s|* w e ge t

i ,m^n. /27rm 2 \ 2 2
/(«)/(* + «) = E E ( ™ ) - * ( 7 m +0(A T-i).
l<m,n<T
Hence
4
/ = J2Y1 c(m,n)(mn)-i+0(A T^)

where
/my* (2irrri 2\ T
dtdu

-!/;0"«(T>^) dt
550 24 . T H E C R I T I C A L ZERO S O F T H E R I E M A N N ZET A F U N C T I O N

with \(x) = x ~ 2 ( s i n : r ) 2 . I f ra ^ n , w e integrat e b y p a r t s gettin g

m( sin | A l o g ( 7 r m 2 / T ) ^ ^ f sin \ A l o g ( 2 7 r m 2 / T ) \ '

c(ra, n) lo g — <
log(7rm /T)2
+ log(27rm 2 /T) /
/•2T
/sin|Alog(27rm2/t)
+ \ log(27rm 2
/t)
27rm2 -2
log 7rm
< mi in
n ^I LA
+ log

If m = n , i t i s easie r t o integrat e ove r t firs t an d the n ove r u gettin g

2 - 2 ^ fnm 2\ 2
c(m,m) = 2 T / (A-H) -. du
J-L_A 2 — zu \ 1
27rra2'-1
< A T mi n ^ 1 , log 7rm
T + log

Using thes e estimate s on e derive s t h a t

^ ^ c(m,n)(mn) * < A T

which complete s th e proo f o f (24.1 0) .

•
2 4 . 2 . A p o s i t i v e p r o p o r t i o n o f critica l z e r o s .

In 1 942 , A . Selber g modifie d th e argument s o f Hard y an d Littlewoo d showin g

t h a t a positiv e proportio n o f zero s o f £(s ) lie s o n th e lin e R e s = | . H e di d no t giv e
the percentag e number . A ver y stron g resul t i s du e t o B . Conre y [Co2] , namel y
that

(24.12) N0(T) > |iV(T )

for sufficientl y larg e T. Conrey' s approac h i s base d o n Levinson' s m e t h o d (se e

[Lev]) whic h i s quit e differen t fro m th e on e use d b y Selberg .

Selberg's refinemen t o f t h e Hardy-Littlewoo d argument s appear s i n mollifica -

tion o f £(s ) b y a Dirichle t polynomial , it s rol e bein g t o diminis h t h e contributio n
of larg e value s o f ((s). Thi s ide a wa s first exercise d b y Boh r an d L a n d a u [BL] ,
however, les s effectivel y t h a n i n t h e hand s o f A . Selber g [S8] . Selber g consider s

(24.13) m \ga+iu)\
C(h+iu)\<p(l+iu)

where (p(s) i s a Dirichle t polynomia l o f typ e

S
(24.14) ^)=£^V
and h(x) i s a real , continuou s functio n o n [0,1 ] suc h t h a t

(24.15) h(x) = l + 0(x )

(24.16) h(x) < 1 -x
 24. T H E C R I T I C A L ZERO S O F T H E R I E M A N N ZET A F U N C T I O N 551

while th e coefficient s 7 ^ hav e arithmetica l natur e (the y ar e als o real an d bounded) .

Clearly f(u) i s rea l an d even . Th e reaso n tha t \ip(\ + iu)\ 2 i s applie d rathe r tha n
<p(| + iu) i s t o ensur e tha t an y sig n chang e o f f(u) a t u = 7 come s fro m th e sig n
change o f £ ( | +iu) becaus e \<p(\ +iu)\ 2 ^ 0 . Th e mollifie r \ip(\ +iu)\ 2 give s som e
extra zero s t o /(it) , bu t the y ar e no t counte d b y th e sig n changes !
A reasonabl e gues s o f th e coefficient s 7 ^ come s b y tryin g ip(s) t o approximat e
£(s)~2. Indee d Selber g choose s 7 ^ t o b e exactl y th e coefficient s i n th e Dirichle t
series representatio n o f th e Eule r produc t

(24.17) cw-^nC 1 -£)*=£-***"'•

Vd

We hav e fo r s > 1 ,

V ~ p*J ^ V ~ ~ ps) ^ \ ~ p*J '

whence |7^ | ^ |7^ | ^ 1 , where 7 ^ ar e th e coefficient s o f
-
(24.18) cw^n^-ir^E™* '-
Pd

As i n th e Hardy-Littlewoo d proo f w e shall compar e th e integral s

pt+A
(24.19) I(t ) - J f(u)du,
pt+A
(24.20) J ( t ) = / \f{u)\du

in th e rang e 0 ^ t ^ T . Du e t o th e mollification , w e shal l b e abl e t o wor k wit h

smaller A ,
1
(24.21) A x (logT)"
where th e implie d constan t i s absolute . W e shal l prov e thre e estimate s
LEMMA 24.3 . Let ( l o g T ) - 1 < A ^ 1 and D = T e with 0 < 6 < ^ . Then we
have

(24.22) / |/(t)|<t t
Jo
T
(24.23) / [T|\f(t)\
/(t)f<ft<T ,
Jo
/o

I
2
(24.24) f |/(t)| d t < A T ( l o g T )1- ,

where the implied constant depends only on 0.

Prom thes e estimate s w e derive Selberg' s boun d (24.3 ) a s follows. Le t £ b e th e
subset o f [0,T ] i n whic h \I(t)\ < J{t). W e obtai n

A= J J(t)dt> f(J(t)-\I(t)\)dt= [ (J(t)-\I(t)\)dt = B-C

J£ J£ JO
552 24 . TH E CRITICA L ZERO S O F TH E RIEMAN N ZET A FUNCTIO N

say. B y th e Cauchy-Schwar z inequalit y an d b y (24.23 ) w e deriv e

A2^\£\f J 2
(t)dt=\£\ f (f \f(t + v)\dvYdt
pT pA /*T+ A
< A\£\ / / \f(t + v)\ 2dvdt ^ 2A 2 |£| / \f(t)\ 2
dt < A 2
\£\T.
Jo Jo Jo
On th e othe r hand , w e have b y (24.22 )
pT pt+A pT
B= / \f(u)\dudt > A / \f(u)\du > AT ,
Jo Jt JA
and b y (24.24 )
\I(t)\ 2dt < A T ^ l o g T ) " 1 .
C2 ^T f
Jo
Combining the above estimates we obtain A T < C A2T(logT)~2+A|£ 12T2, whenc e
\£\ ^ > T , provide d A l o g T i s sufficientl y large . Th e se t £ contain s a sequenc e o f
points {ti, ... , tR\ whic h ar e A-space d o f lengt h R ^ A _ 1 | £ | > T l o g T . Fo r eac h
tr ther e i s a sig n chang e o f f{u) i n (£ r ,£ r 4 - A), s o ther e i s a zer o o f £ ( | + iu) i n
this interva l becaus e \<p{\ 4 - iu)\2 i s non-negative. Henc e w e conclud e (24.3) .
It remain s t o prov e th e estimate s fo r th e thre e integrals . Th e firs t on e (24.22 )
is easy. W e hav e

2
[ \f(t)\dt>\[ {(±+it)tp (±+it)dt
Jo ] IT/2
JT/2
By th e approximatio n (24.1 1 ) an d th e trivia l boun d <p(s) < C D* fo r R e (5) = ^ w e
get
C{s)^2{s)=YJ^n-s^O{DT-^)

where a\ = 1 and \a n\ ^ r 3(n) fo r n ^ N = £> 2T. Assumin g D < T i ( l o g T ) - 1 w e

obtain

I/ r
/ 22
((s)<p2(s)dt>^-2 £
2 ^
|a
Vlog
n |n-i(logn1
)- +o(^)>|
:

if T i s sufficiently large , whic h prove s (24.22) .

The othe r tw o estimate s (24.23 ) an d (24.24 ) ar e muc h harde r t o establis h be -
cause we have to appeal to the nature o f the mollifie r <p 2(s). We split th e argument s
into severa l group s an d evaluat e a fe w integral s o f th e zet a functio n whic h ar e o f
independent interest .
T H E O R E M 24.4 . Let si = \ + ivi,s2 = \+iv 2 with \v\\ < 1 , \v2\ ^ 1 . Let a, b
be positive integers with (a , b) = 1 . We have

(24.25) J C(s 1 +it)as2+it)a^+H^-itdt = TP v(^\+0((ab)iTl(logT)

6
)^

where v = v 2 — v 1and P V(X) is defined by

(24.26) P t,(X)=C(l-w) + C( l + w ) T - — •
1 + IV
 24. TH E CRITICA L ZERO S O F TH E RIEMAN N ZET A FUNCTIO N 55 3

The implied constant in (24-25) is absolute.

Notice tha t b y th e Lauren t expansio n o f £( 1 + iv) w e ge t

(24.27) P V(X) = : + 27 - 1 + 0(\v\)

iv
which i s quit e goo d i f \v | i s small. I n particular , w e hav e
(24.28) P 0{X) = logX + 2-y-l.
COROLLARY 24.5 . If (a , 6) = 1 , we have

(24.29) [ U\ + it)?QUdt = -^ (lo g JL + 2 7 - l ) + O iabTl (IogT)« )

w/iere £/* e implied constant is absolute.
We begin the proof of Theorem 24. 4 by evaluating the following partia l integral s

2 fll it
(24.30) S = / C(si + **)C(* 2 + i * ) a " " 6"Sa+it*

for T < T i < T 2 ^ 2T . Changin g th e variabl e t -> • t - v 1 w e ge t

2
5 = {ab)~h iv f C(| + it)C( | + i v + i t ) ( - ) * * * + £ ,

where E 1 is the non-overlappin g par t o f th e integral , an d i t satisfie s

(24.31) £ < ( a & ) - * r i
by applyin g th e convexit y boun d £(s ) < C |s|* o n th e lin e R e s = ^ .
Next w e approximat e £(s ) b y it s partia l su m

(24.32) Cx(s )= 5 > " ' .

Recall tha t fo r R e 5 = a ^ | an d |s | ^ nX w e hav e

(24.33 ) C W = Cx(« ) - ~ + 0(X-") .

1— 5
In particular , fo r 5 with R e 5 = \ an d T < C |s| < TTT we hav e

(24.34) C(« ) = Cr(a ) + 0 ( r - i ) .

Hence

5 = (o6)"H i t ; / 2
C T (| + « ) C T ( | ~iv- H) (ly'dt + E + E^

where E^ account s th e contributio n o f th e erro r ter m i n (24.34) , thu s i t satisfie s

(24.35) E o o < ( a 6 ) - i ( T l o g r ) i
by virtu e o f th e followin g estimat e (obtaine d fro m (7.52 ) b y Cauchy' s inequality )
3T
|C(!+it)|d*<T(logT)i.
Jo
55 4 24. TH E CRITICA L ZERO S O F TH E RIEMAN N ZET A FUNCTIO N

Inserting th e Dirichle t polynomial s (24.32 ) w e ge t

(24.36) S= f 2 ^^(am)-i'i\bn)-i+iv^itdt + E^E 00.

First w e pul l ou t th e term s o n th e diagona l am = bn. Sinc e a, b are co-prim e

it follow s tha t vn — bl an d n — at wit h t ^ L = Tc~ l wher e c = max(a , b). Henc e
these term s contribut e
1 +iv
s0 = (T 12-T )(ab)- {L(i-iv).
Introducing th e approximatio n (24.33 ) w e arriv e a t

(24.37) S 0 = (T 2 - H X a & r W c a - iv) + —) + O(l) .

V iv J

Now we consider th e contributio n t o the integra l (24.36 ) o f the term s am ^ 6n,

say <S* . B y integratio n w e ge t
(24.38) 5 * = 5 ( T 2)-5(Ti),

where

(24.39) S(t) = i ]T ^ ( a m ) - i - ^ ( 6 n ) " 4 + " + i t (lo g °^) ~ \

am^bn

We shal l sho w tha t th e mai n contributio n t o S(t) come s fro m th e term s nea r th e
diagonal, i.e. , fro m m,n with 1 — £ < ^ < 1 + £ fo r som e smal l e to b e chose n
later. Pu t
1
(24.40) S e(t)=i £ £ {amr ^\bnr^+it{\og1
a
^y .
0<\am—bn\<£bn

For th e remainin g term s w e hav e | log ^ | > |, and sinc e lo g x i s monotonic , w e

can appl y partia l summatio n t o estimat e th e contributio n o f thes e term s b y

e-\ab)-X2 (\ 2
ogT) \U\ + it)Cy(k ~ iv
' **) l
for som e x,y wit h 1 < x,y < T. No w b y Weyl' s subconvexit y boun d Cx(s) < ^
|s|i(log3|^|) 2 , whic h i s valid fo r s with R e s = \ an d x < 4|s|2 (se e (8.22)) , i t
follows tha t
1 1
(24.41) S(t)=S e(t) + 0(e- (ab)- *T*(\ogTf).

For th e term s i n S £(t) w e write am — bn + h wit h O < \h\ < ebn, an d us e th e

approximations

(£)"-(' + £)""--"""^(sk)
getting
1
Se(t)= ^ ] T ih" (bn)lve-ith/bn + E1
am—bn=h
0<\h\<ebn
 24. TH E CRITICA L ZERO S O F TH E RIEMAN N ZET A FUNCTIO N 55 5

where E\ account s fo r th e contributio n o f th e abov e erro r terms . W e hav e

\am—bn\<ebn

< ( 1 + st) ^(enba- 1 + l ) ( 6 n ) - x < C (1 + e r ) ( e a_ 1r + b'1 l o g T ) .

n

Clearly w e ca n interchang e a and b (see th e firs t inequality) , s o combinin g bot h

estimates w e deriv e a symmetric boun d
2
(24.42) E x <C (ab)-^(sT)
provided eT ^ l o g T , whic h conditio n w e hencefort h assume .
Now w e writ e S e(t) a s
ih
S£(t)= J2 'X J2 {bn) iv
e-ith'bn
1 ^E
+
H~<h<H ' Ni<n<N 2
h^O n= — hb(mod a)

where
H+ = min{ebT, e(l + e^aT}
1
H~ = - min{£6T , s(l-e)~ aT}
1
Ni = |/i|/e6 , 7V 2 = min{T , (a T - / i ^ " }.
For t h e inne r su m w e appl y t h e followin g formul a

£e (/H) = -/ e{f(x))dx + 0{\)

Nl<n<N2 a j N
i
n=a(mod a)

which hold s tru e fo r an y smoot h functio n / suc h t h a t \f f\ ^ (2a)~ l an d / " ^=- 0 in

the interva l [iVi , JV2] with N\ < N2 (thi s follow s fro m Propositio n 8.7) . Assumin g
2
(24.43) 2e abT ^ 1
we obtai n
1 fN 2
iv
J 3 {bn) e'ith'bn = - (bx) iv
e-ith'hxdx + O(l).
n= — hb(mod a)

Here w e replac e N 2 b y T m i n { l , a / 6 } makin g a n erro r 0(\h\/ab). T h e n w e chang e

the variabl e x int o xt/2-Kb gettin g
1 + i v £t 2n h
1/ + \ r / \ \ /IAi l

Ni<_n<N2 x 7
^t/znai
n— — hb(mod a)

where d = min{a, 6} . Insertin g thi s int o S £(t) we ge t

2 l
5 e ( t ) = Xr (-pj r x- "^( J ] (2irih)- e{-hx))dx + E 1 +E2
a
\ KJ Jt/2-KdT 0<\h\<et/2nx
H~<h<H+
556 24. TH E CRITICA L ZERO S O F TH E RIEMAN N ZET A FUNCTIO N

where

(24.44) E 2 < Yl ( ^ + I) < (ofc)"^ r + loS T -

0</i<2edT
The inne r su m ove r h in the abov e integral i s a partial su m in the Fourie r expansio n
for —i[>{x) = [x] — x + | . Applyin g th e approximatio n
1
-^(x) = ^ (27ri/i)- e(/ix) + 0 ( ( 1 + ||x|| # r 1 )
-//i<^i</f2
MO
where H = min{ifi , i/2 } (compar e thi s wit h (4.1 8) ) w e ge t

S
f / t \ livv /.0
r c0 2 iv
e(t) = - — / x- ~ il>(x)dx + E 1 +E2 + E3
tt& V27T / J t/2ndT
J hi
where
2
£ 3 <-r/( l + e * ^) x dx= — [\\x +-r )—
ab K J ab K J
Jt/2irdT % £ Jt/27rdT ^ X

eab

We simplif y thi s boun d t o

(24.45) £ 3 < e " 1 ^ ) " ^ lQ gT-

Finally w e use th e formul a

which hold s fo r 0 < y < 1 and R e w > —1 , gettin g

m . I f f )"flL± M - < <5£ + iffl!^ + Bl + * + «

ao \2TTJ I +IV ab iv ab 1 +z v

Now addin g u p th e relevan t formula s w e obtai n

T2P
S = {ab) — l-\-iv
\^)-TlP\*tb) +R
where P V{X) i s defined b y (24.26 ) an d R i s the total error term. B y (24.31 ) , (24.35) ,
(24.37), (24.42)-(24.45 ) w e ge t

J 2 « (afc)-*(T * + e - 1 : H +£ 2 T 2 )(logT) 6 + (logT) 2

provided eT ^ log T an d 2e 2abT < 1 . W e choos e e = T~ 3 gettin g

6
JR«(a6)-5Tf(logT)

provided 2a 6 ^ TB. W e als o hav e th e trivia l boun d

2
<S<(a&)-5 / | C ( i + i < ) | cte«(a6)-iTlogT.
./o
 24. T H E C R I T I C AL ZERO S O F T HE R I E M A NN ZET A F U N C T I O N 55 7

Combining bot h result s w e get rid of the condition 2ab < T Q claimin g a weake r
result

S = (ab) — l-\-iv WAW„(^)

' \27rabJ
+ 0((ab) 2 Tl (log T)6 ).

This yield s (24.25 ) b y addin g th e results for dyadic point s Tj.

R E M A R K S . Th e erro r term s i n (24.25) an d (24.29) coul d b e improved substan -
tially. Fo r example, i f a = b = 1 , our Corollary 24. 5 yield s

2
(24.47) J \C(l +lt)\ dt = T(log^-+2j-l^E(T)

with E(T) < C T§ log6 T, whil e i t is known tha t E(T) < T&+ £ . I t is conjecture d
that (24.47 ) hold s wit h E(T) < Ti+ £ (se e [Iv]).

Theorem 24. 4 can be easily generalize d t o integrals o f type

GV(T) = J C(l+ tf)C(| -it- iv) (l) ltg{t)dt

where g(t) is a smooth functio n o n [0,T]. Fo r g(t) = 1 we get by (24.25)

if (a , 6) = 1 , \v\ ^ 1 and T ^ 2 , the implied constan t bein g absolute . Henc e we

derive by partial integratio n tha t i n general

biv f T / / t \ iv\ «
Gv(T) = -j=j « , ( i ) ( c ( i - i „ ) + < ( i + it,)( —) )d * + 0(a6GT5 (log T) 6 )

where

c-tofOI + f 1 /(01 ^ .
For #(£ ) = (2TT/£)™/ 2 thi s yield s
COROLLARY 24.6 . Let (a, 6) = 1 , |v| ^ 1 and T ^ 2 . TAer a we Aave

/* T / a \ i t
/2TT\ Q
(24.48) I C (|+ii)C(i_rt-iv)(-) ( _ ) d t

W f^+uO / T w + C^) / Tx* | + 0 ( a 6 T l ( )7 )

v
^fab \ 2 + iv \27ra 2J 2 - iv \27rb 2
JJ \ B JJ

where the implied constant is absolute.

Now w e procee d t o the proof o f (24.23). Le t a^ be the coefficients of
2
^
558 24 . T H E C R I T I C A L ZERO S O F T H E R I E M A N N ZET A F U N C T I O N

where /3d = 7d/i(logd/logD) , thu s OLI < C r(£). W e hav e b y Corollar y 24. 5

fT\f(t)\2dt= f T
\t(l+lt)y\\+it)\Ht
Jo Jo
2
= TT*aa b(ab)-^ [ \a±+it)\ (a/bfdt
J
ab °
8
= AT(log ^- + 2 7 - l ) - BT + 0(D 6T9 (logT) )

where

(24.49) A =
J2J2J2 "adotbdiabd)- 1 ,
d (a,b)=l
1
(24.50) B = ^2J2^2a adabd(abd)- \ogak
d (a,6 ) =l

Therefore i t suffice s t o prov e tha t

(24.51) A < ( l o g I } ) - \
(24.52) B < 1 .

First w e treat A i n considerable detai l an d the n w e modify th e argument s t o reduc e

the cas e o f B t o tha t o f A.
We begi n b y diagonalizin g th e quadrati c for m A a s follow s

d6 a

= E d ~l (E / w1) ( E a ^a~1 )2 = E *(<*M2,

d S\d a d

where
a
Ad= Yl «1
a_ =
X I Pd 1 Pd2(did2)~1
a=0(mod d ) a=0(mo d d)
2
for d ^ D . Se t d x = 5 xk an d 5 2 = < M with 5i5 2 | d°°, d | 5i<J 2, (JM, d) = 1 getting

M2ld°°
d|M2

where
Ad(S)= J2 ^kk~\
(fc,d)=i
From no w o n w e tak e h{x) = 1 — x , s o tha t

These ar e the origina l coefficient s o f Selberg, howeve r on e could choos e h{x) almos t
arbitrarily subjec t t o th e condition s (24.1 5 ) an d (24.1 6) . I n thi s particula r cas e we
 24. TH E CRITICA L ZERO S O F TH E RIEMAN N ZET A FUNCTIO N 559

have a simpl e Melli n integra l representatio n

+
(24.54) log x = ^- [ xss 2 ds.
(e)

Now

Ad(S)logD= £ 76kk -Hog+^-=l6^-. [ Z(s + l)(D/5) s

s-2ds
m
{k,d)=i ^

where Z(s) i s th e correspondin g zet a functio n

z{s)= E i kk-2 = J{{i-p-^=Us)k{s)-^

(fe,d) = l p\d

with
CdW = II(1-P"T 1-
P\d

We hav e ( d{s + 1 ) < Cd(l ) = d/<p(d) , CO 5 + 1 ) > H " 1

an d (D/(5) s < 1 on th e lin e
R e s = £ with e — 1/ log D. W e als o hav e

2
/ \s\-^ ds^e-^2 = (logD)-2.
J(e)

Hence w e deriv e A d(S) < C |7a|(d/<p(d)logD)3. Thi s yield s A d < \(d)d/ip(d) log D,
where

A
(^) = X ! Z l i 7 ^ 7 ^ l ( ^ 1 ^ 2 ) ~ 1 < X ^ r_1
S Ttfi^ a = —T^v -
M 2 |d°° Hd° ° M 2=r ^ '
d\5\52 d\r

because \^s\ < 7 5 an d 7 • 7 = 1 . Therefor e w e hav e

2
(24.55) A d4Zd(<p(d))- (log D)- 1 .

Finally insertin g thi s t o th e diagona l for m o f A w e argu e a s follows :

1
A(\ogDf« £ dV3(d) = E ^ I K 2 2
-1 )"3
d^D d^D p| d^
1
« E^Ift +*"*)< E ^ E - ^
d ^ D 2 p| d d^D 2
™| d
_1
^c(|) E ^ «iog^
d^D2

which complete s th e proo f o f (24.51 ) .

For th e proo f o f (24.52 ) w e write logaf r = logafrd 2 — 21 og d an d w e spli t B ac -
cordingly, sa y B = B' — 2B". B y the argument s whic h le d u s to th e diagonahzatio n
of A w e deriv e simila r expression s fo r B' an d B". Thu s w e hav e
B
' = Yl ^( d ) Yl 5 Z ^aOLbiflb)- 1
log ab = 2 J2 P( d A
) dBd
d a,6=0(mo dd)d
560 24 . T H E C R I T I C A L ZERO S O F T H E R I E M A N N ZET A F U N C T I O N

where

a = 0 ( m o d d) q

by th e formula L — A • 1 . Applyin g (24.55 ) w e derive

Bd « ( Y.(d,q)A(q)q-1)d^-2(d)(\ogD)-1 « d^ 2
(d),
q^D*

This boun d i s larger tha n tha t fo r Ad by factor log/} , henc e the former argument s
yield B' < C 1 as required fo r (24.52) . Moreover , w e have
B =
" £ S S ^d^(a^)-1logd - ^^(d)4|,
d (a,6 ) =l d

where

<5|d S\d Sq\d

= ^A(q)qv(d/q) < 2 V (rf)^A(<;) = 2^(d)logd .

g|d qr| d

Hence B" ^ 4Alog£ > <C l by (24.51 ) . Thi s complete s th e proof o f (24.52).
Finally w e sketch a proof o f (24.24) . W e have
rT, /> A
2
1= \I(t)\ dt= / / f(t + u)dv dt
Jo Jo Uo
2
= / ( A - H ) / f(t)f(t + u)dtdu + 0(D Ti)
J - A JG

and b y the argument (Stirling' s formula ) whic h wa s applied i n the proof o f Lemma
24.2 w e ge t

/(()/(< + «) = ca + i()«l -"- * V ( ! + !<>*>2a - « - <»)(f)* + ° ( f f )

=ss> t .)cu-«-<.)(;r(^ + <)'
a 6
Integrating ove r t b y Corollary 24. 6 we ge t

1
/ f(t)f(t + iz)d * = 2 T ^ ^ J2 aadotbdiabd)-
^° d (o,b)= l

{^(VS"+^(Wfn+o<^s<^-
Now integratin g ove r u we ge t

(24.56) 7 = 2rVJ^^aodaM(a6d)-{$(^#) + * ( ^ )}
d (a,b)= l

+ 0(D 4 Tt (log T)7)

 24. T H E C R I T I C A L ZERO S O F T H E R I E M A N N ZET A F U N C T I O N 561

where
rA
v C(i + ^ )
A2 + iu
/ l I I - H1
7/ . I \
iu
= / (A-u)(X -X-iu)p-+ [ (A-\u\)X iu C
( ^u) - - M dl£
J0 v A ;
2 m 7_A V ' " V 2 + iu 2m /

Putting

(24.57) w(z) = /° ° f ^ ) d u

we writ e $ ( X ) = § A - Aw(AlogX ) + 0 ( A / l o g X ). Insertin g thi s fo r X

(T/2ir)i(abd)-1 int o (24.56 ) w e arriv e a t

I = -2ATA' + 0(ATA" + ATA'"(logX)' 1 + £> 4 Ti(logT) 7 )

where

A
' = E E E ^ « M W ^ ( A lo§ ^^).
d (a,b ) =l

d (a,b ) =l

^" = EEEia^awi(aM)_1-
d (a,6 ) =l
The su m A" ca n b e estimate d i n th e sam e wa y a s A,
A" ^ d Y [ ( l + ^)Al«(logD)-^d^l[(l + ^y «(lo gD)-K
d p\d d p\d

The su m A'" look s lik e A except tha t it s term s ar e take n wit h absolut e value .
Consequently, i n th e analyti c argumen t applie d fo r A'" th e Rieman n zet a functio n
occurs a s ({s) 1 /2 rathe r tha n C( 5 ) - 1 / ' 2 resultin g i n a loss o f facto r log D i n th e
bound fo r A!" b y compariso n t o A, i.e. , on e get s A'" < C 1.
For th e su m A' on e derive s th e sam e boun d a s fo r A , namel y A' < C ( l o g D ) - 1 ,
since th e functio n uj(A\ogX) i s bounded an d i t ha s nic e derivatives . Fo r example ,
one ca n us e th e Melli n transfor m o f uj(AlogX) i n X which bring s a sligh t shif t i n
the argumen t o f the involve d Dirichle t series , an d thi s make s n o essential chang e i n
the res t o f th e proof . Anothe r approac h woul d b e t o reduc e A' t o A b y arithmeti c
arguments a s w e reduced B' t o B. T o thi s en d writ e th e powe r serie s

" (-!)*+ ! (2s)2*+ l

^ ( 2 f c + 2) ! 2f c + l '
562 24 . T H E C R I T I C A L ZERO S O F T H E R I E M A N N ZET A F U N C T I O N

For z = A log X = I A log(T/27r ) — A log(a6d) thi s reduce s th e proble m t o estimat -

ing sum s o f typ e
A(k)
= ^SS^adaftdCaMJ-HlogaM)* .
d (a,6 ) =l

Writing
(logabd)k = Y, A*(« )
q\abd

where A ^ i s the vo n Mangold t functio n o f degree k w e arrange A( k' i n term s o f A^

and us e th e boun d (24.55) .
Now collectin g th e abov e estimate s fo r A'\ A", A'" w e ge t
/ < AT(lo g £>)_1 + L> 4 T§(logT) 7 .
Taking D = T^o w e complete th e proo f o f (24.24) .
R E M A R K S . Selberg' s mollification metho d work s well for countin g zero s on an d
near th e critica l line . H e showe d tha t
(24.58) N{\ + 4(5 , T) < T 1 '6 log T
uniformly fo r 5^0. Henc e i t follow s tha t almos t al l zero s o f ((s) li e i n th e regio n

where rj(t) i s an y positiv e functio n whic h increase s t o infinity .

 http://dx.doi.org/10.1090/coll/053/26

CHAPTER 2 5

THE SPACIN G O F ZERO S O F

THE RIEMAN N ZETA-FUNCTIO N

25.1. Introduction .
Throughout thi s chapte r w e assume th e validit y o f the Rieman n Hypothesi s fo r
C(s). Thi s allow s u s t o pu t th e critica l zero s p — | + r y i n a n increasin g sequenc e
according t o thei r ordinate s

(25.1) . . . ^ 7- 2 ^ 7- 1 < 0 < 7 1 < 7 2 < • • •

where 7_ n = — 7n . Recal l tha t th e countin g functio n N(T) = \{n | 0 < 7 n ^ T}\
satisfies

(25.2) N(T) = £ lo g ^ + 0(lo g T)

for T ^ 2 (se e Theore m 5.24) . Henc e i t follow s tha t

(25.3) 7 n ~ 27rn(logn) _1 , a s n - > cxo,

so th e number s

(25.4) C n= ^7nlog|7n |

have uni t mea n spacing , i.e. , ( n ~ n a s |n |— > oo . Actuall y (25.2 ) i s muc h mor e
precise tha n (25.4) , stil l i t i s no t satisfactor y fo r understandin g th e subtl e aspect s
of th e distributio n o f 7 n 's, i n particular , i t tell s ver y littl e abou t gap s betwee n
consecutive zeros . Th e Rieman n Hypothesi s alon e doe s no t she d enoug h ligh t o n
the proble m (i t doe s improv e th e erro r ter m i n (25.2 ) onl y b y facto r lo g log T) .
Our firs t goa l woul d b e t o understan d th e statisti c o f the sequenc e o f number s
£ n . On e expect s tha t th e consecutiv e spacing s

(25.5) 6 n = Cn+ i - C n
are no t purel y rando m (i.e . ar e no t poissonian) , a s expecte d fo r th e normalize d
distances betwee n primes , o r fo r th e normalize d distance s betwee n eigenvalue s o f
the Laplac e operato r fo r th e modula r grou p (se e th e final remark s o f Chapte r 1 0
and [Sa4 ] respectively) , bu t rathe r the y follo w th e Gaussia n Unitar y Ensembl e
distribution fro m rando m matri x theory . W e say that a sequence 0 < C i ^ C 2 ^ • • •
follows GU E i f
1 f°°
(25.6) - ] T f(6 n) ~ / f(s)P(s)ds

563
564 25. T H E SPACIN G O F ZERO S O F T H E R I E M A N N Z E T A - F U N C T I O N

for an y nic e functio n / : M +— > C (sa y o f Schwar z class) , wher e P(s) i s th e limi t
density distributio n o f consecutive spacin g o f eigenvalues o f random unitar y matri -
ces (th e limit wit h respec t t o the rank) . Explicitl y thi s distributio n wa s determine d
by Gaudi n an d Meht a [GM] , an d i s given b y P(s) = det( 7 — Q s), wher e Q s i s th e
integral operato r o n L 2([—1 ,1 ]) whos e kerne l i s
sin^(z-y)
(25.7) Qs(x,y)
TT(X - y)
Specifically th e densit y ca n b e expresse d b y th e followin g infinit e produc t
oo
(25.8) PO O = ]7(1 -A,-(a) )
0

where 1 ^ Xo(s) ^ Ai(s ) ^ .. . ar e th e eigenvalue s o f th e integra l operator .

In genera l th e GU E la w a s state d abov e i s very har d t o establish , becaus e har -
monic analysis cannot contro l consecutive points of a sequence, bu t i t can locate th e
points i n smal l domains . Therefor e mor e tractabl e problem s ar e abou t correlatio n
of sets o f points i n question . W e begin b y presentin g th e origina l wor k o f H. Mont -
gomery [Mo4 ] o n th e pai r correlatio n o f zero s o f th e Rieman n zet a function , an d
then w e stat e mor e genera l result s an d conjecture s fo r th e n- level correlatio n (se e
definitions i n Sectio n 25.3) .
25.2. Th e pai r correlatio n o f zeros .
The mai n too l fo r dealin g wit h th e correlatio n problem s i s the explici t formul a
of Rieman n (o r it s variations ) whic h connect s a su m ove r th e zero s o f £(s ) wit h a
sum ove r prim e numbers . Genera l version s ar e prove d i n Chapte r 5 , bu t her e w e
use a formul a i n somewha t customize d form . Le t g G C£°(R) an d le t
iur
(25.9) h(r) = / g(u)e du.
J — oo

Put TR(S) = 7r~ s / 2 r(«s/2), the local factor a t th e infinite plac e for th e Euler produc t
of C(s) . The n

EM7)-»(i)+*(-5) + s£*">{S( i + i r ) + S ( 1- i r ) }*
(25.10)

- £ ^ ( < ?n( l o g n ) + s ( - l o g n ) ) .
Vv
This i s th e cas e o f Theore m 5.1 2 fo r th e Rieman n zet a function .
The strateg y goe s a s follows . Firs t w e localiz e th e 7-terms , sa y 7 nea r t, b y
choosing a suitabl e tes t functio n h(r) = /i(r , £), wher e t i s a paramete r a t ou r
disposal (h(r) ha s t o b e a n entir e function) . O f course, b y th e uncertaint y principl e
of harmoni c analysis , suc h a localizatio n canno t b e exact ; a t bes t on e ca n se e r
near t a t th e distanc e c/ lo g t, wher e c is a positive constant . Computin g a properl y
weighted Z/2-nor m o f suc h 7-su m wit h respec t t o th e paramete r t involve d i n th e
test functio n ^(7 , t) on e pick s u p term s whic h ar e clos e t o th e diagonal , i.e. , on e
gets a su m ove r pair s o f zero s 7,7 ' wit h 7 — 7 ' bein g quit e small . Then , b y a n
approximation argument , o r b y Fourie r inversio n i n anothe r paramete r whic h i s
 25. T H E SPACIN G O F ZERO S O F T H E R I E M A N N Z E T A - F U N C T I O N 565

built i n th e tes t function , on e create s a su m whos e term s ar e th e desire d functio n

of th e differenc e 7 — 7'.
The origina l approac h o f Montgomer y wa s somewha t special , an d i t i s stil l
being use d b y man y researchers . Sinc e hi s choic e o f tes t function s i s quit e natura l
we use them first. Montgomer y [Mo4 ] introduce d th e function ( a Fourier transfor m
of th e gap s )

(25.11) F(a,T) = — J- J2J2 Ml-l')T^-^

0< 7 ,7'^T

for an y rea l a an d T ^ 2 , where w(u) i s a suitabl e localizin g function . Specificall y

he take s

(25.12) w(u) = 4 ( 4 + i^2 )\ - l

Note tha t F ( a , T) i s real an d F(a, T ) = F(—a, T ) ^ 0 , becaus e th e Fourie r trans -
form o f w i s positive. Precisely , w(v) = 27re _47r'vi, s o
2 2
(25.13) F(a,r) = - ^ re - M |£ e ^ + Q . o S T ) dv.
/ i o y1 1 _ _
0<7^T

By trivia l estimation , usin g (25.2) , on e derive s

(25.14) F(a,T) < F(0,T ) < logT .
Our goa l i s to giv e a n asymptoti c formul a fo r F ( a , T) a s T— > 0 0 which i s unifor m
in a i n range s a s larg e a s possible .
THEOREM 25. 1 (MONTGOMERY) . For 0 ^ a ^ 1 an d T ^2 we have
(25.15) F ( a , T) = a + T~ 2 a log T + 0 ( a T a ~ 1 + T ~ a log2T a + (logT)- 1 )
where the implied constant is absolute.
REMARKS. Fo r a wit h 2(lo g logT)/ log T ^ a ^ 1 - l/(loglogT ) lo g T w e have
(25.16) T(a , T) - a , a s T - > oc .
The asymptoti c formul a (25.1 5 ) hold s for al l a ^ 0 but i t lose s its meaning i f a ^ 1 ,
because th e erro r ter m 0(aT oc~1 ) exceed s th e leadin g ter m a.
In provin g Theore m 25.1 , Montgomery appeale d t o th e explici t formul a
(25.17) L(x,t) = R(x,t)
where L(x,i) i s a specia l su m ove r th e zero s o f £(s ) localize d nea r t , namel y
ri7
(25.18) L(x,t) = 2Y,j
+ (*-7)2'
and R(x, t) i s the correspondin g su m over primes, includin g th e infinit e plac e terms ;
see Sectio n 5.5 . W e hav e
00
A ( \ it
(25.19) R( X)t) = - Y ^ £ l ( -) minf-,- ) + x~ 1 +lt log(t + 2) + E(x,t)
566 25. T H E SPACIN G O F ZERO S O F T H E R I E M A N N Z E T A - F U N C T I O N

where th e las t ter m satisfie s

(25.20) £ ( M ) < -+ - ^
xt +V
if x ^ 1 and t > 0 , th e implie d constan t bein g absolute . Thi s particula r equatio n
(25.17) ca n b e deduce d fro m th e formul a

(25.21) £ A(„)„ - = -L W = l - Z1 - + X*
n^x p 1

which holds for x > 1 , x ^ p n an d s ^ 1 , p, —2n, or directly by a contour integratio n

of

s
(25.22) - ^ ( s )= ^A(n)rT
^l

and th e functiona l equatio n

(25.23) - £ ( S ) - £ ( 1 -s) = ^(s) + ^ ( 1- S ) = log(| S | + 3 ) + 0(1 ) .

We squar e th e equatio n (25.1 7 ) an d integrat e (thi s i s a wa y o f pickin g smal l

differences 7 — 7')? gettin g

2 2
(25.24) I \L(x,t)\ dt = [ \R{x,t)\ dt.
Jo Jo
The lef t sid e i s equa l t o

[T \L(x,t)\ 2dt = ^ ^ x ^ -^[ T

(l + (t->y) 2)-1 (l + (t-<y') 1
2
)- dt
Jo 7 y Jo

where the summatio n i s over al l the critica l zero s p — \ +27 , p' = \ Jri^' (wit h 7,7 '
positive an d negative) . Usin g (25.2 ) w e reduce th e abov e expressio n a s follows :

4 1
E E xih
~Y) A + (* - ^) 2)'1 (1 + (* - V) 2)_1 dt + 0(log3 T)

OO

(1 + (t - 7) 2 )" 1 (1 + (t - ify'dt + 0(log 3 T)

/
u^7 ,7'^i -°°
i(7 7 )
= 2T T ^^ x - ' tt;(7-7/) + 0(log 3r).
0<7,7'^ T

Putting x = T a wit h a ^ O w e arriv e a t

a
(25.25) / \L(T ,t)\2dt = F(a,T)T\ogT + 0(\og 3
T)
Jo
where th e implie d constan t i s absolute .
 25. T H E SPACIN G O F ZERO S O F T H E R I E M A N N Z E T A - F U N C T I O N 56 7

Next w e evaluat e th e othe r sid e o f (25.24) . T o thi s en d w e us e th e followin g

form o f Parseval' s formul a fo r Dirichle t serie s (se e [MVl] , o r refin e th e proo f o f
Theorem 9.1 )

u 2
(25.26) / \Va nn- \ dt = Y\{T + 0{n))\a n\ .

Jo I 1 I !

Write (25.20 ) a s R(x,t) = C(t) + D(t) + E(t), respectively . The n

\R(x,t)\2 = \C(t)\ 2 + \D(t)\ 2 + \E(t)\ 2 + 0(\C(t)D(t)\ + \D(t)E(t)\ + \E(t)C(t)\).

Integrating i n 0 < t ^ T w e deriv e b y th e Cauchy-Schwar z inequalit y

(25.27) / \R(x, t)\ 2dt = C 2 + D 2 + E 2 + 0{CD + DE + EC),
Jo

with C 2 = / \C(t)\ 2
dt an d D 2,E2 define d similarly . B y (25.26 ) w e obtai n
Jo

C--E(r + o(»„^nta»(=,H).

By th e Prim e Numbe r Theore m I/J(X) = x - f 0{x/\og2x) thi s is

2
(25.28) C = Tlogx + 0 ( T + xlogx) .

For £> 2 w e ge t simpl y

(25.29) L > 2 = £ - 2 / log 2

(t + 2)d t = x - 2 T l o g 2 T + 0 ( x - 2 T l o g T ) ,
Jo

and fro m th e erro r ter m (25.20 ) w e ge t

2 2
(25.30) E <^x~ T-{-x.

Inserting (25.28) , (25.29) , (25.30 ) int o (25.27 ) w e arriv e a t

2
(25.31) / \R(x, i) \2dt = T log x + x~ 2T log T + B(x,T)
/o
Jo
where

(25.32) B(x,T) <z:T + xlogx + x-- i l

(log2x)T\ogT.

Comparing (25.31 ) fo r x = T a wit h (25.25 ) w e complet e th e proo f o f (25.1 5) .

Now w e deriv e a fe w consequence s o f Theore m 25.1 .

568 25. TH E SPACIN G O F ZERO S O F TH E RIEMAN N ZETA-FUNCTIO N

THEOREM 25.2 . Let f(x) be a Schwartz function on R such that its Fourier
transform
oo
f(x)e(-xy)dx
/ -OO

is of class C1 with s u p p / C (—1 ,1 ) . Then

£$>(7-Y)/((7-y)^)
0<
(25.34) ™'^T

ii
f(a)\a\da +
}
f(0)}— lo
2?r
g T + O(T)

w/iere £/i e implied constant depends only on f.

PROOF. Fo r an y / o f Schwart z clas s w e hav e

(25.35) 5 ^ X T M 7 - 7 ' ) / ( ( 7 - 7 ' ) - ^ - ) = ^ ( l o g T ) / f(a)F(a,T)da.

Here th e integra l equal s

/ ( / > ) + / V ^ ) ) {« + T -2 a l o g T 4 - 0 ( T - - 1+ T - - l o g 2 r a + (logT)- 1 )}da.

JO
The firs t ter m agree s wit h tha t o n th e right-han d sid e of (25.34). Th e secon d ter m
is

2(logT) / (/(0 ) + 0{a))T-2«da = /(0) + O^logT)- 1

Jo
and th e erro r ter m i s triviall y estimate d b y 0 ( 1 / log T). Thi s complete s th e proo f
of (25.34) . •

In particular , choosin g f(x) = ((sin7rax)/7rax)2 w e ge t

COROLLARY 25.3 . IfO < a < 1 is fixed, then as T -^ oo,

0<7,7'^T \ w / /2o /

From thi s Montgomer y deduce d a lower boun d fo r th e numbe r o f simpl e zero s

ofC(s),
(25.37) JVi(T ) = |{ 0 < 7 ^ T ; p = | + 27 i s simple}| .
(recall w e assum e th e Rieman n Hypothesis) .
COROLLARY 25. 4 (MONTGOMERY) . As T -^ 00, we have

(25.38) j Vl (T)>(?+o(l))^logT,

i.e., at least 2/3 of the zeros are simple.

 25. T H E SPACIN G O F ZERO S O F T H E R I E M A N N Z E T A - F U N C T I O N 569

PROOF. Discardin g al l term s i n (25.36 ) excep t fo r 7 = 7 ' w e ge t

E „ ; < (i + i + < ) ( 1 ) )£ l o g r ,

where ra 7 i s the multiplicity of p = \+i')- Thi s holds for any fixed a wit h 0 < a < 1 ,
the bes t boun d bein g | + o(l ) fo r a — > 1 —. Hence w e deriv e tha t

Wi(T)^ J ] ( 2 - m 7 ) m 7 > ( 2 - | + o(l))^log T

completing th e proo f o f (25.38) . •

REMARK. Th e tes t functio n whic h wa s use d i n (25.36 ) i s no t optima l fo r th e

purpose o f estimating N\(T) alon g th e abov e lines , s o the percentag e 2/ 3 o f simpl e
zeros ca n b e slightl y increased .

Another consequenc e o f Theore m 25. 2 (whic h w e d o no t dra w here ) i s tha t

7n+i — In ca n b e smalle r tha n it s averag e 27r/log7 n b y a facto r A < 1 infinitel y
often, i.e. , limin f 5 n ^ A . Montgomer y succeede d i n showin g thi s wit h A = 0.68 .
The bes t publishe d resul t is A = 0.51 7 1 whic h i s du e t o Conrey , Ghos h an d Gone k
[CGG], whil e a ne w refinemen t point s t o A = 0.51 69... . Recentl y i t wa s show n
that i f 5 n ^ \ — £ sufficiently often , the n th e clas s numbe r h(—D) o f the imaginar y
quadratic field Q(y/—D) i s bounde d belo w b y c{e)D^ (log D)~2 fo r som e constan t
c(e) effectivel y computabl e i n term s o f e\ se e [CI2] .
The formul a (25.1 6 ) hold s uniforml y fo r e < a ^ 1 — e. Wit h a littl e mor e
effort on e coul d sho w tha t (25.1 6 ) hold s uniforml y fo r e ^ a ^ 1 . On e ma y wonde r
what i s th e tru e behavio r o f F(a,T) fo r a ^ 1 ? Montgomer y showe d tha t (25.1 6 )
cannot hol d fo r a > 1 . Moreover , h e describe d heuristi c argument s whic h sugges t
that
(25.39) F(a, T) ~ 1 , a s T - > 00
uniformly i n bounde d interval s 1 ^ a ^ A. B y applyin g thi s t o (25.35 ) wit h
appropriate tes t function s / o f Schwartz class , Montgomery wa s led to the followin g
PAIR CORRELATIO N CONJECTURE . For a < f3 put

iV(a,/3;T) = | { m ^ n ; 0 < 7 m,7n^T,^ < lrn - ln < |^}| .

Then, as T—> • 0 0 we have

2.
(25.40) N(a, /? ; T) ~ N{T) j f ( l - ( ^ )) du.

REMARKS. Th e assertions (25.39 ) for a ^ 1 together with (25.1 6 ) for 0 < a ^ 1

are essentiall y equivalen t t o (25.40) . B y (25.39 ) th e asymptoti c formul a (25.36 )
extends t o al l a ^ 1 with th e functio n a~ 1 + a / 3 replace d b y 1 + l / 3 a 2 . Henc e
letting a —> 0 0 i t follow s b y th e forme r argumen t tha t almos t al l zero s o f £(s ) ar e
simple.
570 25. T H E SPACIN G O F ZERO S O F T H E R I E M A N N Z E T A - F U N C T I O N

25.3. Th e n-leve l correlatio n functio n fo r consecutiv e spacing .

The Pai r Correlatio n Conjectur e o f Montgomer y indicate s tha t th e zero s o f
((s) behav e muc h lik e the eigenvalue s o f large comple x Hermitia n rando m matrice s
whose statistica l distributio n i s known a s the Gaussia n Unitar y Ensemble . Furthe r
support fo r thi s behavio r wa s provide d b y Z . Rudnic k an d P . Sarna k [RS ] wh o
extended Montgomery' s result s t o th e n-leve l correlatio n sum s fo r th e normalize d
zeros o f £(s) . I n thi s sectio n w e formulate th e mai n result s o f Rudnic k an d Sarna k
and w e give a fe w comment s abou t thei r arguments .
Recall tha t ( m denot e th e normalize d ordinate s o f th e zero s p m = | + ij m o f
((s) give n by (25.4) , they ar e ordered i n increasing sequenc e . . . < £_ 2 ^ C- i < 0 <
Ci ^ C 2 ^ • • •, an d C n hav e uni t mea n spacing . Th e Pai r Correlatio n Conjectur e
answers th e questio n o f ho w ofte n th e difference s £ m i — £ m2 fal l i n a give n interval .
One may inquir e abou t pair s o f differences ( mi — C™25 Cm 2~~ Cm3 falling int o a give n
rectangle, th e tripl e o f difference s i n a box , an d s o on . T o contro l th e proble m w e
restrict ou r consideration s t o th e firs t M point s £ m , 1 ^ m < M , wher e M i s a
large number . Le t n ^ 2 and B C R n _ 1 b e a bo x o f dimensio n n — 1 . Pu t

Rn(M]B) = — :|{1 ^ m i , . . . , m n ^ M al l distinct ;

( C m i~~ ~ Qm2i Cm 2~ ~ Cm 3 • • • > Qm n-1 ~ Qm n) £ B\\.

For n = 2 and B = (a , /3) this i s simpl y

R2{M-B) = ^ | { 1 ^ rm ^ m 2 < M ; a < £ m i - C™ 2 < P}\.

For M = N(T) thi s i s very clos e t o N(a, /? ; T)/N(T). A s i n (25.41 ) w e expect tha t
there i s a functio n R n(B) suc h tha t

(25.41) R m(M; B) ~ R n(B), a s M - > o o .

Our goa l i s t o determin e R n(B) fo r an y n ^ 2 .

To simplify th e Fourie r analysi s her e we introduce tes t function s / ( # i , . . . , x n)
of th e followin g type :
TF1. / ( # i , . . . , £n ) i s symmetric ,
TF2. / O n + £ , . . . , s n + * ) = / ( * ! , . . . ,x n) fo r an y t e R ,
TF3. / ( # i , . . . , xn )—> > 0 rapidly a s |x |— > 0 on th e plane s x\ + • • • -f - x n = 0 .
REMARKS. Th e conditio n TF 2 say s tha t / ( # i , . . . , x n ) depend s onl y o n th e
successive differences . Thi s togethe r wit h TF 3 implie s tha t / ( # i , . . . ,x n ) decay s
rapidly o n an y hyperplan e x\-\ \-x n = t. Fo r n = 2 such a functio n i s given b y
f{x\,X2) — f(x\ — X2) where f(x) i s a n eve n functio n o n R whic h decay s rapidl y
to zer o a s \x\ —> 00 .

For an y tes t functio n a s abov e w e pu t

(25.42) R n(M;f) = ^ £ • • • £ /(U>-- - >CmJ ,

all distinc t
 25. TH E SPACIN G O F ZERO S O F TH E RIEMAN N ZETA-FUNCTIO N 571

and see k a n asymptoti c formul a

(25.43) R n(M; f) ~ R n(f), a s M - + oo.

The multipl e su m Rn(M; f) ca n be smoothed furthe r b y replacing th e range of

summation 1 < m i , . . . , mn ^ M b y a smoot h (rapidl y decreasing ) cut-of f function .
Thus w e also conside r
1 1
(25.44) R n(T; / , ft) = £ \ • • E Kl^T' )... h ^ T " ) / ^ ^ , .. . , 7 m nL )
mi,... ,m n
all distinc t

where h(y) i s the cut-off functio n i n question and

(25.45) L - ^ l o g T .

REMARK. Ther e i s no reason t o use a cut-off functio n h(yi, ... , yn) mor e gen -
eral tha n th e produc t h{y\) ... h(y n) becaus e w e wish th e point s 7 m i , . . . , 7mn t o
be clos e t o each other .

In 1 962 , Dyson [D ] considered a statistical distributio n o f energy level s of some

complex system s i n whic h contex t h e determine d th e n-leve l correlatio n densit y
Wn(x\,... , xn) fo r the GUE model. H e showed tha t
(25.46) W n(xu • • • , xn) = det(K(x 2 - Xj )),
where

(25.47) K(x) = ^ .
7TX
One ca n show th e following propertie s o f W n(x):

(25.48) 0 < W n(x) < 1 ,

(25.49) W n{x) = 0 if and only i f xi = Xj fo r som e i ^ j ,
n
(25.50) W n{x) = 1 if and only i f x G Z an d x^ 7^ Xj for i ^ j .

Let u s compute th e density functio n fo r n = 2 . W e get

/ /i f (0) i f (xi - s 2 ) \ = / s i n 7 r ( x 1 -x2)N 2

W 2 (xi,x 2 ) = de t „ , v „ ,v 1 = - .- -

which agree s wit h tha t i n (25.40) .

T H E O R E M 25. 5 (RUDNICK-SARNAK) . Suppose that the cut-off function h(r) is
given by the Fourier integral (25.9) with some smooth compactly supported function
g(u). Suppose the Fourier transform of the test function f{x),

HO = I f{x)e{S • *)dx
is supported in the polyhedral domain |£i | + • • • + |£ n | ^ 2 . Then, as T—> • o c we
have

(25.51) R n(T;f,h) ~ N(T)( [°° h(r) ndr) [ f(x)W n(x)s(-(x

1 + -^x n))dx
572 25. T H E SPACIN G O F Z E R O S O F T H E R I E M A N N Z E T A - F U N C T I O N

where 5(x) is the Dirac distribution at point 0 .

By makin g an appropriate approximatio n t o the cut-off functio n on e can deriv e

from (25.51 ) th e following
COROLLARY 25.6 . Suppose the test function f(x) has its Fourier transform
/(£) supported in |£i | H + |£ n | ^ 2 . Then we have (25.43) with

(25.52) R n(f)= l f(x)W n(x)6(-(x

1 -h-^xn))dx

where x = (xi , ...x n). This can be written as

(25.53) R n(f) = n / f(x 11,...,xn1

)Wn(x 1
,...,xn)dx ...dxn-
JRn-l
where x n — —x\ — .. . — xn.
REMARKS. I n particular, w e hav e

Mf) = 2J f(x,-x)W 2(x,-x)dx = Jf(x)(l- ( ~ ^ ) )dx

where f(x) = / ( x , 0 ) , whic h agree s wit h th e Montgomery pai r correlatio n result .

25.4. Low-lyin g zero s o f L-functions .

A somewha t differen t proble m fro m th e consecutive spacin g o f zeros o f an L-
function concern s th e zeros nea r th e central point , i.e. , in diminishing distanc e t o
s = | . Suc h a problem , o f course, depend s o n the scaling o f zeros. Conside r a n
L-function o f degree d (stil l satisfyin g GRH ) a s in Chapter 5 , given by
oo

(25.54) L ( / , 5) = ^A/(n)n- s,
l

with conducto r Cf > 1 (see (5.7), (5.8)) . Th e number o f zeros of L ( /, s)

(25.55) Pf = \+ilf
with 1 7/ 1 ^ T satisfie s
dT
(25.56) N (T,f)~— logc fT

as T—> • o o by Theorem 5.8 . Thi s suggests that th e proper scalin g factor fo r the low-
lying zero s o f L(/, s) shoul d b e ^ logc/ . No w th e question is : ho w does N(T, f)
behave fo r small T?
To make the question mor e interesting we take a test functio n 0 in the Schwart z
class on R and define th e low-zeros sum

(25.57) W* ) = I>(7 'llr)'

If

where 7 / run s ove r th e ordinates o f zero s counte d wit h multiplicity . Fo r</> with
rapid deca y thi s su m concentrates o n zero s whic h ar e within 0 ( 1 / log Cf) o f the
point s = \.
 25. T H E SPACIN G O F ZERO S O F T H E R I E M A N N Z E T A - F U N C T I O N 573

Because, fo r a singl e L-function , w e ar e effectivel y countin g ver y fe w zeros ,

classical harmoni c analysi s canno t detec t thei r distribution . Therefor e w e mus t
take a sufficientl y larg e famil y o f L-function s int o consideration . Suppos e T i s a n
infinite famil y ordere d b y th e conductor . Pu t
(25.58) T{Q) = {feT;c f^ Q}.
Now w e want t o kno w th e asymptoti c behavio r o f th e averag e densit y

(25.59) AD(F-A,Q) =-±— £ D(f,</>).

If J 7 i s a complet e famil y i n a certai n spectra l sense , on e shoul d expec t tha t

(25.60) AD(T; 0 , Q) ~ / <j>(x)W(T)(x)dx

JR
where W(J r)(x) i s a densit y functio n characterize d b y J-'.
Some investigation s i n thi s directio n wer e undertake n i n th e 1 970' s (se e M .
Jutila [Ju5] , H . L . Montgomer y [Mo4] ) fo r th e famil y o f Dirichle t L-function s
L(s,x)- However , th e recen t studie s b y N . Kat z an d P . Sarna k [KS1 ] , [KS2 ]
expanded fa r beyon d thi s family . B y examinin g variou s familie s T the y reveale d
that th e low-lyin g zero s are alway s governed b y a symmetry (o r monodromy) grou p
G{F) associate d wit h T. Thei r assertion s ar e particularl y convincin g fo r zet a an d
L-functions ove r finite fields, no t onl y becaus e th e Rieman n Hypothesi s i s know n
to b e true , bu t als o du e t o connection s wit h geometr y whic h provide s additiona l
intuition (th e zeros are then eigenvalues of the Frobenius operator, se e Section 1 1 .1 1
and [KS1 ]) .
There ar e als o attractiv e result s fo r familie s o f global L-functions , bu t progres s
is no t ye t a s dee p a s i n th e cas e o f loca l L-functions . I n thi s sectio n w e presen t a
few suc h results .
First w e inspec t th e famil y o f Dirichle t L-function s wit h rea l primitiv e char -
acters Xd- Her e L(s,Xd) ha s th e conducto r Cd = \d\. Sinc e L(s,Xd) hav e eve n
functional equations , w e ma y pu t thei r zero s Pd \ + i^ int o th e sequenc e

(25.61) . . . < 7 !f2) ^ 7^< 0 < 7 (J) < 7 f< • • •

with 7^ ~ — 7^ j fo r £ G Z\{0} (i t i s believed tha t L ( | , Xd) i s never zero) . Le t

(25.62) F{Q) — {d fundamenta l discriminant ; \d\ ^ Q}.

CONJECTURE ( K A T Z - S A R N A K ) . AS Q -> oo, we have

1 V
^ ; | der(Q)t>o JR

where the density function is given by

(25.64) W s p ( ; E ) = i_!^ .
574 25. TH E SPACIN G O F ZERO S O F TH E RIEMAN N ZETA-FUNCTIO N

REMARKS. Kat z an d Sarnak [KS2 ] proved thi s conjecture fo r any test functio n
(j) whos e Fourie r transfor m

(25.65) 0 ( 0 = / <l>(x)e(-£x)dx

is supporte d i n (—2 , 2) (o f course, subjec t t o the validity o f the GRH).

The cas e o f L(s,Xd) * s obtained b y twisting ((s) wit h Xd- On e may generaliz e
this cas e b y startin g wit h an y L(f, s) i n plac e o f £(s ) an d takin g th e famil y o f
L-functions L(f 0 Xd, s) twiste d b y the real character s Xd> Fo r example, choos e a n
elliptic curv e E/Q, an d le t E^ /Q b e th e correspondin g quadrati c twists . The n
the associate d L-function s L(E^ d\s) ar e obtaine d fro m L(E,s) b y twistin g wit h
Xd> The correspondin g conjectur e o f Kat z an d Sarna k predict s tha t th e densit y
function fo r the low-lying zero s o f L(E^ d\ s) i s given by

TTr+/ v . . sin27r x
(25.66) W+ (x) = 1 + — .
Z7TX

Many interesting examples of families o f L-functions ar e offered b y the theory of

automorphic forms , an d a variety o f these ar e studied b y Iwaniec, Lu o and Sarna k
[ILS]. W e are going t o describ e som e o f their results . Le t T = 5Z/2(Z ) an d k ^ 2
be a n eve n integer . Le t Sk(T) b e the linea r spac e o f cusp form s fo r T o f weight k.
Let iffc(r ) b e the Hecke basi s o f primitive form s o f Sfc(r). T o any
oo

we associat e th e Hecke L-functio n

L(f, s) = £ \ f(n)n-s = JJ(1 - \ f(p)p-> + p-

1p

As describe d i n Chapte r 1 4 , the completed functio n

A(/ ) S ) = (27r)-T( S + ^ i ) L ( / , S )

is entir e an d satisfie s th e functiona l equatio n A(/ , s) = £/A(/ , 1 — s) wit h roo t

number give n b y ej — i k — ± 1 (not e Sf depend s onl y o n k but no t on / ). Becaus e
the sig n of the functional equatio n ha s some impact o n the distribution o f low-lying
zeros w e separate th e families o f even an d odd L-functions . Pu t
H
(25.67) M+(tf ) = J2 \ kF)\,
k^K
/c=0(mod 4 )
ff
(25.68) M~(K)= J2 l *(r)l-
k^h
k^K
k=2(mo 4 )
/c=2(mod

The correspondin g densitie s tur n ou t t o be

s
(25.69) W +(x) = 1 + E^i,
Z7TX
 25. T H E SPACIN G O F ZERO S O F T H E R I E M A N N Z E T A - F U N C T I O N 57 5

(25.70) W -(x) = i-*^E +So(x).

Zirx

THEOREM 25. 7 (IWANIEC-LUO-SARNAK) . Suppose </> is a smooth function on

R compactly supported in } — 2 , 2[. Then, as K —> oo , we have
(25.71)
77^ E E D{fA)~ U{x)W+{x)dx,
\ * h^ k^Kif feH
fan. fr\ ^^
k(T)
fc=0(mod 4 )

(25.72)
=7jn E E D{fri)~ I <l>{x)W-{x)dx.
M- R
y ) iris
k^K feH tJrr^s
fe=2(mod 4 )
k(T) JR

Let L(Sy m / , s) b e th e symmetri c squar e L- function associate d wit h / . Recal l

from Sectio n 5.1 2 tha t i t i s the L-functio n o f degre e 3 given b y
oo

L(Sym 2 /, s) = ((2s) £ A/(n 2 )n" s

1

(25.73) = FJ( 1 - A/(P2)P-S + Xf( P2)p-2s - p-'T 1

-
P

As show n b y Shimur a [Sh2] , th e complete d functio n

2
(25.74) A(Sym /,s) = ^ 3 ^T{^)T(^^)T(^)L(Sym2 f,s)

is entire an d satisfie s th e functiona l equatio n A(sym 2 /, s) = A(Sym 2 / , 1 — s) (not e

that th e roo t numbe r i s alway s one) . I n thi s cas e w e hav e
THEOREM 25. 8 (IWANIEC-LUO-SARNAK) . Suppose 4> is a smooth function on
E compactly supported in } — | , | [ . Then, as K— > oo ,

(25J5)
MK) ^ feH£ (T) D{Sym2
^ ' fc^tf k
JR
/; 0)
~ / ^^M*)***
/e=0(mod 2 )

w/iere Af(if ) = M + ( X ) + M~(K) and W sp(x) is given by (25.64).

REMARKS. W e emphasiz e tha t th e condition s o n th e tes t functio n 0 i n The -
orem 25. 7 an d Theore m 25. 8 allo w th e suppor t o f th e Fourie r transfor m <j) (se e
(25.65)) t o b e large r tha n [—1 ,1 ] . T o appreciate thi s fac t w e writ e

/ <t>{x)W{x)dx = / 4){y)W(y)dy

by Plancherel' s theorem . Her e th e Fourie r transform s o f th e correspondin g densit y

distributions are :

W+(y) = S0(y) + lv(y),

W-(y)=50(y)-^(y) + h
wsp(y) = s0(y) -\ n{y)
576 25 . T H E SPACIN G O F ZERO S O F T H E R I E M A N N Z E T A - F U N C T I O N

where So(y) i s th e Dira c distributio n an d rj(y) i s th e characteristi c functio n o f th e

segment [—1 ,1 ] . Thes e Fourie r transform s hav e discontinuit y a t y — 1 and y = — 1,
therefore th e result s ar e distinguishabl e onl y i f th e suppor t o f 4> cover s th e point s
y = 1 and y = — 1.
 http://dx.doi.org/10.1090/coll/053/27

CHAPTER 2 6

CENTRAL VALUE S O F L-FUNCTION S

26.1. Introduction .
We hav e alread y mentione d man y time s th e Gran d Rieman n Hypothesi s fo r
//-functions an d discussed man y aspect s o f the fascinating zero s of L-functions (se e
Chapters 5 or 2 4 for instance ) i n various aspects .
Recently, muc h o f the analyti c theor y o f L-function s focuse d o n estimate s fo r
the value s o f L-function s a t th e centra l critica l poin t s — |. Thes e specia l value s
appear i n various context s an d vanishing o r non-vanishing ar e the mai n questions .
For example , algebrai c technique s develope d b y Mazu r t o stud y rationa l point s
on certai n modula r curves , wit h application s t o diophantin e equations , requir e fo r
full effec t t o prov e tha t ther e exist s a modula r for m o f a certai n typ e whic h doe s
not vanis h a t th e centra l poin t (se e [Ell ] fo r a recen t example) . Anothe r ver y
different situatio n i s the surprising lin k discovere d b y Iwaniec an d Sarna k betwee n
the proportio n o f non-vanishin g o f variou s familie s o f L-function s an d th e Gaus s
Class Numbe r Proble m (se e Theorem 26. 1 below).
Here i s a shor t panoram a o f selected accomplishment s i n recent years , empha -
sizing th e strongest result s fro m analyti c numbe r theory .
The firs t i s the Iwaniec-Sarnak resul t mentione d abov e [IS2] . Amon g differen t
variants, w e choose th e following :
THEOREM 26.1 . For k = 0 (mo d 4), let Hk be the set of weight k Heche forms
on 5L(2,Z) . For any real primitive character if; modulo q with q ^ k°, we have

(26.1) \{f€Hk | |L(/®V,I)|^^-i^}|^(i- 0 (l))|Ffc|

as k -» +oo and i?—> 0.

They als o sho w tha t

J2 L(/,|)L(/®x,|)~i(x , 1 )1 ^ 1
feHk
for an y rea l primitiv e characte r x modul o D , uniforml y fo r D < k' 0 a s k - » +o o
and # —> 0. Hence , i f the constant \ i n (26.1 ) fo r ip — 1 could b e replaced b y \ + S
for an y 5 > 0, with a n effectiv e implie d constant , i t woul d follo w tha t

L ( 1 x ) > >
' ( ! ^
for an y real primitiv e characte r x modul o D wit h a n absolute an d effective implie d
constant.
577
578 26. C E N T R A L VALUE S O F L - F U N C T I O N S

Note i t i s importan t t o hav e a lowe r boun d a s i n (26.1 ) instea d o f simpl y

•^(/J \) 7 ^ 0> so i t i s conceivabl e tha t al l specia l value s ar e non-zero , ye t to o smal l
to solv e th e Clas s Numbe r Proble m alon g thes e lines . Fro m th e poin t o f vie w o f
analytic numbe r theory , a lowe r boun d a s i n (26.1 ) an d simpl e non-vanishin g ar e
usually indistinguishable .
In par t becaus e o f th e relationshi p t o th e Birc h an d Swinnerton-Dye r Conjec -
ture, th e orde r o f vanishing o f weight 2 cusp forms a t 5 = \ ha s attracte d particula r
attention. Recal l tha t i n on e instanc e i t i s easy t o guarante e tha t th e centra l poin t
is a zero , namel y i f L(/ , s) i s a self-dua l L-functio n an d th e sig n o f th e functiona l
equation i s —1 . I t i s reasonabl e t o assum e tha t th e orde r o f vanishin g wil l b e th e
smallest compatibl e wit h thi s restriction , namel y a "generic " eve n for m wil l hav e
L{f, \) 7 ^ 0 an d a generi c od d for m ! / ( / , \) ^ 0 . Brume r (an d independentl y
Murty) conjecture d thi s an d use d GR H t o giv e som e evidence .
Iwaniec an d Sarna k prove d tha t hal f th e specia l value s o f eve n form s d o no t
vanish (an d ar e no t to o small , a s i n (26.1 )) . Kowalsk i an d Miche l [KM1 ] , an d
independently VanderKa m [vdK ] considere d highe r derivatives . W e quot e tw o
results:
T H E O R E M 26. 2 (Kowalski-Michel) . Let 5 2 (g)* be the set of primitive weight 2
forms of level q. We have

\{f e S 2(q)* I e f = - 1 and L'(f, \) + 0} | > ( ^ - o(l) ) |5 2 (g)*|

for primes q —> +00 .

THEOREM 26. 3 (Kowalski-Michel-VanderKam) . For any fixed k ^ 0 we have
\{f € S 2{qY I e f = (-l) f e and L^ k\f, § ) ^ 0} | > |(p fc - o(l))\S 2(q)*\

for primes q — > -hoO ; where po = 1 /4 , p\ — 1 /l§, P2 — 0.4825; p% = 0.495 ; and

moreover, p k = \ - ^ + O(^) .
Although th e erro r ter m depend s o n /c , it i s possible to combine this result wit h
an averag e boun d
2
Y, (ordL(/,,)) «l
s
fes2(q)* =2
for q prime an d deriv e (se e [KMV3 ] fo r bot h results )
COROLLARY 26.4 . We have
Y oTdL(f,8)^(c + o(l))\S 2(qy\
fes2(qys=2
for primes q —>• +0 0 with c — 1.1891.
This i s particularl y interestin g becaus e 1 .1 89 1 < | , whic h wa s th e valu e ob -
tained b y Brume r o n GRH .
In th e followin g section s w e will giv e considerabl e detail s o f the proo f o f Theo -
rem 26.2 . Thi s result expand s a n earlier on e of B. Duke [Du3 ] where the proportio n
was g/(logg) 4 .
Those non-vanishin g result s hav e nic e interpretatio n i n term s o f arithmeti c
geometry, a s alread y mentione d i n Sectio n 5.1 4 . Indee d associate d t o th e curv e
Xo(q) = To(q)\M i s a n abelia n variety , it s jacobia n variety , denote d Jo(q), whic h
 26. CENTRA L VALUE S 579

is o f dimensio n di m Jo (g) = di m S2(q) = \S2(q)*\ (th e latte r onl y fo r q prime) .

The jacobian i s defined ove r Q an d it s grou p o f rational point s i s finitely generate d
by th e Mordell-Wei l theorem . Eichle r an d Shimur a compute d th e Hasse-Wei l zet a
function o f Jo(q) i n term s o f modula r forms : fo r q prime i t take s th e for m

L(J0(q),s)= J ]L (f>8)'
fes2(q)*
If th e Birc h an d Swinnerton-Dye r Conjectur e fo r abelia n varietie s hold s fo r J 0 (g) ,
we hav e
r a n k J 0 ( g ) ( Q ) = or d L(J 0(q),s)= V or d L(/,s) ,
/
s=l 2 —^ s=l/2
so lower bound s fo r th e orde r o f vanishin g giv e lowe r bound s fo r th e ran k o f Jo{q)-
In fact , b y th e genera l for m o f th e Gross-Zagie r formul a (Theore m 23.4 ) fo r a n
arbitrary weigh t 2 cusp form , on e show s tha t i f th e orde r o f vanishin g i s exactl y 1 ,
then / contribute s 1 to th e ran k (b y conjugate s o f Heegne r points) . S o Theore m
26.2 implie s
THEOREM 26.5 . For q prime, q -> +oc , we have

rank J 0(q)(Q) > ( — - o(l) ) di m J 0(q).

On th e Birc h an d Swinnerton-Dye r Conjecture , Corollar y 26. 4 give s a corre -

sponding fairl y goo d uppe r boun d fo r ran k Jo(q)- Ther e i s currently n o comparabl e
unconditional bound .
Recall fro m Propositio n 5.2 1 that , o n GRH , th e maxima l orde r o f vanishing o f
an L-functio n o f degre e d wit h conducto r q a t th e centra l critica l poin t i s abou t
(log g) (log | l o g g ) _ 1 . Assumin g th e Birc h an d Swinnerton-Dye r Conjecture , thi s
translates t o a conjectural uppe r boun d fo r th e maximum ran k o f an abelia n variet y
over Q . On e ca n se e i t i s bes t possibl e b y takin g power s o f a n ellipti c curv e wit h
positive rank . However , mor e convincing i s the cas e of Jo(q) (whic h i s conjecturall y
"almost" irreducible) . Fo r q prim e on e get s b y GR H an d th e Petersso n formul a
the slightl y bette r estimat e (5.94) . Th e conducto r i s N q = g dimJ o(g) 5 henc e b y
dim Jo(q) ~ #/1 2 w e hav e di m Jo(q) ~ (logN q)(\og\ogNq)'1 . S o Theore m 26. 5
shows tha t th e uppe r boun d o n GR H canno t b e improve d (excep t fo r th e implie d
constant).
Of cours e on e ca n als o stud y classica l Dirichle t L- functions. Miche l an d Van -
derKam [MvdK ] studie d th e analogu e o f Theorem 26. 3 for th e famil y o f primitiv e
characters modul o q. A different challeng e i s to conside r onl y rea l character s mod -
ulo q ^ Q. Ther e Soundararaja n [Sou ] ha s prove d
THEOREM 26.6 . Let Q be the set of odd squarefree integers. For q E Q, let
n n
X8q( ) — (&q/ ) be the even quadratic character modulo Sq. Then

\{q<Q\ q£QandL(lx8 q)^0}\^Q-o(l))\{q^Q \ q € Q}\

as Q —> • + 0 0 .

Notice the coincidence of the proportio n 7/ 8 wit h Theorem 26. 2 (wher e a facto r
I reall y come s fro m consideratio n o f od d forms) . I n fact , Soundararajan' s proo f
580 26. CENTRA L VALUE S O F L-FUNCTION S

uses a mollifie r M{\) a s i n Theore m 26. 2 (se e Sectio n 26. 2 an d following) , an d

remarkably fo r a mollifier o f length Q A , the proportion obtaine d i s 1 — (2 A + 1)~3
exactly a s in Theorem 26.9 ! Th e Katz-Sarnak philosoph y o f monodromy group s of
a famil y o f L-functions ca n nicely explai n thi s phenomenon .
This concep t o f monodrom y grou p o f families , althoug h lackin g a forma l def -
inition yet , ha s alread y bee n use d t o explai n ver y strikin g results . Fo r example ,
Conrey an d Soundararaja n [CoSo ] hav e bee n abl e t o prov e fo r th e first tim e tha t
there ar e infinitel y man y L-function s o f rea l character s wit h n o rea l zero s i n th e
critical strip (no t only at s — \). Th e success of their approac h (whic h depends on a
constant turnin g ou t t o be positive) wa s "predicted " b y the conjectured symmetr y
of th e family o f real characters .
THEOREM 26.7 . Let Q be as in Theorem 26.6. We have

\{q^Q\ L(x-8 9, s) ± 0 for s G [0, 1]}| ^ Q - o(l))\{q ^ Q \ q G Q}\

as Q -» -foe.

We conclud e wit h a stil l challengin g problem : prov e that , give n a modula r

form / (sa y of weight 2 and level q), there exist s a positive proportio n o f quadrati c
characters \ suc h tha t L(f 0 x ? |) 7 ^ 0- Thi s i s a conjectur e o f Goldfel d tha t stil l
resists proof . Th e best know n resul t i s due to Ono [Ono] , usin g algebrai c method s
(congruences an d Fourier coefficient s o f half-integral weigh t modula r forms) , losin g
a facto r (logQ) 6 fo r som e S G (0,1). Analytically , usin g Heath-Brown' s larg e siev e
inequality fo r rea l character s (Theore m 7.20) , Perell i an d Pomykal a [PP ] have th e
best result , wit h a facto r Q~ e lost .
REMARK. Othe r specia l point s o n th e critica l lin e ar e thos e Sj = \ + itj as -
sociated t o a n eigenvalu e o f th e Laplac e operato r o f a Maas s cus p for m <pj fo r a
congruence subgroup . Th e deformatio n theor y o f Phillip s an d Sarna k show s tha t
the vanishing of the special value of the Rankin-Selberg L- function L(f<g>(pj, \+itj)
(where / i s a weigh t 4 holomorphic cus p form ) ha s an interpretatio n fo r th e ques -
tion o f th e existenc e o f cus p form s i n non-arithmeti c groups . W . Lu o [Luo ] has
shown tha t fo r fixed / G S^p)*, wher e p i s prime, a positiv e proportio n o f specia l
values L(f <g > ipj, \ - h itj) ar e non-zero.
26.2. Principl e o f th e proo f o f Theore m 26.2 .
Since q is prime and k — 2 < 1 2 , we can apply the Petersson formula t o the basis
5*2(q)* (properl y normalized ) o f primitive form s o f £2(3), a s describe d i n Corollar y
14.23 and afterwards . Le t / G S2(q)*• W e denote it s Fourier expansio n

f(z) = X^M^vW^)
and recal l its multiplicative propertie s (1 4.50) , (1 4.51 ) . T o prove Theore m 26.2 , w e
first deriv e th e weighted analogue :
THEOREM 26.8 . We have

feS2(q)* od d
L/(/,l/2)^0
 26. CENTRA L VALUE S 581

for primes q —> +00 .

Recall fro m (1 4.60 ) tha t th e h i n superscrip t indicate s tha t th e normalizin g

factor (47r) _ 1 ||/|| - 2 i s inserted .

The metho d o f proo f i s base d o n compariso n o f th e firs t an d secon d moment s

of specia l value s Z/(/ , 1 /2) . Sinc e "larg e values " o f ! / ( / , 1/2) hav e a larg e effec t o n
the secon d moment , w e us e mollification , a s i n th e proo f o f Selberg' s theore m i n
Chapter 24 . Le t
M
i = E * W-ef)MU)L'U,\)
f£S2(q)*
and
k 2
M2= J2 !(l-e/)|M(/)LU|)|
f£S2(q)*

for som e M(f) G C. Compariso n o f a n uppe r boun d fo r M 2 an d a lowe r boun d fo r

Mi yield s b y Cauchy' s inequalit y a lowe r boun d fo r th e quantit y w e stud y

Ml
(26.2 ) E* 1> M
2
*/ = -!
£'(/,§)*o

In orde r t o achiev e th e bes t possibl e estimate , w e will fin d asymptotic s fo r M i

and M 2 . Not e tha t i f th e mollifie r i s ignore d (tak e M(f) = 1 ) , a facto r log< ? is los t
in th e fina l estimate .
To ge t sum s amenabl e t o Petersson' s formula , w e loo k fo r mollifier s a s linea r
forms i n A/(m) :

A
(26.3) M(/ ) = J2 % /M

for som e M = q A wit h 0 < A < \ an d rea l coefficient s x m supporte d o n squarefre e

numbers < M , suc h tha t

(26.4) x^ <^r(m) (logqf

for som e absolut e implie d constant .

Theorem 26. 8 follow s b y lettin g A —> • \ i n th e followin g

THEOREM 26.9 . For 0 < A < 1 /2 , we have

1 > 1
^ 2\ (2 A + 1)0
fes2(qy od d
L'(/,l/2)=#)

for all primes q large enough in terms of A.

582 26. CENTRA L VALUE S O F L-FUNCTION S

26.3. Formula s fo r th e firs t an d th e secon d moment .

The moment s M i an d M 2 ca n bot h b e reduce d t o combinations o f averages o f
! / ( / , \) o r Z/(/, \) 2 twiste d b y Heck e eigenvalues . Fo r m ^ 1 , le t

(26.5) D(m)= Yl k W-e f)Xf(m)L\f^),

fes2(q)*
(26.6) H(m)= Y? |(l-^/)A / (m)L / (/,i) 2 .
fes2(q)*

We hav e

(26.7) M l =£^kD(m)

and b y (26.3 )

A
M2= E " L'u,\f E 7 = = /^)A/^)
Sf = — 1 rai,ra2 ^

(26.8) = V j V ^bm^brn , R( y
b mi,m2

We will obtain asymptoti c formula s fo r D(m) an d H(m) valid if ra is small enoug h

with respec t t o q.
We first us e contour integratio n an d th e functiona l equatio n t o expres s L'(f, | )
and ! / ( / , | ) 2 a s sums o f rapidly convergen t series . Thi s coul d b e derived fro m
(5.12) b y differentiation , bu t w e redo th e computation s anyway .
Let N ^ 2 and G be a fixed polynomia l suc h tha t G(—s) — G(s), G(—N) =
. . . = G ( - l) = 0 an d G(0 ) = 1 . I n particular, G'(0 ) = G< 3>(0) = 0. Conside r

/=
i/A^+^f-
(2)

We ca n shif t th e contou r o f integration t o the lin e R e (s) = —2, picking u p a

double pol e a t s = 0, henc e

ds
I=resA(f,s+^ + ^- J A(f,s+l)G(s
(-2)

By th e functiona l equatio n (Theore m 1 4.1 7) , wit h sig n—7 ? = sj, th e integral on

Re (s) = — 2 is equal to £//, so

(l-£/)/=resA(/)S + i ) ^ .

Computing th e residu e b y writin g th e Taylo r expansion s aroun d 0 , we obtai n

l2
2{l-ef)I = {l-e s)(^) ' L'{f,\).
 26. C E N T R A L VALUE S 583

On the othe r hand , w e can comput e / by expanding th e L- function i n an absolutel y

convergent Dirichle t serie s o n th e lin e R e (s) = 2 , which give s

'-(£)"*£
V2W
U ^WV £)
^
where
3ds

(3/2)
Hence, comparin g bot h expressions , w e hav e

(26.9) ( i _ e / ) L U ! ) = 2 (l- e/ )g^v(^).

We estimat e V easil y b y shiftin g th e contou r t o th e left , o r right : w e hav e

(26.10) V(y) = -logy- 7 + 0(y),
J
(26.11) V(y)<^y~ forallj^l .
Similarly, w e conside r th e integra l
J
'hSK{,-'+^G{4
(2)

and procee d t o evaluat e i t i n th e sam e fashio n a s fo r I. Shiftin g th e contou r t o

Re (s) — —2 and applyin g th e functiona l equatio n fo r th e squar e o f th e L-functio n
A(/,S)2 = A ( / , 1 - S ) 2
(where th e sig n i s ei = 1), we ge t

2J=resA(/,s+i)2^.
5 = 0 S°

Further, fro m th e multipHcativit y o f th e Heck e eigenvalue s (se e (1 4.50) ) w e

derive th e Dirichle t serie s expansio n

L(/,s)2 = C 9(25)^r(n)A/(n)n-s,

so th e ter m b y ter m integratio n yield s

27r
SI ^ v
«J
where
2
(26.12) W(y) = ± J ( q(l + 2s)T(s) G(s)y-s^.
(1/2)

Hence

n>l
584 26 . C E N T R A L VALUE S O F L - F U N C T I O N S

For od d /, w e have L ( / , \) = A(/ , \) — 0 so the above residu e equal s A'(/ , \) 2 =

2%Lf(f, \)2 fro m th e Taylor expansion s of A(/, s + \) an d G(s). Therefor e fo r odd
forms w e have
2
(26.13) ! U i ) = 2 j : f r W p ) .

The functio nV K satisfies th e bound

(26.14) y l
WU)(y) < log 3(?/ + y" 1 ), fo r all i^j^0,
(i)
(26.15) 2/W (y) < C y_J ', fo r all i ^ 0 , j^ 1 .
(the implie d constan t dependin g o n z and j) an d there exist s a polynomia l P ,
independent o f g, of degree a t most 2 , such tha t

(26.16) W(y) = -^(logy) 3 + P(logy) + O ^ - ^ l o g y ) 2 + y).

This las t fac t follow s b y writing

G W r w ^ i + i!;)

by shiftin g th e contour , an d computing th e residue.

Using (26.5 ) an d (26.9) w e get

where
A , (I,m) = ^ ( l - ^ ) A / ( 0 A / ( m ) .
/
We no w appeal t o (1 4.65 ) o f Corollar y 1 4.2 6 which show s tha t A ' i s a stron g
approximation t o the Kronecker delt a

A%rn) = S(lM + 0^^(^)\lo gq)\).

One coul d us e also (1 4.64) , keepin g th e Kloosterman sums , bu t this wil l no t b e

necessary fo r the application t o M\ an d Theorem 26.2 . Henc e w e have b y (26.10)
and (26.1 1 )
3 81 4
D(m) = 4=VC-^) +0(g- / m / (logg)2)
with a n absolute implie d constant . B y (26.7), (26.1 0 ) an d (26.4) w e get
PROPOSITION 26.1 0 . We have for m^l,

(26.17) D{m) = - i= (log ^-) + 0(<T3/8 m 1/4 (log q) 2 + mq'^ 2

)

where q = y /g/(27re7), and if M = qA, then

s 7
(26.18) M 1 = ] T ^(logl)+0( q- (logq) ).
 26. CENTRA L VALUE S 585

where 6 = | ( \ - A) .
In th e following , whe n w e write a n erro r ter m o f the for m 0(q~ 8)1 i t i s implie d
that 5 > 0, 6 depends onl y o n A , an d th e valu e o f S may chang e fro m lin e t o line .
We no w procee d t o get a n expressio n fo r M 2 a s a quadrati c for m i n th e xm.
The ide a i s th e sam e a s above , bu t w e canno t simpl y estimat e th e contributio n o f
the su m o f Kloosterma n sum s i n th e Petersso n formul a b y Weil' s bound : w e ope n
the Kloosterma n sum s an d transfor m furthe r t o deriv e th e require d expression .
Let m ^ 1 be fixed, m ^ q 2A < q. W e ar e no w read y t o comput e M 2. B y
(26.6) an d (26.1 3 ) w e hav e

H(m) = £ T -§W(^) ^ ( 1 -e f)Xf{m)Xf(n)

r(n) /47r. 22,n\ ,

E - WT T — )A'(m,n) . Q

Our previou s approximatio n t o A'(ra , n) b y 5(m, n) i s not sufficientl y strong . Pro m

Corollary 1 4.2 6 w e have th e mor e precis e formul a

A , (m,n) = ( 5 ( m , n ) - - ^ ] T -S(mq,n]r)J 1 ^J'^>j

v (r,g ) =l
1
+ 0 (r3((m,n))^)i(logg))

Hence b y (26.1 4) , (26.1 5 )

r(m) /47r 2mN

(26.19) H( ? n) = ^ H / ( ^ ! ! ! ) + X ( m . ) + o((^)i(log^)
9
where X(m) i s a su m o f serie s o f Kloosterma n sum s

1
(26.20) X{m) = - ^ ^ T r- Xr(m),
^ (r, 9)=i

(26.21) X r (m) = - £ ^ S ( m 5 , n ; r ) j1 ( ^ ^ ) W ( ^ ) f ( „

For technica l reason s w e have inserte d i n th e summatio n th e facto r £(n ) wher e £ is

a fixed functio n £ : R+ — » [0,1 ] , whic h i s C°° an d satisfie s
£(z) = 0 , 0 ^ x ^ \, £(x) = 1 , x ^ 1.
Obviously thi s extr a facto r doesn' t affec t th e su m (al l positiv e integer s ar e a t leas t
1!), bu t i t wil l b e usefu l t o gai n convergenc e i n som e serie s appearin g later .
We first estimat e th e contributio n o f thos e term s wit h r > R i n (26.20) , wher e
R > 0 will b e chose n late r (R = q 2). Usin g J\(x) < C x, Weil' s boun d fo r Klooster -
man sums , (26.1 4 ) an d (26.1 5) , w e ge t
2T T

(r,g) = l
586 26 . C E N T R A L VALUE S O F L - F U N C T I O N S

We denot e no w X f(m) th e remainin g par t o f th e su m i n X(ra) , namel y

r1
_
(26.22) X'(ra ) = ^F Yl ^(m).
(r,g) = l

Let r ^ R. Th e expressio n X r(m) ca n b e compute d usin g th e summatio n

formula (4.56 ) fo r sum s o f Kloosterma n sums . Thi s yield s

2 C°° fx
XJm) = — 5 ( m , 0 ; r ) / (lo g ^— + i)t{x)dx
r Jo r

+ — 2jr(/i)S(/i<7-m,0;r ) / * o ( — Jt(x)d x

- - 5 ^ T ( / i ) S ( / i g + m,0;r ) / ° ° lfo( ft *)*(s)ris

ft^i

where
2
47r jmx\ (4TT X\ £(x)
(26.23) t(l )= J l ( _ ^ _ ) W ( _ ) .
r V q J \ q / y/x

Hence

(26.24) X'(m ) = ^ £ ^S(m,0;r)L(r )

(r,<jr) = l

+ ^FE \Y,T(h)S(hq-m,0;r)y(h)
(r,<j) = l

- ^ E ^E^)^+™>^ w + o ( ^)
(r,g) = l

with

(26.25) L(r ) = y o (lo g ^- ^ (- y j T ) W (—)-/= l

(26.26) y(/i ) - J Yo(—y-)t(x)dx,

(26.27) k{h)= / Xo(^^)t(^)dx ,

the erro r ter m comin g fro m removin g th e functio n £(x) fro m th e firs t ter m usin g
t(x) < (logqfx- 1 /2 fo r 0 < x < 1 , and |S(ra,0;r) | ^ r .
 26. CENTRA L VALUE S 58 7

Let X"(m) b e th e firs t ter m i n X'(m). W e hav e thu s b y changin g th e integra -

variable x into ij^r ,

X"(m) = -2 £ \s{m^T) j ^ {\og^-+ 1 )J1 {2V^)W{r2x)~,

(r,q) = l
=
in J Z m(l + 25)Cg(l + 25 ),- 1r( 5 ) 2 G( 5 )L( 5 )^,
(1/2)

by (26.1 2) , wit h

Z*(s) = ] T 5 ( m , 0 ; r ) r " s ,

(r, 9 ) = l

1 8 /2
L(s) = -2 / (lo g ^+ 7) J 1 (2y/mx)x'- - dx.
Jo 2 7 r

Extending th e su m ove r r to al l r ^ 1 , we ge t

d|ra

if R e (s) — 2 by th e formul a (3.2 ) fo r th e Ramanuja n sum . Moreover , fo r al l s with

\ < Re (s) < 1 , we hav e

(26.28) L(s) = m s " 1 / 2 r ( - 5 ) r ( 5 ) - 1 flo g — + 27 + ^( 1 + s) + ^(1 - 5) )

\m /
where ^ = T'/ T an d Q = q/An 2. Indee d formul a 1 4.561 .1 4 o f [GR] give s (fo r
- 2 <Re(s ) < - | )

r°° r r(i(l + f)
y j!(x)^ x = 2s ^ 2
/° ° (log*) Mx^dx = 2 ' B i ± || (lo g 2 + ^ ( l + '-) + i t f (l - 0)

and a simple computatio n give s (26.28) . Consequentl y w e hav e

x {m) =
" 1
hi I (- 2)^-2sM5-1 r(5)2G(.)L(^ + o ( ^ i o g g )
(1/2)

since (o n R e (5) = -|)

C,(l + 2S )r( S ) 2 G( S )L( S ) « \r(s)r(-s)s- 1 \(logq+ |^( 1 + S) | + Ml - a )|).

The functio n F(s) bein g integrate d ca n b e writte n

9.
F(s) - m - 1 / 2 5 r / s ( m ) - 1 G ( 5 ) r ( 5 ) r ( - s ) flo g " + 27 + ^( 1 + 5) + ^(1 - s))
m
where
V.(m)= J2 (£)*
ab=m
588 26. C E N T R A L VALUE S O F L - F U N C T I O N S

(Fourier coefficient s o f the non-holomorphic Eisenstei n serie s fo r 5X(2,Z)). Thus ,

F(s) i s an odd function o f s. Moreover , i t is holomorphic i n the strip |Re(s) | < 1,
except fo r a tripl e pol e a t s = 0 , an d decrease s exponentiall y i n vertica l strips .
Shifting th e contour t o Re (s) — —\ and then changin g s into — s, w e conclude tha t
r(ra )
X"(m) = f r e s F ( 8 ) + O p p i o g g )

Expanding aroun d s = 0 we have

1
s- r(s)r(-s) = -s"3 + (72 - r"(i)) s-1 + o(s),
2
G(s ) = l + iG"(0)s + O(s 3 ),
2 7 + ^ (1 + s) + V(l - s) = V>"(0)s 2 + 0 ( s 4 ) ,
2 3
% ( m ) = -r(m ) + | T ( m ) s + 0 ( s )

where

(26.29) T(m)= £ (log^)* .

Combining those , w e obtain therefor e

(26.30) X"( TO ) = -^(log^) +a^(log^)+o(^log,),

4y/m V m) y/m \ m/ \ R /
where a = | ( 7 2 - r " ( l) - §G"(0 ) - V"(0)).
Coming to the other tw o terms i n (26.24), we claim tha t fo r q < m an d R ^ q 2

x
(26.31) " 7 E* 4 E r(h)S(hq - m, 0; r)y(ft) « m '\~^\

(26.32) ±- Y^ ^ Er(ft)S(fe<7 +™> 0; r)k(h) « (m 1 /^" 1 + <T1/2)<f

for an y £ > 0, the implied constan t dependin g onl y o n e.

We choose R = g2. B y (26.20) , (26.22) , (26.24) , (26.30) , (26.31 ) , (26.32 )
together, w e deduce

(26.33) X M = - ^ ( l o g 5 )+ a I ^ ( I o g 4 ) + o ( ( ^ + 4 = y ) .
4>/m \ mJ y/m V m /V V q yfqJ J
Inserting thi s i n (26.19) an d appealing t o (26.1 6) , we obtain
PROPOSITION 26.1 1 . The twisted mean-value (26.6) is given by
(26.34)
1 r(m) / Q \ 3 \T(m) ( Q\ r(m) / Q\ 1 / 4 _ 1 / 4 +££
H{m) = —-^jJ-[log--) - - — )=+ lo g — ) + — =• Pi lo g — )+0{ml/*q i / 4 +
12 y/m V m / 4 y/m \ mJ y/m \ mJ

where Q = q/Air 2 and P^X) = P(X) + aX for 1 ^ m ^ q2A with A < 1 /2 . The
implied constant depends only on e and A.
 26. C E N T R A L VALUE S 589

R E M A R K . I n particular , fo r m = 1, (26.1 7 ) an d (26.34 ) giv e asymptoti c for -

mulas fo r th e firs t an d secon d moment s o f the specia l value s o f the derivative s
£'(/>!):
k
(26.35) J2 \{1 -ef)L'U,\) = \\ogq-1\og2ir + 0{q-ll2),
/eS2(g)*

(26.36)
J2k |( 1 - ef)LU \f = ±(log ^ )3 + A (log J L) 4 - 0(g-^«).

Finally, th e expression fo r the second momen t o f special value s a s quadrati c

form i n the mollifier coefficient s follow s fro m (26.8 ) an d (26.34).
THEOREM 26.1 2 . Assume M = qA with A < \. Then
s
M2 = —M 2i - -M 22 + M2 3 + 0(q- )

where M21 , M22 and M23 ar e quadratic forms in the variables x m given by

3
(26.37) M 12 = j: \ £ ^ ^ ^ (lo g - $ - ) ,
1
1
6 771,77 2

T
(26.38) M 22 = Vj V (mim2)x^A - Q - ) ,
2 L
^ 0 ^ vfi\Vfi2 V mim 2/
6 77li,77l 2

(26.39) M 23 = £ I £ I^l!^) XbBiStmajPi (l og _ « _ ),

1
1
6 771,77 2

and 6 is a positive number depending only on A. Recall that Q = g/(47r 2).

R E M A R K . I n fact ou r proof i s complete onl y for A < |. Thi s limitatio n come s
from th e error ter m i n (26.1 9) , al l other part s bein g adequat e fo r A < \. Th e full
extension t o this cas e ca n be established b y arguments essentiall y th e same a s i n
the othe r parts , s o we omit th e details.

The proof s o f (26.31) an d (26.32) requir e onl y uppe r bound s fo r the integrals
y(h) an d k(h). Thes e ar e estimated b y standard integratio n b y parts (a s in Section
4.4). Combinin g wit h the bound for Ramanujan sum s \S(hq±m, 0 ; r)| < (hq±m, r)
one derive s th e results (fo r details se e the original pape r [KM1 ]) .
26.4. Optimizin g th e mollifier .
Now (26.1 8 ) expresse s Mi as a linear form , an d Theorem 26.1 2 expresse s M 2
as a quadratic form , i n the coefficients x m. Th e optimal choic e of xm t o maximiz e
M1/M2 (se e (26.2)) i s not easy to express explicitl y becaus e th e quadratic for m is
not diagonal . Ou r strategy wil l be to decompose M 2 as a sum of quadratic forms ,
one o f which, sa y M' 2, is easily diagonalized . W e then choos e x m t o maximize
M1/M2, an d simpl y evaluat e th e othe r quadrati c form s fo r this chose n vector .
One coul d justif y afterward s tha t thi s i s asymptotically bes t possibl e eve n fo r the
full quadrati c form . I t turns ou t that bot h M 2\ an d M2 2 in Theorem 26.1 2 hav e
contributions o f the same orde r o f magnitude, wherea s M23 is smaller.
590 26. C E N T R A L VALUE S O F L - F U N C T I O N S

We separat e rai an d 777, 2 hi (26.37 ) b y mean s o f th e formul a

r r
r(m1m2) = ^ ^ W 7 W)
a|(mi,7n2)

getting (afte r relaxin g th e resultin g coprimalit y conditio n b y mean s o f th e Mobiu s

function)

r l r ^ aV ^ r(mi)r(m 2) A < 2V
6a mi,rri2

Expanding th e logarithm , M21 is a linea r combinatio n o f th e quadrati c form s

n(t,«,«,«;) = (iogQ) » 5>(%fc",H) y*

0,(™)
u

where
r m
(*) V ^ ( )n \ i
m

,<*,. \ Y. «£W.
ab=k

and t, u, v an d it ; ar e non-negativ e integer s suc h tha t t-\-u + v + w = 3. I n fac t

M23 is also such a linea r combinatio n wit h parameter s suc h tha t t + u + v + w ^ 2 ,
however.
For th e chose n vecto r (x m) i t wil l b e clea r tha t

n ( t , t i , ^ ) < < n ( o , « , ^ ) ^ —
logg
by direc t estimation , s o w e ca n restric t ou r attentio n t o U(u^v,w) = 1 1 (0 , u, v,w).
Accordingly we write i/(k) fo r uo(k). Not e that fo r k ^ M , w e have v(k) = cp(k)k~ 2.
The contributio n t o M21 of thos e II(n , v, w) i s the quadrati c for m
m 2 1 = n(3,0,0 ) - 6n(2,1 ,0 ) + 61 1 (1 ,1 ,1 ) + 61 1 (1 ,2,0 ) - 61 1 (0,1 ,2 ) - 21 1 (0,0,3 )
(using th e obviou s symmetr y IL(u, v, w) = n(u , w,v)). W e selec t th e on e quadrati c
form I I = 1 1 (3,0,0 ) a s reference an d w e choose (x m ) whic h maximize s M f /II. The n
we evaluate th e othe r quadrati c form s II(u , v, w) an d M2 2 fo r tha t vector .
By definition , I I i s alread y diagonalize d

n = (iog£) 3l>(%fc-
k

The x m ar e recovere d b y th e formul a

(26.40) x m = £ ^T-Vk,
k

where g = [i* fi i s the Dirichle t convolutio n invers e o f r. Henc e th e linea r for m M i

can b e writte n a s
M1 = £ —log-^E^ *
 26. C E N T R A L VALUE S 591

where
*(*) = £ £*(«)(logf).
ab=k

Since

we hav e J2j(k)k~s = C(s + l)- 2((log<z)C(* + 1 ) + C'(* + 1))

fc>i
logg r/ ._lv
( r ) ' ( * + i)
<(* + !)
which show s tha t

(26.41) j(k) = ^-{logqk).

By Cauchy' s inequalit y w e have

with equalit y i f y& = ^ 4 fo r k ^ M . Henc e th e best choic e o f mollifier i s

^\ 0 ,i f k > M.
In particula r y^ , an d s o als o # m (se e (26.40)) , i s supporte d o n squarefre e num -
bers an d th e growt h conditio n (26.4 ) i s immediatel y verified . Sinc e j(k) i s abou t
(log k)/k an d v i s abou t A: -1 , it i s quit e clea r fro m lookin g a t th e variou s expres -
sions involve d tha t w e will get a positive (harmonic ) proportio n i f M — q A fo r any
A > 0 , with bot h M x2 and M 2 bein g o f size (logg) 6 .
The result s o f evaluating M\ an d M 2 i, M22 are as follow s
PROPOSITION 26.1 3 . Let M = q A and xm be as above. We have

Mx = A (^ + I + \)(logq) 3 + 0((\ogq) 2),

M21 = (JA 6 + f A 5 + f A 4 + f A 3 + 2A2 + \A)(\ogqf + 0((\ogqf),

M23<(logg)5(loglogg)4.

PROOF. Fo r Mi, by definition o f yk an d (26.41 ) , w e have

«, = „og0)-n = i;f^ = i:^f(^«2

and th e result follow s b y partial summatio n fro m

KJ
k<K ^
592 26. C E N T R A L VALUE S O F L - F U N C T I O N S

For M21 , we hav e

)
U(u,v,w) = (logQr^(k)y^ y^)
k

and w e expres s y£ i n term s o f yk usin g th e highe r vo n Mangold t functio n A ^ =

// • (log) 2. Precisel y w e hav e

T
(26.42) y?= Y, -f^)yke-
i^M/k

See Sectio n 1 . 4 ((1 .43 ) an d after ) fo r th e propertie s o f A * that w e need .

The sum in (26.42) is thus a sum over squarefree I having at mos t i prime factor s
(i ^ 3) . T o evaluate we can find the associate d Dirichle t serie s or more elementaril y
separate th e su m int o th e part s wit h a fixed numbe r o f prim e factors . Thos e ar e
multiple sum s ove r primes , an d the y ar e o f Merten s typ e sinc e T{£)£~ 1 = 2H~ X
for suc h £ with LJ(£) = j prim e factors , henc e eas y t o evaluat e asymptotically . W e
quote th e resul t (se e [KM1 ] , 2.5.2 , fo r details) : on e finds

2/il) = Q ^ (lo g f y(log<f +

W k ) + 0(^(lo gqy(loglogq))

for i = 1 , 2 , 3 wit h c\ — — 1, C 2 = \ an d C 3 = 0 . The n th e computatio n o f th e

various n(u , v, it;) i s a simpl e matte r o f summatio n b y parts . Fo r instanc e w e hav e

11(0,1,2) = - J £ ^!(i og ^)3(iog93M2fc)(log^MA;)

+ 0((log< ?)5(loglog<?)3)

=/ (log- ) ( l o g . f M V K l o g . f M z ) — + 0((logc ? ) 5 (lo g log< ? ) 3 )

JY\ X / X
AogM
3 5
=/ y {3logqM-y)(2\ogqM-y)dy + 0((\ogq) (log\ogq)3)
Jo
= ( ~ A 6 - ^A 5
- ^(logg) 6
+ 0((log< Z ) 5 (loglog (? ) 3 )

(recall log g = lo g y/q + O(l)) . Whe n everythin g i s computed , th e valu e o f M21 i s

obtained a s claimed . D

There remain s t o trea t th e quadrati c for m M22 defined b y (26.38) . Recal l tha t
T(n) i s define d b y (26.29) . I t follow s tha t

(2,43, • « . E ^mm)H^ym d\(mi,m,2)

 26. CENTRA L VALUE S 59 3

We us e this t o separat e m i a n d 7712, which yield s t h e expressio n

z
^-^o^^ a ^ ^ mim 2V o r 77117712/
ba 77li,77l 2

0 ^m T r(mi)T(m 2) / Q \

mi,m2

- 4 V ] l / i ( f e ) V " —X 1km Xkm2'

^ *-^ mim 2

This i s rewritte n a s

(26.44) M 22 = 2(n(l,0,0)-n(0,l,0)-n(0,0,l))-4^i/ 1 (%fc^

where

Z
_ (o ) _ V ^ T ( m )
k — zk ~ 2_ > m XkT
m
T
(l) V ^ ( m ) /i \

and

ft(a,6,c) = (lo g Q ) a X ; K % i 6 ) ^ •fe

it

As befor e w e express t h e new variables i n term s o f yk b y

(26.45) Zk = 2j2 ±^y ke,

l^M/k
X)
(26.46) 4 = ]T -AJ_±i Zfc ,+ £ " ^ ^ V
£^M/k i^M/k

To se e this, notic e t h a t b y (26.40 ) w e hav e

*-E(E^-)>
rn _ , . .to-
T(m)

The Dirichle t generatin g serie s fo r t h e coefficien t o f £ is L(s + 1 ) wher e

i(5) = c(S )- 2 E T ( n ) n " s -

71

It i s easy t o se e t h a t

^2T(n)n-s = 4<C" - 2(CC') ' = 2(CC" - (C) 2),

594 26 . C E N T R A L VALUE S O F L - F U N C T I O N S

so L(s) = 2(C /C~1 )/, whic h give s (26.45) . The n w e derive (26.46 ) a s follow s

m £b=m £ m^M/£

^ r(l)A(l )^ T(£)A(£)
£^M/k £^M/k

by (26.43) .
By (26.45 ) an d (26.46 ) w e can evaluate Zk and z\. ' muc h a s before wit h y^\
and w e get

*--5^(*T)'™,""*>+0<4(1°")')
»i2)+0(-|y(l089)2(l«6l»g9)).

From thi s w e easily evaluat e th e quadratic form s tha t occu r i n (26.44 )

11(1,0,0) = - ( l o g Q ) 5 > ( % f c 2 / l 2 ) + 0((log<?) 5) = - n ( l , 2 , 0 ) + 0((log<z) 5),

k

n(o, i, o) = - £ " W ^ S f + o((iogq)5) = -n(o, 1,2) + o((iog<?)5),

k

6(0,0,1) « (lo g qf.

Moreover, b y direct estimatio n

^2^i{k)ykzk < (logg) 5 (loglogg) 3 .
k

Now (26.44 ) yields :

PROPOSITION 26.1 4 . Let m = qA. For x m as above, we have

M22 = ( - ^ A 6 - ^ A 5
- ^A 4
- i A 3 ) ( l o g g ) 6 + 0((logg) 5 (loglo g ( ? ) 3 ).

that Prom
fo r ATheorem
<\ 26.1 2 , Proposition 26.1 3 and Proposition 26.1 4 , we can comput e

M L = I f i1
_ , 0 /(loglogg) 4 \
\
3
M 2 2 V (2 A + i ) ; ^ v iog g ;
(where th e very simpl e fractio n appear s quit e miraculousl y b y partial fractio n ex-
pansion). Inequalit y (26.2 ) the n implie s Theore m 26.9 , and therefor e Theore m
26.8.
 26. C E N T R A L VALUE S 595

EXERCISE. Sho w b y simila r method s tha t

fes2(qy
L(/,|)/0

for prime s q —>• +00 .

26.5. P r o o f o f t h e o r e m 26.2 .
Passing fro m Theore m 26. 8 (non-vanishin g i n "harmonic " average ) t o Theore m
26.2 (non-vanishin g i n "natural " average ) ca n b e don e i n differen t ways . W e us e
Shimura's formul a
2
(26.47) 4^||/|| = ^ | p L ( S y m 2 / , 1 ) + 0((\ogqf) (fo r q prime )

relating the Petersson nor m t o the specia l value of the symmetric squar e L- function
of / a t s = 1 . Th e latte r i s given b y

L(Sym 2 /,*) = £ Pf(n)n-° = C q(2s) ]T X 2

f(n )n-'

for R e (s) > 1 ; se e Sectio n 5.1 2 . A t s = 1 th e serie s barel y fail s t o converg e

absolutely, bu t o n average i t ca n be represented b y a very short sum . Pro m [KM2] ,
we quote a general useful resul t base d o n the large sieve type estimate for symmetri c
square //-function s (se e Theore m 7.28) .
PROPOSITION 26.1 5 . Let aj, for f e 5 2 (g)* ; be complex numbers such that

(26.48) a f < q l~5 for some S > 0,

(26.49) Y^ \&\f <: (logq) A for some A >0,

f
Then we have

with x = q K, where K is any positive number and 7 is a positive number which

depends only on 5 and K. The implied constant depends only on K, 5 and A.

We us e a mollifie r M ( / ) a s i n (26.3 ) o f lengt h M = q A wit h A < | , definin g

N,= }2 M(/)L'(/,l/2) ,
2
N2= Yl \M(f)L\fA/2)\
e / =-i

and appl y Propositio n 26.1 5 wit h a f = M(f)U(f,\) an d a = | M ( / ) L ' ( / , \)\ 2

ly/4 2
respectively. Th e convexit y boun d ! / ( / , \) < C g (logg) (us e (26.9 ) t o prov e i t
again) yield s easil y (26.48 ) an d (26.49) .
59 6 26. C E N T R A L VALUE S O F L - F U N C T I O N S

It follow s tha t

* = ^ff1 E * E ^f(*mf)i>v, h)+ ofa1 -4),

1 £
' / = ~1 ^ 2<C*
(*,<?) = !

2
C(2) ^ ^ d*
(*,<?) = !
K
with x — q wher e K > 0 can b e chose n arbitrarily .
Inserting th e formula s (26.9 ) an d (26.1 3) , an d usin g result s o f Sectio n 26.3 ,
we ar e le d t o a linea r for m expressio n fo r TV i an d a quadrati c for m fo r N2 whic h
are ver y simila r t o thos e o f Theore m 26.1 2 . I t i s importan t t o b e abl e t o tak e K
arbitrarily smal l t o retai n a n asymptoti c formula . Th e sam e principl e a s in Sectio n
26.4 applie s t o fin d th e mollifie r an d evaluat e N\ an d A^ , bu t ther e ar e man y
technical complication s du e t o lac k o f complet e multiplicativit y o f th e coefficient s
involved.
REMARK. I f on e i s willing t o accep t a slightl y weake r resul t
(26.50) | { / e S 2(q)* \e f = -l an d L'(f, \) + 0} | » \S 2(q)*\

for q prime, with n o explicit constan t (th e analogu e o f Selberg's theore m o n critica l
zeros of C( 5))? some shortcuts ar e available. First , ther e i s no reason t o optimize th e
mollifier a s in Section 26.4, and the computations ther e can be simplified quit e a bit .
Then instea d o f appealing t o Propositio n 26.1 5 and performin g th e computatio n o f
Ni an d N2 on e ca n argu e b y mean s o f Cauchy' s inequalit y

1 = 2 l
E £ 4d|/|| ^ V2^i67r 2||/||4J V E )'
fes2(qy od d fes 2(qy od d/ fes 2(q)* od d
L'(/,l/2)^0 L'(/,l/2)^ 0 L'(/,l/2)# 0
Then th e asymptoti c formul a

<26 5i) c
- £»~ «2
as q — » -foo fo r som e constan t c > 0 togethe r wit h Theore m 26. 8 impl y (26.50) .
Formula (26.51 ) i s a specia l cas e o f a resul t o f E . Roye r [Roy ] abou t moment s o f
L(Sym 2 /, 1 ) ove r / G 52(g)*. Not e tha t b y (26.47) , thi s amount s t o th e negativ e
second momen t o f L(Sym 2 /, 1 ) . Th e proo f o f (26.51 ) use s a densit y theore m fo r
zeros o f familie s o f L-function s t o restric t th e summatio n t o / satisfyin g a "quasi -
Riemann Hypothesis" .
REMARK. Beside s bein g use d t o buil d mollifiers , a s w e hav e done , th e twiste d
mean-values o f typ e (26.5) , (26.6 ) fo r familie s o f L-function s hav e a powerfu l us e
for buildin g a n "amplifier" , i.e . a su m o f th e typ e
Ak = y ZA(f)\L(f,±)\k

where A(f) > 0 , sa y

A(f) = \Y,c e\f(e)
 26. C E N T R A L VALUE S 597

but thi s tim e A(f) is meant t o b e larg e fo r f = fo e T, some fixed form, s o that b y
positivity on e deduce s fro m a n uppe r boun d fo r th e mean-valu e Ak a n individua l
upper boun d
/k
(26.52) \L(f 0,±)\^Al A(f0)-\
In th e bes t circumstances , th e boun d fo r Ak ca n b e achieve d i f L is a small powe r
of |.F| , and matche s tha t fro m Lindelo f Hypothesi s an d orthogonalit y

^«m 1 + £ (£n 2 ) 1 / 2
i
so that (26.52 ) is a gain if the mean-value of Q is comparable in size to the peak value
A(fo). Thin k o f the cas e of primitive Dirichle t character s o f conductor q where on e
can tak e Q = x(fy ( s e e Propositio n 1 4.2 2 for a suitable choic e fo r classica l modula r
forms).
This amplificatio n metho d ha s bee n th e mos t successfu l fo r th e proo f o f sub -
convexity bound s fo r L- functions o f degre e ^ 2 , particularl y i n conducto r aspec t
where othe r technique s (Wey l shifts , reducin g th e averagin g family ) ar e inoperant .
The implementatio n is , however , quit e delicat e i n al l case s (requirin g off-diagona l
analysis) s o we refer t o paper s lik e [DFI2] , [DFI3] , [KMV2 ] an d th e surve y [M2 ]
for mor e informatio n o n thi s importan t subject .
This page intentionally left blank
 Bibliography

[ASl] A . Adolphso n an d S . Sperber , On twisted exponential sums, Math . Ann . 29 0 (1 991 ) ,

713-726.
[AS2] A . Adolphso n an d S . Sperber , Character sums in finite fields, Compositi o Math . 5 2
(1984), 325-354 .
[AS3] A . Adolphso n an d S . Sperber , Newton polyhedra and the total degree of the L-function
associated to an exponential sum, Invent , math . 8 8 (1 987) , 555-569 .
[Ahl] L . Ahlfors , Complex Analysis, McGra w Hill , 1 978 .
[AM] M . Atiya h an d I.G . MacDonald , Introduction to commutative algebra, Addison-Wesley ,
1969.
[AL] A.O. L Atki n an d J . Lehner , Heche operators on To(m), Math . Ann . 1 8 5 (1 970) , 1 34 -
160.
[At] F . V . Atkinson , The mean value of the zeta-function on the critical line, Proc . Londo n
Math. Soc . (2 ) 4 7 (1 941 ) , 1 74-200 .
[B] A . Baker , Transcendental Number Theory, Cambridg e Univ . Press , 1 990 .
[BH] R . Bake r an d G . Harman , The difference between consecutive primes, Proc . Londo n
Math. Soc . 7 2 (1 996) , 261 -280 .
[Ba] W . Banks , Twisted symmetric-square L-functions and the nonexistence of Siegel zeros
on GL(3) , Duk e Math . J . 8 7 (1 997) , 343-353 .
[Bar] M . B . Barban , The "large sieve" method and its application to number theory, Uspeh i
Mat. Nau k 2 1 (1 966) , 51 -1 02 ; Englis h transl . i n Russia n Math . Survey s 2 1 (1 966) ,
49-103.
[BG] J . Bernstei n an d S . Gelbart , An introduction to the Langlands program, Birkhauser ,
2003.
[BL] H . Boh r an d E . Landau , Sur les zeros de la fonction f(s ) de Riemann, Compt e Rendu s
de l'Acad . de s Science s (Paris ) 1 5 8 (1 91 4) , 1 06-1 1 0 .
[Bol] E . Bombieri , Maggiorazione del resto nel "Primzahlsatz" col metodo di Erdos-Selberg.,
1st. Lombard o Accad . Sci . Lett . Rend . A 9 6 (1 962) , 343-350 .
[Bo2] E . Bombieri , Le grand crible dans la theorie analytique des nombres, S.M.F , 1 974 .
[Bo3] E . Bombieri , Counting points on curves over finite fields (d'apres S. A. Stepanov),
Lecture Note s i n Math . 38 3 (1 974) , 234-241 .
[Bo4] E . Bombieri , On exponential sums in finite fields, II, Invent , math . 4 7 (1 978) , 29-39 .
[Bo5] E . Bombieri , On the large sieve, Mathematik a 1 2 (1 965) , 201 -225 .
[BD] E . Bombier i an d H . Davenport , Some inequalities involving trigonometrical polynomi-
als, Ann . Scuol a Norm . Sup . Pis a (3 ) 2 3 (1 969) , 223-241 .
[BFI] E . Bombieri , J . Priedlande r an d H . Iwaniec , Primes in arithmetic progressions to large
moduli, Act a Math . 1 5 6 (1 986) , 203-251 .
[BI] E . Bombier i an d H . Iwaniec , Some mean-value theorems for exponential sums, Ann .
Scuola Norm . Sup . Pis a CI . Sci . 1 3 (1 986) , 473-486 .
[Bou] J . Bourgain , Remarks on Montgomery's conjectures on Dirichlet sums, Lectur e Note s
in Math . 1 46 9 (1 991 ) , 1 53-1 65 .
[BDCT] C . Breuil , B . Conrad , F . Diamon d an d R . Taylor , On the modularity of elliptic curves
overQ: wild 3-adic exercises, J . Amer . Math . Soc . 1 4 (2001 ) , 843-939 .
[BC] W . E . Brigg s an d S . Chowla , On discriminants of binary quadratic forms with a single
class in each genus, Canad . J . Math . 6 (1 954) , 463-470 .
[Bru] R . W . Bruggeman , Fourier coefficients of cusp forms, Invent , math . 4 5 (1 978) , 1 -1 8 .
[Brl] V . Brun , Uber das Goldbachsche Gesetz und die Anzahl der Primzahlpaare, Archi v fo r
Math.og Naturvid . B 3 4 (1 91 5) , no . 8 .

599
600 B I B L I O G R A P H Y

[Br2] V . Brun , Le crible d'Eratosthene et le theoreme de Goldbach, C . R . Acad . Sci . Pari s

168 (1 91 9) , 544-546 .
[Bu] D . Bump , Automorphic forms and representations, Cambridg e Univ . Press , 1 996 .
[BDHI] D . Bump , W . Duke , J . Hoffstei n an d H . Iwaniec , An estimate for the Hecke eigenvalues
of Maass forms, Internat . Math . Res . Notice s 4 (1 992) , 75-81 .
[Burl] D . A . Burgess , On character sums and L-series, I, Proc . Londo n Math . Soc . (3 ) 1 2
(1962), 1 93-206 .
[Bur2] D . A . Burgess , On character sums and L-series, II, Proc . Londo n Math . Soc . (3 ) 1 3
(1963), 524-536 .
[BH] C . J . Bushnel l an d G . Henniart , An upper bound on conductors for pairs, J . Numbe r
Theory 6 5 (1 997) , 1 83-1 96 .
[Byk] V . A . Bykovsky , Spectral expansion of certain automorphic functions and its number-
theoretical applications, Proc . Steklo v Inst . (LOMI ) 1 3 4 (1 984) , 1 5-33 ; Englis h transl .
in J . Sovie t Math . 3 6 (1 987) , 8-21 .
[Car] F . Carlson , Uber die Nullstellen der Dirichletschen Reihen und der Riemannschen
(-Funktion, Arkiv . fu r Mat . Astr . oc h Fysi k 1 5 (1 920) .
[CF] J.W. S Cassel s an d A . Frolich , Algebraic Number Theory, Academi c Press , 1 990 .
[Chi] Chamiz o an d H . Iwaniec , On the sphere problem, Rev . Mat . Iberoamerican a 1 1 (1 995) ,
417-429.
[Ch] J . R . Chen , On the representation of a larger even integer as the sum of a prime and
the product of at most two primes, Sci . Sinic a 1 6 (1 973) , 1 57-1 76 .
[CS] S . Chowl a an d A . Selberg , On Epstein's zeta-function, J . Rein e Angew . Math . 22 7
(1967), 86-1 1 0 .
[Col] B . Conrey , Zeros of derivatives of Riemann's ^-function on the critical line, J . Numbe r
Theory 1 6 (1 983) , 49-74 .
[Co2] B . Conrey , More than two fifths of the zeros of the Riemann zeta function are on the
critical line, J . Rein e Angew . Math . 39 9 (1 989) , 1 -26 .
[CGG] B . Conrey , A . Ghos h an d S . M . Gonek , A note on gaps between zeros of the zeta
function, Bull . Londo n Math . Soc . 1 6 (1 984) , 421 -424 .
[CI1] B . Conre y an d H . Iwaniec , The cubic moment of central values of automorphic In-
junctions, Ann . o f Math . (2 ) 1 5 1 (2000) , 1 1 75-1 21 6 .
[CI2] B . Conre y an d H . Iwaniec , Spacing of zeros of Hecke L-functions and the class number
problem, Act a Arithmetic a 1 0 3 (2002) , 259-31 2 .
[CoSo] B . Conrey an d K . Soundararajan , Real zeros of quadratic Dirichlet L-functions, Invent .
math. 1 5 0 (2002) , 1 -44 .
[Corl] J.G . va n de r Corput , Zahlentheoretische Abschdtzungen, Math . Ann . 8 4 (1 921 ) , 53-79 .
[Cor2] J.G . va n de r Corput , Verscharfung der Abschdtzungen beim Teilerproblem, Math . Ann .
87 (1 922) , 39-65 .
[Cor3] J . G . va n de r Corput , Sur Vhypothese de Goldbach pour presque tous les nombres
premiers, Act a Arithmetic a 2 (1 937) , 266-290 .
[Cox] D . Cox , Primes of the form x 2 + ny 2, Wiley , 1 989 .
[Cra] H . Cramer , On the order of magnitude of the difference between consecutive prime
numbers, Prac e Mat.-Fiz . 4 5 (1 937) , 51 -74 .
[Dal] H . Davenport , On some infinite series involving arithmetical functions. II, Quart . J .
Math. Oxf . 8 (1 937) , 31 3-320 .
[Da2] H . Davenport , Analytic methods for Diophantine equations and Diophantine inequali-
ties, An n Arbo r Publishers , 1 963 .
[DH1] H . Davenpor t an d H . Halberstam , The values of a trigonometrical polynomial at well
spaced points, Mathematik a 1 3 (1 966) , 91 -96 .
[DH2] H . Davenpor t an d H . Halberstam , Primes in arithmetic progressions, Michiga n Math .
J. 1 3 (1 966) , 485-489 .
[DHe] H . Davenpor t an d Heilbronn , On the class-number of binary cubic forms, I, II, J .
London Math . Soc . 2 6 (1 951 ) , 1 83-1 92 , 1 92-1 98 .
[Del] P . Deligne , La conjecture de Weil, I, Inst . Haute s Etude s Sci . Publ . Math . 4 3 (1 972) ,
206-226.
[De2] P . Deligne , La conjecture de Weil, II, Inst . Haute s Etude s Sci . Publ . Math . 5 2 (1 980) ,
137-252.
 BIBLIOGRAPHY 601

[De3] P . Deligne , Cohomologie etale, SGA 4-\, Lectur e Notes . Math . 569 , Springe r Verlag ,
1977.
[DeSe] P . Delign e an d J-P . Serre , Formes modulaires de poids 1 , Ann . Sci . Ecol e Norm . Sup .
(4) 7 (1 974) , 507-530 .
[DI] J.M . Deshouiller s an d H . Iwaniec , Kloosterman sums and Fourier coefficients of cusp
forms, Invent , math . 7 0 (1 982/83) , 21 9-288 .
[DS] H . Diamon d an d J . Steinig , An elementary proof of the prime number theorem with a
remainder term, Invent , math . 1 1 (1 970) , 1 99-258 .
[Dir] P.G . Dirichlet , Demonstration d'une propriete analogue a la loi de reciprocity qui existe
entre deux nombres premiers quelconques, J . Rein e Angew . Math . 9 (1 832) , 379—389 .
[Dul] W . Duke , The dimension of the space of cusp forms of weight one, Internat . Math .
Res. Notice s (1 995) , 99-1 09 .
[Du2] W . Duke , Hyperbolic distribution problems and half-integral weight Maass forms, In -
vent. math . 9 2 (1 988) , 73-90 .
[Du3] W . Duke , The critical order of vanishing of automorphic L-functions with large level,
Invent, math . 1 1 9 (1 995) , 1 65-1 74 .
[Du4] W . Duke , Some problems in multidimensional analytic number theory, Act a Arith -
metica 5 2 (1 989) , 203-228 .
[Du5] W . Duke , Elliptic curves with no exceptional primes, C . R . Acad . Sci . Pari s Ser . I
Math. 32 5 (1 997) , 81 3-81 8 .
[DFIl] W . Duke , J . Friedlande r an d H . Iwaniec , A quadratic divisor problem, Invent , math .
115 (1 994) , 209-21 7 .
[DFI2] W . Duke , J . Friedlande r an d H . Iwaniec , Bounds for automorphic L-functions, II,
Invent, math . 1 1 5 (1 994) , 21 9-239 .
[DFI3] W . Duke , J . Friedlande r an d H . Iwaniec , The subconvexity problem for Artin In-
junctions, Invent , math . 1 4 9 (2002) , 489-577 .
[DFI4] W . Duke , J . Friedlande r an d H . Iwaniec , Bilinear forms with Kloosterman fractions,
Invent, math . 1 2 8 (1 997) , 23-43 .
[DFI5] W . Duke , J . Friedlander , H . Iwaniec, Equidistribution of roots of a quadratic congruence
to prime moduli, Ann . o f Math . (2 ) 1 4 1 (1 995) , 423-441 .
[Dull] W . Duk e an d H . Iwaniec , Bilinear forms in the Fourier coefficients of half-integral
weight cusp forms and sums over primes, Math . Ann . 28 6 (1 990) , 783-802 .
[DuI2] W . Duk e an d H . Iwaniec , Estimates for coefficients of L-functions, I, Automorphi c
forms an d analyti c numbe r theor y (Montreal , P Q , 1 989) , CRM , 1 989 , pp . 43-47 .
[DuI3] W . Duk e an d H . Iwaniec , Estimates for coefficients of L-functions, II, Proceeding s
of th e Amai n Conferenc e o n Analyti c Numbe r Theory , Univ . Salerno , Salerno , 1 992 ,
pp. 71 -82 .
[DuI4] W . Duk e an d H . Iwaniec , Estimates for coefficients of L-functions, IV, Amer . J . Math .
116 (1 994) , 207-21 7 .
[DK] W . Duk e an d E . Kowalski , A problem of Linnik for elliptic curves and mean-value
estimates for automorphic representations, Invent , math . 1 3 9 (2000) , 1 -39 .
[D] F . J . Dyson , Statistical theory of the energy levels of complex systems, I, J . Mathemat -
ical Phys . 3 (1 962) , 1 40-1 56 .
[Ell] J . Ellenberg , Galois representations attached to Q- curves and the generalized Fermat
equation A A + B 2 = C p, Amer . J . Math , (t o appear) .
[El] P.D.T.A . Elliott , On inequalities of large sieve type, Act a Arithmetic a 1 8 (1 971 ) , 4 0 5 -
422.
[EK] P . Erdo s an d M . Kac , The Gaussian law of errors in the theory of additive number
theoretic functions, Amer . J . Math . 6 2 (1 940) , 738-742 .
[Est] T . Estermann , On Goldbach's Problem: Proof that Almost All Even Positive Integers
are Sums of Two Primes, Proc . Londo n Math . Soc . (2 ) 4 4 (1 938) , 307-31 4 .
[Fon] J.M . Fontaine , II n'y a pas de varietes abeliennes sur Z , Invent , math . 8 1 (1 985) ,
515-538.
[FK] E . Fouvr y an d N . Katz , A general stratification theorem for exponential sums, and
applications, J . Rein e Angew . Math . 54 0 (2001 ) , 1 1 5-1 66 .
[FI] E . Fouvr y an d H . Iwaniec , Gaussian Primes, Acta . Arithmetic a 7 9 (1 997) , 249-287 .
[FoM] E . Fouvr y an d P . Michel , Sur le changement de signe des sommes de Kloosterman
(preprint).
602 BIBLIOGRAPH Y

[Fri] J . Priedlander , Primes in arithmetic progressions and related topics, Analyti c Numbe r
Theory an d Diophantin e problems , Birkhaiiser , 1 987 , pp . 1 25-1 34 .
[FG] J . Friedlande r an d A . Granville , Limitations to the equi-distribution of primes, III,
Compositio Math . 8 1 (1 992) , 1 9-32 .
[FI1] J . Friedlande r an d H . Iwaniec , The polynomial X 2 + Y 4 captures its primes, Ann . o f
Math. (2) 1 4 8 (1 998) , 945-1 040 .
[FI2] J . Friedlande r an d H . Iwaniec , Summation formulae for coefficients of L-functions (t o
appear).
[FI3] J . Friedlande r an d H . Iwaniec , A Note on Character Sums, Contemporar y Math . 1 6 6
(1994), 295-299 .
[FI4] J . Friedlande r an d H . Iwaniec , Incomplete Kloosterman sums and a divisor problem,
Ann. o f Math . 1 2 1 (1 985) , 31 9-350 .
[Gal] P.X . Gallagher , A large sieve density estimate near a — 1 , Invent , math . 1 1 (1 970) ,
329-339.
[Ga2] P.X . Gallagher , Primes in progressions to prime-power modulus, Invent , math . 1 6
(1972), 1 91 -201 .
[Ga3] P . X . Gallagher , BombierVs mean value theorem, Mathematik a 1 5 (1 968) , 1 -6 .
[Ga4] P . X . Gallagher , The large sieve and probabilistic Galois theory, Proc . Sympos . Pur e
Math., Vol . XXIV , Amer . Math . S o c , 1 973 , pp . 91 -1 01 .
[GM] M . L. Gaudin an d M . Mehta , On the density of eigenvalues of a random matrix, Nuclea r
Phys. 1 8 (1 960) , 420-427 .
[Ge] F . Gert h III , Extension of conjectures of Cohen and Lenstra, Expositione s Math . 5
(1987), 1 81 -1 84 .
[GeJ] S . Gelbar t an d H . Jacquet , A relation between automorphic representations of GL(2)
and GL(3), Ann . Sci . Ecol e Norm . Sup . (4 ) 1 1 (1 978) , 471 -542 .
[GJ] R . Godemen t an d H . Jacquet , Zeta functions of simple algebras, Lectur e Note s Math .
260, Springe r Verlag , 1 972 .
[Gol] D . Goldfeld , A simple proof of SiegeVs theorem, Proc . Nat . Acad . Sci . U.S.A. 7 1 (1 974) ,
1055.
[Go2] D . Goldfeld , The class number of quadratic fields and the conjectures of Birch and
Swinnerton-Dyer, Ann . Scuol a Norm . Sup . Pis a CI . Sci . (4 ) 3 (1 976) , 624-663 .
[GS] D . Goldfel d an d P . Sarnak , Sums of Kloosterman sums, Invent , math . 7 1 (1 983) , 243 -
250.
[GHL] D . Goldfeld , J . Hoffstei n an d D . Lieman , Annal s o f Math . (2 ) 1 4 0 (1 994) , 1 61 -1 81 .
[GHB] D . Goldsto n an d D . R . Heath-Brown , A note on the differences between consecutive
primes, Math . Ann . 26 6 (1 984) , 31 7-320 .
[GR] I.S . Gradhstey n an d I.M . Rizhik , Table of integrals, series and products, 6t h Edition ,
Academic Press , 2000 .
[Gl] S . Graham , An asymptotic estimate related to Selberg's sieve, J . Numbe r Theor y 1 0
(1978), 83-94 .
[G2] S . Graham , On Linnik's constant, Act a Arithmetic a 3 9 (1 981 ) , 1 63-1 79 .
[GK] S. W Graha m an d G . Kolesnik , van der Corput's method of exponential sums, Cam -
bridge Univ . Press , 1 991 .
[GRi] S . W . Graha m an d C . J . Ringrose , Lower bounds for least quadratic nonresidues, An -
alytic numbe r theor y (Allerto n Park , IL , 1 989) , Birkhauser , 1 990 , pp . 269-309 .
[Gra] A . Granville , Unexpected irregularities in the distribution of prime numbers, Pro -
ceedings o f th e Internationa l Congres s o f Mathematicians , Vol . 1 , 2 (Zurich , 1 994) ,
Birkhauser, Basel , 1 995 , pp . 388-399 .
[Gr] G . Greaves , Sieves in number theory, Ergebniss e de r Mathemati k un d ihre r Grenzge -
biete (3) , vol . 43 , Springe r Verlag , 2001 .
[GZ1] B . Gros s an d D . Zagier , Heegner points and derivatives of L-series, Invent , math . 8 4
(1986), 225-320 .
[GZ2] B . Gros s an d D . Zagier , Points de Heegner et derivees de fonctions L, C . R . Acad . Sci .
Paris Ser . I Math . 29 7 (1 983) , 85-87 .
[HaTu] G . Halas z an d P.Turan , On the distribution of roots of Riemann zeta and allied func-
tions, I, J . Numbe r Theor y 1 (1 969) , 1 21 -1 37 .
[Hal] H . Halberstam , On the distribution of additive number-theoretic functions, II, III, J .
London Math . Soc . 3 1 (1 956) , 1 -1 4 , 1 4-27 .
 BIBLIOGRAPHY 60 3

[HaRi] H . Halbersta m an d H . E . Richert , Sieve methods, Academi c Press , 1 974 .

[HaLa] G . H . Hard y an d E . Landau , The lattice points of a circle, Proc . Roya l Soc . A 1 0 5
(1924), 244-258 .
[HL1] G . H . Hard y an d J . E . Littlewood , Some Problems of 'Partitio Numerorum. 7 III. On
the Expression of a Number as a Sum of Primes., Act a Math . 4 4 (1 922) , 1 -70 .
[HL2] G . H . Hard y an d J . E . Littlewood , The zeros of Riemann 's zeta function on the critical
line, Math . Z . 1 0 (1 921 ) , 283-31 7 .
[HR] G.H . Hard y an d S . Ramanujan , Asymptotic Formulae in Combinatory Analysis, Proc .
London Math . Soc . 1 7 (1 91 8) , 75-1 1 5 .
[HW] G.H . Hard y an d E.M . Wright , An introduction to the theory of numbers, 5t h edition ,
Oxford Universit y Press , 1 979 .
[HT] M . Harri s an d R . Taylor , The geometry and cohomology of some simple Shimura vari-
eties, Princeto n Univ . Press , 2002 .
[Ha] R . Hartshorne , Algebraic Geometry, Grad . Text s i n Math . 52 , Springe r Verlag , 1 977 .
[HB1] D.R . Heath-Brown , A mean value estimate for real character sums, Act a Arithmetic a
72 (1 995) , 235-275 .
[HB2] D.R . Heath-Brown , An estimate for Heilbronn's exponential sum, Analyti c numbe r
theory, Vol . 2 (Allerto n Park , IL , 1 995) , Birkhauser , 1 995 , pp . 451 -463 .
[HB3] D . R . Heath-Brown , Prime numbers in short intervals and a generalized Vaughan
identity, Canad . J . Math . 3 4 (1 982) , 1 365-1 377 .
[HB4] D . R . Heath-Brown , Zero-free regions for Dirichlet L-functions, and the least prime in
an arithmetic progression, Proc . Londo n Math . Soc . (3 ) 6 4 (1 992) , 265-338 .
[HB5] D . R . Heath-Brown , Lattice points in the sphere, Numbe r theor y i n progress , Vol . 2 ,
de Gruyter , Berlin , 1 999 , pp . 883-892 .
[HB6] D . R . Heath-Brown , Cubic forms in ten variables, Proc . Londo n Math . Soc . (3 ) 4 7
(1983), 225-257 .
[HBP] D.R . Heath-Brow n an d S.J . Patterson , The distribution of Kummer sums at prime
arguments, J . Rein e Angew . Math . 31 0 (1 979) , 1 1 1 -1 30 .
[Heel] E . Hecke , Uber eine neue Art von Zetafunktionen, Math . Zeit . 6 (1 920) , 1 1 -51 .
[Hee] K . Heegner , Diophantische Analysis und Modulfunktionen, Math . Z . 5 6 (1 952) , 227 -
253.
[H] D . Hilbert , Beweis fur die Darstellbarkeit der ganzen Zahlen durch eine feste Anzahl
n-ter Potenzen (Waringsches Problem), Math . Annale n 6 7 (1 909) , 281 -305 .
[Hil] A . Hildebrand , An asymptotic formula for the variance of an additive function, Math .
Z. 1 8 3 (1 983) , 1 45-1 70 .
[Hi2] A . Hildebrand , On the constant in the Poly a-Vinogradov inequality, Canad . Math .
Bull. 3 1 (1 988) , 347-352 .
[Hof] J . Hoffstein , On the Siegel-Tatuzawa theorem, Act a Arithmetic a 3 8 (1 980/81 ) , 1 67 -
174.
[HL] J . Hoffstei n an d P . Lockhart , Coefficients of Maass forms and the Siegel zero, Annal s
of Math . 1 4 0 (1 994) , 1 61 -1 81 .
[Ho] G . Hoheisel , Primzahlprobleme in der Analysis, S.-B . Preuss . Akad . Wiss . Phys.-Math .
Kl. (1 930) , 580-588 .
[HI] Loo-Ken g Hua , Introduction to number theory, Springer , 1 982 .
[H2] Lo o Ken g Hua , On Waring's problem, Quart . J . Math . Oxfor d 9 (1 938) , 1 99-202 .
[Hub] H . Huber , Zur analytischen Theorie hyperbolischen Raumformen und Bewegungsgrup-
pen, Math . Ann . 1 3 8 (1 959) , 1 -26 .
[Hul] M . N . Huxley , The large sieve inequality for algebraic number fields, Mathematik a 1 5
(1968), 1 78-1 87 .
[Hu2j M . N . Huxley , Large values of Dirichlet polynomials, A.cta Arithmetic a 2 4 (1 973) ,
329-346.
[Hu3] M . N . Huxley , On the differences between consecutive primes, Invent , math . 1 5 (1 972) ,
164-170.
[Hu4] M . N . Huxley , Area, lattice points, and exponential sums, Th e Clarendo n Press , 1 996 .
[HuW] M . N . Huxle y an d N . Watt , Exponential sums and the Riemann zeta function, Proc .
London Math . Soc . 5 7 (1 988) , 1 -24 .
[Ing] A . E . Ingham , On the estimation of N(a,T), Quart . J . Math . 1 1 (1 940) , 291 -292 .
604 BIBLIOGRAPH Y

[IR] K . Irelan d an d M . Rosen , A Classical Introduction to Modern Number Theory, 2n d

Edition, Grad . Text s i n Math . 84 , Springer-Verlag , 1 990 .
[Iv] A . Ivic , The Riemann zeta-function, Theory and applications, Dove r Publications ,
2003.
[II] Iwaniec , Character sums and small eigenvalues for TQ (p), Glasgo w Math . J . 2 7 (1 985) ,
99-116.
[12] H . Iwaniec , On zeros of DirichleVs L series, Invent , math . 2 3 (1 974) , 97-1 04 .
[13] H . Iwaniec , Spectral theory of automorphic functions and recent developments in ana-
lytic number theory, Proceeding s o f th e IC M Berkele y 1 986 , Amer . Math . S o c , 1 987 ,
pp. 444-456 .
[14] H . Iwaniec , Topics in Classical Automorphic Forms, A.M.S , 1 997 .
[15] H . Iwaniec , Introduction to the spectral theory of automorphic forms, 2n d edition ,
A.M.S an d R.M.I , 2002 .
[16] H . Iwaniec , Fourier coefficients of modular forms of half-integral weight, Invent , math .
8 7 (1 987) , 385-401 .
[17] H . Iwaniec , The spectral growth of automorphic L-functions, J . Rein e Angew . Math .
4 2 8 (1 992) , 1 39-1 59 .
[18] H . Iwaniec , The half-dimensional sieve, Act a Arithmetic a 2 9 (1 976) , 69-95 .
[19] H . Iwaniec , A new form of the error term in the linear sieve, Act a Arithmetic a 3 7
(1980), 307-320 .
[110] H . Iwaniec , Almost primes represented by quadratic polynomials, Invent , math . 4 7
(1978), 1 71 -1 88 .
[III] H . Iwaniec , Small eigenvalues of Laplacian for T o (AT), Act a Arithmetic a 5 6 (1 990) ,
65-82.
[112] H . Iwaniec , Prime geodesic theorem, Journa l Rein e Angew . Math . 3 4 9 (1 984) , 1 36-1 59 .
[113] H . Iwaniec , Nonholomorphic modular forms and their applications, Modula r form s
(Durham, 1 983) , Horwood , 1 984 , pp . 1 57-1 96 .
[114] H . Iwaniec , The lowest eigenvalue for congruence groups, Topic s i n geometry , Birk -
hauser, 1 996 , pp . 203-21 2 .
[ILS] H . Iwaniec , W . Lu o an d P . Sarnak , Low-lying zeros of families of L-functions, Inst .
Hautes Etude s Sci . Publ . Math . 9 1 (2001 ) , 55-1 31 .
[IM] H . Iwanie c an d P . Michel , The second moment of the symmetric square L-functions,
Ann. Acad . Sci . Fenn . Math . 2 6 (2001 ) , 465-482 .
[ISl] H . Iwanie c an d P . Sarnak , Perspectives on the analytic theory of L-functions, GAF A
Special V o l u m e G A F A 2 0 0 0 (2000) , 705-741 .
[IS2] H . Iwanie c an d P . Sarnak , The non-vanishing of central values of automorphic L-
functions and Landau-Siegel zeros, Israe l J . Math . 1 2 0 (2000) , 1 55-1 77 .
[JS] H . Jacque t an d J . Shalika , On Euler products and the classification of automorphic
representations, I and II, Amer . J . Math . 1 0 3 (1 981 ) , 499-588 , 777-81 5 .
[JPS] H . Jacquet , I . Piatetskii-Shapir o an d J . Shalika , Rankin-Selberg convolutions, Amer . J .
Math. 1 0 5 (1 983) , 367-464 .
[Jul] M . Jutila , Lectures on a method in the theory of exponential sums, Springe r Verlag ,
1987.
[Ju2] M . Jutila , On character sums and class numbers, J . Numbe r Theor y 5 (1 973) , 203-21 4 .
[Ju3] M . Jutila , On large values of Dirichlet polynomials, Topic s i n numbe r theor y (Proc .
Colloq., Debrecen , 1 974) , North-Holland , 1 976 , pp . 1 29-1 40 .
[Ju4] M . Jutila , Zero-density estimates for L-functions, Act a Arithmetic a 3 2 (1 977) , 55-62 .
[Ju5] M . Jutila , Statistical Deuring-Heilbronn phenomenon, Act a Arithmetic a 3 7 (1 980) ,
221-231.
[Kl] N . Katz , Twisted L-functions and m,onodromy, Princeto n Univ . Press , 2002 .
[K2] N . Katz , Exponential sums over finite fields and differential equations over the complex
numbers: some interactions, Bull . Amer . Math . Soc . (N.S. ) 2 3 (1 990) , 269-309 .
[K3] N . Katz , Gauss sums, Kloosterman sums and monodromy groups, Princeto n Univ .
Press, 1 988 .
[K4] N . Katz , Sommes exponentielles, S.M.F , 1 980 .
[K5] N . Katz , Sums of Betti numbers in arbitrary characteristic, Finit e Field s Appl . 7
(2001), 29-44 .
 BIBLIOGRAPHY 605

[KS1] N . Kat z an d P . Sarnak , Random matrices, Frobenius eigenvalues, and monodromy,

A.M.S, 1 999 .
[KS2] N . Kat z an d P . Sarnak , Zeroes of zeta functions and symmetry, Bull . Amer . Math .
Soc. 3 6 (1 999) , 1 -26 .
[Kh] A . Y . Khinchine , Three pearls of number theory, Dove r Publications , 1 998 .
[KSa] H . Ki m an d P . Sarnak , Refined estimates towards the Ramanujan and Selberg conjec-
tures, J . Amer . Math . Soc . 1 6 (2003) , 1 75-1 81 .
[KSh] H . Ki m an d F . Shahidi , Cuspidality of symmetric powers with applications, Duk e Math .
J. 1 1 2 (2002) , 1 77-1 97 .
[Klo] H.D . Kloosterman , On the representation of numbers in the form ax 2 + by 2 + cz 3 -\-dt 2,
Acta Math . 4 9 (1 926) , 407-464 .
[Kol] E . Kowalski , Analytic problems for elliptic curves (2001 ) (preprint) .
[Ko2] E . Kowalski , On the "reducibility" of arctangents of integers, Amer . Math . Monthl y
111 (2004) , 351 -354 .
[KM1] E . Kowalsk i an d P . Michel , A lower bound for the rank of Jo(q)j Act a Arithmetic a 9 4
(2000), 303-343 .
[KM2] E . Kowalsk i an d P . Michel , The analytic rank of Jo(q) and zeros of automorphic L-
function, Duk e Math . J . 1 0 0 (1 999) , 503-542 .
[KMVl] E . Kowalski , P . Miche l an d J . VanderKam , Mollification of the fourth moment of
automorphic L-functions and arithmetic applications, Invent , math . 1 4 2 (2000) , 9 5 -
151.
[KMV2] E . Kowalski , P . Miche l an d J . VanderKam , Rankin-Selberg L-functions in the level
aspect, Duk e Math . J 1 1 4 (2002) , 1 23-1 91 .
[KMV3] E . Kowalski , P . Miche l an d J . VanderKam , Non-vanishing of high derivatives of au-
tomorphic L-functions at the center of the critical strip, J . fu r di e Rein e un d Angew .
Math. 52 6 (2000) , 1 -34 .
[Kor] N . M . Korobov , Estimates of trigonometric sums and their applications, Uspeh i Mat .
Nauk. 1 3 (1 958) , 1 85-1 92 .
[Kubl] J . Kubilius , Probability methods in number theory, Usp . Mat . Nauk . 6 8 (1 956) , 31 -66 .
[Kub2] J . Kubilius , Sharpening of the estimate of the second central moment for additive
arithmetical functions, Litovsk . Mat . Sb . 2 5 (1 985) , 1 04-1 1 0 .
[Kuz] N . V . Kuznetsov , The Petersson conjecture for cusp forms of weight zero and the
Lmnik conjecture, Mat . Sb . (N.S. ) I l l (1 980) , 334-383 ; Math . USSR-S b 3 9 (1 981 ) ,
299-342.
[LO] J . Lagaria s an d A . Odlyzko , Effective versions of the Chebotarev Density Theorem,
Algebraic numbe r fields : L-function s an d Galoi s propertie s (Proc . Sympos. , Univ .
Durham, Durham , 1 975) , Academi c Press , 1 977 , pp . 409-464 .
[Lai] E . Landau , Uber die Einteilung der positiven ganzen Zahlen in vier Klassen nach der
Mindestzahl der zu ihrer additiven Zusammensetzung erforderlichen Quadrate, Arch .
der Math . u . Phys . (3 ) 1 3 (1 908) , 305-31 2 .
[La2] E . Landau , Bemerkungen zum Heilbronnschen Satz, Act a Arithmetic a 1 (1 936) , 1 —1 8 .
[La3] E . Landau , Uber die Nullstellen der Dirichletschen Reihen und der Riemannschen
C-Funktion, Arkiv . fu r Mat . Astr . oc h Fysi k 1 6 (1 921 ) .
[La] S . Lang , Algebraic Number Theory, 2n d edition , Grad . Text s i n Math . 1 1 0 , Springer -
Verlag, 1 994 .
[LT] S . Lan g an d H . Trotter , Frobenius distribution in GLi extensions, Lectur e Note s i n
Math. 504 , Springe r Verlag , 1 976 .
[Lau] G . Laumon , Exponential sums and l-adic cohomology: a survey, Israe l J . Math . 1 2 0
(2000), 225-257 .
[Lav] A . F. Lavrik , Approximate functional equations of Dirichlet functions, Izv . Akad . Nau k
SSSR Ser . Mat . 3 2 (1 968) , 1 34-1 85 .
[Leb] N . N . Lebedev , Special functions and their applications, Dove r Publications , 1 972 .
[Lev] N . Levinson , More than one-third of the zeros of the Riemann zetafunction are on
a = 1 /2 , Adv . Math . 1 3 (1 974) , 383-436 .
[L] W . Li , L-series of Rankin type and their functional equations, Math . Ann . 24 4 (1 979) ,
135-166.
[Lil] Yu . V . Linnik , The large sieve, Dokl . Akad . Nau k SSS R 3 0 (1 941 ) , 292-294 . (i n
Russian)
606 BIBLIOGRAPH Y

[Li2] Yu . V . Linnik , The dispersion method in binary additive problems, AMS , 1 963 .
[Li3] Y u V . Linnik , Additive problems and eigenvalues of the modular operators, Proc . In -
t e r n a l Congr . Mathematician s (Stockholm , 1 962) , pp . 270-284 .
[Li4] Yu . V . Linnik , On the least prime in an arithmetic progression, I. The basic theorem,
Rec. Math . [Mat . Sbornik ] N.S . 1 5(57 ) (1 944) , 1 39-1 78 .
[Li5] Yu . V . Linnik , On the least prime in an arithmetic progression. II. The Deuring-
Heilbronn phenomenon, Rec . Math . [Mat . Sbornik ] N.S . 1 5(57 ) (1 944) , 347-368 .
[Li6] Yu . V . Linnik , On Dirichlet's L-series and prime-number sums, Rec . Math . [Mat .
Sbornik] N.S . 1 5(57 ) (1 944) , 3-1 2 .
[Li7] Yu . V . Linnik , New versions and new uses of the dispersion methods in binary additive
problems, Dokl . Akad . Nau k SSS R 1 3 7 (1 961 ) , 1 299-1 302 .
[LR] J . H . va n Lin t an d H . -E . Richert , On primes in arithmetic progressions. Act a Arith -
metica 1 1 (1 965) , 209-21 6 .
[Lit] J . E . Littlewood , On the zeros of the Riemann Zeta-function, Cambridg e Phil . Soc .
Proc. 2 2 (1 924) , 295-31 8 .
[LRW] J . va n d e Lune , H.J. J t e Riele , D.T . Winter , On the zeros of the Riemann zeta function
in the critical strip, IV, Math . Comp . 1 7 4 (1 986) , 667-681 .
[Luo] W . Luo , Nonvanishing of L-values and the Weyl law, Ann . o f Math . (2 ) 1 5 4 (2001 ) ,
477-502.
[LRS] W . Luo , Z . Rudnic k an d P . Sarnak , On Selberg's eigenvalue conjecture, Geom . Funct .
Anal. 5 (1 995) , 387-401 .
[Ma] H . Maass , Uber eine neue Art von nichtanalytischen automorphen Funktionen und
die Bestimmung Dirichletscher Reihen durch Funktionalgleichungen. Math . Ann . 1 2 1
(1949), 1 41 -1 83 .
[Mai] H . Maier , Primes in short intervals, Michiga n Math . J . 3 2 (1 985) , 221 -225 .
[Mar] G . Margulis , Discrete Subgroups of Semisimple Lie Groups, Ergebniss e de r Math , un d
ihrer Grenzgebiet e 68 , Springe r Verlag , 1 991 .
[MSD] B . Mazu r an d P . Swinnerton-Dyer , Arithmetic of Weil curves, Invent , math . 2 5 (1 974) ,
1-61.
[Mes] J-F . Mestre , Formules explicites et minorations de conducteurs de varietes algebriques,
Compositio Math . 5 8 (1 986) , 209-232 .
[Ml] P . Michel , The subconvexity problem for Rankin-Selberg L-functions and equidistribu-
tion of Heegner points, Annals , o f Math , (t o appear) .
[M2] P . Michel , Analytic number theory and families of automorphic L-functions (t o ap -
pear).
[M3] P . Michel , Autour de la conjecture de Sato-Tate pour les sommes de Kloosterman, I,
Invent. Math . 1 2 1 (1 995) , 61 -78 .
[MvdK] P . Miche l an d J . VanderKam , Non-vanishing of high derivatives of Dirichlet L-func-
tions at the central point, J . Numbe r Theor y 8 1 (2000) , 1 30-1 48 .
[Mi] T . Miyake , Modular forms, Springe r Verlag , 1 989 .
[MW] C . Moegli n an d J-L . Waldspurger , Poles des fonctions L de paires pour GL(N), app.
to Le spectre residuel de GL(N), Ann . Sci . EN S (4em e serie ) 2 2 (1 989) , 605-674 .
[Mol] G . Molteni , Upper and lower bounds at s = 1 for certain Dirichlet series with Euler
product, Duk e Math . J . I l l (2002) , 1 33-1 58 .
[Mol] H.L . Montgomery , The analytic principle of the large sieve, Bull . Amer . Math . Soc .
84 (1 978) , 547-567 .
[Mo2] H.L . Montgomery , Topics in multiplicative number theory, Lectur e Note s i n Math . 227 ,
Springer Verlag , 1 971 .
[Mo3] H . Montgomery , Zeros of L-functions, Invent , math . 8 (1 969) , 346-354 .
[Mo4] H . L . Montgomery , The pair correlation of zeros of the zeta function, Analyti c numbe r
theory (Proc . Sympos . Pur e Math. , Vol . XXIV) , Amer . Math . S o c , 1 972 , pp . 1 81 -1 93 .
[MV1] H . L . Montgomer y an d R . C. , Vaughan , The large sieve, Mathematiki a 2 0 (1 973) ,
119-134.
[MV2] H.L . Montgomer y an d R.C . Vaughan , Hilbert's inequality, J . Londo n Math.Soc . (2 ) 8
(1974), 73-82 .
[MV3] H.L . Montgomer y an d R.C . Vaughan , The exceptional set in Goldbach's problem, Act a
Arithmetica 2 7 (1 975) , 353-370 .
 BIBLIOGRAPHY 60 7

[Mori] C . Moreno . Prime number theorems for the coefficients of modular forms, Bull . Amer .
Math. Soc . 7 8 (1 972) , 796-798 .
[Mor2] C . Moreno , Algebraic curves over finite fields, Cambridg e Univ . Press , 1 991 .
[Mor3] C . Moreno , Analytic proof of the strong multiplicity one theorem, Amer . J . Math . 1 0 7
(1985), 1 63-206 .
[Motl] Y . Motohashi , Spectral theory of the Riemann zeta-function, Camdridg e Univ . Press ,
1997.
[Mot2] Y . Motohashi , An induction principle for the generalization of Bombieri's prime num-
ber theorem, Proc . Japa n Acad . 5 2 (1 976) , 273-275 .
[Nak] H . Nakazato , Heegner points on modular elliptic curves, Proc . Japa n Acad . Ser . A
Math. Sci . 7 2 (1 996) , 223-225 .
[Ne] J . Nekovar , On the parity of ranks of Selmer groups, II, C . R . Acad . Sci . Pari s Ser . I
Math. 33 2 (2001 ) , 99-1 04 .
[Od] A . M . Odlyzko , Some analytic estimates of class numbers and discriminants, Invent .
math. 2 9 (1 975) , 275-286 .
[Oej J . Oesterle , Nombres de classes des corps quadratiques imaginaires, Asterisqu e 1 21 —
122 (1 985) , 309-323 .
[Ono] K . Ono , Nonvanishing of quadratic twists of modular L-functions and applications to
elliptic curves, J . Rein e Angew . Math . 53 3 (2001 ) , 81 -97 .
[PP] A . Perell i an d J . Pomykala , Averages over twisted elliptic L-functions, Act a Arith -
metica 8 0 (1 997) , 1 49-1 63 .
[PSa] Y . Petridi s an d P . Sarnak , Quantum unique ergodicity for SL2(0)\M.2> and estimates
for L-functions. Journa l o f Evolutio n Equation s 1 (2001 ) , 277-290 .
[Ph] E . Phillips , The zeta-function of Riemann; further developments of van der CorpuVs
method, Quart . J . Math . 4 (1 933) , 209-225 .
[PS] R . Phillip s an d P . Sarnak , On cusp forms for co-finite subgroups of PSL(2,R), Invent .
math. 8 0 (1 985) , 339-364 .
[Pi] N . Pitt , On shifted convolution of C s(s) with automorphic L-functions, Duk e Math . J .
77 (1 995) , 383-406 .
[Poi] G . Poitou , Sur les petits discriminants, Seminair e Delange-Pisot-Poitou , 1 8 e anne e
(1976-77).
[Pol] G . Polya , Uber die Verteilung der quadratischen Reste und Nichtreste, Nachr . Konigl .
Gesell. Wissensch . Gottingen , Math.-phys . Klass e (1 91 8) , 21 -29 .
[Pos] A . G . Postnikov , On Dirichlet L-series with the character modulus equal to the power
of a prime number, J . India n Math . Soc . (N.S. ) 2 0 (1 956) , 21 7-226 .
[PR] A . G . Postniko v an d N . P . Romanov , A simplification of A. Selberg's elementary proof
of the asymptotic law of distribution of prime numbers, Uspeh i Mat . Nau k (N.S. ) 1 0
(1955), 75-87 .
[R] H . Rademacher , On the Partition Function p(n), Proc . Londo n Math . Soc . 4 3 (1 937) ,
241-254.
[Ra] D . Ramakrishnan , Modularity of the Rankin-Selberg L-series, and multiplicity one for
SL(2), Annal s o f Math . (2) 1 5 2 (2000) , 45-1 1 1 .
[Ral] R . A . Rankin , Van der CorpuVs method and the theory of exponent pairs, Quart . J .
Math. (2 ) 6 (1 955) , 1 47-1 53 .
[Ra2] R . A . Rankin , The difference between consecutive prime numbers, V, Proc . Edinburg h
Math. Soc . (2 ) 1 3 (1 962/1 963) , 331 -332 .
[Ra3] R . A . Rankin , Contributions to the theory of Ramanujan's r function and similar
arithmetical functions, II, Proc . Camb . Phil . Soc . 3 5 (1 939) , 351 -372 .
[Re] A . Renyi , On the representation of an even number as the sum of a single prime and
single almost-prime number, Izvestiy a Akad . Nau k SSSR . Ser . Mat . 1 2 (1 948) , 57-78 .
[Rey] E . Reyssat , Quelques aspects des surfaces de Riemann, Progres s i n Math . 77 , Birk -
haiiser, 1 989 .
[Rie] B . Riemann , Uber die Anzahl der Primzahlen unter einer gegebenen Grosse, Monats -
ber. Berlin . Akad . (1 859) , 671 -680 .
[RV] F . Rodriguez-Villegas , Square root formulas for central values of Hecke L-series, II,
Duke Math . J . 7 2 (1 993) , 431 -440 .
[Rot] K.F . Roth , On the large sieves of Linnik and Renyi, Mathematik a 1 2 (1 965) , 1 -9 .
608 BIBLIOGRAPH Y

[Roy] E . Royer , Statistique de la variable aleatoire L(Sym 2 / , 1 ) , Math . Ann . 32 1 (2001 ) ,

667-687.
[Ru] W . Rudin , Real and complex analysis, McGraw-Hill , 1 987 .
[RS] Z . Rudnic k an d P . Sarnak , Zeros of principal L-functions and random matrix theory,
Duke Math . J . 8 1 (1 996) , 269-322 .
[Sal] H . Salie , Uber die Kloostermanschen Summen S(u,v,q), Math . Z . 3 4 (1 931 ) , 91 -1 09 .
[Sal] P . Sarnak , Estimates for Rankin-Selberg L-Functions and Quantum Unique Ergodicity,
J. Funct . Anal . 1 8 4 (2001 ) , 41 9-453 .
[Sa2] P . Sarnak , Class Numbers of indefinite binary quadratic forms, Journa l o f Numbe r
Theory 1 5 (1 982) , 229-247 .
[Sa3] P . Sarnak , Some applications of modular forms, Cambridg e Univ . Press , 1 990 .
[Sa4] P . Sarnak , Arithmetic quantum chaos, Bar-Ila n Univ. , 1 995 .
[Sch] W . Schmidt , Equations over finite fields. An elementary approach, Lectur e Note s i n
Math. 536 , Springe r Verlag , 1 976 .
[ST] D . B . Sear s an d E.C . Titchmarsh , Some eigenfunction formulae, Quart . J . Math . Ox -
ford 1 (1 950) , 1 65-1 75 .
[SI] A . Selberg , The general sieve method and its place in prime number theory, Proc . ICM ,
vol. 1 , Cambridge , MA. , 1 950 , pp . 286-292 .
[S2] A . Selberg , On the estimation of Fourier coefficients of modular forms, Proc . Sympos .
Pure Math. , Vol . VIII , Amer . Math . S o c , 1 965 , pp. 1 -1 5 .
[S3] A . Selberg , Lectures on sieves, Collecte d Paper s Vol . II , Springe r Verlag , 1 991 , pp. 6 6 -
247.
[S4] A . Selberg , On the zeros of Riemann's zeta-function, Skr . Norsk e Vid . Akad . Osl o I
(1942).
[S5] A . Selberg , Bemerkungen uber eine Dirichletsche Reihe, die mit der Theorie der Mod-
ulformen nahe verbunden ist, Arch . Math . Naturvid . 4 3 (1 940) , 47-50 .
[S6] A . Selberg , Harmonic analysis and discontinuous groups in weakly symmetric Rie-
mannian spaces with applications to Dirichlet series, J . India n Math . Soc . (N.S. ) 2 0
(1956), 47-87 .
[S7] A . Selberg , On the estimation of Fourier coefficients of modular forms, Proc . Symp .
Pure Math . 8 , 1 965 , pp . 1 -8 .
[S8] A . Selberg , On the zeros of the zeta function of Riemann, De r Kong . Norsk e Vidensk .
Selsk. Forhand . 1 5 (1 942) , 59-62 .
[Sel] J-P . Serre , Cours d'arithmetique, 2n d edition , P.U.F , 1 977 .
[Se2] J-P . Serre , Modular forms of weight one and Galois representations, Algebrai c numbe r
fields: L-function s an d Galoi s propertie s (Proc . Sympos. , Univ . Durham , Durham ,
1975), Academi c Press , 1 977 , pp . 1 93-268 .
[Se3] J-P . Serre , Congruences et formes modulaires (d'apres H. P. F. Swinnerton-Dyer),
Seminaire Bourbaki , 24 e anne e (1 971 /1 972) , Exp . No . 41 6 , Lectur e Note s i n Math .
317, Springe r Verlag , 1 973 , pp . 31 9-338 .
[Se4] J-P . Serre , Corps locaux, Hermann , 1 968 .
[Se5] J-P . Serre , Representations lineaires des corps finis, Hermann , 1 971 .
[Se6] J-P . Serre , Quelques applications du theoreme de densite de Chebotarev, Publ . Math .
IHES 5 4 (1 981 ) , 1 23-201 .
[Se7] J-P . Serre , Minorations de discriminants, Oeuvres , Vol . Ill , Springer-Verlag , 1 986 ,
pp. 240-243 .
[Se8] J-P . Serre , Letter to J.M. Deshouillers.
[Shah] F . Shahidi , Symmetric power L-functions for GL(2), Ellipti c curve s an d relate d topics ,
CRM Proc . Lectur e Note s 4 , Amer . Math . S o c , 1 994 , pp . 1 59-1 82 .
[Sha] D . Shanks , Class number, a theory of factorization, and genera, 1 96 9 Numbe r Theor y
Institute (Proc . Sympos . Pur e Math. , Vol . XX) , Amer . Math . S o c , 1 971 , pp. 41 5-440 .
[Shi] G . Shimura , On modular forms of half-integral weight, Annal s o f Math . 9 7 (1 973) ,
440-481.
[Sh2] G . Shimura , On the holomorphy of certain Dirichlet series, Proc . Londo n Math . Soc .
(3) 3 1 (1 975) , 79-98 .
[Sh3] G . Shimura , Introduction to the arithmetic theory of automorphic functions, Princeto n
Univ. Press , 1 971 .
 BIBLIOGRAPHY 609

[Siel] C . L . Siegel , Uber die Classenzahl quadratischer Zahlkorper, Act a Arithmetic a 1 (1 936) ,
83-86.
[Sie2] C.L . Siegel , On the theory of indefinite quadratic forms, Ann . o f Math . 4 5 (1 944) ,
577-622.
[Sie3] C . L . Siegel , Lectures on quadratic forms, Tat a Institute , 1 967 .
[Sil] J . Silverman , The arithmetic of elliptic curves, Grad . Text s i n Math . 1 06 , Springer -
Verlag, 1 986 .
[Sou] K . Soundararajan , Nonvanishing of quadratic Dirichlet L-functions at s = ^ , Ann . o f
Math. (2 ) 1 5 2 (2000) , 447-488 .
[Stl] H . Stark , A complete determination of the complex quadratic fields with class-number
one, Michiga n Math . J . 1 4 (1 967) , 1 -27 .
[St2] H . Stark , A transcendence theorem for class-number problems. II, Annal s o f Math . (2)
96 (1 972) , 1 74-209 .
[St3] H . Stark , Some effective cases of the Brauer-Siegel theorem, Invent , math . 2 3 (1 974) ,
135-152.
[Ste] S . A . Stepanov , The number of points of a hyperelliptic curve over a finite prime field,
Izv. Akad . Nau k SSS R Ser . Mat . 3 3 (1 969) , 1 1 71 -1 1 81 .
[Tal] J . Tate , Fourier analysis in number fields and Heche's zeta functions, Algebrai c Num -
ber Theory , Academi c Press , 1 990 , pp . 305-347 .
[Tat] J . Tate , Number theoretic preliminaries, Proceeding s o f Symposi a i n Pur e Math . 33 ,
vol 2 , A.M.S , 1 979 , pp . 3-26 .
[TW] R . Taylo r an d A . Wiles , Ring-theoretic properties of certain Hecke algebras, Ann . o f
Math. (2 ) 1 4 1 (1 995) , 553-572 .
[Tchu] N . G . Tchudakov , Sur le probleme de Goldbach, C . R . (Dokl. ) Acad . Sci . URSS , n . Ser .
17 (1 937) , 335-338 .
[Th] A . Thue , Uber Anndherungswerte algebraischer Zahlen, J . Rein e Angew . Math . 1 3 5
(1909), 284-305 .
[Tl] E.C . Titchmarsh , The theory of functions, 2n d Edition , Oxfor d Univ . Press , 1 939 .
[T2] E . C . Titchmarsh , The theory of the Riemann zeta-function, 2n d edition , Oxfor d Univ .
Press, 1 986 .
[T3] E . C . Titchmarsh , Eigenfunction expansions associated with second-order differential
equations, Clarendo n Press , 1 962 .
[Tot] A . Toth , Roots of quadratic congruences, Internat . Math . Res . Notice s (2000) , 71 9-739 .
[Tul] P . Turan , Uber einige Verallgemeinerungen eines Satzes von Hardy und Ramanujan,
J. Lond . Math . Soc . 1 1 (1 936) , 1 25-1 33 .
[Tu2] P . Turan , Uber die Primzahlen der arithmetischen Progression, Act a Litt . Sci . Szege d
8 (1 937) , 226-235 .
[vdK] J . VanderKam , The rank of quotients of J 0(N), Duk e Math . J . 9 7 (1 999) , 545-577 .
[VdP] M . va n de r Put , Grothendieck's conjecture for the Risch equation y' = ay + b, Indag .
Mathem. N.S . 1 2 (2001 ) , 1 1 3-1 24 .
[Va] R . C . Vaughan , Sommes trigonometriques sur les nombres premiers, C . R . Acad . Sci .
Paris Ser . A- B 28 5 (1 977) , A981 -A983 .
[VI] I . M . Vinogradov , A new estimate for £( 1 + it), Izv . Akad . Nau k SSSR , Ser . Mat . 2 2
(1958), 1 61 -1 64 .
[V2] I . M . Vinogradov , On Weyl's sums, Mat . Sborni k 4 2 (1 935) , 521 -530 .
[V3] I . M . Vinogradov , A new method of estimation of trigonometrical sums, Mat . Sborni k
(1) 4 3 (1 936) , 1 75-1 88 .
[V4] I . M . Vinogradov , Perm . Univ . Fiz.-Mat . ob.-v o Z h 1 (1 91 8) , 1 8-24 .
[V5] I . M.Vinogradov , Some theorems concerning the theory of primes, Math . Sb . 2 4 4
(1937), 1 79-1 95 .
[V6] I . M . Vinogradov , Representation of an odd number as a sum of three primes, Dokl .
Akad. Nau k SSS R 1 5 (1 937) , 291 -294 .
[Vi] A . I . Vinogradov , The density hypothesis for Dirichet L-series, Izv . Akad . Nau k SSS R
Ser. Mat . 2 9 (1 965) , 903-934 .
[Vor] G . Voronoi , Sur une fonction transcendante et ses applications a la sommation de
quelques series, Ann . Sci . Ecol e Norm . Sup . (3) 2 1 (1 904) , 207-267 , 459-533. .
[W] L . Washington, Introduction to cyclotomic fields, Grad . Text s i n Math . 83 , 2nd edition ,
Springer Verlag , 1 997 .
610 BIBLIOGRAPH Y

[Wa] T . Watson , Rankin triple products and quantum chaos, PhD thesis, Princeton Univer-
sity, 2001 .
[Wei] A . Weil , On some exponential sums, Proc . Nat . Acad . Sci . U . S . A . 3 4 (1 948) , 204-207 .
[We2] A . Weil , Uber die Bestimmung Dirichletscher Reihen durch Funktionalgleichungen,
Math. Ann . 1 6 8 (1 967) , 1 49-1 56 .
[Wl] H . Weyl , Uber die Gleichverteilung von Zahlen mod. Eins, Math . Ann . 7 7 (1 91 6) ,
313-352.
[W2] H . Weyl , Zur Abschatzung von C( l + ti), Math . Zeit . 1 0 (1 921 ) , 88-1 01 .
[W] A . Wiles , Modular elliptic curves and Fermat's last theorem, Ann . o f Math . (2 ) 1 4 1
(1995), 443-551 .
[Wil] E . Wirsing , Das asymptotische Verhalten von Summen uber multiplikative Funktionen,
II, Act a Math . Acad . Sci . Hungar . 1 8 (1 967) , 41 1 -467 .
[Wi2] E . Wirsing , Elementare Beweise des Primzahlsatzes mit Restglied, II, J . Rein e Angew .
Math. 2 1 4 / 2 1 5 (1 964) , 1 -1 8 .
[Wi3] E . Wirsing , Growth and differences of additive arithmetic functions, Topic s i n classica l
number theory , Vol . I, I I (Budapest , 1 981 ) , Colloq . Math . Soc . Jano s Bolyai , 34 , North -
Holland, Amsterdam , 1 984 , pp . 1 651 -1 661 .
 Index

A-process, 204 , 21 1 , 21 5 Bilinear form , 1 69 , 1 74 , 1 86 , 21 8 , 326 , 340 ,

^-process, 204 , 21 1 , 21 6 343, 346 , 350 , 42 0
L-functions o f varieties , 9 6 Binary quadrati c forms , 498 , 50 3
A 2 -sieve, 1 60 , 43 0 Birch an d Swinnerton-Dye r Conjecture , 1 48 ,
^-adic cohomology , 26 9 367, 520 , 538 , 578 , 57 9
£-adic cohomology group s with compac t sup - Brun-Titchmarsh inequality , 1 67 , 42 0
port, 30 4 Buchstab identity , 34 9
£-adic sheaf , 31 0 Burgess' bound , 53 6
j-invariant, 36 4
Canonical height , 54 1
Abel-Jacobi Theorem , 299 , 54 3 Cauchy's inequality , 1 62 , 1 69 , 22 0
Abelian variety , 1 45 , 1 4 6 Central character , 54 0
Absolute logarithmi c height , 54 1 Central critica l point , 57 7
Absolute values , 54 1 Central value , 51 4 , 52 9
Additive binar y problem , 46 8 Character sum , 74 , 1 01 , 365, 422 , 42 9
Additive character , 44 , 1 75 , 1 81 , 271 , 467 , Characters, 43 , 48 7
475 Chebotarev Densit y Theorem , 1 43 , 48 9
Additive function , 9 , 2 8 Circle method , 1 71 , 315, 44 3
Additive reduction , 36 5 Class group , 58 , 51 0
Adjoint square , 1 37 , 37 3 Class grou p character , 59 , 1 31 , 1 34 , 361 ,
Algebraic curves , 1 4 5 369, 51 6
Algebraic surface , 31 0 Class number , 58 , 1 25 , 368 , 402 , 504 , 523 ,
Almost primes , 1 5 9 537, 56 9
Ambiguous classes , 50 9 Class Numbe r Formula , 38 , 1 24 , 402 , 51 3 ,
Ambiguous form , 50 7 516
Amplification method , 379 , 59 7 Class numbe r on e problem , 1 24 , 57 7
Amplifier, 59 6 Closed point , 29 7
Analytic L-function , 1 41 , 1 43 , 1 4 9 Cohen-Lenstra heuristics , 51 0
Analytic conductor , 9 5 Combinatorial sieve , 1 6 4
Approximate functiona l equation , 97 , 25 7 Companion sums , 273 , 27 8
Arithmetic etal e fundamenta l group , 30 0 Complete L-function , 9 4
Artin L-functions , 96 , 1 26 , 1 41 , 356 Complete exponentia l sum , 45 8
Artin-Shreier covering , 30 2 Complete family , 57 3
Automorphic L-functions , 84 , 93 , 37 5 Complete sums , 269 , 31 8
Automorphic form , 61 , 93, 1 3 1 Completing technique , 31 8
Auxiliary polynomial , 28 2 Complex multiplication , 367 , 36 9
Averaging technique , 357 , 38 7 Conductor, 45 , 60, 93, 94, 1 09 , 1 30 , 1 49 , 1 90 ,
247, 365 , 539 , 572 , 579 , 59 7
Bad reduction , 36 4 Conductor exponent , 36 5
Bernoulli numbers , 6 7 Congruence subgroup , 35 4
Bernoulli polynomial , 66 , 48 4 Conjugacy classes , 39 4
Bessel function , 72 , 73 , 90, 91 , 245, 258, 358, Constant ter m i n th e Fourie r expansion , 39 0
386, 408 , 41 1 , 454, 51 2 Continuous spectrum , 7 5
Betti numbers , 305 , 30 7 Converse theorems , 37 7
612 INDEX

Convexity bound , 1 01 , 1 1 9 , 1 37 , 244 , 535 , Euler Pentagona l Number s Theorem , 449 ,

549, 553 , 59 5 456
Critical line , 1 1 3 , 547 , 56 2 Euler product , 1 1 , 41 , 45 , 60 , 94 , 1 1 3 , 1 46 ,
Critical strip , 96 , 1 01 , 1 05 , 1 1 3 , 1 4 5 297, 366 , 371 , 493, 521 , 529 , 551 , 56 4
Critical zero , 547 , 56 3 Euler-Maclaurin formula , 66 , 68 , 76 , 206 ,
Cubic Gaus s sum , 49 1 483, 48 5
Cusp, 355 , 38 7 Exceptional character , 37 , 43 4
Cusp form , 61 , 65 , 83 , 1 36 , 1 86 , 356 , 368 , Exceptional eigenvalues , 390 , 399 , 41 0 , 49 7
479, 57 4 Exceptional zero , 93, 111 , 112, 1 21 , 140, 390,
428, 43 4
Dedekind et a function , 450 , 51 6 , 54 4 Exceptional zer o repulsion , 42 8
Dedekind multiplie r system , 45 0 Exclusion-inclusion, 1 45 , 1 56 , 1 71 , 308 . 33 9
Dedekind sum , 450 , 456 , 51 6 Explicit formula , 1 08 , 1 1 8 , 1 27 , 265 , 337 ,
Dedekind zet a function , 1 25 , 51 3 410, 440 , 56 4
Degree, 9 4 Exponent pair , 21 4
Deligne bound , 291 , 357, 379 , 493 , 53 2 Exponent Pai r Hypothesis , 86 , 21 4
Density conjecture , 232 , 249 , 26 5 Exponential integrals , 206 , 20 8
Density theorem , 420 , 547 , 59 6 Exponential sums , 1 97 , 225 , 443 , 45 6
Determinant equation , 38 3
Deuring-Heilbronn phenomenon , 42 8 Family o f L-functions , 96 , 1 75 , 573 , 58 0
Difference betwee n consecutiv e primes , 26 6 Family o f exponentia l sums , 31 2
Differencing process , 201 , 204, 21 1 , 212, 220 , Farey sequence , 451 , 458, 469 , 47 7
Ford circle , 45 2
330, 46 6
Four square s theorem , 46 8
Diophantine approximation , 28 2
Fourier coefficients , 1 7 4
Dirichlet L-functions , 45 , 83 , 84 , 1 82 , 267 ,
Fourier coefficient s o f cus p forms , 31 9 , 40 4
324, 329 , 368 , 388 , 479 , 51 3 , 534 , 573 ,
Fourier coefficient s o f modular forms , 83 , 291,
579
479
Dirichlet approximatio n theorem , 1 99 , 45 7
Fourier expansio n a t infinity , 1 3 2
Dirichlet character , 45 , 1 7 9
Fourier integrals , 20 4
Dirichlet convolution , 1 2 , 59 0
Fricke involution , 83 , 366 , 36 8
Dirichlet diviso r problem , 21 , 79 , 1 98 , 21 5
Frobenius automorphism , 270 , 30 2
Dirichlet polynomial , 1 94 , 229, 253, 429, 549,
Frobenius conjugac y class , 30 2
550, 55 4
Function field, 292 , 30 1
Dirichlet series , 1 1
Functional equation , 81 , 84 , 94 , 244 , 258 ,
Discrete subgroup , 35 4
362, 366 , 368 , 375 , 377 , 51 2 , 530 , 539 ,
Discrete valuation , 292 , 30 1
547, 566 , 58 2
Discriminant, 36 3
Functional equatio n o f Eisenstei n series , 38 9
Dispersion method , 1 7 1
Functor, 30 1
Divisor function , 1 3 , 74 , 86 , 33 4
Fundamental discriminant , 52 , 124, 508, 538,
Divisor o n a curve , 29 3
540, 54 3
Dual sum , 1 70 , 46 8
Fundamental domain , 58 , 354, 396 , 499, 504 ,
Duality principle , 1 70 , 1 74 , 1 84 , 1 89 , 23 4 521
Fundamental Lemma , 1 53 , 1 59 , 43 7
Effective divisor , 293 , 29 8
Fundamental unit , 38 , 402 , 51 6
Eichler-Shimura theory , 36 7
Eisenstein series , 80 , 357, 360, 369 , 388 , 411, Gamma factor , 9 4
479, 491 , 51 1 , 58 8 Gauss circl e problem , 20 , 73 , 1 98 , 21 5
Elliptic curve , 1 45 , 1 90 , 269 , 31 3 , 363 , 403 , Gauss sum , 1 8 , 47 , 49 , 60 , 79 , 84 , 1 1 9 , 1 79 ,
520, 529 , 57 4 192, 1 99 , 239 , 274 , 321 , 347 , 456 , 474 ,
Elliptic differentia l operator , 38 6 478, 488 , 491 , 49 5
Elliptic functions , 54 3 Gauss-Bonnet formula , 35 5
Elliptic motion , 39 5 Gaussian prime , 53 , 1 30 , 29 0
Epstein zet a function , 51 3 Gaussian Unitar y Ensemble , 563 , 57 0
Equidistribution, 1 30 , 1 37 , 1 98 , 31 3 , 352 , Genus character , 362 , 51 6
487 Genus theory , 480 , 506 , 54 0
Etale covering , 30 0 Geometric Frobenius , 302 , 30 4
Euler characteristic , 30 7 Geometric fundamenta l group , 30 4
Euler Idonea l Numbe r Problem , 52 0 Goldbach problem , 1 71 , 33 9
 INDEX 613

Good reduction , 36 4 Jacobian variety , 1 47 , 57 8

Grand Densit y Conjecture , 25 0
Grand Rieman n Hypothesis , 1 01 , 1 1 3 , 1 36 , Kloosterman fractions , 1 85 , 48 1
175, 1 82 , 1 94 , 249 , 324 , 41 9 , 57 7 Kloosterman sum , 1 8 , 78, 1 85 , 187, 278, 308,
GRH, 1 1 3 313, 322 , 382 , 404 , 469 , 476 , 481 , 492 ,
Gross-Zagier curve , 36 8 584
Gross-Zagier formula , 57 9 Kloosterman sum s zeta-function , 41 2
Group la w o n ellipti c curves , 29 1 Kloosterman zet a function , 27 8
Kloosterman-Salie sum , 281 , 35 8
Haar measure , 48 7 Kronecker Limi t Formula , 51 6 , 52 4
Hadamard an d d e l a Vallee Poussin , 41 , 101, Kronecker symbol , 52 , 57 , 1 24 , 506 , 508 , 53 0
306 Kuznetsov formula , 65 , 1 86 , 38 0
Hankel transform , 71 , 99
Harish-Chandra/Selberg transform , 393 , 397, Langlands functoriality , 96 , 369 , 49 3
501 Langlands program , 1 3 2
Harmonic polynomial , 36 2 Laplace operator , 89 , 1 31 , 1 32 , 1 86 , 362 ,
Harmonics, 9 , 1 74 , 239 , 31 9 , 356 , 360 , 38 3 404, 563 , 58 0
Hasse bound , 1 93 , 26 9 Large sieve , 1 64 , 1 67 , 1 74 , 1 94 , 239 , 420, 59 5
Hasse derivative , 28 2 Large siev e inequality , 1 25 , 423 , 58 0
Hasse-Weil zet a function , 99 , 1 45 , 1 46 , 272 , Lattice points , 1 9 7
366, 369 , 539 , 541 , 57 9 Least quadrati c non-residue , 1 8 2
Hecke L-function , 60 , 86 , 366 , 368 , 374 , 379 , Lefschetz trac e formula , 305 , 31 0
511, 57 4 Legendre formula , 1 55 , 34 1
Hecke basis , 57 4 Legendre symbol , 27 1
Hecke character , 56 , 59 , 1 29 , 1 41 , 1 42 , 302 , Length o f close d geodesies , 41 0
362 Length spectrum , 40 1
Hecke congruenc e groups , 35 4 Levinson's method , 55 0
Hecke eigenvalues , 1 34 , 1 92 , 51 3 , 58 2 Lindelof Hypothesis , 1 01 , 116, 1 86 , 1 94 , 214,
Hecke form , 37 2 235, 243 , 256 , 59 7
Hecke operator , 80 , 1 86 , 36 7 Linear equivalenc e classes , 29 8
Heegner point , 494 , 542-544 , 57 9 Linear forms , 497 , 49 9
Heilbronn sums , 30 0 Linear fractiona l transformation , 353 , 50 4
Higher vo n Mangold t function , 1 6 , 59 2 Linearly equivalen t divisors , 29 3
Hilbert Clas s Field , 54 4 Liouville function , 1 4
Hilbert inequality , 1 7 5 Lisse ^-adi c sheaf , 30 1
Hilbert's Theore m 90 , 29 6 Local Langland s conjecture , 1 3 8
Hybrid larg e sieve , 1 8 3 Local roo t number , 1 3 6
Hyperbola method , 22 , 31 , 37, 42 0 Local roots , 9 4
Hyperbolic conjugac y classes , 40 1 Local zet a function , 36 5
Hyperbolic geodesies , 38 4 Log-free zero-densit y estimate , 42 8
Hyperbolic laplacian , 38 5 Low-lying zero , 57 4
Hyperbolic measure , 354 , 51 4
Mobius function , 32 , 36 , 42 , 345 , 421 , 430 ,
Hyperbolic motion , 39 5
Hyperbolic plane , 38 4 443, 446 , 59 0
Hyperelliptic curve , 281 , 28 7 Mobius inversion , 1 3 , 44 , 46 , 1 5 4
Mobius Randomnes s Law , 338 , 44 4
Idoneal number , 50 8 Maass form , 1 31 , 38 7
Imaginary quadrati c field , 1 7 , 56 , 361 , 503 , Major arc , 46 7
508, 54 2 Matrix coefficients , 48 7
Incomplete characte r sum , 31 7 Mellin transform , 61 , 90, 25 7
Incomplete Eisenstei n series , 387 , 38 9 Mertens formula , 3 4
Incomplete gamm a function , 51 3 Minimal model , 36 4
Incomplete Kloosterma n sum , 47 1 Minor arc , 457 , 462 , 464 , 466 , 46 7
Invariant integra l operator , 392 , 39 3 Mixed o f weight s < w, 30 5
Isoperimetric inequality , 39 8 Modular curves , 57 7
Modular form , 1 45 , 35 6
Jacobi inversio n formula , 47 3 Modular for m o f weigh t one , 35 6
Jacobi sum , 49 , 31 1 , 49 1 Modular group , 50 3
Jacobi symbol , 52 , 47 3 Modular interpretation , 54 2
614 INDEX

Modular parameterization , 542 , 543 , 54 5 Principal divisors , 29 3

Modularity conjecture , 1 91 , 366 Principal genus , 50 6
Mollification, 550 , 562 , 58 1 Pure o f weigh t w, 30 5
Mollifier, 25 1
Monodromy group , 31 3 , 57 3 Quadratic character , 1 8 4
Mordell-Weil theorem , 57 9 Quadratic Reciprocit y Law , 5 1
Multiple Kloosterma n sum , 308 , 49 2 Quantum uniqu e ergodicit y conjecture , 49 4
Multiplicative characters , 34 , 271 , 308
Radius o f convergence , 28 9
Multiplicative function , 1 54 , 1 65 , 339 , 52 1
Ramanujan A function , 360 , 367 , 493 , 51 6
Multiplicative inverse , 1 9 , 5 5
Ramanujan r function , 37 2
Multiplicative reduction , 36 5
Ramanujan sum , 1 8 , 44 , 48 , 1 79 , 280 , 323 ,
Multiplicity on e principle , 37 3
325, 461 , 472 , 478 , 481 , 483 , 488 , 587 ,
Near-orthogonality, 1 70 , 1 9 2 589
Nebentypus, 1 3 1 Ramanujan-Petersson conjecture , 95 , 100, 101,
Negative curvature , 38 4 115, 1 31 , 1 37 , 1 46 , 378 , 391 , 48 0
Newton polyhedron , 308 , 30 9 Random matri x theory , 56 3
Norm map , 27 0 Rankin's trick , 341 , 349, 52 5
Numerus idoneus , 50 8 Rankin-Selberg L-function , 97 , 1 1 8 , 1 89 , 375,
378, 535 , 58 0
Order o f vanishin g o f a n L-function , 1 1 7 Rankin-Selberg convolution , 1 4 , 97, 1 06 , 1 1 0
Orthogonality o f characters, 35 , 46, 1 21 , 31 1 , Real character , 46 , 4 9
318, 425 , 482 , 49 2 Real primitiv e character , 38 , 46, 57 , 84 , 1 22 ,
Orthogonality relations , 44 , 45 , 1 79 , 271 , 573, 57 7
511 Real quadrati c field, 38 , 51 6
Reflection method , 243 , 25 7
Polya-Vinogradov inequality , 325 , 32 6 Regular discriminants , 51 0
Pair Correlatio n Conjecture , 266 , 56 9 Regulator, 1 2 5
Parabolic motion , 39 5 Residual spectrum , 38 8
Peter-Weyl theorem , 48 7 Resolvent o f th e laplacian , 390 , 40 5
Petersson formula , 1 36 , 1 87 , 1 88 , 380 , 404 , Riemann Hypothesis , 305 , 306, 309, 320, 337,
411, 579 , 58 0 427, 51 4 , 51 8 , 524 , 547 , 563 , 56 8
Petersson inne r product , 357 , 37 1 Riemann Hypothesi s fo r curve s ove r finite
Petersson norm , 1 38 , 542 , 59 5 fields, 1 46 , 329 , 46 3
Poincare duality , 30 5 Riemann Hypothesi s fo r Dirichle t L-functions ,
Poincare metric , 353 , 38 4 443
Poincare series , 357 , 404 , 50 0 Riemann Hypothesi s fo r ellipti c curve s ove r
Poincare uppe r half-plane , 353 , 38 4 finite fields, 36 6
Point-pair invariant , 39 1 Riemann zet a function , 1 2 , 1 1 9 , 1 97 , 204 ,
Poisson distribution , 26 6 216, 388 , 56 1
Poisson summatio n formula , 61 , 69 , 75 , 99 , Riemann-Roch Theorem , 294 , 296 , 29 8
204, 31 9 , 359 , 391 , 398, 404 , 47 3 Root number , 94 , 1 42 , 1 45 , 367 , 377 , 538 ,
Polya-Vinogradov inequality , 523 , 52 4 574
Positivity, 41 , 1 05 , 1 57 , 1 80 , 37 7 Roots o f a n exponentia l sum , 27 4
Primary element , 5 4 Roots o f quadrati c congruences , 5 5
Prime Numbe r Theorem , 25 , 31 , 1 1 0 , 264 ,
401, 41 9 , 427 , 446 , 448 , 489 , 527 , 56 7 Salie sum , 272 , 323 , 476, 477 , 49 5
Primes i n arithmeti c progression s t o larg e Sato-Tate Conjecture , 1 37 , 289 , 324 , 49 3
moduli, 26 6 Sato-Tate measure , 49 2
Primes i n shor t intervals , 264 , 26 6 Scaling matrix , 355 , 38 7
Primes splittin g completely , 52 1 Scattering matrix , 388 , 39 7
Primitive character , 46 , 48 , 70 , 79 , 85 , 1 1 9 , Selberg Eigenvalu e Conjecture , 96 , 390 , 41 5
179, 244 , 422 , 491 , 508 , 59 7 Selberg sieve , 1 6 0
Primitive conjugac y classes , 396 , 40 2 Selberg trac e formula , 65 , 39 1
Primitive cus p form , 99 , 1 34 , 370 , 373 , 374 , Self-dual L-function , 95 , 1 06 , 57 8
513, 52 9 Semistable, 36 5
Primitive Heck e character , 6 0 Separation o f variables , 326 , 35 0
Primitive ideal , 57 , 50 8 Series o f Kloosterma n sums , 38 0
Primitive root , 27 1 Shifted primes , 30 , 1 5 3
 INDEX 615

Short characte r sums , 289 , 32 6 Zero-density theorem , 249 , 26 4

Siegel mas s formula , 48 1 Zero-detecting polynomials , 25 2
Siegel's bound , 1 24 , 479 , 51 4 Zero-free region , 1 05 , 1 1 0 , 1 28 , 1 35 , 1 38 ,
Siegel-Walfisz Theorem , 41 9 , 427 , 48 9 249, 264 , 336 , 347 , 428 , 42 9
Sieve dimension , 1 5 7 Zeta functio n o f a n exponentia l sum , 27 3
Sieve methods , 25 , 1 53 , 1 71 , 265 , 340 , 349 ,
430, 49 3
Sieve problem , 2 7
Sieve weights , 33 8
Sieving level , 34 9
Sifted sum , 1 5 3
Sifting range , 1 6 1
Sign o f th e functiona l equation , 95 , 53 7
Singular integral , 46 0
Singular series , 460 , 466 , 47 8
Smoothing, 40 , 73 , 74 , 1 1 1 , 1 69 , 239 , 49 7
Special value , 1 94 , 577 , 59 5
Spectral decomposition , 403 , 406, 50 1
Spectral theory , 1 9 , 34 4
Spectral theor y o f Kloosterma n sums , 40 3
Square roo t cancellation , 306 , 31 2 , 31 4
Strict diviso r function , 34 2
Strong multiplicit y on e principle , 1 39 , 37 5
Subconvexity bound , 1 01 , 204, 329, 494, 554,
597
Subdivision method , 23 6
Sums o f Kloosterma n sums , 41 0 , 58 6
Symmetric square , 1 4 , 1 37 , 1 89 , 532 , 535 ,
575, 59 5

Tate twists , 30 1
Tate-Shafarevitch group , 36 7
Tempered diviso r function , 7 5
Theta function , 1 8 , 61 , 85 , 361 , 456 , 473 ,
479, 51 3 , 54 1
Theta multiplier , 1 8 6
Trace map , 27 0
Trivial sheaf , 30 6
Trivial zeros , 9 6
Twisted modula r form , 1 3 3

Uncertainty principle , 56 4
Unramified, 94 , 1 1 8
Upper-bound sieve , 43 0

Von Mangold t function , 1 5 , 31 , 42

Voronoi formula , 39 8

Waring problem , 45 6
Weierstrass equation , 363 , 54 3
Weil bound , 79 , 1 30 , 269, 288, 381, 403, 413,
472, 58 5
Weil conjectures , 37 8
Well-factorable function , 42 0
Well-spaced, 1 75 , 21 8 , 23 2
Weyl Law , 1 86 , 39 1
Weyl shift , 21 6 , 31 4
Weyl sum , 1 98 , 201 , 33 5
This page intentionally left blank
Titles i n Thi s Serie s
53 H e n r y k Iwanie c an d E m m a n u e l Kowalski , Analyti c numbe r theory , 200 4
52 D u s a McDuf f an d D i e t m a r Salamon , J-holomorphi c curve s an d symplecti c topology ,
2004
51 A l e x a n d e r Beilinso n an d Vladimi r Drinfeld , Chira l algebras , 200 4
50 E . B . D y n k i n , Diffusions , superdiffusion s an d partia l differentia l equations , 200 2
49 Vladimi r V . C h e p y z h o v an d Mar k I . Vishik , Attractor s fo r equation s o f
mathematical physics , 200 2
48 Yoa v B e n y a m i n i an d J o r a m Lindenstrauss , Geometri c nonlinea r functiona l analysis ,
Volume 1 , 200 0
47 Yur i I . M a n i n , Frobeniu s manifolds , quantu m cohomology , an d modul i spaces , 1 99 9
46 J . Bourgain , Globa l solution s o f nonlinea r Schrodinge r equations , 1 99 9
45 Nichola s M . K a t z an d P e t e r Sarnak , Rando m matrices , Frobeniu s eigenvalues , an d
monodromy, 1 99 9
44 M a x - A l b e r t K n u s , A l e x a n d e r Merkurjev , an d Marku s R o s t , Th e boo k o f
involutions, 1 99 8
43 Lui s A . Caffarell i an d Xavie r Cabre , Full y nonlinea r ellipti c equations , 1 99 5
42 V i c t o r Guillemi n an d S h l o m o Sternberg , Variation s o n a them e b y Kepler , 1 99 0
41 Alfre d Tarsk i an d S t e v e n Givant , A formalizatio n o f se t theor y withou t variables , 1 98 7
40 R . H . Bing , Th e geometri c topolog y o f 3-manifolds , 1 98 3
39 N . Jacobson , Structur e an d representation s o f Jorda n algebras , 1 96 8
38 O . O r e , Theor y o f graphs , 1 96 2
37 N . J a c o b s o n , Structur e o f rings , 1 95 6
36 W . H . Gottschal k an d G . A . H e d l u n d , Topologica l dynamics , 1 95 5
35 A . C . SchaefFe r an d D . C . Spencer , Coefficien t region s fo r Schlich t functions , 1 95 0
34 J . L . Walsh , Th e locatio n o f critica l point s o f analyti c an d harmoni c functions , 1 95 0
33 J . F . R i t t , Differentia l algebra , 1 95 0
32 R . L . W i l d e r , Topolog y o f manifolds , 1 94 9
31 E . Hill e an d R . S . Phillips , Functiona l analysi s an d semigroups , 1 95 7
30 T . R a d o , Lengt h an d area , 1 94 8
29 A . Weil , Foundation s o f algebrai c geometry , 1 94 6
28 G . T . W h y b u r n , Analyti c topology , 1 94 2
27 S . Lefschetz , Algebrai c topology , 1 94 2
26 N . Levinson , Ga p an d densit y theorems , 1 94 0
25 Garret t BirkhofF , Lattic e theory , 1 94 0
24 A . A . A l b e r t , Structur e o f algebras , 1 93 9
23 G . Szego , Orthogona l polynomials , 1 93 9
22 C . N . M o o r e , Summabl e serie s an d convergenc e factors , 1 93 8
21 J . M . T h o m a s , Differentia l systems , 1 93 7
20 J . L . Walsh , Interpolatio n an d approximatio n b y rationa l function s i n th e comple x
domain, 1 93 5
19 R . E . A . C . P a l e y an d N . W i e n e r , Fourie r transform s i n th e comple x domain , 1 93 4
18 M . Morse , Th e calculu s o f variation s i n th e large , 1 93 4
17 J . M . W e d d e r b u r n , Lecture s o n matrices , 1 93 4
16 G . A . Bliss , Algebrai c functions , 1 93 3
15 M . H . S t o n e , Linea r transformation s i n Hilber t spac e an d thei r application s t o analysis ,
1932
14 J . F . R i t t , Differentia l equation s fro m th e algebrai c standpoint , 1 93 2
 TITLES I N THI S SERIE S

13 R . L . M o o r e , Foundation s o f poin t se t theory . 1 93 2

12 S . Lefschetz , Topology , 1 93 0
11 D . Jackson , Th e theor y o f approximation , 1 93 0
10 A . B . Coble , Algebrai c geometr y an d thet a functions , 1 92 9
9 G . D . Birkhoff , Dynamica l systems , 1 92 7
8 L . P . Eisenhart , Non-Riemannia n geometry , 1 92 7
7 E . T . Bell , Algebrai c arithmetic , 1 92 7
6 G . C . Evans , Th e logarithmi c potential , discontinuou s Dirichle t an d Neuman n problems ,
1927
5.1 G . C . Evans , Functional s an d thei r applications ; selecte d topics , includin g integra l
equations, 1 91 8
5.2 O . Veblen , Analysi s situs , 1 92 2
4 L . E . Dickson , O n invariant s an d th e theor y o f number s
W . F . O s g o o d , Topic s i n th e theor y o f function s o f severa l comple x variables , 1 91 4
3.1 G . A . Bliss , Fundamenta l existenc e theorems , 1 91 3
3.2 E . Kasner , Differential-geometri c aspect s o f dynamics , 1 91 3
2 E . H . M o o r e , Introductio n t o a for m o f genera l analysi s
M. M a s o n , Selecte d topic s i n th e theor y o f boundar y valu e problem s o f differentia l
equations
E. J . Wilczynski , Projectiv e differentia l geometry , 1 91 0
1 H . S . W h i t e , Linea r system s o f curve s o n algebrai c surface s
F . S . W o o d s , Form s o n noneuclidea n spac e
E. B . Va n Vleck , Selecte d topic s i n th e theor y o f divergen t serie s an d o f continue d
fractions, 1 90 5