Progress in Mathematics

Vol. 35

Edited by
J. Coates and
S. Helgason

Arithmetic and
Geometry
Papers Dedicated to
I.R. Shafarevich on the Occasion
of His Sixtieth Birthday
Volume I Arithmetic
Michael Artin,
John Tate,
editors

Birkhauser
Boston - Basel. Stuttgart Birkhauser
Boston Basel Stuttgart
Editors:
Michael Artin John Tate
Mathematics Department Mathematics Deparhnent
Massachusetts Institute of Technology Harvard University
Cambridge, MA 02139 Cambridge, MA 02138
Igor Rostislavovich Shafarevich has made outstanding contribu-
This book was typeset at Stanford University using the TEX document preparation tions in number theory, algebra, and algebraic geometry. The flour-
system and computer modern type fonts by Y. Kitajima. Special thanks go to ishing of these fields in Moscow since World War I1 owes much to his
Donald E. Knuth for the use of this system and for his personal attention in the de- influence. We hope these papers, collected for his sixtieth birthday,
velopment of additional fonts required for these volumes. In addition, we extend will indicate to him the great respect and admiration which mathema-
thanks to the contributors and editors for their patience and gracious help with im- ticians throughout the world have for him.
plementing this system. Michael Artin
Igor Dolgachev
Library of Congress Cataloging in Publication Data John Tate
Main entry under title: A.N. Todorov
Arithmetic and geometry.

(Progress in mathematics ; v. 35-36)

Contents: v. 1. Arithmetic - v. 2. Geometry.
1. Algebra- Addresses, essays, lectures. 2. Geo-
metry, Algebraic- Addresses, essays, lectures.
3. Geometry - Addresses, essays, lectures. 4. Shafare-
vich, I. R. (Igor' Rostislavovich), 1923-
I. Shafarevich, I. R. (Igor' Rostislavovich), 1923-
11. Artin, Michael. 111. Tate, John Torrence,
1925- . IV. Series: Progress in mathematics
(Cambridge, Mass.) ; v. 35-36.
QA7.A67 1983 513'.132 83-7124

ISBN 3-7643-3132-1 (v. 1)

ISBN 3-7643-3133-X (v. 2)

CIP-Kurztitelaufnahme der Deutschen Bibliothek

Arithmetic and geometry : papers dedicated to

I. R. Shafarevich on the occasion of his 60.
birthday 1 Michael Artin ; John Tate, ed. -
Boston ; Basel ; Stuttgart : Birkhauser
(Progress in mathematics ; ...)

NE: Artin, Michael (Hrsg.); Safarevic, Igor' R.:

Festschrift
h
Vol. 2. Geometry
All rights reserved. No part of this publication may be reproduced, stored in a re-
trieval system, or transmitted, in any form or by any means, electronic, mechani-
cal, photocopying, recording or otherwise, without prior permission of the copy-
right owner.

@ Birkhauser Boston, Inc., 1983

ISBN 3-7643-3132-1
Printed in USA
I.R. Shafarevich, Moscow, 1966
 Volume I Arithmetic Volume I1 Geometry
N. Aoki and T. Shioda, Generators of the NCron-Severi Group of a V.I. Arnold, Some Algebro-Geometrical Aspects of the Newton
Fermat Surface Attraction Theory
S. Bloch, p-adic Etale Cohomology M. Artin and J. Denef, Smoothing of a Ring Homomorphism Along a
J.W.S. Cassels, The Mordell-Weil Group of Curves of Genus 2 Section
G.V. Chudnovsky, Number Theoretic Applications of Polynomials with M.F. Atiyah and A.N. Pressley, Convexity and Loop Groups
Rational Coefficients Defined by Extremality Conditions H. Bass, The Jacobian Conjecture and Inverse Degrees
J. Coates, Infinite Descent on Elliptic Curves with Complex R. Bryant and P. Griffiths, Some Observations on the Infinitesimal
Multiplication Period Relations for Regular Threefolds with Trivial Canonical Bundle
N.M. Katz, On the Ubiquity of Pathology in Products H. Hironaka, On Nash Blowing-Up
S. Lang, Conjectured Diophantine Estimates on Elliptic Curves F. Hirzeb~ch,Arrangements of Lines and Algebraic Surfaces
S. Lichtenbaum, Zeta-Functions of Varieties over Finite Fields at s = 1 V.G. Kac and D.H. Peterson, Regular Functions on Certairi Infinite-
B. Mazur and J. Tate, Canonical Height Pairings via Biextensions dimensional Groups
J.S. Milne, The Action of an Automorphism of C on a Shimura Variety W.E. Lang, Examples of Surfaces of General Type with Vector Fields
and its Special Points Yu.1. Manin, Flag Superspaces and Supersymmetric Yang-Mills
N.O. Nygaard, The Torelli Theorem for Ordinary K3 Surfaces over Equations
Finite Fields B. Moishezon, Algebraic Surfaces and the Arithmetic of Braids, I
A.P. Ogg, Real Points on Shimura Curves D. Mumford, Towards an Enumerative Geometry of the Moduli Space
1.1. Piatetski-Shapiro, Special Automorphic Forms on PGSp, of Curves
M. Raynaud, Courbes sur une variCtC abClienne et points de torsion C. Musili and C.S. Seshadri, Schubert Varieties and the Variety of
A. Weil, Euler and the Jacobians of Elliptic Curves Complexes
A. Ogus, A Crystalline Torelli Theorem for Supersingular K3 Surfaces
M. Reid, Decomposition of Toric Morphisms
M. Spivakovsky, A Solution to Hironaka's Polyhedra Game
A.N. Tjurin, On the Superpositions of Mathematical Instantons
A.N. Todorov, How Many Kahler Metrics Has a K3 Surface?
0 . Zariski, On the Problem of Irreducibility of the Algebraic System
of Irreducible Plane Curves of a Given Order and Having a Given
Number of Nodes
 Generators of the NCron-Severi Group
of a Fermat Surface
Noboru Aoki and Tetsuji Shioda

To Z.R. Shafarevich

1. Introduction

The Nkron-Sevcri group of a (nonsingular projective) varicty is, by

definition, the group of divisors rnodulo algebraic equivalence, which is
known to be a fiuitcly generated abelian group (cf. [2]). Its rank is called
the Picard number of the variety. Thus the Nkron-Severi group is defined
in purely algebro-geometric terms, but it is a rather delicate invariant of
arithmetic nature. Perhaps, because of this reason, it usually requires some
nontrivial work before one can determine the l'icard number of a given
variety, let alone the full structure of its N6ron-Severi group. This is the
case even for algebraic surfaces over the field of complex numbers, where
it can be regarded as the subgroup of the cohornology group I12(X, Z)
characterized by the LefscheLz criterion.
NOWthe purpose of the present paper is to find certain explicitly defined
curves on the complex Fermat surface

whose cohomology classes (equivalently, algebraic equivalence classes) form

(a part of') gcrlcrators of the Ntkon-Severi group NS ( X & ) @Q. The Pirard
number p ( X k )has recently been determined and is given by the following
for~nula:

where
 AOKI AND SHIODA GENERATORS OF THE N ~ R O N - S E V E R IGROUP 3

L= { 1
0
(m :
(m :
even)
odd)
Futhermore, it is expected that a similar approach should be applicable
to Fermat varieties of higher dimension for the explicit construction of

(m/3)* = { :I3 (if 3 1 m)

(if 3 m )
algebraic cycles corresponding to given Hodge classes. Combined with the
method based on the inductive structure of Fermat varieties (31, this might
lead to the verification of the Hodge Conjecture for all Fermat varieties.
(m/2)* = { :I2 (m : even)
(m : odd)
and where c(m) is a bounded function of m which is expressed as a certain 2. Preliminaries
sum over divisors d of m such that (dl 6) # 1 and d 5 180 (see Shioda
+ +
[4], Aoki[l]). It is known that 3(m - l ) ( m - 2) 1 6, is the rank of the First recall that the Hodge classes on the Fermat variety
subspace of NS (XL) which is spanned by the lines of the ambient space Xk : zy = 0 are described in terms of the characters of the group
P 3 lying on X L ; in particular, if (m, 6) = 1, then NS ( X & ) @ Q is spanned G k = (I.~,n)n+2/(diagonal)acting on Hn(Xz,C) (see [3,Thm. I]). In the
by the classes of these lines (cf. [4,Thm.7]). special case n = 2, this gives
The main results of this paper can be stated roughly as follows:

(i) One half of the term 48(m/2)* in the above formula of p ( X L )

corresponds to the subspace of NS ( X L ) @ Q spanned by the classes of
curves lying on (the intersection of XL with) certain quadric surfaces of where the index set 8; =
the form
z2-c.xy=0
(up to permutation of coordinates x, y, z , w).
(ii) The term 24(m/3)* corresponds, in a similar sense, to curves lying
on cubic surfaces of the form is naturally regarded as a subset of the character group of G L and V(a)
(resp. V ( 0 ) )is the eigenspace of H 2 ( X % ,C) with character a (resp. trivial
w3 - c . xya = 0 . character) which is known to be 1-dimensional. In particular, the Picard
number of the (complex) Fermat surface X: is
(iii) Another half of the term 48(m/2)* corresponds similarly to curves
lying on quartic surfaces of the form

Next we recall the following structure theorem of B&, which has been
formulated and partially proven in Shioda [4] and recently fully proven by
More precise statements will be given in $3. Aoki [I]. The formula (1.2) is a consequence of this result in view of (2.3).
Concerning the structure of the Ndron-Severi group of the (complex) +
We call an element a = (ao, a l , az, a a ) E B:n decomposable if a; a j =
Fermat surface X$,, the following problems remain to be studied: 0 (mod na) for some i # j , and indecomposable otherwise.

i) to find curves corresponding to the "exceptional" term ~ ( m ) ; T h e o r e m (8%). (i) If (m, 6) = 1, then 8; consists of decomposable
ii) to find genwators of NS ( X k ) over P. elements.
4 AOKI AND SHIODA

Observe t h a t w c is a multiple of p,(cl.(C)) such t h a t

(ii) If (m,6 ) > 1, then every indecomposable element a = ( a o ,a l l an, a s )
of 8% with G C D ( u , ) = 1 i.7 equal ( u p to permutation) to one of the "stan-
dard" elements a,, Pi or 7 j below, except for finitely m a n y "exceptional"
elements which exist only for m <180: as is clear from the invariance of the intersection number under an automor-
phism.
a) m = 2d, a i = ( i ,d -t i, m - 2 i , d ) , 1 5 i < d , (i,d ) = 1,
i # m/4. Theorem 0. For any m > 1, let L denote a line of P3 lying o n X:.
b) m = 2 4 Pi = ( i ,d + i , d + 2 i , m - l i ) , 1 <
i < dl (i,d ) = 1, T h e n I represents a decomposable element o f BL,
and conversely every
i # m/3,m/4,m/6. decomposable element is represented by some lines. Further
c) + +
m =; 3d, y j = ( j , d j , 2d j , m - 3 j ) , 1 5 j < d l ( j ,d ) = 1,
3
j #4 6 . (2.8) WL. GL = -m .

If a = ( a o ,a l , a 2 , a 3 ) is an element of 8 % with G C D ( a i ) = d > 1, Proof. See the proof of [4, Thm. 71.
set a: = a i / d and m' = mid. Then a' = (ab,a:,aL,a:) is a n element

f: X % -
of Bk, with GCD(a:) = 1, and one checks easily that the morphism
X L , given by ( x i ) H ( 2 9 ) induces the map
f *: H 2 ( x & , ) - + H 2 ( x : )
In the next section, we shall exhibit certain curves on X & which rep-
resent the "standard" indecomposable elcments of 8 % stated in Theorem
( 8 % ) . Note that, for this purpose, it sufices to consider the case a = a l l
P I , and 7 1 ,because in general we have V(a)" = V ( t a )for any automor-
such t h a t f * V ( a l )= V ( a ) .Thus, in order to construct explicit curves on phism a of C inducing at : < H ct on the subfield of m-th roots of unity
X k corresponding to V ( a ) ,it sufices to consider the case G C D ( a ; ) = 1. ( t E ( Z / m ) X )while
, cl.(C)O = c l . ( C O ) ,C 0 being the conjugate of C urder
Now define, for any character a of G k , o (this niakes sense since our surface X : is defined over Q).

3. Main Results
Then p, is the projector of H;,.,,,(Xk) to V ( a ) .In particular, if a E BL
and if E = cl.(C) is the class of an algebraic cycle C on X;, then p,(E) ia Theorem 1. Let m = 2 d ( d > 1 ) and a1 = ( 1 , d + 1, m - 2, d). Let
a n ulgebraic cohomology class generating the subspace V ( a )provided that C denote the curve of degree m i n P3 defined by
p,(t) # 0. Note that p a ( ( ) # 0 if and only if the "intersection number"
p a ( t ) . E Z I # 0. (3.1) z2 - q i x y = 0 , xd + yd + G w d = 0.
T h e n C is a nonsingular irreducible curve lying o n X L which represents
Definition. Given a E 8% and a curve C on X L , we say (for short) t h a t the Hodge class a1 and which satisfies
C represents the IIodge class a if the following two conditions are satisfied.
Let G = G&.

Theorem 2. Let m = 3 d ( d 2 1 ) and 71 = ( 1 , d + 1,2d + 1 , m - 3).

def
Let C denote the curve of degree m i n P3 defined by
(2.6) wc = a ( y ) c l . ( g ( C ) )# 0 , i.e., w c - t ~#
c 0.
(3.3) w3 - m x y z =0, x d + yd + zd = 0.
gEG/Gc
6 AOKI AND SFIIODA

T h e n C is a nonsingular irreducible curve lying o n X& which represents 4. Proof of Theorems 1 and 2
4
the IIodge class 71 and which satisfies
First of all look at the polynomial identities:

Theorem 3. Let m = 2 d ( d > 2) and = (1, d + 1,d + 2, m - 4).

(i) I n case d is odd, let C denote the curve of degree 2m o n Xk defined
(in a d d h i o n t o the Fermat equation) by .
Replacing x, y, . . by xd, yd, .. . , it is clear that the curve C defined by (3.1)
or (3.3) in P3 actually lies on the Ferrnat surface X L . It is easy to check
that C is nonsingular. Since a nonsingular complete intersection curve in
+ +
P3 of rnultidegree (d,e) has the genus 1 de(d e - 4)/2, we have

T h e n C is a nonsingular irreducible curve which represents the Hodge class By the adjunction formula, the self-intersection number of C is given respec-
pl and which satisfies tively by

In what follows, we shall prove only Theorem 1, since Theorem 2 can be

proven in the same way.
(ii) I n case d is even, let C' denote the curve of degree 2 m i n P3 defined For g = [1: (1 : 52: (31 E G = G
,: let
by

Then a1 = (1,d + 1, m - 2, d ) E Bk is expressed as p l ( g ) = u(g) . ~ ( g ) .

For si~nplicity,we set 5; = ~ ; ( i= 1,2,3) so that ud(g) = 1, p(g) = 3
T h e n C' ia a singular irreducible curve lying o n Xk which represents the
; f l ) . Now the curve g(C) ig defined by the following
) t l e 3 ( ~=
and ~ ( g =
Hodge class P1 and which satisfies
equations:

Hence g stabilizes C if and only if a(g) = 1 = 3 = 1. Thus

R e m a r k . (a) The value of fi, m,
etc. in the defining equations of
C are fixed once and for all. (b) Theorems 1,2,3 (and Theorem 0 in 52)
remain valid for the Fermat surface x;(~) in characteristic p, provided p
does not divide m. This will be clear from the proof. On the other hand, we see easily:
8 AOKI AND SIIIODA

Since Gc is contained in I'Cer(al), we can define the element w c of Now we are ready to compute w c . ~ j cby (2.6). Let K ( a ) denote the
NS(Xk) @I C m in (2.6): image of Ker(u) in GIGc, and let K(u)* -- K ( u ) - {id.). Similarly we
define K(r), K(p), etc. Then

In order to compute w c - B c by (2.7), we need to know the intersection

numbers (C . g(C)) for g E Ker(u) U Ker(r) U I<er(p). We distinguish the
Note that a(= al) and a . u d induce nontrivial characters on each of the
three cases: (a) g E Ker(a), g @ Ker(r), (b) g @ Ker(u), g E K e r ( ~ ) ,
subgroups K ( u ) , K ( r ) , K(P), and K(T)n K ( p ) of G / G c , and so we have
and (c) g @ Ker(u) U Ker(r), g E Ker(p).
CK(o). a(g) = -1, etc. Hence the above is equal to
Case (a). In this case, we have 1 = 1, 3 = -1. By (4.6), C n g ( C )
consists of the 2d points PC= (1, c2,24/2~,0)(c2d = -1). For each P = PC,
let O p denote the' local ring of X k a t P, regarded as a localization of the
afline ring k[y, z, w] of Xk - {x = 0) (x = 1). The intersection multiplicity Since [G : Gc] = 2m, this implies
of C n g(C) a t P is equal to the length of the quotient ring

completing the proof of Theorem 1. Q.E.D.

Therefore we have 5. Proof of Theorem 3

First we note that the curve C' in P 3 defined by (3.7) actually lies on
X& in case m = 2d with d even. This follows from the elementary identity
Case In this case, we have a(g) f 1, 1 = 3 = *I. The inter-
(b).
section C n
g(C) consists of the 2d points P,, = (0, q, 0 , l ) and
-m)
if 1 -
Qq = (q,0,0,1) (qd = if 1 = I , and of the d points P,, only
-1. The intersection multiplicity of C n g(C) a t each P,, or Qq
is computed as before and is equal to 2. Therefore (C . g(C)) = 4d or 2d, by substituting z, . . . by xdI2,.. . . The curve C' has the d singular points
according as 1 = I or -1. P,, = ( l , q J0,O) where 9d/2 = (- 1 & -)/a. To find the normalization
of C', observe that xy is a square on C' as the first equation of (3.7) shows.
Case (c). In thiscase, w e h a v e & ) # 1, 1 = -1 a n d c 3 = 1. Suggested by this, we consider the following curve 6' in P4:
There are d intersection points (1,0,0, c) (cd = -1) in C n g(C), and the
multiplicity a t each of them is equal to 2. Thus we have (C - g(C)) = 2d.

Since el = u ~ ( ~the
) , case (b) and (c) can be summarized as follows: It is easy to check that C is nonsingular and that the projection of P4
to Pn (x, y, z , w , t ) I--+ (2, y, z, w) defines a birational morphism of ?I onto
Z'
C' in case d is even, and an isomorphism of to C in case d is odd. As a
10 AOKI AND SNIODA

nonsingular complete intersection curve in P4, the genus of 6 is given by Therefore, with the same notation as in the proof of Theorem 1 (54),
+ +
d . 2 . 2 . (d + 2 + 2 - 5)/2 1 = 2d2 - 2d I . Therefore we have by the we have (write p = PI):
adjunction formula

and

[ pa(C1) = 2d2 - d +I (pa : arithmetic genus)

(5.3)

The rest of the proof is simillar t o that of Theorem 1 given in $4. We

shall sketch it only for the case d odd, omitting details.
For g = [l : (1 : 52 : c3] E G, we set

By (2.7) and (5.5), we have

Then we have
This proves part (i) of Theorem 3. Part (ii) is similar (and even simpler),
so it will be omitted. Q.E.D.

For each g E G - Gc, the intersection of the curve C and its transform
g(C) is described as follows:

(1) Assume g &/Ker(a). (1) If g E Ker(r), there are d points (0, y, z,O),
yn + a z d = 0, with intersection multiplicity 4. (2) If g E Ker(adr),
there are d points (x, 0, z, 0), xd +
-zd = 0, with multiplicity 4.
(3) If g E ~ e r ( o or
~ )g E Ker(a2), then there are m points (x, y, O,O),
+
xm ym = 0, with multiplicity 2. (4) If g E G - Ker(ad) U Ker(u2), there
+
are m points (x, y, O,O), xm ym = 0, with multiplicity 4.
(11) Assume g E Kcr(a), g &/Ker(r). Then there are m points (x, y,O, O),
+
xm ym = 0, with multiplicity 2, and m2 points of the form (x, y, z , w),
+
xd yd = 0, ZW f _ 0, with multiplicity 1.
 AOKI AND SI-IIODA

References p-adic Etale Cohomology

[I] Aoki, N.: Properties of Dirichlet characters and its application to
Fermat varieties. Master Thesis, University of Tokyo, 1982 (in S. Bloch
Japanese).
[2] Lang, S.: Diophantine geometry, Intersc. Publishers, New York-
London, 1962. To I.R. Shafarevich
[3] Shioda, T.: The IIodge Conjecture for Fernlat varieties, Math. Ann.
245 (1979), 175-184. What follows is a report on joint work with 0. Gabber and K. Katof
[4] Shioda, T.: On the I'icarcl number of a Fermat surface, J. Fac. Sci. A manuscript with complete proofs exists and is currently being revised.
Univ. Tokyo, Sec. Ih, 28 (1982), 725-734. For co~npellingphysical reasons (vis. time, space, and distance) however,
I will give here only statements of results; and my coauthors have not
had the opportunity to correct any stupidities which may have slipped in.
Received June 18, 1982 The conjectures in $3 are my own. I like to think that this research has
been strongly influenced by the work of Shafarcvich, both by his work on
N. Aoki, Department of Mathematics algebraic geometry in characteristic p and by his work on a r i h n c t i c a l al-
Faculty of Science gebraic geometry. In fact, recently Ogus has used these results to apply the
University of Tokyo basic Rudakov-Shafarevich result on existence and smootliness of moduli
Hongo, Tokyo 113, Japan for K3 surfaces in charactcristic p to the study of the moduli space when
p=2.
Professor Tetsuji Shioda
Department of Mathematics
Faculty of Science
$0 Global Results
University of Tokyo
Hongo, Tokyo 113, Japan
For purposes of this report let the phrase "p-adic field" mean a field K of
charactcristic 0 which is the quotient field of a henselian discrete valuation
ring A with perfect residue field k of characteristic p > 0. For example, Qp
is a p a d i c field. Let V be a smooth, complete vnriet,y over a p a d i c field
K . Let R be the algebraic closure of K , and write G -- Gu!(R/K). The
ktale cohomology groups

are representation spaces for G which are known to give the correct bctti
numbers for V.
Concerning the structure of these cohomology groups one has s basic
conjecture formulated by Tate and proved by him for II' when V has good
reduction. Let C be the completion of W .By colitinuity G a c k on C and
 p-ADIC ETALE COIIOMOLOGY 15
14 S. BLOCH

over k. Associated to 7 one has crystalline cohomology groups H&JP/ W)

hence G acts diagonally on
[I] and dellham-Witt cohomology groups 1Z6(F, Wflt) [l4]. Both are W(R)-
modules with given u-linear frobenius endomorphisms f , where W ( R ) is
the ring of p-Witt vectors over R and o : W ( R ) --+ W ( R ) is induced by
the p-th power map on ff, There is a spectral sequence ("slope spectral
sequencen)
C o n j e c t u r e ( T a t e [25]). There is a canonical isomorphism o f
G-modules

which degenerates at El up to torsion [2], [14].

It turns out that the "p-geometry" of the closed fibre P as expressed in
the deRham-Witt cohomology profoundly influences the structure of the
For a G-module M, we write M ( m ) for the G-module obtained by G- modules g r n H X ( v , Q,).
twisting M by the m-th power of the dual of the cyclotomic character. Also
Rm denotes the sheaf of Kahler m-forms for Vc over C. Recent results on Case 1. Some ZIt(y,WRS) is not finitely generated over W(K). In
Tate's conjecture have been obtained by Fontaine [g].
this worst (or "most interesting" in the sense of the old Chinese proverb)
We now fix a model X of V over S = S P A , i.e. we fix a diagram
case, we don't even formulate a conjecture about g r n H & ( v ,Q,).
i
v - x , 3 x- X.=Y
Case 2. All Ilt(Y, WRS) are finitely generated. Y is said to be Hodge-
(0.2) t I C W i t t (op. cit.). In this case the slope spectral sequence (0.5) is known to
SpK =q -+ S= S p A t s = Sp.k
degenerate at El [15]. The conjecture (discussed in more detail in $3 below)
with all vertical arrows flat and proper. A bar will either indicate algebraic is that writing d = dim Y,
or -integral closure (viz.
- r,
h,K )or base extension (7= X3, V = Vr,
i, j,etc.). There is a spectral sequence
is the Tate module associated to a lifting over S of the p-divisible formal
group with covariant Dieudonnk module H ~ - ~ ( Y ,
WRd+n-q)/(tors).

inducing a G-stable filtration Case 3. Let Bn C R;Lr be the sheaf of locally exact Kahler n-forms on
Y. Y is said to be ordinary if Hq(Y, Bn) = (0) for all q and n. Roughly
(0.4) FoI3;,(V, Z,), grnfIq = FnHq/~ ~ " 1 1 ~ . speaking, "ordinary" is equivalent to "IIodge-Witt" plus equality of Hodge
and Newton polygons. For a more careful discussion see $2 below. Let
Since the pcoho~~~ological
dimension of Y is equal to d i m P [l L] one has
' ) { X E H!,JS(P/W) I / X = p i X ) . Note G acts on this
H R ~ ~ ( Y / W ) (=
group via the action of G a l ( x / k ) on IZ~,,,(r/W).

T h e o r e m (0.6). Let notation be as above and assume Y is ordinary.

Then for all integers q, i there ezist functorial G-isomorphisms
This construction is analogous to one used in the study of vaqishing cycles
[12] and degenerating families of varieties [23]. We assume now and hence-
forth that X is smooth over S, so V has good reduction and Y is smooth
16 S. BLOCH p-ADIC ETALE COIIOMOLOGY

(1.2) LL = Zariski sheaf associated to the presheaf

The spectral sequence (0.3) degenerates up to torsion a t E p .

For example, the generic stalk of L; is the galois cohomology of the quotient
field of the (non-strict) henselization of X a t the generic point of Y.
C o r o l l a r y (0.7). When Y is ordinary, Tate's conjecture (0.1) holds, ML is the sheafification of LL for the ktale topology.
r.e. The structure of LL is closely related to the K-thoery of Ov in the sense
Hit(V, C ) @ EI'(VC, Rm)(-m). of Milnor (i.e. using symbols [20]). the Kummer sequence
l+m=n
The corollary follows from (0.G) and the fact proved in [25] that

induces an exact sequence of Zariski sheaves on Y

for i # 0, so there are no extensions between C ( r ) and C(s) for r # s.

For local sections 21, . .. ,x, of i* j, 0 ,; let

8 1 Local Results

Let notation be as in (0.2), and define ktale sheaves on Y

be the local section of L; defined by the cup product of the images of the
xi in Lk. These "symbols" satisfy the I<-theory type relations

To analyse the spectral sequence (0.3) it is necessary to understand the 0 = (2, -2) = {x, y) + {y, 2) = {2,1 - 1)
local structure of the sheaves ML. Our results in this area rest on Kato's
calculation of the Galois cohomology of quotient fields of henselian discrete for local sections x, y, z such that 1 - z is invertible.
valuation rings [17] and Gabber's theorem [lo]. We need Y to be smooth Define UmLA for m 2 1 to be the image
but make no hypothesis about Y ordinary. Note that the geometric stalks
of ML are the ktale cohomology groups

where x E A is now (and ever shall be) a uniformizing clement. Define

where A is the strict henselization of the local ring on X a t a point y E Y.

For example, taking y to be a geometric generic point, A is a henselian
discrete valuation ring with quotient field A[:] and the result can be read Note L t = (Z/pnZ)= and UmLt = (0) for rn 2 1.
off from [17]. One wants to understand the grFL',. At the moment this is done only
In fact it is possible and occasionally technically convenient to work with in the two cases n .= 1 (which together with the case m = 0 suffices for
a finer object. we view it a s a functor from &ale sheaves on X to Zarieki the proof of (O.G), but probably will noL suffice to understand the "non-
sheaves on Y, and we define ordinary" casc) or m 5 3
where e = ord,(p) is the index of ramification.
18 S. BLOCH p-ADIC ETAlAE COHOMOLOGY

For simplicity I will restrict attention to the case n = 1. Define {n, 21,. . . , x,-l) H dlog(zl) A - . A d l o g ( ~ ~ E- ~R;);tog
(1.4)(iii) (1 + nmy, X I , . . . ,xr-l) +-+ pdlog(~1A
) . - .A dlog(zr-l)
B' = Im(d : n;-I -+ n;)
(1.4)(iv) (1 + xmy, XI,. . . , x,-1) t-+ dB A dlog(z1) A . . A dlog(~,-1) E Br

{x, 1 + r m y , x l , . . . ,z,-1)
C: 2' 4 Qr Cartier operator [5] djj A dlog(T1) A . A dlog(;z',-z) E Br-'
(1.4)(v) (1+ xmy, . . ,~ ~ - 1 )jjdlog(~1)A . A d l o g ( ~ , - ~ )
XI,. +-+

{r,1 + n my, xl, . .. ,xr-2) dlog(f1) /\ . - A dlog(f

H
T h e o r e m (1.4). (i) The sheaves L', are generated Zariski locally by
symbols (1.3).
(ii) 9r$(L;) = q, 63 log %;og Note that the ktnle sheaves M & have similar filtrations U'M', and the
(iii) If 0 < m< 5 and (m, P) = 1, then grF(L;) E n;-' structure of gri;M', is deduced from that of gri;L& by passage t o the &ale
topology. For example
(iv) If 0 < m < 5 and ~ l m then
Br-l
, grF(L1) B; @ y
(v) Assume e' = 3 is an integer, and let a E k be the residue clam
of pn-". Then
The following result (due to Kato) should be important in studying the
non-ordinary case.

(iv) UmL; = (0) for m > el. T h e o r e m (1.7). There exists a unique homomorphism of sheaves
ML -, (R>ls)/p"(R>l,) given by { X I , . . . ,2,) H dlog(x1) A . . . A dlog(x,)
Remark (1.5). The deItham-Witt complex (141 is a complex of pro- for xi local sections of 0; and trivial on symbols with one element constant
sheaves W'RQ = {WnRQ)n21. The pro-object WebZglOgis defined to be (ie. in K).
the kernel of 1-frobenius on W'Rq. Multiplication by pn induces exact
sequences (op.cit.) Applied to the spectral sequence (0.3), this result gives a map

When Y is ordinary, tensoring on the left with C yields the isomorphism

The isomorphism in (1.4) (ii) above extends to (0. G)(iii).
In addition to the Galois cohomological results of Kato, the main idea
in the proof of (1.4) is a general theorem of Gabber [lo].

T
The maps in (1.4) can be explicated in terms of symbols as follows (here T h e o r e m (1.8). Let M- S be a smooth morphism of schemes,
xi E 02, y E Ox, F = ~ ( m o d x ) ) : s E S , 9 a bounded below complex of torsion e'tale sheaves on a retrocom-
pact open set V C S. Write
20 S. BLOCH p-ADIC ETALE COHOMOLOGY 21

where j: x-'(V)+ M . Let g t a r ( K ) be th,e Zariski sheaf o n M , associated is zero and then one argues using the structure theory (1.4). An analysis of
to the presheaf the structure of IIm(H, uK) in the non-ordinary case is the key to further
U H d ( u i t ,K lu). progress in this area.

For z E M,, let i, : Sp(lc(x)) 4 M , and write L; for the Zariski sheaf

$2 Ordinary Varieties

Recall the variety Y is ordinary if Hm(Y, Bn) = (0) for all m and n.
T h e n the Cousin complex gives a flasque resolution
The following proposition overlaps with results of Illusie, and we would like
(1.8.1) O+zd( K )
-zar 5 Lo-+L1-+L~-+... to acknowledge considerable inspiration from conversations with him. The
formulation here is due to Kato.
In our situation, if A is the henselization of the local ring of X a t a point
y of Y and E is the quotient field of the henselization of X a t the generic Proposition (2.1). The following conditions (1)-(5) are equivalent.
point of Y, then the injectivity of 6 in (1.8.1) implies (1) Y i s ordinary.
(2) H4(Y, Rk, BFPK HYT, 0%) is a n isomorphism for a n y q, 7 .
(3) IZq(Y,WnRjqlOg) @z/pnzWn(E) H q ( Y , WnR') is a n isomorphism
In this way, the stalks of M L a t points other than the generic point can be for any q, r , n.
controlled. (4) Hq(Y, WRb, ,og) @ z pW(K) -+ Hq(P,WR') is a n isomorphism
Concerning the proof of (0.6), we have by (1.6) for any q, r.
( 5 ) T h e frobenius f : HQ(Y,WR;)-+ Hq(Y, WR;) is a n isomorphism
for any q, r.
Moreover, Y is ordinary and H&,(Y/W) is torsion-free for all q i f and
adjoining
-
ap-" to A "kills" symbols {x, z l , . . . ,zr-1)) so writing only i f the following holds:
M: = {M:)n21 for the pro-system associated to the limiting situation
(6) For any q, the Newton polygon defined by the slopes of frobenius o n
x / S , we get an exact sequence (defining UM:) of pro-systems of ktale H&y,(Y/W) coincides with the Hodge polygon defined by the h q - ' ( ~ , R k l k ) .
shkaves on-P,
If IZq(Y, WR') is torsion-free for a n y q , r , the conditions (1)-(6) are also
o -+ uM: -+ M: -+ w.fl$,,og -+ o. equivalent t o
The key result is then ( 7 ) For any q, the slopes of frobenius o n H&ys(Y/W) are all integers.

Proposition (1.9). Assume P is ordinary. Then By way of example, it is possible to show that ordinary hypersurfaces
of degree d fill out a non-empty open set in the moduli space. This result
( U ) =( 0 all m, r. is due to Deligne (unpublished). As a second example, a general polarized
abelian variety is ordinary. (Gabber points out that no one has published a
proof of connectedness for the moduli space of polarized abelian varieties.
One reduces to showing the map Failing this, the above assertion is to some extent conjectural.)

R m ( Y , u'M;) 4 Hm(p,
u r l )
 S. BLOCH

$3 Open Questions and Conjectures where = Image (B,',,p,c,(M - cusps) -+ H1(M)). Let 7'1 be the Hecke
operator corresponding t o 1. Let g be a modular form of weight k 2 for +
Here are some directions one might try t o explore. I' which is an eigenfunction for Hecke and suppose T1(g) = alg for 1 / N
with p / a p . Let w k ( g ) = 0, Ker(Tl - a[) C wk.
The idea is t h a t w k ( g )

-
p-adic functions on moduli: It follows from (0.6) (i) that when Y is should correspond to a submotive of
ordinary, the G-module gr;IIQ(V, Q p ) is independent of the lifting X of
Y.The extensions k times
H k + l ( E x M ...xME)= Hk+'(Ek)

do, however, depend on the lifting. Assume for simplicity k = E. We

+ + +
which has IIodge type (k 1,0) (0, k 1) and for which the characteristic
polynomial of the conjugacy class of inverse frobenius [4p1] c G a l ( Q / Q )
obtain a map is
{liftings of Y )-,~ x t ' ( ~ r ~ ,gIr >i ~+ l,l ~ q ) +
T~ - a p T pk+l.
The hypothesis p up implies that the IIodge and Newton polygons as-
sociated to this motive coincide. The challenge now is to establish some
version of (0.6) involving projections, together with an appropriate modifi-
cation of (2.1) (6) or (7) sufficient t o prove
where (K* = I@ K * /K *pn ). Note that the group FIom is independent of
lifting. choosing a basis for the dual Horn*, we get functions Conjecture. Let ~ ~ ( g ) , ,=
~ , nKer(Tl - a,) on H : & t ( E k / ~ ( E ) ) .
A Then E"EI~+'(E~,Qp)induces a filtration F'wk(g) with
{liftings of Y)4 K * @ Q.

The subgroup 1 + ?rA C K * is padically c o ~ p l e t e . I would conjecture

that these functions take values in 1 + ?rhC K * @ Q . This, in any event,
is what happens for abelian varieties and K3 surfaces [16], [6]. It would
be very interesting t o have a general theory of these "functions on p a d i c
I n particular, the p-adic galois representation Wk(g) should have a Hodge-
moduli".
Tate decomposition as in (0.1).

Galois representations associated to modular forms [7]:I am endcbted

Structure of H * ( V ,Qp) when Y is Hodge- Witt: To begin with, we con-
t o K. ltibct for suggesting the following possible application of the theory.
jecture in this case that the spectral sequence (0.3) degenerates up to torsion
Let M be the cornpactified modular curve associated to the principal con-
a t E2. If so, we obtain from (1.6) a map
gruence subgroup J' C S L 2 ( Z ) of some level N > 3. We view M as a
f
scheme over Qp for some p prime to N. Let E -+ M be the universal
curve with level structure. For some integer k > 0, consider
 S . BLOCI-I p-ADIC ETALE COHOMOLOGY
24

We obtain
References

[I] Berthelot, P., Cohomologie cristalline des schkmas dc caractkristique

p > 0. Lecture Notes in Math. 407, Springer-Verlag (1974).
Let d = dim 7,so [2] Bloch, S., Algebraic K-theory and crystalline cohomology, Pub.
Math. I.H.E.S. 47, 187-268 (1974).
[3] Bloch. S., Some formulas pertaining to the K-theory of commuta-
tive group schemes, J. Algebra, 53, 304-326 (1978).
[4] Bloch, S., Gabber, O., and Kato, K., p-adic ktale cohomology
The maps q'8above are compatible with cup product, from which one
(manuscript).
deduces that under cup product
[5] Cartier, P., Une nouvelle opkration sur les formes diffkrentielles,
C. R. Acad. Sci. Paris 244, 426-428 (1957).
[6] Deligne, P., (with Illusie, L.) Cristaux ordinaires et coordonnks
canoniques, in Surfaces algkbriques, Lecture Notes in Math. $68,
so there is a pairing
Springer-Verlag (1981).
[7] Deligne, P., Formes modulaires e t reprksentations I-adiques, Skm.
Bourbaki, 1968/69, exp. 355, Lecture Notes in Math. 179, Springer-
Verlag (1971).
which we conjecture t o be perfect. Let C be the category of pointed artinian
Dwork, B., Normalized period matrices, Ann. Math., 94, 337-388
local A-algebras, and let gndenote the Zariski sheaf associated t o the
(1971).
functor K , of Quillen. (Maybe it would be better t o consider the Milnor
~onta'ine,J. M., Formes diffbrentielles e t modules de Tate . . . , Inv.
K,.) Define for A E ob C
Math. 65, 379-409 (1982).
Gabber, O., Gersten's conjecture for some complexes of vanishing
cycles (manuscript).
Grothendieck, A., et.al., SGA4, Lecture Notes in Math. 269, 270,
Conjecture. The functor H m ( K n ) : C 4 (abelian groups) has (under 305, Springer-Verlag (1972-73).
the hypothesis Y Hodge- Witt) a pro-representable Kpiece" (for more details Grothcndieck, A., Deligne, P., Katz, N., SGA7, Lecture Notes in
on this point, see [24]) which is a p-divisible formal group with Dieudonnk Math. 288, 340, Springer-Verlag (1972-73).
module H m ( Y , WW"")/tors and Tate module g r ~ H m i n - l ( ~ Qp(n)).
, Hartshorne, R., Residues and Duality, Lecture Notes in Math. 20,
Springer-Verlag (1966).
* Added in proof: 0. Gabber has decided not t o collaborate, so these Illusie, L., Complexe de deRham-Wit t et cohomologie cristalline,
results will be given in a joint paper with K. Kato. Also, Ogus asks me Ann. Sci. Ec. Norm. Sup. 4e skr. 12, 501-661 (1979).
to stress t h a t his result mentioned below, which is given in his paper in Illusie, L., Raynaud, M., Les suites spcctrales assocides au complexe
these volumes, came in response t o a question of Shafarevich, and t h a t the de deRham-Witt (manuscript).
results on the moduli of K3 surfaces are due t o Rudakov-Shafarevich. Katz, N., Serre-Tate local moduli, in Surfaces alge'briques, Lecture
Notes Math. 868, Springer-Verlag (1981).
Kato, K., Galois cohomology of complete discrete valuation fields
(~re~rint).
 S. BLOCH

Kato, K., A generalization of local class field theory by using K - The Mordell-Weil Group
groups. I, 11, J. Fac. Sci. Univ. Tokyo, 26, 303-376 (1979), and 27,
603-683 (1980). of Curves of Genus 2
Koblitz, N., p a d i c variation of the zeta function over fanlilies of
varieties defined over finite fields, Comp. Math. 31, 119-218 (1975).
Milnor, J., Algebraic K-theory and quadratic forms, Inv. Math. 9, J. W. S. Casselv
318-344 (1970).
Oort,F., Lenstra, H., Simple abelian varieties having prescribed
To Z.R. Shafarevich
formal isogeny type, J. Pure Appl. Algebra 4, 47-53 (1974).
Serre, J-P., and Tate, J., Good reduction of abelian varieties, Ann.
Math. 88, 492-517 (1968). 0.0 Introduction.
Steenbrink, J., Limits of Hodge structures, Inv. Math. 31, 229-257
(1975). In 1922 Mordell [2] proved Poincark's conjecture that the group of ra-
Stienstra, J., The formal completion of the second chow group; a tional points 011 an abelian variety of di~nension1 (= elliptic curve with
K-theoretic approach, in Proceedings of the Rennes Conference in rational ~ o i n t )is finitely generated. His proof was somewhat indirect. In
Algebraic Geometry. 1928 Weil [5] in his thesis generalized Mordell's result t o abelian varieties
Tate, J., pdivisible groups, in Proceedings of a Conference on local of any dimension and to any algebraic number field as ground field. At
fields, Springer-Verlag (1967). the same time, Weil [6] gave a very simple and elegant proof of Mordell's
original result. I observed some time ago t h a t Wcil's simple proof admits a
further simplification. He uses the explicit form of the duplication and ad-
dition theorems on the abelian variety, but these can be avoided by, roughly
Received July 19, 1982
speaking, the observation t h a t every element of a cubic field-extension k(0)
Supported in part by N.S.F.
+ +
of a field k can be put in the shape (a0 b)/(cO d) (a, b, c, d E k). This ad-
ditional simplification has little interest in itself, but suggested that similar
Professor Spencer Bloch
ideas might be usefully exploited in investigating the Mordell-Weil group
Department of Mathematics
of (the Jacobians of) curves of genus greater than 1. This is done here for
University of Chicago
curves of genus 2.
Chicago, Illinois 60637
The literature dealing with individual curves of genus 1 is vast. Very
little, if anything, has been done with curves of higher genus. In fact the
machinery which has been developed, while admirable for deep and general
inveutigatiorls, appears (pace Mumford [3,4]) little suited to the explicit
treatment of explicitly given curves. I cannot cven find in the literature
an explicit set of equations for thc Jacobian of a curve of genus 2 together
with explicit exprcssioris for the group operation in a forrn amenable t o
calcula.tion: nor do I give one, since it is not needed for my present purposes.
Although I do not give any nu~nericalexamples, I hope and believe that the
theorems proved here will make it practicable to compute the Mordell-Weil
group of explicitly given curves of genus 2.
 THE MORDELL-WEIL GROUP 20
28 J. W. S . CASSELS

where F(X) E k[X] and k is the ground field. When P has degree 5 or
0.1 It is useful to recall the main steps of Weil's simple proof of Mordell's
6, the genus of C is 2. Lines X = const. meet C in effective divisors of
theorem in a shape suited to later generalization. The ground field is k,
degree 2. They are all linearly equivalent, and their equivalence class will
and an abelian variety of dimension 1 is given by
be denoted by 0.The Mordell-Wcil group $ is defined to be the group of
divisor classes of degree 0 defined over k . By the Riemann-Roch theorem,
every element of $ contains a divisor a - @, where a is effective, defined
where over k, of degree 2. The divisor a is uniquely determined by the element of
9 , except that the whole divisor class QD corresponds to the zero element.
We shall show that there is a map p from 9 to an abeliari group M giving
does not llwe multiple factors. There is a single point 0 a t infinity. Every a situatiori very similar to t h a t used by Weil [6].
divisor class of degree 0 defined over k contains a uniquely defined divisor In a sense, the case when F ( X ) is of degree 5 is very special. A bilinear
8 - 0, where R is a point on C defined over k. The group 9 of divisor classes transformation of X and the appropriate change of Y takes a curve (0.2.1)
of degree 0 defined over k (the Mordell-Weil group) is thus also the group with F of degree 5 into one with F of degree 6. IIowever, the case of degree
of points on C defined over k with the induced law of conlposition [C is its 5 is somewhat simpler, so we treat it separately first.
own Jacobian]. The case when F ( X ) has degree 4 also fits into the picture, a t least when
Weil's proof that 9 is finitely generated when k is an algebraic number the coefficient of X4 is 1. Then C is of genus 1 and is birationally equivalent
field falls into three parts: to a C with F of degree 3. We shall, however, give this case separately, as
it motivates some features of the degree 6 case and because of the unified
(I) IIe constructs an abelian group M (depending on C and k) and picture presented by degrees 3, 4, 5, 6 together.
a honlo~riorphism
0.3 It is convenient to introduce here notation which will be common to
all four cases. There follows a synopsis which will bring out the common
pattern, although detailed enunciations will have to await the relevant later
whose kernel is precisely 2 9 . This step is valid for any field k of charac-
sections.
teristic 0.
The ground field is any field k of characteristic 0. Its algebraic closure
(It) If k is an algebraic number field, the image CL(G) is a finite group. is % We are concerned with curves
Hence 5 / 2 5 is finite. ("Weak Mordell- Weil Theorem.")
(111) Finally, the proof that 9 is finitely generated uses an "infinite
descent."
where

In this paper we are concerned with the generalization of (I), and so our
ground field is any field k of characteristic 0. Given the definition of p,
the generalization of (11) is immediate: we do not prove it because it can has no multiple factors. We write
be confidently left to the reader. Finally, the generalization of (111) in a
shape amenable to practical calculation might well be difficult: I have not
yet attempted it. where O is the image of the indeterminate T in the quotient ring, so

0.2 We shall be concerned with curves

30 J. W. S. CASSELS
THE MORDELL-WEIL GROUP

is the direct sum of fields K j corresponding to the irreducible factors

1.0 Genus 1, Degree = 3.
of k'(X).
Let k[O]* be the group of invertible elements of k[O]. Our hornomor-
After multiplying X, Y by appropriate elements of k*, we may suppose
phisms p will be into the group
that

= (k[Ol*/k*{k101*f2 ( N even). This is (effectively) the case treated by Weil [6]. For a point a on
C : Y2 = F(X) defined over k we define
More precisely, the image will be in the kernel M of the map

(0.3.6) Norm: L -4k*/(k*)2

(in the notation of $0.3) as follows:
induced in the obvious way by the Norm map of k[O] into k.
When N = 3 or 4, so C is of genus 1, the general element of 9 is given (i) p ( 0 ) = 1, where o is the point a t infinity.
by 8 = (a, d) on C , where a, d E k. If d # 0, then a - O is in k[O]* and we (ii) a = (a, d), where a, d E Ic and d # 0. Then a - 0 E k[O]* and p(a)
define p(fi) to be its image in L. For elements of $ not given by such 8, the is its image in f . We note that
definition of p is given later.
When N = 5 or 6, so C is of genus 2, the general element of 9 is given
by an effective divisor a x)
= {(al, dl), (az, d2)) (al, az, d l , d2 E defined and so p(a) E M, where M is (as always) the kernel of the norm map (0.3.6).
over k. If dld2 # 0, then (al - @)(az- O) E k[O]* and p ( 9 ) is defined to
be its image in L. For elements of 9 not of this type, the definition of p is (iii) = (a,O), where a E k. Then k[O] is a sum (0.3.4) of fields. One of
them, K1 say, is isomorphic to k, the isomorphism being given by O --+a.
given later.
The image of a-O is non-zero in K j ( j # 1). We replace the K1-component
Then in all the four cases listed below, p is a homomorphism of 9 into M:
0 of a - O by an element of k*, leaving the other components unchanged,
in sucli a way that the norm for k[O]/k is in (k*)2. The image in L is then
(A) N = 3, a3 = 1. The zero element of 5is the point a t infinity. unique and is necessarily in M: it is defined to be p(8). We can make this
The kernel is 29.' construction explicit. Let
(R) N = 4, a4 = 1. The double point at infinity has two branches O+,
O w . Either can be taken as zero element of G. The kernel is generated by
2$ and 0+, 0-.
(C) N = 5, a5 = 1. Lines X = const. meet C in linearly equivalent
Then the norm of a - O from ejZI K j is Fo(a) (= F1(a)). Hence the
K1-component of ~ ( a , 0is) defined to be
divisors. Their class 8 is taken as the zero of 9. The kernel is 2.5.
(D) N = 6, a6 E k*. The sero of 9 is as in (C). The kernel is generated
by 2.5 and by divisors linearly equivalent to 2&- 2@, where (a is an effective
divisor of degrce 3 defined over k. For an alternative discussion see $1.3.

1.1 T h e o r e m 1.1. The map

 J. W. S. CASSELS TIIE MORDELL-WEIL GROUP 33

given by the components except, possibly, the K1-components. Since the values of
p are in M,they are determined by the Ki components ( j # I), and so
(1.1.2) (1.1.5) holds also for the K1 components. There remains the case when,
say, d = e = 0 and a # b. Then a, b, t are the intersection of C with Y = 0,
is a homomorphism. and (1.1.5) follows by direct calculation. [Alternative arguments for these
special cases are in 51.3.1
Proof. If a = (a, d), then -EL = (a, -d), so clearly
1.2 T h e o r e m 1.2. The kernel of p: 5 -+ M is precisely 2 5 .
(1.1.3) P(-a) = P(a).
Proof. Since M C L is of exponent 2, the kernel certainly contains 25.
Hence it is enough to show t h a t We have therefore to show that p(a) = 1 implies EL = 26 for some b E 5.
Since 0 = 20, we need consider only

(1.2.1) EL = (a, d), p(8) = 1.

implies

(1.1.5)

If any of a, b, t is 0, the implication is trivial, so without loss of generality

(1.1.6) a = (a, d); b = (b, e); t = (c, f ) ,

where a , b, c, d, el f E k .
(using F(0) = 0), where rz, r l , ro are linear forms in 31, so. We may
The condition (1.1.4) is equivalent to the statement t h a t a,b, t are the
therefore find s l , so not both 0 such that
intersections of C with a line:

Suppose, first, that s l = 0, so so # 0. Then pa = 0 and

This is impossible, since the defining relation for 0 is of degree 3.

and Hence, by homogeneity, we may suppose t h a t 31 = -1, and so by (1.2.2),
(1.2.3) we have
(1.1.9) (a- 0)(b - Q)(c - 0) = (110 +
If def # 0, this gives (1.1.5) a t once by the definition of p. The proof
remains valid when two or all three of a, b, t coincide. On replacing 0 by an indeterminate T we have
If the images of a - 0, b - 0, c - 0 vanish in only one of the fields Kj
of (0.3.4), say they do not vanish in I G ( j # I), then (1.1.5) holds for all
 TIIE MORDE1,L-WEIL GROUP

for w E k. In fact IIence

on comparing coefficients of T 3 . IIence, on changing the signs of T I , TO if Comparison with (1.0.5) shows that the new definition of p(a) coincides with
necessary, the line the old.
We now show how to prove Thcorcm 1.1 using the new defiriition of p.
This time we show directly that the K1-components satisfy (1.1.5). Let
passes through = (a, cl). The residual intersection is 2b, where b = (so, t ) + +
a = (a, 0) be as above, let b # and let t satisfy n b t = 0. Let X be
and t is given by (1.2.9). Hence E 2 5 , as required. a generic point, and work with the ground field k ( ~ ) .As in the proof of
Theorem 1.1, there are 1 1 ( ~ )lo(x)
, E k ( ~ such
) that the line
1.3 The previous treatment of the case when a - O is not invertible was
somewhat ad hoe. We now sketch a rather more highbrow treatment which
has the advantage that it carries over automatically to the cases we shall meets C in x arid b. Let the third intersection be y. Then, as before,
consider later.
Let n = (a,0) and choose the notation in (0.3.4) so that the projection
of k[O] onto the field K 1 is given by O + a. Let ;Y be a generic point of C
over k and choose t to be a local uniforrnizer for X at 8. Then X(X)- a has Let x specialize t o a, so E, specializes to t. The left-hand side has a double zero
a double zero a t x -+ a; and so and so, if we divide by t 2 before specializing, where t is a local uniformizer,
we have
M(a)(b - a)(c - a ) E ( I C * ) ~ .
where M(x) specializes to a finite non-zero M(n) E k*. Replacement of t This is just the assertion of the K1-component of (1.1.5)) as required.
by another local unifor~nizer~nultipliesM(R) by an elcmcnt of We
show that p(n) can be defined by replacing the K1- component 0 of a - O
by M(n) before taking the image in L.
In fact, Y(x) has a simple zero a t 8, and so 2.0 Genus 1, Degree = 4.

Here we consider
where N(B) E k*. We have

(1.3.3) F ( X ) = (X - a)Fo(X),
There is a double point a t infinity. We denote the two branches by O+, 0-,
where where

(1.3.4) Fo(X)Ek[X], fi(a)#O.

On taking the equation Y 2 = F(X) of C, dividing by t2 and specializing X

to R, we get We take O+ as the zero of the Mordell-Weil group 5. Let & be the subgroup
generated by 0-. We shall generally be concerned with GI&,
and so the
dislinction between O+ and 0- disappears.
 TIIE MOILDELL-WEIL GROUP 37

Here the definition of L is different from that in $1. For a E 9 we define Here one of the four intersections is always a t o+. We suppose, firsl, lhat

so O+ is a double intersection. Let n = (a, d) be arbitrary. We can choose

as follows:
bo so that (2.1.4) passes through a. Let b = (b, e) be the fourth intersection.
Then
(i) /do+)= p ( ~ = ) 1.
(ii) If n = (a, d) with d # 0, so a - O E k[O]*, then p(a) is the image of
a - O in L. As before,
for some 1 E k t . On putting X -+ O and recalling the definition (2.0.3)
of L , we obtain

where M C L is the kernel of the norm map (0.3.6).

(iii) If n = (a,O), we define p(a) either as in (iii) of $1.0 or, equivalently, a t least when de # 0, so a - O, b - O are invertible. The truth of (2.1.8)
as in $1.3. when de = 0 can be verified either by an a d hoe argument using (2.0.4)
as in $1.1 or by the technique of $1.3. We have ignored the possibility
2.1. Theorem 2.1. The m a p t h a t b = o+: but then the right-hand side of (2.1.7) becomes 1(X - a), so
p(a) = 1 = p(o+). The divisor of poles of

given b y
is 20-, so 0 + b = 20- in 9, and hence
(2.1.2)

is a group homomorphism. The kernel contains the subgroup & generated

by 0-.
From (2.1.3) and (2.1.10) we have

Proof. Let A = (a, d). The function X - a has zeros a and a' = (a, 4)
and poles O+, ti-. Hence a' + a = 0- in 9. Clearly p ( d ) = p(a), so and so p(a) depends only on n niodulo &.
We now abandon the assumption (2.1.6). Let a=(a, d) and
b = (b, e) # (a, -d). Then we can choose b l J bo so that (2.1.4) meets C in
fi and b [a double intersection a t a if b = a]. Let t = (c, f ) be the residual
Now consider the intersections of intersection. Then

(x2 + blX + b ~ -)F (~X ) = l(X - a ) ( X - b)(X - c).

with C. They are given by On putting X -+ O we have

38 J. W.S. CASSELS THE MORDELL-WEIL GROUP

a t least if def
$1.3. Since
# 0.
A

we have completed the proof t h a t /L

-
When def = 0, one can either argue ad hoc or as in

+ b + t = 20- 0 (mod E),

is a homomorphism.
where t3, . . . , t o are linear in

(2.2.7)

Then we get s l
-?Q, 8 1 . We may choose

t3 = 0.

# 0 by (2.2.5) and, without

31, 30 so that

loss of generality,

2.2. T h e o r e m 2.2. The kernel of p : 9 3 M is the subgroup generated (2.2.8) 31 = -1.

by 2 9 ur,d E.
Substituting (2.2.6) into (2.2.2) we have
Proof. The kernel contains & by Theorem 2.1. It contains 2 9 since M
is of exponent 2. We have therefore only to prove that the kernel is not
larger than stated.
Here tz # 0 because the defining relation for O is of degree 4, and so,
) 1. If n = 0, or 0-, the conclusion holds trivially, so we may
Let p ( ~ =
without loss of generality, t 2 = 1.
suppose t h a t
Then

(2.2.10) (T2 + t l T + + 1(T - s o ) 2 (-~ a) = F(T).

By hypothesis,
Hence A = (a, d ) is a zero of

(2.2.11) Y * ( X 2 + t l x + to)
for some I E k* and some c3, c2, C I , CQ E k. ,We cannot have c j = c2 = 0,
since the defining relation F ( O ) = 0 of Q is of degree 4. for one choice of sign. There is a double zero a t (a0, r ) for appropriate
Suppose, first, that c3 = 0. Then without loss of generality c2 = 1 and r E k. The remaining zeros and poles are at O+, O-. Hence the co~~clusion
of the theorem holds.

Hence a = (a, d) lies on 3.0 Genus 2. Generalities.

A curve
for orle choice of sign. The only points other than occurring as zeros or
poles of the left-hand side of (2.2.4) arc 0+ and 0-. Hence the conclusion
of the theorem holds.
where F has degree 5 or 6, is of genus 2. Any line X = constant cuts C
We are left with
in a divisor which is linearly equivalent to the divisor cut out by the line
a t infinity. This class of divisors of degree 2 defined over k will be denoted
by @.
For 31,30 E k we have The differentials of the first kind are
40 J. W.S. CASSELS THIS MOILDELL- WEIL GROUP

Hence the divisors in @ are the only divisors which can be the zeros of Then
such a differential. It follows from the Riemann-Roch theorem that every Y ; = ( 7 2 1 x 1 + 722)6J'((711X1 + r12)/(721X1 + 7 2 2 ) )
divisor class of degree 2 defined over k, other than 8, contains precisely one = Fl(X1) (say)
effective divisor, and it is defined over k. An effective divisor a of degree E k[Xl].
2 defined over k is an unordered pair of points {a1, a2) such that either (i)
81 and R2 are each defined over k or (ii) al, a2 are defined over a quadratic
If F is of degree 5, then in general Fl is of degree 6. Coriverscly, If F is
extension of k and conjugate over k. Either a1 or 82 may be a t infinity, and I of degree 6 and has a zero in k, then we can use such a transformation to
a2 = R1 is permitted. obtain r of degree 5.
The Mordell-Wcil group 5 is, by definition, the group of divisor classes
3.2 If a, b, E k with a # b are zeros of F ( X ) ,then the divisor {(a, 0), (b, 0))
of degree 0 defined over k. If a E 5, a # 0, we represent it by the unique
effective divisor g in a + 8. The set of such together with 8 thus give is of order 2 in the Mordell-Weil group since
an isomorphic image of 9. Here

and both the divisors on the right-hand side are in 8. If E' is of degree 5,
where a,
@, (a are effective of degree 2, means that
necessarily effective) divisor in the class 8. If
a+@- -- is a (not
we may take the point 0 at infinity instead of one of (a, 0), (b, 0).
If F splits completely over k (or if we work over the algebraic closure q,
we thus get

then, clearly,
divisors of order 2. We have therefore obtained them all.
(3.0.5) -g = {(a, -4, (b, -f 1). I

We shall often by abuse of language write 8 for any effective divisor in 8.

They all correspond to the zero clement of 9.
4.0 Genus 2. Quintic Case.
We note that the divisor class @ is canonically distinguished. This is
contrary to the genus 1 case. Here we consider

3.1 The curve (3.0.1) may be transformed birationally into other curves
of the same shape but with in general different E'. Let
where, after multiplying X, Y by elements of k* if necessary, we have
I

Put
There is a single point a t infinity. The poles of X, Y are 20 and 50,
x = (711x1 + ~12)/(721X1+ 722) respectively. We retain the notation of 50.3. In particular,
and
YI = ( +
T ~ ~ xT~ ~ ~ ) ~ Y .
42 J. W. S. CASSELS THE MORDELL-WEII, GROUP

For an effective divisor a of degree 2 dcfined over k, we define p ( 8 ) f where

as follows:
(4.0.8) &(X) k[X].
(i) p{0,0)=1.
(ii) a = {o, a), where a = (a, d) and d # 0. Then p ( a ) is the image of We may take Y(xI) = t t , when ( X ( X ~-) a l ) / t f takes the value
& - - @ i n f.
(iii) a = {al,a2), where aj = (aj, di) and dld2 f 0. Then {(a1 - a2)&&)]-l

a t XI = R1. Hence (4.0.6) may be replaced by

and p(%) is its image in L. (4.0.9) -X~J'o(al) = -X2l;0(a2) = -Fo(0)

There remain the cases like (ii),(iii) but where d = 0 or dldz = 0,
as the missing component.
respectively. Here one must use the expression (0.3.4)of k[O]as a direct
sum of fields K j . When K j = k: we can proceed as in 51.3, arid the details
are left to the reader. When the degree of KJ over k is 2 3, the projection 4.1 Theorem 4.1. The map
of a - Q or (al - @)(a2- 0) into K j is non-zero, and so no difficulty arises.
The only case left is when K I (say) is of degree 2 and ( a l - @)(a2- 0) (4.1.1) p:g+ M
projects into 0 in K1. Then k(al) = Ic(an) = kl (say), dl = d2 = 0, and
there are two isomorphisms Xi of kl onto K1 given by defined by

(4.1.2) a -+ ~ ( a )

Adjoin two independent generic points X i , X2 over k1. We extend the non- iu a group homomorphism.
trivial automorphism a (say) of k l / k to k l ( ~ 1 , ~ 2
by) u X ~ = X2. Let t l
be a local uniforrnizer for xl a t = (al,O), so a t l = t 2 (say) is a local
Proof. We note, first, that the definition of p gives p ( a ) = 1 for all
uniforrnizer for X2 a t all1 = 8 2 .
Now
a @.
The automomorphism Y -+ -Y of k(X, Y) over k(X) clearly maps a
into - a . Hence

where Nj(at, a2) E k;. Clearly uNl(al, 82) = N,(n,, a,), and so
(4.1.3)

Now let
This provides us with the missing K1-component. Clearly choice of another
local uniformizer instead of t l multiplies (4.0.6) by a11 element of (K;)2,
(4.1.4) a={(a1,d1)1(a21dz))1 @ = {(blle~),(b2,e2))
and so CL(%)is well defined.
We can give an explicit form of (4.0.6). Let
be defined over k. If, say,

(4.1.5) al=bl, dl=-el,

 THE MORDELL-WEIL GROUP
44 J. W. S. CASSELS

Now suppose that

then

Then
(h2x2 + h l X + h0)2- F ( X )
= -(X - a l ) ( X - a 2 ) ( X - bl)(X - bz)(X - c),
directly from the definition of p. Hence we may suppose t h a t the situation where c E k. Then 6 = (0, (c, f)), where f is given by (4.1.7), and (4.1.13)
exemplified by (4.1.5) does not occur. We may then find h3, h2, h l , ho E k follows agan in the obvious way.
such that There remains the case when a or @ contains O. If o occurs twice, we
argue as for (4.1.5). If 0 occurs only once, we make the obvious modifications
(4.1.7) Y = h 3 x 3 + h 2 x 2+ h l X + hO to the above argument with hg = 0. This concludes the proof.
meets C in a
and @. The total intersection is of degree 6, and so the
4.2 T h e o r e m 4.2. The kernel of
residual intersection (n; (say) is of degree 2. Clearly (a is defined over k, and

(4.1.8) Q[+@+(a=O.

Suppose, first , that

Proof. Since M is of exponent 2, the kernel contains 25. We have t o
show that it is not any bigger.
Then a
We shall say that the divisor is in "general position" if

and
for some cl,cz E F, and

(4.1.11) a = ( ( ~ 1f,l ) , ( ~ 2 f12 ) h We have t o show that

where f l , f 2 are determined'by (4.1.7). If

(4.1.12) d l d z e ~ e z f l f z# 0, implies a E 25, and will do this first for IS[ in general position.
Put

where T is an indeterminate. then (4.2.3) implies that

directly by substituting 0 for X in (4.1.10). If (4.1.12) fails, we still get
(4.1.13) by arguing in the separaLe fields K j as in 51.3.
 TIIE MORDELL-WEIL GItOUP
J. W. S. CASSELS

Then (4.2.13) implies

where

since both sides have degree 5 4. Hence G(T)is a perfect square, a case
Suppose, first, t h a t we have already eliminated.
There is, however, a pair (I,, M ) = (L1, M I ) (say) for which 12 = 0, since
this is a single condition. Then m3 # 0 and, by homogeneity, we may take
m3 = 1. Similarly, there is a pair (L, M) = (L2, M 2 ) for which m3 = 0
Then and l2 = 1. IIence in general

(4.2.16) L = uL1 + vL2, M = uM1 + vM2, (U,VE k),

since the defining r.elation for O is of degree 5. Hence r2 = 0 and a1 = a2 E k. where

+
There are now two cases: (i) dl d2 = 0. Then 9 E @ E 2 g . (ii) dl = d2.
Then a1 = an = a (say), and L1 = degree 1
M1 = T 3 + degree 2
(4.2.17)
(4.2.9) El = {a, a) = 2{0,a) E 25. L2 = T2 + degree 1
M2 = degree 2.
We shall therefore suppose that (4.2.7) does not hold.
Now (4.2.13) gives
We now consider polynomials

for some w = w(u,v) E k.

(4.2.11) +
M ( T ) = m 3 ~ 3 m2T2 + m l T + mo E k[T] Put
such t h a t

where we consider both choices of sign together for the moment. Then
The condition that the coefficient of CI4 in (4.2.12) vanishes is a single
homogeneous linear condition on 12,I l l lo, which is rlon-vacuous since (4.2.7)
does not hold. Hence the set of pairs {I,, M ) is a 2-dimensional k-linear
space. On eliminating p between (4.2.5) and (4.2.12), they all satisfy
1 If w = 0, then G(T) is a square, which we have already excluded. Hence

(4.2.21) w E k*.

Now specialize T -+ a l . Then G(al) = 0, J'(al) = d:; and so

Suppose, first, that, we can find L # 0, M such that
48 J. W. S. CASSELS TIIE MORDELL-WEIL GROUP 49

Here dl # 0 by general position, and so IIitherto we have allowed both choices of sign in (4.2.19). We should have
reached the comclusion of the theorem if (4.2.24) holds for either choice of
(4.2.23) 0 # Jw E k(al, d l ) = k(a2, d2). sign. Hence we are left with the situation when(4.2.30) holds for u = +1
and both v = +I and v = -1. It follows from (4.2.16) and (4.2.30) t h a t
Suppose, first, that

(4.2.24) w E (k*)2.
On substituting u = 1, v = 0, T = a1 in (4.2.18), we get
Then we can choose s E k with 32 = w so that

passes through 81. We subdivide cases.

where w = w(1,O) E k. Here dl # O by general position, so

(i) If (4.2.25) also passes through R2, the intersection of (4.2.24) with C is
9 + 2@, where @ is given by L ( X ) = 0 and (4.2.25). IIence the conclusion a contradiction. This concludes the proof when 9 is in general position.
of the theorem holds. [We note in passing that we should also have reached a contradiction from
(ii) If (4.2.25) does not pass through 82, then it passes through supposing that (4.2.24) holds for both choices of sign in (4.2.19).]
4 = (a2, -d2). This can happen only if 81,82 are not conjugate over k, Now suppose that p ( 9 ) = 1 b u t is not in general position. Suppose
and so are each defined over k. Arguing as in (i), we have t h a t there is a @ such that

However, is in general position. Then p ( a ) = 1, and the general position argument

a
shows that E 2 5 . Hence 9 E 2 5 , and we are done. Unfortunately there
may not exist such a @, and we must extend the ground field.
Let X1,X2 be independent generic points of C over k and let & be the
Here {o,o) and {a2, a;) are both in 8, and again the conclusion of the
subfield of k(xl, x?) fixed under the interchange of X 1 and X2. Then
theorem holds.

If (4.2.24) does not hold, then by (4.2.23)

(4.2.28) k(a1, d l ) = k(a2, d2) - k ( d 9 ) say1 9 55 (k*I2

Put

and
There is a natural injection of L into L, and we regard L as a subset of
(4.2.29) w = g s ,2 sEk*.
-
L. We denote by p the analogue of p, but for the ground field k-. If 9 is
defined over k, then clearly
Then by (4.2.20) and (4.2.23) we have

(4.2.30) M(al) = qdldg, 9E k.

50 J. W. S. CASSELS TIIE MORDELL-WEIL GROUP

Let 9 be the Mordell-Weil group for L.

On arguing as for (4.2.34) but given by
- { x ~ , x ~we
with @= a -
) , obtain E 2 9 . Hut 2 9 n 9 = 29, and we arc done.
- (5.1.2)

is a group homomorphism.
5.0 Genus 2. The Sextic Case.
Proof. The proof is very similar to that of Theorem 4.1 for the quintic
case, so we only briefly indicate how the different definition of L comes in.
Let

There is a double point a t infinity, with two branches O+, 0- (say) defined
ovcr z.
If a6 E (k*)2,then O+, 0- are ixidividually dcfined over k, but in where, for simplicity, a l l a2, bl, b2 are distinct. There are h3, hp, hl, ho E k
any case {o+, 0-1 is defined over k. such that
VJe retain the notation of $8.3. In particular, now

passes through a and @. Let 6 be the residual intersection, so

For any effective divisor a of degree 2 defined ovcr k, we define
If hz # a6, we have

(i) p(0+,0-) = 1. Further, p(0+,0+) = ~ ( o - , o - ) = 1 if O+, 0- are

deiined over k.
(ii) If 0 = o+ or 0- and a = (a, d) with d # 0 are defined over k, then for some n E k* and cl,c2 E %. Here
p(0, 8) is the image of u - O in 1.
a
(iii) If = {(al,dl), (a2,d2)) with dld2 # 0, then p ( a ) is the image of
for appropriate fl, fa. On putting .X = O in (5.1.6), we get

as required.
When d = 0 in (ii) or dld2 = 0 in (iii), the definition is extended precisely The remaining cases are left to the reader.
as in the quintic case [$4.0]. As always, p ( g ) E M.
5.2 The kernel of p contains 2 9 as usual. In general it contains other
5.1 Theorem 5.1. The map elements. Let

(5.2.1) 8 = {a, a),

52 J. W. S. CASSELS THE MORDELL-WEIL GROUP 53

where the point is defined over k. Then p ( 8 ) = 1. Two divisors g and k. In particular, none of the t j are at infinity, and the t j are distinct, where
= {b, b) of this type are in the same class of 5 modulo 2 5 since t j = (cj1 fj).
We may choose i l l i o , h 4 , . . . ,ho E k, not all 0, such that
(5.2.2) {a, a) + {It, b) = 2{a, It).
One has curves C for which (5.2.1) is in 2 g and others for which it is not.
Indeed, let as = 1 and let 0 be one of the two points a t infinity defined over meets C in 26; where
k. Then it is easy t o see that {o, O) E 2 5 precisely when

for some Ciearly not both i l l io vanish. The intersection of (5.2.6) with C has degrce
8, and so the residual intersection is B.
Suppose, first, that both points of B = {(dl,gl),(d2,62)) are finite.
Then
and H ( X ) ,G(X) are of degree 3 and 5 2, respectively.
It is possible for there to be no points on C defined over k, and so no
divisors of type (5.2.1). We now describe further elements of the kernel. Let
a be an effective divisor of degrce 3 defined over k. Then by Riemann-Roch for some I E k*. Hence p(B) = 1. If a point of B is a t infinity, the
there is an effective divisor B of degree 2 defined over k such that modification to the proof is obvious.

5.3 T h e o r e m 5.3. The kernel of p : 9 -+ M is generated b y 2 5 and the

divisors described in Lemma 5.2.
We show below that p(B) = 1. An argument similar to (5.2.2) shows that
different B lie in the same class of 5 modulo 2 5 and also in the same class Proof. Let a be a divisor with
a
as the of type (5.2.1), if they exist.
It follows that the ?B of type (5.2.5) are in the kernel whenever there are
points a on C defined over k. We intend to show that the B are in the kernel
even if no such a exist. One can construct a proof by adjoining a generic We suppose, first, that a is in "general position," by which we mean
point much as a t the end of $4, but we give a different argument, since we
need the details later.

L e m m a 6.2. The of type (5.2.1) and the B given by (5.2.5) are in with
the kernel. They all lie in the same class of 5 modulo 2.9, which may or
may not be 2 5 itself.

Proof. After what has been said above, all that is needed is to show Put
that B is in the kernel. We may suppose that there are no points on C
defined over k, and so bhat 6 = {tl, f 2 , f a ) where the t i are conjugate over
 THE MORDELL- WEIL GROUP
J. W. S. CASSELS

Then (5.3.1) implies that

(5.3.5) G(O) = np2

We now distinguish two cases:
for some n E kt and some
(I) Suppose that

There are
Then for one choice of u with w = nu2 the curve

(5.3.8) +
M ( T ) = m 3 ~ 3m 2 T 2 + m l T + mo E k[T]
passes through R1. Suppose, first, that it passes through 82. Then, by a
such that +
now familiar argument, the total intersection of (5.3.16) is a 2@, where
@ is given by L ( X ) = 0 and (5.3.16). Hence a E 29.
If (5.3.16) does not pass through 82, then it passes through & = (a2, -d2).
This can happen only if k ( a l , d l ) = k(a2, d2) = k, since otherwise, al,a2
since the absence of terms in 05, O4 imposes two linear conditions on arc conjugate over k. The preceding argument shows that
12,11,10. Hence

Now
and so

(5.3.11) ~ M ( T ) '- L(T)~G(T)= wF(T)

for some w E k. If w = 0, then G(T) is a square contrary to the general since {a2,8:) E 8 : and again is in the specified kernel.
position hypothesis, so (11) If (5.3.15) does not hold, we must have

On comparing terms in T 6 in (5.3.11), this gives (using a1 f a2 by "general position"), and

(5.3.20) w/n = p2, u E k*.

We note for later use that the substitution of a1 for T in (5.3.11) gives
On substituting a1 for T , wc have
56 J. W. S. CASSELS THE MORDELLWEIL GROUP 57

for appropriate choice of u. We now go over to the (XI,Yl) coordinates and, for simplicity of nota-
We shall show that we now have the situation described in the proof of tion, drop the suffix "I ." It is clear that the change of coordinates does not
Lemma 5.2. Our first objective is to define a bilinear transformation of alter the definition of the p function. Further, by (5.3.5) and (5.3.25) we
X which, as it will turn out, will ensure that I ( X ) (in the notation of the may now suppose that
proof of the lemma) can be taken to be a constant.
On comparing the coeficients of T6 in (5.3.11) and using (5.3.20)) we
have
After a substitution X -+X + constant, we may eliminate the linear term
(5.3.22) nm32 - I: = w = gnu 2 . in G(X), and so

Hence n is a norm for k(dg)/k and so by (5.3.18) is representable by the

homogenized version of G(T), say
We now consider polynomials
(5.3.23) n = ( a - yal)(a - r a z ) = a
2
+ g l a r + g27 2
( a , 7 E k).
(5.3.32) L'(T) = 1:T3 + 1',T2 + 1:T + 1; # 0, E k[T],
Choose P , 6 E k so that a b # p7, and put (5.3.33) M1(T) = mkT4 + . - + m; E k[T]
(5.3.24) x = ( a x 1 + P)/(rXi + 6). such that

Then (5.3.34) L1(8)p = Mt(Q)

and, in addition,

for some

There are three linear conditions on the four 1:) so an L'(T) f 0 certainly
exists.
On putting If m i = 1; = 0, we could take L = L', M = MI in (5.3.11) and then
- 1,I contradicts (5.3.12 bis), (5.3.30). Hence, by homogeneity, we may
(5.3.27) y1 = Y / ( ~ X I+ q3, suppose that m&= 1; = 1, and so
the equation of C becomes
(5.3.36) Lt(T) = T 3+ u T 2 + linear E k[T],
(5.3.28) Y: = Fi(X1) E XI], (5.3.37) M1(T) = T 4 + u T 3 + quadratic E k[T],
for some a E k. On squaring (5.3.34) and using (5.3.30) we have

(5.3.38) L'(~)~G(Q
=) M'(Q)~,

Here Fl is of degree 6 in general, but is of degree 5 ' 3 F ( a l 7 ) = 0. We and so

can avoid this exceptional case, since (5.3.23) has a 1-parameter family of
solutions (a,7).
58 J. W. S. CASSELS THE MORDELL-WEIL CIlOUP

for some w' E k, since terms in T8, T7 cancel by (5.3.33)) (5.3.36),(5.3.37). IIere dl # 0 by "general position," so
IIere wt = 0 is eliminated by the usual argument, and so

(5.3.40) w' E k*.

contrary to (5.3.19). Hence (5.3.43) cannot occur. This concludes the proof
Following the by now familiar routine, we substitute a1 for T and obtain of the theorem for lir in general position.
The completion of the proof when g is not in general position is similar
t o that for the quintic case a t the end of $4.2 but requires some extra twists.
Let Xl1X2 be independent generic points of C and let 3 as before, be the
If w' E (k*)2,we may choose v with w' = v2 SO that subfield of k(xl, x2) I i e d under the interchange of XI and x 2 . We denote
by ,L. 8,- -p, 5 the analogues over & of L , M, p, 5. P u t 3 = ( x ~~, 2 ) .Let
/A(d) = 1, so
passes through al. = (al,dl). Since a2 is conjugate to a1 over k, (5.3.42) -p ( a + 2X) = 1

also passes through an. IIence the total intersection of (5.3.42) with C is and + % is in general position. Hence, by what has already been proved,
d + 2a1 where (a: is the divisor of degree 3 given by L 1 ( X )= 0 and (5.3.42). +
either d + 2x E 2 5 or d 2x is of the type described in Lemma 5.2
Hence lir is in the specified kernel. (but with & as grouKd field).
There remains the possibility +
Arguing as at the end of $4.2, if a 2x E 2 9 then d E 29.
e
Suppose now that d + 2X 2 5 but that theye is a B of the type (5.2.5),
+
where 6 of degree 3 is defined ov& k. Then B g+2x E 2 5 , so B +a E 2 9
-
and the conclusion of the theorem holds.
We shall show that this leads to a contradiction. By (5.3.39) we may suppose
that
In fact if e
+ 2X 25, then such a B always exists. It is enough to
5
show that if there is a dzisor of degree 3 defined over then there is
(5.3.44) M1(al) = vdlJg. a divisor 6 of degree 3 defined over k, and this follows by a specialization
argument. Let b be any element of k and let e E satisfy e2 = F(b). Then
With L, M given by (5.3.7), (5.3.8), we have, by (5.3.9), (5.3.34), that -
k specializes to k under XI + (6, e); X2 + (b, -e). We can take for any
5
specialization of over this.
(5.3.45) {L1(0) + L(O))p = M t ( 0 ) + M(O)

and so, by (5.3.30)) as usual,

References
(5.3.46) {M1(T) + M(T))~ - {L1(T) + L(T))~G(T)= (PT + q)F(T)
[I] J. W. S. Cassels. Diophantine equations, with special reference
for some p, q, E k. Here to elliptic curves. J. London Math. Soc. 41 (1966), 193-291.
Corrigenda 42 (1967)) 183.
[2] L. J. Mordell. On the rational solutions of the indeterminate equa-
by (5.3.12 bis). On substituting a1 for T in (5.3.46) and using (5.3.21), tions of the third and fourth degree. Proc. Camb. Phil. Soc. 21
(1922), 179-102.
(5.3.44), we get
[3] D. Mumford. On the equations defining abelian varieties I. Invent.
(5.3.48) (u +~ ) ~ g=
d ; (pal + q)d?. Math. 1 (1966), 287-354.
 J. W. S. CASSELS

[4] D. Mumford. On the equations defining abelian varieties II. Invent. Number Theoretic Applications
Math. 3 (1967), 75-135.
[5] A. Weil. L'arithmktique sur les eourbes algklriques. Acta Math. of Polynomials with Rational Coefficients
5 2 (1928), 281-315.(= CEuvres Scicntifiques I, 11--45).
[GI A. Weil. Sur un thkort?me de Mordell. Bull. Sci. Math. (11) 54 Defined by Extremality Conditions
(1929), 182- 191. (=CEeuvres Scientifiques I, 47-56).
G. V. Chudnovsky
Received April 23, 1982
To Z.R. Shafarevich
Professor J. W. S. Cassels
Department of Pure Mathematics and Mathematical Statistics Introduction
16 Mill Lane, Cambridge CB2 lSB, England
It is well known that classes of polynomials in one variable defined by
various extremality conditions play an extremely important role in complex
analysis. Among these classes we find orthogonal polynomials (especially
classical orthogonal polynomials expressed as hypergeometric
and polynomials least deviating from zero on a given continuum (Chebicheff
Orthogonal polynomials of the first and second kind appear
as denominators and numerators of the Pad6 approximations to functions of
classical analysis and satisfy familiar three-term linear recurrences. These
poly~iomialswere used repeatedly to study diophantine approximations of
values of functions of classical analysis, especially exponential and logarith-
q [2], [3], [4], [5]. The methods of
mic functions a t rational points x = ~ / [I],
Pad6 approximation in diophantine approximations are quite powerful and
convenient t o use, since they rcplace the problem of rational approxima-
tions to numbers with the approximations of functions. There arc, however,
arithmetic restrictions on rational approximations to functions if they are
to be used for diophantine approximatioris. The main restriction on poly-
nomials here is to have rational integer coeflicients or rational coeficients
with a controllable denominator. Such arithmetic restrictions transform a
typical problem of classical analysis into an unusual mixture of arithmetic
and analytic difficulties. For example, recurrences defining orthogonal poly-
nomials must be of a special typc to guarantee that their solutiorls will
have bounded denominators. In this paper wc consider various classes of
polynomials generat,ed by imposing arithmetic restrictions on classical ap-
proximation theory problems (orthogonal or Chebicheff
62 G . V. CHUDNOVSKY NUMBER TIIEORETIC APPLICATIONS 63

In the first part of the paper we consider polynornials with rational While classes of polynomials with rational integer coefficients studied in
coeficients arising from rational approximations t o logarithmic and al- Chapter I arise from the orthogonality condition and the Pad6 approxima-
gebraic functions. 'These polynomials are defined through linear recur- tion, polynomials studied in Chapter I1 can be characterized .w Chebicheff
rences with integer coefficients, and they have denonlinators growing not polynomials with rational integer coefficients. 'rhey can be characterized
faster than in geometric progression. Together with the ordinary three- as polynomials with rational integer coefficients least deviating from zero
term linear recurrences, we consider more general matrix linear recurrences on a continuum E. Number-theoretic interest in these polynomials was
that are more convenient for the determination of arithmetic properties of originated in 1936 by the famous Gelfond-Schnirelman approach to the
polynomials generated by them. All the recurrences under consideration prime number theorem, in which deviations from zero of Chebicheff poly-
are representations of the so-called contiguous relations, i.e., linear rela- nomials with rational integer coefficients were directly related to the dis-
tions I~etweensystems of analytic functions having the same monodromy tribution of pri~nes.We refer the reader to a recent book by Ferguson [lo],
properties [2], [6], [7]. The transformatio~~ from one contiguous system of where this approach and related results are presented.
functions to another, described by linear recurrences, is broadly called the We study the original Gelfond-Schnirelman approach in the context of
Backlund transformation because its appearance in completely integrable the "integer" transfinite diameter of a continuum E in the cases E = [O, 11
systems as a canonical transformation [6], [7]. In our case, contiguous rela- and G = [O, l/4], which are important for the prime number theorem.
tions are analyzed in the simplest cases of Fuchsian linear differential equa- We include an exposition of the results of Fekete and the recent results of
tions with regular singularities a t O,1, oo, and several apparent singularities Aparicio on upper and lowcr bounds on the "integer" transfinite diameter.
(generalizations of Gauss hypergeometric functions). Though the one-dimensional version of the Gelfond-Schnirelman
The new recurrences are applied to logarithmic and inverse trigonometric approach is obviously insufficient, we extend the remarks of Trigub and
functions, and lead t o new LLdense"sequences of rational approximations examine the nlultidinie~isionalgeneralization of the Gelfond-Schnirelman
t o logarithms of rational numbers. Using these "dense" sequences of ap- method. Analysis of the simplest cases of polynornials with rational integer
proximations (see Lemma 1.1), we obtain new improved measures of ir- coefficients. least deviating from zero on [0, lId shows that this approach
rationality of particularly important numbers like In 2, n/&, r,fi.These works. Interesting developments here (such as an approximation of $J(x)by
measures are considerably better than those [I] furnished by the ordinary solutions of singular integral equations) coming from statistical mechanics
Pad6 approximation technique. We present here only those applications and the many-body problems are presented.
of the method of Backlund transformations [8] where all the recurrence It is quite remarkable that the a~lalyticmethods employed here are
relations are given explicitly. essentially those of the Petersburg school (Chebicheff, Markov, Bernstein),
Another series of recurrences arises from Thue polynomials introduced with some arithmetic adjustments.
by Thue [9] a s a means of approximation of algebraic numbers a of degrees This work was supported by the U.S. Air Force under Grant AFOSR-
n 2 3. These polynomials P,(x), Q,(x) with rational integer coefficients of 81-0190 and the John Simon Guggenheini Memorial Foundation.
degree r n + [f]are defined by the "Padk-type" relations

References
for a polynomial S ( x ) and arbitrary r 2 1. We pay particular attention t o [l] G . V. Chudnovsky, C. R. Acad. Sci. Paris, v. 288 (1979), A-607-
the cubic a, when the three-term recurrence connecting P,(x), Q,(x) is given A-609; C. It. Acad. Sci. Paris, v. 288 (1979), A-965-A-967.
explicitly in terms of invariants and covariants of the cubic polynomial P(x) [2] G. V. Chudnovsky, in Bifurcation phenomena in mathematical
whose root is a . Applications to measures of diophantine approximations of physics and related topics, D. Reidel, Boston, USA, 1980, 448-510.
cubic irrationalities (including cubic roots m, D > 0 and solutions f (a), [3] K. Alladi, M. Robinson, Lecture Notes in Mathematics, Springer,
1980, v. 751, 1-9.
d < 0 of Weber's modular equation) are presented.
 G . V. CIIUDNOVSKY NUMBER TIIEORETIC APPLICATIONS

F. Beukers, Lecture Notes in Mathematics, Springer, 1981, v. 888, 1 - alb. Namely, (Lemma 1.1) the measure of irrationality is
90-99.
G. V. Chudnovsky, Lecce Lectures, Lecture Notes in Physics,
Springer, v. 120, 1980, 103-50. for all c > 0 and rational integers p,q; 191 >
D. V. Chudnovsky, G. V. Chudnovsky, Lett. Math. Phys., 4 (ISSO), "Dense" sequences of rational a.pproxirnations are established using linear
373-80. recurrences generated by contiguous relations. Asymptotics of solutions of
D. V. Chudnovsky, G. V. Chudnovsky, J. Math. Pures Appl., Paris, these recurrences define "density constants1' and measures of irrationality.
France, 61 (1982)) 1-16. We start with examples of Padk approximants and corresponding orthogonal
D. V. Chudnovsky, Festschrift in honor of F. Giirsey, 1982 (in polynomials. We examine Pad6 approximations to binomial functions and
press). powers of logarithmic functions. Here the recurrences are consequences
A. Thue, J. Reine Angew. Math. 135 (1909), 284--305. of contiguous rclations between generalized hypergeometric functions. We
Le Baron 0.Ferguson, Approximation by polynomials with integral describe separately Thue polynomials generating "dense" sequences of
coeficients, Mathematical surveys #17, American Mathematical rational approximations to algebraic numbers. This way we obtain new
Society, Providence, 1980. effective measures of diophantine approximations to cubic irrationalities.
In $4 we study new, rapidly convergent, improved Padk approximations.
Improved Pad6 approximations provide new bounds for measures of ir-
rationality of values of the logarithmic (inverse trigonometric) functions.
Chapter I. Padbtype Approximtions and Measures Particular attention is devoted to two special numbers In 2 and TI&. For
of ~ r r i t i o n a l i tof~ Logarithms and Algebraic Numbers these two numbers we present linear recurrences defining new "denseJ' se-
quences of rational approximations. Better "density constants" for these
We study diophantine approximations to numbers that are values of the sequences considerably improve measures of irrationality for ln2, ?r/d
logarithmic (or inverse trigonometric) or algebraic functions using polyno- given in [2], [3].
mials with rational coeflicients that are denominators and numerators or
rational approximations to corresponding functions. The main tool here
is given by recurrences and linear ordinary differential equations (1.o.d.e.) $1. Proofs of irrationality of the number 0 follow basically the same pattern:
with rational function coefficients satisfied by these polynomials. Linear
one establishes the existence of approximations p/q to 0 such that
recurrences themselves are reformulations of contiguous relations [I]
between systems of functions with the same monodromy group. They were
for
analyztd in the papers of D. V. Chudnovsky and G. V. Chudnovsky, see
e.g., [16]-[19], as Backlund transformations.
and $ ( N ) / N + co for N -+ co. Moreover, if we approximate 0 by a
These contiguous relations between approximants and functions furnish
"dense" sequence of rational numbers, then we obtain a simple method for
US with "denseJJsequences of rational approximations pn/gn to numbers a
finding the measure of irrationality of 0.
that are values of functions (at rational points). These "denseJ' sequences
We state the quite obvious and repeatedly reproduced
of approximations to a allow us to estimate the measure of irrationality
of a. The "dense" sequence in this paper means the following:
Lemma 1.1. Let us assume that there exists a sequence of rational
integers Pn,Q , such that

with a > 0, b < 0 as n -+ oo. Under these assumptions the exponent

of irrationality of a can be expressed in terms of the "density constantJJ
 NUMIIER TIIEORETIC API'LICATIONS

as n 4 ca and T h e n there are m linearly independent solutions x;) .

: j = 1 , . . , m of
(1.3) such that

as n 4 oo, where 7; < 0. T h e n the number 0 i s irrational and, moreover,

and there is only one (up t o a scalar multiplier) solution xno f (1.3) such
that
loglXnI -nloglArnl-
Though recurrences (1.3) are extremely important for the construction of
for all rational integers p, q with Iql >_ q o ( ~ ) . diophantine approximations, they do not solve the problem of "arithmeticJ1
If some family of rational approximations P n / Q n to 0 is found, then the asymptotics, since for a,(n) being polynomial in n, the expressions for
determination of the sizes of Pn and Q, as i n (1.1), i s called a n "arith- Xn tend to have complicated denominators. Only in very special cases do
metic" asymptotic, while the error of approzimation, as i n (I.!?), i s called these expressions have the denominator growing not faster than geometric
a n "analytic" asykzptototic of rational approximation. progression.
The essential point i n the correct determination of "analytic" and "arith-
metic" asymptotics of rational approximations is to establish the (linear)
recurrence formulas connecting successive P, and Q,.
52. One of the situations, in which we know both the "arithmetic" and
"analytic"asymptotics, is the case when the numbers under consideration
Knowing them, it is rather easy to determine "analytic" asymptotics of
are values of (generalized) hypergeometric functions.. The corresponding
P,, Q,, using the following l'oincark lemma: family of rational approximalions is a specialization of a system of Padk
approximations to (generalized) hypergcometric functions [5], [6]. The
L e m m a 1.2. Let "analytic" asymptotic is determined using the Riemann boundary value
problem that is associated with the monodromy group of the corresponding
differential equation. From the point of view of applications, we present
the corresponding results for systems of functions with only the simplest
singularities. For completeness we remind readers of the definition of the
be a linear recurrence with coeficients depending o n n such that Pad6 approximation (cf. [g]):

ai(n) -+ ai when n -+ oo. Definition 2.1. Let fl(x), . . . , fn(x) be functions analytic at
x = 0 and let m l , . . . , m, be non-negative integers (called weights). Then
Suppose the roots of the "limit" characteristic equation polynomials Pl(x), . .. ,P,(x) of degrees a t most m l , .. . ,mn are called Pad6
approzimants to f l(z), . . . ,fn(x) if

are distinct i n absolute values:

has a zero a t x = 0 of order > x y = l ( m i + 1) - 1. The function R(x) is
called the remainder function.
68 G . V. CHUDNOVSKY NUMBER THEORETIC APPLICATIONS

For systems of functions with the simplest monodromy (like an Abelian Then for any x # O,1, oo where r,(x) < r$(x), we have
group as a monodroniy group) wc know explicit expressions for the remaind-
er function and the Pad6 approximants [5]. These expressions are given as
simple contour integrals in the complex plane or multiple integrals in the
real domain. Thesc explicit expressions can be used in order to obtain an
asymptotics of the remainder function and Pad6 approximants by means of
the Laplace (steepest descent) method. Alternatively, asymptotics can be
i = 1,.. .,n as M
2 x m . j
obtained from contiguous relations between solutions of 1.o.d.e. satisfied by -+ m. Here 1; = exp(_).
the remainder function. The corresponding 1.o.d.e. and contiguous relations
themselves can be obtained as well from integral representations.
We start with the asymptotics of the remainder function and Padk Following [5] we present the contour integral representation for the
approximants in the Pad6 approximation problem for binomial, logarithmic remainder function and Pad6 approximants in the most general case iii).
and similar hypcrgcometric functions, see [I], [5], and then present explicit Moreover, our expressions include the inhomogeneous case of n 1 func- +
integral representations. :
tions fo(x), f ~ ( x ) ,. . ,fn(x):

Theorem 2.2. Let w l , . . . ,w, be distinct (mod 2 ) complex numbers.

Let fl(x), . . . ,fn(x) be one of the following systems of functions:
The normalized remainder function R(x) for the Padk approximation
problem to the syste~nof functions fo(x), . . . ,fn(x) with weights mo, . . . ,mn
i) fi(x) = log(1 - x)'-' : i == 1,.. . ,n; has the fol!owing form:
ii) f i ( x ) = ( l - x ) " i : i = 1 , ...,n;
.
iii) fi(x) = 2F1(1;wi; c; x) : i = 1, . . , n.

for

be the remainder function in the Pade' approximation to fl(x), . . . ,fn(x)

with weights m l , . . . ,m, at x = 0.
and the closed contour C encircling all zeros of Fo(s). Similar integral rep-
Let resentations hold for Pad6 approximants P;(x) : i = 0,. . .,n, but with
the contour C replaced by C;, where C; encircles the zeros of
s ( , . ? ?n + k - wi) but none of the other zeros of Fo(s), i = 0,. . . , n and we
formally put wo = 0 and for nzo > 0 assume that w; - w j # 0 (mod 2 ) for
Then the asymptotics of IR(x)l and IPi(x)l is determined everywhere in C i f j.
using the following notations: Explicit expression for Pad6 approximants and the remainder function
in cases i) and ii) can be deduced from Hermite's expression [4] of Pad6
~ , ( x ) = m i n { l l - - ~ ~j =~ 0I , : ...,n - 1 ) , approximants for syst,ems of exponential functions, see [3]. We present the
multiple integral representation for the remainder function in problcms i)
and ii) following the Hermite method. Similar expressions can be given in
70 G. V. CIIUDNOVSKY NUMBER TIIEORE'I'IC APPLICATIONS 71

problem iii) its well. In problerii iii) we put: analyzed. These recurrences for systems of functions satisfying Fuchsian
linear differential equations are called continguous relations, following
Riemann 111. We concentrate here only on the case of logarithmic functioris
as a particular case of the Gauss hypergeometric function.
The recurrences that are consequences of Gauss contiguous relations
between p F l functions can be prcserited in the form:

F ( m + 1,1, k;z) = F ( m , 1, k - 1; z)+ zF(m, I, k; z);

This integral representation, unlike tlie contour integral representation F(m11 + 1, k; z) = F ( m , 1, k - 1; z) + (Z - l)F(m, 1, k; z).
(2.1), can be directly used to obtain asyniptotics for the remainder func-
tion. Formulas similar to (2.2) (see [20]) can be given for the Pad6 ap- The specification of initial conditions F(1, I, k;z) gives us Pn(z), Qn(z),
proximants Pi(x) : i = 1,. . . , n if tlie domain of integration is moved to Rn(z) in the Pad6 approximation problem for In (1 - i):
the complex space. From the integral representation (2.2) o ~ i ccan obtain
an integral representation for the Pad6 approximation problem i), if one
formally puts wi = 0. Apparently a simpler expression can be obtained if
in i) one considers instead the Pad6 approximation problem Ibr functions
where R,(z) = O(Z-~-') as lzl -r oo, Pn(z) and Qn(z) are polynomials of
hi-'z : i = 1 , . . . , n at z = 1 with equal weights m 1 = . . . = rn, = rn.
degrees n and n - 1, respectively.
The expression for the remainder function R ( z ) in this problem is the fol-
lowing,
n-1
i) If Fl(l,l,k; z) = & {(-z)~-' - (1 - z ) ~ - ~for
) k # 2 and
(ti - I)" F l ( l , l , 2; z) = ln(1 - $), then
R )= .. . - tl. * .tn-l)m dti 1
D z i=l tyfl def
Rn(z) = I;;(n+ 1 , n + 1 , n + 2;z);
where the domain D, of integration is (for real z > 1) the following
one: D, = { ( t l , . . . ,tn-1) E Rn-l : 1 < ti < z, t l , . . . ,t n _ l < z). ii) if F2(1,1,k; z) = Sk2, then
There is an interesting similarity bctweeri tlie expression (2.2) for the Pad6
approximation to n binomial functions and the Hermite formula for the def
remainder function in the Pad6 approximation problem to n exponential
Pn(z) = F 2 ( n - t 1 , n + 1 , n + 2; z);
functions eWoZ, .. . ,eWn-lZa t x = 0 with weight-s mo,. . . ,mn-1 [21]:
iii) if F3(l,1, k; z) = & { ( - z ) ~ - ~ -(I - z ) ~ - ~ ) F3(1,
, 1,2; z) = 0, then

def
&,(a) = F 3 ( n + l , n + 1 , n + 2;z).

These recurrences are usually replaced by a single three-term recurrence

Unlike "analytic" asymptotics, denominators of coeacients of Pad6

approximants Pi(z) cannot be easily determined. To find der~oniinatorsof
Pad6 approximants and to determine "arithmetic" asymptotics, the recur- satisfied by Xn = Pn,Qn or R,. Nevertheless, the recurrence (2.3) conceals
rences relating Pad6 approxirnants with contiguous weigl~tsshould be the most remarkable arithrrietical properties of Pn(x) arid Q,(x):
72 G.V. CHUDNOVSKY NUMBER TIIEORKTIC APPLICATIONS 73

iv) coefficients of Pn(z) are rational integers; and rational numbers. For this we siniply use Lemma 1.1. The "arithmeticJ' and
v) cocficients of Qn(z) are rational numbers with the common denomi- "analytic" asymptotics are givcn by the properties iv)-v) and thc recurrence
nator dividing the least common multiplier of 1,. . . , n, denoted by (2.3), together wit11 Lemma 1.2. \Ne start with the number ln2, which
lcm{l, . . . ,n) (i.e., growing not faster than e('+O(l))n as n -r m). corresponds in the above-menticned scheme to z = -1. Lemma 1.1 on
"dcnse" sequences of approximations immediately gives us, as in [2], [8],
Statements iv), v) do not follow immediately from (2.3), since one may [15]:
-3.660137409...-c
suspect the denominator to grow like n!, because of the division by n 1+ IqIn2-pl > q
in order to find Xn+l. However, previous matrix recurrences together for the rational intcgers p, q, providcd that 191 2 ql(t) for any E > 0.
with (2.3) immediately imply iv), v). Hence one needs all the recurrences Similarly, Padk approximations to the logarithmic function a t points of
simultaneously. the Gaussian field Q ( i ) give us a measure of the diophantine approximation
Comparing (2.3) with the classical recurrence for Legendre polynomials
to ../&[21, [81, [151.
[7], an immediate relation is established:

where P n ( x ) is the Legendre polynomial of degree n:

for rational integers p, q, provided that Iq( >
q2(c) for any 6 > 0. Moreover,
results of computer experiments show that [lo]:

Similarly, -
for all rational integers p, q, with 191 > 2'. Historically the first explicit
Rn(z) . z-" = Qn(x), x = 1- 2z-', published result on the measure of irrationality of belongs to Dariilov
[Ill. Later, similar bounds with dinerent exponents had been obtained
where g,(x) is a Legendre function of the second kind and, consequently, independently by several rescarchcrs (including Wirsing and Reukcrs), see
[2], [8]. All these results were based on the same system of Pad4 approxima-
tions of the logarithmic function and, hence, on the Hermite technique.
The difference in the exponent is explained by different accuracy in the
Because of the connection with the classical tlioery of orthogonal poly- computation of asymptotics. We prcsent now the recurrence leading to the
nomials one can write exact asymptotics together with explicit solutions to the recurrence: the
exponent 7.309.. . in the nieasurc of irrationality of r / f i is connected with
the following nice three-tcr~nlincar recurrence:

This shows that coefficients of Q,(z) do have denominators, but they are
of the form : m = 1,.. . ,n, which proves the properties of Pn,Q,. The
asymptotics of Pn, Qn follows from (2.3) according to Poincark's lemma. + +
As n -+ oo, x 2 72 1 = 0, where Xn -
xn as n -+ oo.
The properties i)-v) of Padk approximants and the remainder function Then there are two solutions p, and q,, of this recurrence such that
in the Pad6 approximation problem for the logarithmic function enable us
to estimatc the measure of diophantinc approximations of the logarithms of gn E Z for all n
74 G. V. CHUDNOVSKY NUMBER TIIEORETIC APPLICATIONS 75

and diophantine approximations to n / & and ln2. These new measures are
p, - l c m { l , . . . , 2 n ) E Z for alln; remarkable because for the first tirnc they do not make use of the Hermite
technique of 1873 [4].
by the Poincark theorem Ncverthcless, IIermite's technique of Pad6 approximations to powers of
the logaritlirnic function provides a nontrivial bound for the measure of
irrationality of n. For this wc use Pad6 approximiations to the system of
functions in i) of Theorem 2.2 for n = 6, cf. [3]. The bound obtained so
Then far is the following:
IqT
- > lql-18.8899414 ...
The expression for q, (E Z) is rather simple: for rational integers p, q with Iq1 2 q4(6) and any 6 > 0.

53. In the study of diophantine a.pproximations to algebraic numbers,

we have a class of polynomials with rational coefEcients similar to the Pad6
This number is indeed an integer, and it can be represented in a number of approximants that were introduced by Thue [22]. Let a be an algebraic
different ways since it is connected with tlie values of Legendre polyno~nials. riumber of degree n > 2 which is a root of an irreducible polynomial
Namely, P(x) E Z[x] of degree n. Then for every integer r 2 0 there exist polyno-
mials A,,(x), BT(x) with rational integer coefficients of degree of at most
r n + [g] such that

Similarly, an expression exists for a "near integer" solution p,: where C,(x) is a polynomial. These polynomials provide a "dense" sequence
of rational approximations p,/q, to a, if one starts with an initial good ra-
tional approxirnatiori x = plq to a, and puts p,/q, = B,(p/q)/A,(p/q). In
particular, if the initial approximation p/q is too good, then the exponent
of the measure of irrationality of a will be less than n, i.e., we obtain an
for o(m) = 1 +4 + + $. This explains why the denominator of pn improvement over the Liouville theorem. It is in this way that A. Thue
divides lcm{l, . . . ,2n). [22] proved his first ineffective theorem on diophantirie approximations to
The asymptotics of p,,, q, give a weak measure of irrationality of TI&: algebraic numbers according to which there are only finitely many rational
approximations lo a with an exponent of irrationality larger than n/2. This
method of Thue ay well as his theorem was subsequently generalized to the
multivariable case by Siegel, Schneider, Gelfond, Dyson, and fi~iallyRoth
[23]. Itoth's theorem states that there are only finitely many approxirna-
for 191 L 93(6) and X = In
In (( 22+- ffii))++1l = -7.3099863. +
tions to a with an exponent of irrationality larger than 2 E for every
6 > 0. However, this theorem as well as the original one of Thue does not
-
Thounh the r>ossibilitiesof l'add approximations
-- to the logarith~nicfunc-
give an cffective bound on the size of these approximations. Only Baker's
tion are co~npletelyexhausted by the bounds above, the improved rational
approximatioris to the logarithmic function give new measures of irrational- theorem gives us a slight effective improvement over tlie Liouville theorem,
ity. In 54 we present new recurrcIices and new bounds of the measure of and since any effcctivci~nprovcmcntin the exponent is important, attention
76 G . V. CHUDNOVSKY NUMBER THEORETIC APPLICATIONS 77

to the old Thue method has been revived in the hope that it can give new, We present recurrences and expressions for Thue polynomials associated
l ~,n), = 1, is
powerful results. Typically the case of roots a = ( ~ / b ) ~( m with a in the homogeneous form, when an initial approximation to a is z/y
considered since in this case Thue polynomials are expressed explicitly in (see y = 1 above). The approximants A,(x, y) and Ur(x, y) are homogeneous
terms of Pad4 approxir~lantsX,(z), Y,(z) to the function z", v = m/n, a t polynomials in x, y of degree 37 + 1 with rational coeficiemts and satisfy
z = 1, [24], [25], [26]. The Pad6 approximants X,(z), Y,(z) are polynomials .C,(x, y) for a polynomial C,(x, y). The
odi,(x, y) - B,(x, y) = (x - ay)2ri-1
of degree r and satisfy a simple three-term linear recurrence following three-term linear recurrence

(and the same recurrence is satisfied by Y,(z)), r = 1,.. . . Polynomials - 9(3r + 1 ) ~(x,f Y)~z,-I(x,Y) = 0
Xr(z), Yr(z) are, in fact, expressed as hypergeometric polynomials (Jacobi
has both A,(x, y) and B,(x, y) as its solutions. This enables us to express
We have
A,(x, y) and B,(x, y) in terms of Jacobi polynomials X,(z) from (3.1) for
Y,(z) = z'X,(z-I). v = 113. We put for r 2 0
(3.1) X,(z) = 2F1(-r, -r - v ; 1 - v ;z),

In view of the Kummer formula X,(z), Y,(z) can be expressed as a

hypergeometric polynomial in 1 - z as well: see (3.1) for v = 113. Then two linerly independent solutions of (3.2) are
(2r)!
. 2F1(-r, -I - V ;-2r; 1 - a),
X,(z) = r ! ( l - V). . .(r - v)

Thue polynomials A,(x) = A,(x, I), B,(x) = B,(x, I), similar to other
classical extrernal polynomials, satisfy Fuchsian 1.o.d.e. of the second order
Using properties of polynomials X,(z), Y,(z), many interesting results in addition to the three-term recurrence (3.1). If we put f(x) = f ( x , I),
on the diophantine approximation to roots a = ( ~ / b ) were ~ l ~ obtained then y(x) = A,(x), B,(x) are two linerly indepeudent solutions of the 1.o.d.e.
(251 [26], (31 (see specially [3] for new results on denominators of X,(z)).
Below we consider another class of algebraic numbers: cubic irrationalities,
following [27] (cf. [28]).
We consider cubic irrationalities a that are solutions of the algebraic It is interesting to note that, though f (x) has three regular singularities,
equation f ( a ,1) = 0 for the binary cubic form there are no accessory parameters in this equation.
For small Ix - ayl, polynomials A,(x, y), U,(x, y) are approxirnants t o a.
It, follows from Poincark's Lemma 1.2 that $ loglA,(x, y)l, $loglB,(x, y)l
are asymptotically logI&1, while $ loglA,(z, y)a - B,(x, y)l is asymptoti-
with rational integers a,b, c, d. With this form we associate an invariant
+ +
D (discriminant): D = -27a2d2 18abcd b2c2 - 4ac3 - 4b3d, and two
+
cally logltll where (1, 6 2 are roots of t2 2C(x, y ) t - 27D f(x, y)2 = 0,
(quadratic and cubic) covariants
It1( < I t 2 [ .According to [3] and (271, the common denominator A, of the
coefficients of the polynomials A,(x, y) and B,(x, y) satisfies

1
- log A, -+
4
-
r 6 .
78 G . V. CHUDNOVSKY NUMBER TI-IEORETIC APPLICATIONS

Also there are additional divisibility roper ties [26], [28]: coemcients of Example.
the polynomial X,(1 - 3 3 / 2 z )arc divisible by 3[3'/21. IIence according to 1) D = 2 : ~ = 1 . 4 2 9 7 0 9 ...; 2 ) D = 3 : ~ = 1 . 6 9 2 6 6 1 ...;
Lemma 1.1 we obtain the following 1271: 3) D = 6 : x = l . 3 2 0 5 5 4 ...; 4) D = 7 : ~ = 1 . 7 2 7 5 0 3 ...;
5) D = l O : x = l . 4 1 3 8 8 6 ...; 6 ) D = l 2 : ~ = 1 . 9 0 7 8 4 0...;
Theorem 3.3. Let x , y be rational integers and X / Y be a n approxima-
7) D = 13 : x = 1.824735.. .; 8 ) D = 15 : x = 1.493153.. .;
tion to a : 1x - ayl = min{lx - B y ] : f ( P , 1) = 0 ) and let D l G ( x ,Y ) = G ,
H ( x , y) = H be as above. For a rational integer M such that 9)
11) D = 19 : x -
0 = 17 : x = 1.198220.. . ; 10) D = 18 : x = 1.907841.. . ;
1.269963.. .; 12) D = 20 : x = 1.194764.. .:

The second class of numhcrs is given by values of modular functions and

we put G I = IG/MI, Dl = 27D f 2 / M 2 . There exists a n eflective imyrove- is closely connected to the IIecgner-Stark studies of one-class discriminants.
m e n t over the Liouville theorem o n the diophantine approximations t o a , if These numbers arc very interesting because they admit a few unusually
33'/2 . (G1 + d m ) > Df . e*&i2 where v3(D1)2 i for i = 0, . . . $ 3 . large partical fractions in thc iuitial part of their continued fraction ex-
The exponent i n the measure of irrationality of a is the following. For pansions (Stark [29]). The modular function is f ( z ) = q - 1 / 4 8
qn-1/2),
+ nz=,(l
arbitrary rational integers p and q we have
>
-
= e 2 x i z for Irn z 0. The six appropriate numbers here are f ( & )
for which the imaginary quadratic field Q(&), d < 0 has a class nurn-
ber h = I and Id/ 3 (mod 8 ) : D = -3, -11, -19, -43, -67, -163. In
all these cases we present the corresponding (cffcctive) exponent n < 3
if Iql > q,(a) and q,(a) depends eflectively only o n a and a n arbitrary of the measure of irrationality of algcbrraic numbers a from the field K ,
E > 0. The exponent x is
K = ~ ((A)): f

for rational integers p, q with an effective constant c > 0 depending only

where v3(D1)2 i for 0 < i < 3. on K . Simultaneously we prcserit an cquaiion satisfied by f = f ( & ) [29].

We note that once the bound (3.4) is established for a given cubic I. Let d = -3. Then f - 2 = 0 , so n = 2.429709.. ..
irrationality a , the measure of irrationality with the same exponent is true 11. Let d = -11. Then f 3 - 2 f 2 + 2 f - 2 = O a n d n = 2.326120 ....
for any other irrational number fro111 the cubic field Q ( a ) . Namely for an
111. Let d = -19. Then f - 2f - 2 = 0 and rc = 2.535262.. ..
arbitrary irrational p E Q ( a ) we have
IV. Let d=-43. Then f 3 - 2 f 2 - 2 = O a n d n = 2 . 7 3 8 3 8 7 ....
V. L e t d = - 6 7 . Then f 3 - 2 f 2 - 2 f - 2 = O a n d n = 2 . 8 0 2 3 7 0 ....
VI. Let d = -163. Then f 3 - 6 f 2 + 4 f - 2 = O and n = 2.882945 ....
for an effective constant c , > 0 depending only on c > 0 and the ficld
Q ( a ) . IIence in all examples it is sulficient t o give only an exponcnt for a
singlc generator of the field Q ( a ) . The first series of cxamples we prcscnt 54. Here we present recurrences and their solutions that provide "dense"
gives the measure of diophantine approximations for cubic irrationalities sequences of rational approximations to In 2, x/fi and x with "density con-
from a given pure cubic extension ~(m) for a rational integer D > 0. stants" better than rational approximations given by the Pad6 approximrr-
80 G . V. CHUDNOVSKY NUMBER THEORETIC APPLICATIONS 81

tion t o the logarithmic function. Following Lemma 1.2, one can determine Lemma 1.2 by the roots of the quartic polynomial. Numerically one has
the "analytic" and "arithmetic" asyrnptotics of solutions of these recur-
rences. We follow the gcneral scheme of 52 but with matrix and scalar
recurcnces of a more complicated form reflecting a different monodromy
structure.
The new, better measure of irrationality of In 2 is based on a new set and
of contiguous relations, reflecting the presence of apparent singularities. loglPn In 2 - &,I -- -1.77602924 -n
Thcse recurrences are the following:
as n -+ 00.
G(m + 1, n , k; z) = G(m, n, k - 2; z) + (22. - l)G(m, n, k - 1; z ) Hence one has the following measure of irrationality of In 2:
(4.1) + (z2 - z ) ~ ( mn,, k; z);
G(m, n + 1,k; z) = G(m, n, k - 2; z) + G(m, n , k; z)(z - z2).
Solutions to these recurrences are completely determined by the initial The best exponent in the measure of the diophantine approximation to
conditions G(1,1, k; 2). In 2 we can achieve this way requires more complicated recurrences than
These are two kinds of initial conditions t h a t determine the sequences (4.1), correspponding to more apparent singularities. The properties iii), iv)
P, and Q,: hold, but tlic asymptotics of P,, Q, and R, are substituted by the following
ones:

and
((1 - z)2-k - ( - z ) ~ - ~ )for k # 2,
ii) G 2 ( l l1,k; z) = k-2
dzf
G z ( l , I , 2; z) = 0. Then Q, - G2(N1, N2,N3; -1);
logIP, In 2 - &,I - 1.93766649 . n
N1 = [0.88n], N2 = [0.12n], N3 = n. as n -+ co. Hence the measure of the diophantine approximation of In 2 is:

The specialization z = -1, above, corresponds t o the specializatiou of In 2

as a value of the logarithmic function In (1 - 112). Two solutions P, and
Q, of (4.1) with initial conditions i) and ii) satisfy the following familiar
for rational integers p, q with (ql 2 ql.
properties:
Sometimes new, better measures of irrationality are corinected with
iii) P, are rational integers; new three-term linear recurrences rather than with matrix or multi-term
recurreIices. A new measure of the diophantine approximation for the
iv) Q, are rational numbers whose deriominators divide Icm{l, .. .n ) . number x / a arises this way from thc Padk-type approximation to the
function arctg x/x. We present the corresponding linear recurrences in
Now the remainder R, = Pnln 2 - Q, again arises from the recurrences their general form
(4.1) and is very small, when n -+ oo so t h a t Qn/Pn determines a very good
rational approximation to In 2.
The numbers P,, Q,, R, satisfy a scalar recurrence with coefficients
depending on n which is not a three-term recurrence, but a four-term linear
recurrence. The asymptotics of its solutions is determined according t o
82 G . V. CHUDNOVSKY NUMBER TIIEORETIC APPLICATIONS 83

and for i, j = f l , . . . ,f 3 we denote The approximants x?), x n ( possess

~ ~ the
~following
~ ~ properties:
)

b(1, j) = b(i; 3n + j, 4n + j); I. x:") E Z, denominators of Xponin)

divide lcm{l, . . .,n);
d = cb(2,l)b(2,0) - cb(1, l)b(3,0)eb(2,O)b(3,I), 11. As n -+ m one has
d = c'b(2, -2) - b(2, -3) - c1b(3,-3)b(1, -2) - e1b(2,-2)b(l, -3);

and b(i; n, m) : i = 1,2,3; a, c, el a', c', e' dependent on n and z have the
following form:

Applying Idemma 1.1 to this sequence of approximations, we obtain the

measure of irrationality for .rr/a ,
which is cor~siderablybetter than the
previous one:
Id& - > 141- 4.817441679...
PI
for rational integers p, q with Iq1 qo. >
One can ge~ieralizethe recurrences for Y, in order to obtain the best
measure of irrationality for so far. These three-term linear recur-
rences are very similar to the one given above, and their solutions (denoted
by P, and Q,) satisfy the following properties.

I. The numbers P,, Q, are rational numbers, where P, are integers

and the denominators of Q,, divide Zcm{l,. . . ,n);
11. We have
This single recurrence is in fact a consequence of simple matrix relations
equivalent to contiguous relations and is similar to one given for In 2.
loglPnl}
10glQnl
- 2.191056949 - n
There are two linearly independent solutions of this recurrence: one
"integer" solution ye),
and another "almost integer" yLnonin) (when
as n -+ oo.
111. The numbers Qn/Pn approximate n / a :
the denominators divide Icm{l, . . . ,4n)).
We prescnt here a formula for an "integer" solution of the recurrence
expressed as a sum of products of binomial coefficients:
as n -+ oo.

Properties 1-111 imply, according to Lemma 1.1, the following measure

of irrationality of n/&:

For n / & one should take z = -3, and for normalization we put
xn= 12["/.'1 . y[n/41. for all integers p, q with lql > qo.
84 G. V. CHUDNOVSKY NUMBER THEORETIC APPLICATIONS 85

These and similar measures of the irrationality of n / f i arise from three- Theorem 2.2 are n+l-term linear recurrences with coefficients that are poly-
term linear recurrences similar to the one satisfied by Y, above. Expressions nomials in N and x. They are deduced from matrix contiguous relations
of P, (or y t n ) )are slightly different but are expressed x
i sums of products for the generalized hypergeometric functions. The asymptotics of Padk
of three binomial coeflicients, unlike the case of Pad6 approximations to approximants and the remainder functions which follow from lemma 1.2 is
logarithms, when they are sums of the products of two binomial coefficients presented in Theorem 2.2. Sirnilar recurrences can, in general, be presented
only PI, PI, 1151. for the generalization of Thuc polynomials from $3, which are defined as
New dense sequences of approxi~nationsare found for n 2 , improving on polynomials PN(x, y) E Z [x, y] such that
the measures of irrationality for r2and T . They have expressions like above
but with the "integer" solution being a sum of a large number of binomial
coefficients. Also the three-term linear recurrence is substituted by multi-
term linear recurrences. We obtained the following measure of irrationality
of n 2 (improving that of Apdry) for ilBl+ i282 < N (with parameters 01,02). These Gelfond-Dyson poly-
nomials, as similar Roth polynomials in d variables, lead to better effective
measures of irratibnality of an algebraic number a, provided one knows
d - 1 initial good approximations to a. In particular cases of algebraic
roots a, Gelfond-Dyson polynomials can be determined explicitly through
so that (nq - p( > (ql-16 for rational integers p, q with Iql 2 ql. differential equations and the recurrences which they satisfy. A special
case of Gelfond-Dyson polynomials is the one considered by Siege1 [24],
which can be reduced to the Padk approximations of theorem 2.2. case ii)
with ml = . . . = m, = N , w l = 0, w2 = v ,...,w, = ( n - l ) and ~
$5. For number theoretical applications, it is important to use simul- a = (1 - x)" with x, v E Q; z # 0. The asymptotics of the remainder
taneous approximations to several numbers and, in particular, linear forms function and Padk approxirnants as N -+ co is given in 2.2, but for diophan-
in 1 , 0 , . . . , en-l for a given number 0. For example, to study diophan- tine applications we need to know the common denominator of Padk ap-
tine approximations to a value f(xo) of a function f(x), one has to apply proximants Pi(x) : i = 1, . . . ,n. This denominator was estimated in 53 for
Pad6 approximations to a system of functions 1, f(x), . . . ,f (x)"-' (see 2.1). n = 2 and Y = 113. We now present the bounds on denominators in the
Theorem 2.2.) cases i) and ii), furnish 11swith analytic tools to examine these general case, cf. [W].
Pad6 approximations. Integral representations (2.1) or (2.2) give us infor- For n = 2 and an arbitrary rational I/ = s / r , (a, r) = 1, the common
mation about the arithmetic properties of coefficients of Pad6 approximants. denominator AN of Pad6 approximants Pl(x), P*(x), in Pad6 approxima-
This way we can study rational approximations (or, in general, approxima- tions to (1 - x)" a t x = 0 with weight N, has the following asymptotics
tions by algcbraic numbers of f i e d degree) t o the numbers (1 - x)" or
In (1 - x) for v E Q(v 4 Z) and x E Q, x # 0. Moreover, since weights
in Padk approximations can vary, various similar diophantine approxima-
tion problems can be studied using Pad6 approximations. One of them is
the p-adic diophantine approximation to algebraic numbers and algebraic
roots, and another is the diophantine approximation by numbers whose as N -r co. Similarly, for n > 2 the asymptotics of the common
denominators have primc factors from a fxed set. Padk approximatiorls denominator of coefficients of Padk approximants can be determined ex-
to a system of functions lead to matrix (and linear) recurrences generaliz- 4
plicitly through sums of values of d log I-( with rational z [27]. However,
ing the ones given for n = 1. Linear recurrences connecting Pad6 ap- a simple upper bound can be given for the common denominator A% of
proxirnants and the remainder function for equal weights m l = =
e . .
coenicients of Pad6 approximants Pl(x), . . . ,P,(n) in the Padk approxima-
m, = N for the Pad6 approximation to functions in cases i), ii), iii) of tion to functions 1, (1 - x)", . . . ,(1 - x)(~-')" a t x = 0. For v = s / r ,
86 G . V. CIIUDNOVSKY NUMBEIL THEORETIC APPLICATIONS 87

(3, r) -I 1 and prime r we have D. V. Chudnovsky, Festschrift in honor of F. Giirsey, 1982 (in
press).
F. Bcukers, Lecturc Notes in Math., Springer, 1981, v. 888, 90-99.
D. V. Chudnovsky, G . V. Chudnovsky, Lett. Math. Phys. 4 (1980),
373.
D. V. Chudnovsky, G. V. Chudnovsky, J. Math. Pures Appl. 61
as N -+ oo. (19821, 1-16.
This lcads to new effective measurcs of irrationality (or measures of D. v.' Chudnovsky, G. V. Chudnovsky, Phys. Lett., 82 A (1981),
approximations by algebraic numbcrs of fixed degree) for algebraic roots 271.
i/rd with D > 0 (see the examples of $3 for r = 3). For example, D. V. Chudnovsky, G. V. Chudnovsky, Phys. I,ctt., 87 A (1981),
estimates of denominators from [27], $6, give uninflated bounds for an 325.
effectivc improvement OF the Liouville thcorem for numbers 0
with a K. Mahler, Math. Ann. 105 (1931), 267-76.
fixed D and r >
ro(D). We mention that for D = 2 one can take ro = 83. K. Mahler, Math. Ann. 168 (1967), 200-27.
A. Thue, J. Reine. Angew. ~ a t h . ~ 1 3(1909),
5 284-305.
W. J. LeVeque, Topics in number theory, Addison-Wesley, v. 2,
1957.
C. L. Siegel, Abh. Prcus. Akad. Wiss. No. 1 (1929).
References C. L. Sicgel, Akad. d. Wiss. in Gijttingen 11, Mat.-Phys. I., 1970,
no. 8, Gott. (1970), 169-95.
A. Baker, Quart. J. Math. Oxford 15 (1964), 375-83.
B. Riemann, Oeuvres Mathematiques, Blanch'ard, Paris, 1968, pp.
G. V. Chudnovsky, Ann. of Math. 117 (1983), No 3.
353-63.
D. K. Faddeev, Studies in number theory, ed. by A. V. Malyshev,
G. V. Chudnovsky, C. R. Acad. Sci. Paris, 288 (1979), A-607-A-609.
N.Y. 1968, 51-56.
G. V. Chudnovsky, C. R. Acad. Sci. Paris, 288 (1979), A-965-A-967.
H. M. Stark, in Applications of computers in algebra, Academic
Ch. IIermitC, C. R. Acad. Sci. l'aris, 77 (1873), 18-24, 74-79, 226.
Press, 1970, 21-35.
G. V. Chudnovsky, Cargesi: lectures in Bifurcation phenomena in
mathematical physics and related topics, D. Reidel, Boston, 1980,
448-510.
G. V. Chudnovsky, J. Math. Pures Appl. 58 (1979), 445-76. Chapter 11. On the Notion of the
G. Szego, Orthogonal polynomials, AMS, Providence, 1938. "Integer" Transfinite Diameter. and the
K. Alladi, M. L. Robinson, Lecture Notes Math., v. 751, Springer, el fond-~chnirelman Method in the Prime Number Theorem
1979, 1-9.
G . A. Baker, Jr., Essentials of Padk approximants, Academic Press,
1975. $1. We remind the readcr of the definition of the transfinite diameter of
C. V. Chudnovsky, Lecture Notes Physics, v. 120, Springer, 1980, a continuum E: defined by Fekete [I]. We put
103-150.
V. Danilov, Math. Zarnetki, 24 (1978), No. 4.
d,(E) =
21,
max J-JI z ~
...,z,EE I.< j
- zj1 2/n(n-1)

A. Y. Khintchine, Continued fractions, University of Chicago Press,

1964. Then d(E) = limn,, d,(E) is called the transfinite diarnter of E. Alterna-
I<. Mahler, Philos. Trans. Roy. Soc. London, 245A (1953), 371-98. tively, the transfinite diametcr can bc described using thc maximum of
88 G . V. CIIUDNOVSKY NUMBER THEORETIC APPLICATIONS 89

the absolute value of a deviation on E of polynomials of growing degree. Hence a . e*(2nN+1) is a non-zero rational integer and a < 62nN, so that
Among all polynomials P,(z) = zn + . + a0 which have the leading ~ ' 7- <~ " ~ or -2nN . log 6
e$(2nN+1) <G(2nN + 1) for an integer N 2 1.
coefficient one, we choose the polynomial with the minimal value of its norm This lower bound for $(x) would lead to the prime number theorem, if as
lPnlE= maxzEEIPn(z)l. The value of this norm is denoted by p", Then Gelfond-Schnireleman had hoped, 6 -+ l / e as n -+ oo. Unfortunately, a s
d(E) = limn,, p,. For an interval of length 1 in the plane, the transfinite Gelfond found much later, 6 > l / e and even 6 > (l+fi)-' # 0.414213.. .
diamter is 114. (more precise results on values of 6 are presented below).
Polynomials least deviating from zero were proposed as an instrument Despite the failure of the straightforward application of the original Gel-
for the elementary proof of the prime number theorem by Gelfond and fond-Schnirelrnan idea, it was soon realized that the same analytic method
Schnirelman in 1936 in their analysis of Chebicheff's proof of the Bertrand could still be used for the proof of thc prime number theorem if one
postulate. An account of this extraordinary attempt can be found in [2] and considered only polynomials with several variables with integer rational
[3]. The modern reader can find an account of the Gelfond-Schnirclman coefficients least deviating from zero on a d-dimensional cube [0, lId.
method for the prime number theorem in the historical section of the book The multidimensional approach has essentially been proposed by
by Le Baron Ferguson [4]. Less readily available, but not less interesting Trigub [5] and Ferguson [4]. Moreover, Trigub [5] has suggested that, per-
are the works of other representatives of the Russian school in constructive haps, an appropriate polynomial P(x1, . . . , xd) can be constructed using an
function theory (see Trigub [5] and Aparicio [ll]). analog of the Roth method by demanding that P(x1,. . . ,xd) have zeros
We repeat briefly the essence of the original simple attempt of Gelfond of different orders at given points inside the cube [0, lId. This avenue
and Schnirelman. The main number theoretical function under considera- has not yet been properly pursued, and instead the examples of polyno-
tion is the Chebicheff function +(x): mials P ( x l , . . . ,xd) considered here are given explicitly, rather than proved
t o exist by the Dirichlet box principle. Results for polynomials arising
from different discriminants and connected with the Selberg integrals are
presented below. Though we do not prove the prime number theorem in
its full generality, our constants seem to be close to the expected 1.
where the summation is extended over primes p sich that pm 5 x. The
prime number theorem itself is equivalent t o the statement +(x) -z for
large x[6]. The function +(x) is a logarithm of a more familiar arithmetic
52. We start with the notion of the integer transfinite diameter of a
function: e*(') = lcm(1,. . . , x ) for an integer x. The idea of Gelfond-
Schnirelman is to estimate +(x) from below by taking an integral of a
continuum E. For a given n >
1 we put
polynomial with rational integer coeficients least deviating from zero on
the interval [0,1]. Namely, let P ( x ) =xy=o pixi be a non-zero polynomial
pn(E) = rnin
f',.(z)EZ[z]
maxl~(x)l'l~,
zEE
of degree n with integer rational coefficients such that
where the nlin is taken over the set of all polyno~nialsPn(x) with integer
rational coeficients and of degree of at most n. Then the quantity

Then 6(E) = lim pn(E)

n-w

is called the integer transfinite diameter of E. The only obvious lower

and a is a rational number whose denominator obviously divides bound on 6(B) is d (.E ),, the transfinite diameter of E difined in 61. There ~ -~

is a lower boaid on 6 ( 6 ) for B = [a,b] given by the Fekete theorem 171, see
Gelfond [8]:
90 G . V. CHUDNOVSKY NUMBER THEORETIC APPLICATIONS Q1

P r o p o s i t i o n 2.1. For b > a, b - a < 1 and E = [a,b] we have using different methods, one of which is based on the transformation of
6(E) 5 d w . cyclotomic polynomials [!I], and another of which is based on rational trans-
formations of an interval into itself [lo].
The first bound for 1(E) for E = [O, 1/41 is from polynomials Pm(x)
The precise value of 6(E) is unknown and it essentially depends on the
arithmctic properties of a and b. It is of utmost interest for us to consider with rational integer coefficients, whose zeros form a complete systern of
corij ugate algebraic numbers:
the case of E = [0, I].

Lemma 2.2 [ 5 ] , [ S ] . We have 6([0,1]) =

for j = 0,. . . ,m - 1 and i = G. The leading coefficient a m of poly-
This lemma is simply established using the quadratic map x --t x (1 - x). nomial Pm(x) E Z[x] that has roots x j can be determined explicitly as
There is a convenient method to improve both the upper and lower
bourids of S(E), using studies of the class 11 of polynornials with rational
a, = nyz: lz?'l. Hence
integer coefficients, all the roots of which are lying in the interval E. In H
we take a subclass HI of polynomials irreducible over Z. Another important
quantity connected with the integer transfinite diameter of E is expressed
in terms of the leading coefficients a ( P ) of polynornials P(x) E I l l . Namely,
we define
1(E) = lim { min ~ a ( P ~ ) l - ~ / ~ } ,
If we denote the root, largest in absolute value, of x2 62 + +
1 = 0,
n--oo P,(Z)EHI -3 - 2 f i by a, then a y m = la1 l(a-(2m+1) +
+ l)/(a-' ~ ) l ' / ~or,
where niin is taken over polynomials Pn(x) E H I of degree of at most n
lim,, a z w ' = la/ = (1+ This gives an immediate lower bound
for 6(E):
and with the leading coefficierit a(P,). Apparently 1(E) is a much better
lower bound for S(E) than that given by an ordinary transfinite diameter 6(E) 2 (1 + &)-I
d(E). This remark is a consequence of the followir~gsimple lernma: for E = [O, 11. This bound is stronger than the one given by the value of the
transfinite diamter of E , and it is this bound that shows the failure of the
Lemma 2.3. Let Pi(%)E Z[x] be a system of polynomials having initial attempt by Gelfond-Schnirelman (because 6 > e-'). This bound
no common zeros such that each polynomial P;(x) has degree n;, all roota was obtained independently in [5] and [!I]. Another transformation, this
in the continuum El and the leading coelficient a;. Then for an arbitrary time rational, leads to a conjectural value of 1(E) and 6(E) for E = [0,1].
polynomial P(x) E Z[x] of degree n, we have The rational transformation for E = [O,l/4] proposed by Aparacio [lo] is
the following one: t -+ t (1 - 4t)/(l - 3t)2. This transformation can be used
m a x l ~ ( x ) l ' />_
~ I& {[a;~-~/"~}. to define, slarting from a given polynomial Q(x) belonging to the class H,
zEE n-oo
another polynomial Ql(x) from the same class H according to the following
rule:
C o r o l l a r y 2.4. For an artibrury continuum E, S(E) 2 1(E).

The lower bound from Corollary 2.4 is in many cases better than that
given by d(E). One may even conjeclure that S(1C) = 1(E) at least for E where deg(Q) = d. In this way one defines polynomials Qn(x) of degrees 2n
being an interval with rational ends, for example for E = [O, 1) or starting from the initial polynomial Qo(z) = 1 - 52. This sequence of
f3 = [O, 1/41. The construction of a polynomial P,(x) satisfing the condi- polynomials can be used via Lemma 2.3 to get a lower bound for 6(E),
tions of Lernnia 2.3 and such that (aillln- is decreasing can be achieved E = [O, 1/41. For this, one has Lo evaluate the lower bound of 1(E) given
92 G.V. CHUDNOVSKY NUMBER THEORETIC APPLICATIONS 93

by a limit of lqn12-" with qn being the leading coeficient of Q n ( x ) . Using Remark 3.1. Let E = where I is an interval. There exists a
an iterative formula for the computations of Q , ( x ) , we obtain a product polynomial P ( x 1 , . . . ,xd) with rational integer coefficients in absolute value
less than 1 in E, if and only if the length of I is less than 4.
expansion [ l o ] for X = limn,,lqn)-2-". This formula, following [ l o ] ,has
the form X = nr=o +
- ~ O0 = 1, dk+' = Ok 6;'. This gives the
0 ~ where
We consider polynomials P ( x 1 , . ..,xd) E Z [ x l , . . . , z d ]having different
following estimate of the constant X :
degrees in different variables in each monomial in P ( x l , . . . ,x d ) . This
approach seems more rewonable since it involves sums of values of the
+function such as S : $(Y) dy, which are typically easier to evaluate than
The constant X-' gives us the lower bound for 1(E) and, hence, for 6 ( E )
$k). - We remind [6] that the prime number theorem is equivalent to
+ ( ~ ) d y x 2 / 2 for large x . The relationship between the values of
with E = [O, 1/41. One may assume that X-' is, in fact, a correct value of the $-function and polynomials least deviating from zero is based on the
b ( E ) for E = [ 0 , 1 / 4 ] . following consequence of the Gelfond-Schnirelman method. Let
The upper bounds on 6 ( E ) that are close to the value X-' are achieved
by considering polynomials of the following form P ( x 1 , . . . ,*d) E XI,. . . ,~ d and
] max
O<Zi<l
XI,.. . ,xd)l < 6.

where a, b are positive and v j are non-zero rational integers. Examples with
J = 0 , 1 , 2 are the following:
where lcm is taken over all d-plets ( i l , . . . , i d ) such that x f ' . . .x y is a
monomial in P ( x l , . . . , x d ) . Formula (3.1) is here the basis for obtain-
ing lower bounds for values of the $-function. Another auxiliary state-
ment deals with the interval [O, 1/41. If P ( x l , . . .,x d ) E Z [ x l , . . . ,x d ] and
max~<z;~l/.ilP(~l,...t~d)l < 1, then <
where Q o ( x ) = 1 - 5 2 ) Q 1 ( x )= 29x2 - l l x + 1. This gives the following up-
< <
per bounds for b ( E ) for E = [O, 1/41; 6 ( E ) 0.2; 6 ( E ) 0.186165509.. .;
where lcm is taken over all d-plets ( j l , . . . ,j d ) such that jk 5 2ik +
6(E) < 0.1840868.. . as compared with the lower bound given by
X-' : S ( E ) 2 0.17701068. Better upper bounds are achieved if J 2 2.
1
for k = 1 , . . . , d and 2 : . . . x y is a monomial in P ( x l , . . . , x d ) . In bounds
(3.1) and (3.2) it is essential that the corresponding i ~ ~ t c g r aare l s non-zero.
Usually it can be achieved by taking an arbitrary polynomial P o ( x l , . . ,xd) .
and taking its powers P F .
$ 3 . The multidimensional generalization of the Gelfond-Schnirelm meth-
The simplest polynomial with integer coefficients least deviating from
od gives the possibility of an alternative analytic proof of the prime num-
zero on I d is suggested by the definition of the transfinite diameter of
ber theorem with the possibility of improving the error term. According
to the method outlined in $1, the main problem is to find polynomials
I : A(x1,. . . ,xd) = ni< j ( x i - xj)'. Here
P ( x l , . . . , x d ) with rational integer coefficients least deviating from zero in
[0, l I d . For one matter, however, this generalization cannot be too much
diITercnt from the one-dimensional case.
94 G. V. CHUDNOVSKY NUMBER THEORETIC APPLICATIONS 95

for an arbitrary N and limd,, vd = length(I)/4(= d(1)). In order to (3.1) as its only arithmetic part. Since we are interested only in the
apply the bound (3.1) for the lower bound of the values of the +-function, asymptotic law of the distribution of primes, instead of estimating the
we express the right-hand side of (3.1) for the polynornial A ( x l , . . . ,xd)N integral
1
So So
. . . 1 P(xl, . . ., xd) dxl.. . dxd for P(xl, . . . ,z d ) E Z [xl, . . . ,xd]
in terms of values of the $-function: least deviating from zero on [0.lId, we consider integrals

loglcm{(il+ I). . .(id + 1)) < 4(2N(d - 1) + I) + +((2d - 3 ) N + 1)

+ +(2N(d - 2) + 1) + ..-+ + ( ( ~ ( -d 1)+ 1).
for a sufficiently large integer N . Hence, according to the Laplace method,
Hence with the help of inequality (3.1) we obtain an obvious, lower bound: the asyrnptotics of this integral is given by the contribution of critical points
of P ( x l , . . . ,xd) in the cube [0, lId. Namely, let zj = ( x i , . . . , x i ) be all
solutions of the equations

with vd -t 4 as d 4 oo. This gives an inflated lower bound for the integral
of the +function: contained in [0, lId.Then the asymptotics of

log 16

is given by a linear combination of P ( F ~ ) ~We. use this approach in the

The quantity log 1 6 / 3 ' = 0.924196241. .: is still below 1 but is better case of d -+ oo when the equations determining the critical points turn into
than the Chebicheff value of 0.921.. . that used ingenious arithmetic con- an integral equation.
siderations.
Obviously the polynomial A(xl, . . . ,xd) is not a polynomial least deviat-
niZl
We start with a polynomial A(xlJ . . . ,xd) d x4(1 - xi)*. From the
symmetry considerations (despite arithmetic objections) it is bktter to con-
ing from zero on I ~ .For 1 = [O, 11 it is more natural to multiply sider the symmetric interval [-I, 11. In this case the asymptotics of the
d
A ( x l , . . . , xd)N by Hi,, xM(1 - xi)M for a certain M > 1. It so happens integral is deter~ninedby the critical points of
that in this particular case there is the famous Selberg explicit expression
for the corresponding integral [12]:

The critical points (maximums) of this expression are determined by Stieltjes'

thcorem [13] as equilibrium configurations of one-dimensional Coulomb par-
ticles with fixed particles a t f l with charges p >
0 and q 0. >
Lemma 3.5. Let p >
0, q 2 0 and let XI,. . . ,xd be numbers from
[-I, 11 such that the function
for Re u > 0, Re y > 0, Re z 2 0. However, since we want to consider
polynomials more complicated than in (3.4), where an explicit expression
in terms of F-factors does not exist, we ignore the explicit formula and
take a purely analytical approach. This approach is based on inequality
96 G . V. CHUDNOVSKY NUMBER THEORETIC APPLICATIONS

attains its maximal value in [-I, lId. Then X I , . . . ,xd are all zeros of the Hence for N 4 oo (sufficiently large with respect to d) we obtain
Jaeobi polynomial Pf")(x) for a = 2p - 1, /3 = 2q - 1. This maximal
value T is:

The maximal value of zI(1.5 + a + 6) for a 2 0 and b > 0 is achieved for

b = 0, and a 0.78, when

We take E = [O, 1/41 (see (3.2)), and in this case the polynomial
P ( x l ) . . . , ~ d ) EZ [ X ~ ~ . , . b
JX~] Hence, in this case the value of "Chebicheff's" constant is 0.990358.. . . It
is, however, unexpected that b = 0 would be an optimal choice in view
of $2, where the corresponding polynomial 1 - 4t for E = [O, 1/41 plays
an important role. This phenomenon can be properly understood only by
analyzing integral equations determining the asymptotics of the integral
(critical point of P(xl,. . . ,xd)) as d -+ oo.
We consider, as above, the polynomial

l xi) E [O, 1/41 with xi E [0, 11.

for yi = ~ ; ( -
The asymptotics for T as d -+ oo is achieved in the case when p and q
grow linearly with d or a = ad, p = bd, for constants a 2 0 and b 2 0
with d -+ oo. From (3.2) we obtain: for Vl > 0, V-1 >0; Vl = vld, V-1 = v-ld for vl 2 0, v-1 > 0. The
system of algebraic equations determining the critical points of

takes the following form

for 2p - 1= ad, 2q - 1 = bd and T,,,,being the maximum given by Lemma
3.5. We apply this formula when d is sufficiently large (d 4 oo), and hence
an asymptotics for A has the following form

- def (1 + a)2 (1 + b)2 i = 1,.. . ,d and xi E [-I, +I]. As d -+ oo we can apply classical methods
A r ~d~ = d2{log(l + a) . - 4 -
+ log(1 + 6) . - 4 -
of statistical mechanics, and for d -+ oo we can consider a distribution
a2 b2 1 function p(a) of xi on [-I, +I]. Classical methods [l4] suggest that x; fill
(34 - - log a - - log b - -(2 + a + b)2 log(2 + a + 6) the subinterval [-a, a] c [-I, I] and that the density p(a) is characterized
4 4 4
as $ ~ f f(xi) = = ~ ,$: f(a)p(a)da + o(1) for d -+ oo and an arbitrary
integrable function f (a). Hence the equations determining the critical
98 G . V. CAUDNOVSKY NUMBER THEORETIC APPLICATIONS

points of Q(xl, . . . ,xd) turn into a singular integral equation: for X E [-a,a] a n d 0 < a < 1,and

for x E [-a, a]. Following the Laplace method, we have An example above (with b = 0) shows that the direct transfer of the
one-dimensional considerations of $2 to multidimensional integrals of $3
is impossible. We present instead a different general scheme for the im-
provement of the "Chebicheff constantJ' and the proof of the prime number
theorem. This scheme is based on the solution of singular integral equa-
tions similar to (3.8) in terms of hyperelliptic integrals. For this we start
with poly~lomialsQl(x), . . . ,Qk(x) with rational integer coefficients having
all their zeros in the interval E = [O, 1/41. Then the d-dimensional integral
that can be used for improvement of the lower bound of the +-function as
in (3.2) is the following one
asd-+mand N + m .
We return now to the solution of the singular linear integral equation
which is a particular case of the Prandtl equation (see [15]). This equation
is of considerable interest for quantum field theory and plasma physics and
the same equation appears in the determination of the asymptotics of Pad6
approximations (161. The solution is given by the following explicit formula: for ti = xi(1 - xi) E [O, 1/41 and xi E [0, I] and vl 2 0,. . . ,vk 2 0. Let
{al,. . .,a,) be a complete set of (distinct) zeros of Ql(t), . . . , Qk(t). As
1 d -r oo, this system of algebraic equations determining the asymptotics
a t-x of (3.12) turns into a single singular integral equation for the distribution
(3.10)
1
function p(a). Let 0 5 a1 < a 2 < . . . < an 5 114. Then there are n 1 +
P= intervals E, = [a,, b,] such that a0 = 0, bo < a l ; an < a,, bn = 114;
a, < a, < .
6, < a,+l (3 = 1, . . ,n). The singular integral equation takes
the following form
where f (x) = - i{vlln--'--
11-21
+
v-llnl)
11+4
+ co, and where the constant co
is determined from the equation S:u
p ( t ) l n l d t = /(x). We present an
It-4
explicit expression for p(x) in the following form:

for z E UF=o [a,,b,], together with the normalization condition

Let us consider a particularly important case when vl = v-1, so that

and the regularity conditions p(a,) = p(b,) = 0, if b, - a, # 0 and w,-1 > 0,

w, > 0.
100 G . V. CHUDNOVSKY NUMBER THEORETIC APPLICATIONS 101

An explicit expression is the following one factorial as a sum of the values of the +function. For an arbitrary x 21
we, following Chebicheff, put T(x) = log{[x]!). Then
(-l)n-#+l
~ ( =4 X
~ 2 m ~ O I-(as)(.
x - b,)l
(3.14)
bn(xr + e=O
~ ( - i ) na, - ~ Jt - b
2 * ~ ~ ] Upper and lower bounds (4.2) for $(x) can be deduced simultaneously from

where
the integral representaion, since (r) 2n -1
and (,) both can be written as
13-function integrals. Namely,

and x E [a,, b,] : s = 0 , . . . , n, and Pn(x) is a polynomial of degree n

(see (171). and

54. Si~nultaneouslywith a lower bound for values of the +function, This example shows how an integral of a polynomial with integer rational
the Gelfond-Schnirelman method gives an upper bound for values of the
+function, and hence for the number of primes. For this we again use coefficients over the interval [-I, 11 can be inverted and represented as an
integrals over the polynomials with integer rational .coefficients, but this integral over a unit circle of a rational function with a singularity inside
time taken over a different path (domain) 'of integration. The essence of a circle. Namely, as in $5 1-2, looking on the term-by-term integration of
the method can be illustated by an example of the integral So
1
xn(l - x ) ~ ~ x ,
+
P ( z ) (1 ~ ) " z - ~with
, d(P) 5 2k - n, we get:
which is an equivalent of P. Erdos' method of obtaining upper and lower
bounds for a(x) used widely in all books on number theory (see [18]). In
this estimate we look a t(2) written as
n
k<pln
p divides
dz

if P(z) E Z [I]. Hence

From this we derive simultaneously upper and lower bounds for values of
the +function:
+
n . log 4 o(n) <
+(2n); for P ( z ) E Z[z], d(P) < 2k - n. In this formula, as above in $2, it is
+(2n) - $(n) 5 n log 4 + o(n). assumed that the integral S P ( z ) ( l + z)n% is non-zero. Since to bound
this integral we are using t k k h t i c a l points method, we can always insure
Thus we obtain the nonvanishing of the integral by considering ~ ( 2 instead ) ~ of P(z), for
a large integer N. We can instead consider integrals

for large n. Identities like (4.1) form a basis for the original Chebicheff-
Sylvester method [19], and they all follow from the representation of a
102 G. V. CHUDNOVSKY NUMBER TIIEORlZTIC APPLICATIONS

for P ( t ) E Z[t], and obtain

En4 z p(n)T(E). Similar to (4.1) we put

TI p 5 max l(t
t[-2,21
+ 2)"I2~(t)l,
k<p<n
2k-n Also we have
P ( t ) E Z[t], d ( P ) 5 7. A similar consideration exists in the multi-
dimensional situation. This way we achieve

(4.3) n
k<pSn
pd < - 2 <max
ti<2 I P ( t ~ , . . . , t d )H ( 2 + t i ) n / 2 1 ,
i=l
Hence, using the identity (4.4) (see the method of [20]), we obtain

where P(t1,. . . ,td) E Z[tl,. . . ,td] and dt,(P) <

k - 5. Hence we have for a large x. This method (due essentially to Landau) clearly shows the
a situation completely parallel to that of (3.1)-(3.2) but now with upper "mock" proof based on a consideration of the following product of rational
bounds for values of the +-function instead of lower bounds. Again, the numbers that are either binomial coeacients or their inverses:
natural choice of the polynomial P ( t l , . . . ,td) is the following one
d
~ ( t , ., . . ,t i ) = fl (ti - 2)W . (ti - tj)lY
i<i
i=l
It remains only to represent each of the binomial coefficients (;!/,$ as
for appropriate w > 0, v > 0. This way we obtain an upper bound one-dimensional integrals over [O,1], if p(n) = -1, or over the unit circle
for +l(x) = $ : +(y)dy of the same degree of accuracy as its lower bound (or [-2,2]) if p(n) = +l.
from $3. When the work on Chapter II was completed, the author found that
It is useful to remark that the methods of both $3 and 54 can be brought the Gelfond-Schnirelman approach (apprently without knowledge of the
t o a uniform form of integration over the same interval (say [0,1]), if one original attempt) was successfully pursued by M. Nair [21], [22]. The results
allows a polynomial with rational coefficients to have denominators which of M. Nair [22] are essentially those of section 3 of Chapter 11.
are powers of 2. In particular, this can provide us with a "mock" proof Also recent work of 11. G. Diamond, K. McCurley (231 and P. Erdos
of the prime number theorein using multiple integrals of polynomials. The revives interest in the Sylvester approach (191 similar to that of $4.
form of a polynomial and the proof itself are, unfortunately, dependent on More work in this direction (approximation by polynomials with integer
an identity which is known to be equivalent to the prime number theorem coefficients, integral estimates and the prime number theorem etc.) has to
itself. This is the identity (see [6]) be mentioned, see D. Cantor [24]. The elementary approach to the prime
number theorem, similar to that of Gelfond-Schnirelmen, was rediscovered
recently by several researchers, cf. the brief discussion of this approach
among others in the new review of 11. Diamond, Bull. Amer. Math. Soc., 7,
(1982), 553 (added in the proof).
where p(n) is the Mobius function (cf. Landau [20] for this and other
methods of derivation of the prime number theorem). We use now the Added in proof. Recently the author, following the discussion of $5,
original Chebicheff method of (4.1), (4.2). Namely, we apply the Mobius Ch. I, proved the following Theorem. Let j(x) be an algebraic func-
inversion formula to the function T(x) = log{[x]!). We thus have +(x) = tion with the Taylor expansion f(x) = C~'onnxn, a, E Q. Then
104 G. V. CHUDNOVSKY NUMBER THEORETIC APPLICATIONS 105

for every 6 > 0 and rational integers a, b with Ibl" > ~ ~ I a 1 ~ ~ ~ , P51 D. V. Chudnovsky, G. V. Chudnovsky, Letters Nuovo Cirnento, 19,
1f ( a / b )- P/Ql > where P and Q are arbitrary rational integers No. 8 (1977), 300-02.
with IQI > QO(e,a, b). Here cl = cl(e) > O and Qo are effective constants. PJ J. Nuttall, ~ifuurcationphenomena in mathematical physics and
related topics, D. Reidel Publishing Company, Doston, 1980, 185-
202.
N. I. Muskhelishvili, Singular integral equations, P. Noordhoff,
References N. V. Croningen, Holland, 1953.
W. J. LeVeque, Topics in number theory, v. 1, Addison-Wesley,
L. V. Ahlfors, Conformal invariants, McGraw-Hill, 1973. 1956.
A. 0. Gelfond, L. G. Schnirelman, Uspekhi Math., Nauk, No. 2, J. J. Sylvester, Messenger. Math. (2)) 21 (1891)) p. 120.
(1936) (Russian). E. Landau, Sitz. Akad. Wissen. Wien, Math-Nat. Klasse, bd. CXVII,
A. 0. Gelfond, Comment "On determining the number of primes Abt. IIa, 1908.
not exceeding the given valuesJ' and "On primesJ' in P . L. Chebicheff, M. Nair, The American Mathematical Monthly, 89, No. 2, (1982),
Complete Works, v. I., Academy, Moscow, 1946, 2nd printing 126-29.
(Russian). M. Nair, J. London Math. Soc. (2), 25 (1982)) 385-91.
Le Baron 0. Ferguson, Approximation by polynomials with integral H. G. Diamond, K. S. McCurley, Lecture Notes Math., v. 899,
coeficients, Mathematical surveys #17, American Mathematical Springer, 1981, 239-53.
Society, Providence, 1980. D. G. Cantor, J. Reine Angew. Math., 316 (1980), 160-207.
R. M. Trigub, Metric questions of the theory of functions and
mappings, No. 2., pp. 267-333, Naukova Dumka Publishing House,
~ i e v 1971
, (Russian). Received August 6, 1082
A. E. Ingham, The distribution of prime numbers, Cambridge Tracts,
#30, cambridge Univ. Press, 1932. Dr. G . V. Chudnovsky
M. Fekete, C. R. Acad. Sci. Paris. 1923, 17. Mathematics Department
A. 0. Gelfond, Uspekhi Math. Nauk. 10 (1955), No. 1, 41-56 Columbia University
(Russian).
New York, New York 10027
E. Aparicio, Revista Matematica Hispano-Americana, 4 Serie, t.
X X X ~ I I NO.
, 6, 259-70 Madrid, 1978.
E. Aparicio, Revista de la Universidad de Santader, Numero 2,
Parte 1 (1979), 289-91.
E. Aparicio, Revista Matematica Hispano-Americana, 4 Serie,
v.XXXVI (1976), 105-24.
A. Selberg, Norsk Mathematisk Tidsskrift, 26 (1944), 71-78.
G. Szego, Orthogonal polynomials, American Mathematical Society
Colloquium Publication #23, Providence, Rhode Island, 1939.
L. HulthCn, Arkiv for Mat. Astron. Fysik, 26A, Hafte 3, No. 11
(1938), 1-106.
 Infinite Descent on Elliptic Curves with
Complex Multiplication
John Coates

To I.R. Shafarevic

1. Introduction

I t is a pleasure t o dedicate this paper t o I. R. Safarevic, in recognition

of his important work on the arithmctic of ellipt,ic curves. As anyone who
has worked on the arithmctic of elliptic curves is acutely aware, it is still
dominated today, despite its long and rich history, by a wealth of tantilining
conjectures, which are convincingly supported by numerical evidence. The
most important amongst these conjectures, a t least from the point of view
of diophantine equations, is the conjecture of Birch and Swirinerton-Dyer,
which grew out of the attempt t o apply to elliptic curves the quantitative
local to global principles employed by Siege1 in his cdebrated work on
quadratic forms. This conjccture is so well known that there is no need to
repeat its precise statement here. However, we do wish to point out t h a t no-
one has yet found a direct and natural link bctwem the existerlce of rational
points of infinite order om an elliptic curve defined over a number field and
the behaviour of its II~sse-WeilLSeries a l the point 8 = 1 in the contplex
plane, as is predicted by ~ h conjecture
e of Birch arid Swinnerton-Dyer.
Guided by Artin and Tate's [15] success with the geometric analogue, most
recent work has attempted t o establish such R corinexion indirectly by p a d i c
techniques which combine the classical infinite dcscent of Mordcll and \Veil
with ideas from Iw=awa's theory of Zp-extensions of number fields. The
firat results in this direction were found by Mazur [lo], who studied descent
theory on abelian varieties over an arbitrary Zp-extension of the base field,
assuming only that the abelian variety has good ordinary reduction a t
every ramified prime for the Zp-ctxtension (we shall describe this by simply
saying that the abeliar~variety is ordinary for the Z,-extension). The
prevent paper, which w ~ ~notivated
? by earlier work of Wiles a ~ l dmyself
[3], pursues the study of this descent theory in a very special caae, which
neverthelrss contains important families olP elliptic curves, whose arithnlctic
108 J. COATES INFINITE DESCENT O N ELLIPTIC CURVES

properties are presumably typical of all elliptic curves. Namely, we consider 1. Descent Theory over the Base Field F
elliptic curves with complex multiplication, arid a certain non-cyclotomic
Zp-extension of the base field, whose existence is closely associated with We begin by recalling that the TatcSafarevic group LUF of E over F is
the hypothesis of complex multiplication, and for which the elliptic curve defined by the exactness of the sequence
is ordinary (see $2 for the precise definition). Much of the material in this
article is not rccent work and was presented in the Hermann Weyl Lectures
a t the Institute for Advanced Study, Princeton, in 1979, and will eventually
form part of a morc detailed set of notes on these lectures. However, the
where the product on the right is taken over all finite places v of F, and Fv
crucial results of $3, which are due to Bernadette Perrin-Riou [12], were
denotes the completion of F a t v . Now take a to be a non-zero element of
only proven after these lectures took place. Finally, there have been two
0 = Endp(E), which is not an automorphism. Taking Galois cohomology
important recent developments on the problems discussed in this paper.
of the exact sequence of G(F/F)-modules
Firstly, P. Schncider (see [13], [l4], and a manuscript in preparation) has
now established the equality of the analytic height and algebraic height for
arbitrary abelian varieties and Zp-extensions of the base field, for which
the abelian variety is ordinary. Secondly, Mazur and Tate (see their article we obtain the exact sequence
in this volume) have found some striking new descriptions of the analytic
height attached to a n elliptic curve over a Zp-extension, subject always to
the hypothesis that the elliptic curve is ordinary for the Zp-extension.
Notation. Let K be an imaginary quadratic field, 0 the ring of integers
of K , and F a finite extension of K. Throughout this paper, E will denote By dclinition, (IIIF), is a subgroup of (H1(F, E)),. The Selrner group sP)
a n elliptic curve defind over F which has complex multiplication by K in of E over F relative to a is defined to be 'the inverse image of (LUF), in
the scnse that its ring of F-e~ldomorphismsis isomorhic to 0 . As usual, we H'(F, E,), under the map a t the right hand end of (2). Thus we have the
fix an isomorphism of 0 with EndF(E) such that the embedding of K in funda~nentalexact sequence
F induced by the action of EndF(E) on the tangent space to E/F a t the
origin is the given embedding of K in F. Note also that the hypotheses
that EndF(E) is the maximal order of K involves no real loss of generality
si11ce every elliptic curve E over F with Q BzEndF(E) isomorphic to K We now introduce two modified forms of the Selmer group, whose defini-
is isogenous over F to one with this property. If N is an extension field of tions are motivated by the Iwasawa theory of 52. Indeed, in both cases,
F, we write E ( N ) for the group of N-rational points on E. If L / N is a we modify the local conditions in precisely the primes of F which are
Galois extension of fields, we write C(Z,/N) for the Galois group of L over ramified in the Zp-extension introduced in 52. We first specify the condi-
N . Letm denote a fixed algebraic closure of N, and A a discrete module tions that will be imposed on the prime number p for the rest of this paper.
Henceforth, we assume that p satisfies:-
for G ( m / N ) . The Galois cohomology groups of G(W/N) acting on A will
be denoted by H i ( N , A); when A = E(N), we simply write H i ( N , E). For
each 0-module D l we put D, for the kernel on D of a non-zero element H y p o t h e s i s on p. (i) p splits in K into two distinct primes, say
a of 0 . If J! is an integral ideal of K , let Dh = naEh D,. As usual, if ( P ) = pp*; (ii) E has good reduction at every prime o/ F above p;
(iii) p # 2.
D = E(F), we simply write E,, and Eb. Finally, if /3 EK, /3* will denote
its conjugate over Q.
As is well known, we note that (i) and (ii) are equivalent to the assertion
that lii has good ordinary reduction a t each prime above p. Also the
110 J. COATES INFINITE D E S C E N T O N ELLIPTIC CURVES

condition p # 2 is not needed in $1, but it is essential for the later parts
of the paper. Write Kp(= Qp) and Op(= Zp) for the completion of K a t
the first being the natural inclusion, and the second being given by multi-
p and its ring of integers, and put Dp = Kp/Op. Now take a # 0 t o be
plication by r*. We now pass to the limit and define
any element of 0 = EndF(E), which is not an automorphism. The first
modification of the Selrner group is t o neglect t o impose local conditions at
the primes above p, as follows. Define LUb by the exactness of the sequence

Here the inductive limits are taken with respect to the maps induced form
where the product is now taken over all finite primes v of F, which do
not divide P. We then define sP) to be the inverse image of (LUb), in
(a), and the projective limits with respect to the maps induced from (9).
We introduce more notation. If X is a finite set, #(X) will denote its
II1(F, E,) under the map on the right of (2). The second modification of
cardinality. Let A be an 0-module. We write
the Selmer group is to impose more stringent local conditions aL the primes
above p. Indeed, if v is any place of F, we have the local analogue of (2),
namely

We also put T,. (A) = A,.-, where the projective limit is taken with
Writing r,,, for the restriction map from EI'(F, E,) to H1(Fv, E,), it is respect to the maps induced by multiplication by T*.
plain from the definition of the Selmer g o u p that the image of SF) under If v is a place of F, write k, for the residue field of v, and Nv for the
cardinality of k,. When l3 has good reduction at v, we denote the reduction
r,,, is contained in the image of E ( F v ) / a E ( F u )under the map on the left
of (5). Hence the restriction map defines an application of E modulo v by E,/k,. Let $F denote the Grossencharacter of E over F
in the sense of Deuring-Weil. Thus $F is a Grossencharacter of F, which is
s,,: s$)- E(Fv)/aE(Fv). unramified precisely a t the places where E has good reduction, and which
has the property that, for each v where E has good reduction, GF(v) is
We now define RF)by the exactness of the sequence the unique element of O = EndF(E), whose reduction modulo v is the
Frobenius endomorphism of $ relative to k,. If a,b are elements of Kp,
we write a -- b if a = ~ bwhere
, E is a unit in Kg; we use the same notation
for elements of K, viewing them in Kp via the canonical inclusion.
where h, is given by the product of the s,,, over all v dividing P.
These modified forms of the Selmer group are only interesting for suitable Lemma 1. For each prime v of F dividing P, we have
choices of the endomorphism a, which we now explain. For the rest of the
paper, we fix an element T of 0 satisfying -
(i) #(Ev(kV)) 1 - $JF(v)/Nv;
(ii) Reduction modulo v induces an isomorphism from

for some integers k 2 1. P u t q = TT* = p k . For each positive integer n,

we have the natural maps
onto the p-primary subgroup of &(kv).
112 J. COATES INFINITE DESCENT ON ELLIPTIC CURVES

Proof. The definition of $JF implies easily that If v is any element of Ex.-, we define

By the theory of complex multiplication, p divides $JF(v) because v divides Plainly w,(u, v ) belongs to the group pqn of qn-th roots of unity, and it is
P, and + F ( ~ ) + F ( ~ )=
* Nu. Hence the right-hand side of the displayed well known that the map
~ , to multiplication by a unit in Kp.
expression is equal to 1 - + b . ( ~ ) / Nup
This proves (i). Let E1,,(FV) denote the group of Fv-rational points in the
kernel of reduction modulo v. Since v lies above p, (7) implies that ?ran
is an automorphism of E1,,(FV), and so reduction modulo v induces an given by pn(v)(u) = w,(u, v ) is an isomorphism of G(P/F)-modules. More-
isomorphism >
over, it is easy t o verify that, for all n l, we have
E(F,)I~*"E(F,) =
2v(kv)/~*nR,(kv).
Assertion (ii) of the lemma now follows, since one sees easily that the
pprimary subgroup of Ev(kv)is killed by T * " for all sufficiently large n. We now begin the proof proper of Proposition 2. Noting that the image
Putting a = A*" in (6)) and passing to the projective limit, we conclude on the right hand side of (1) actually lies in the direct sum rather than the
from (ii) of Lemma 1 that we have an exact sequence product, we define Bn by the exactness of the sequence

where v runs over all finite places of F. Simple diagram chasing shows that
there exits a map
P r o p o s i t i o n 2. We have
gn: H ( H 1 ( ~ vEJ ) ) ~ . -+Bn
v IP
such that the sequence

Proof. We first recall the definition of the Weil pairing, suitably nor-
malized for the proof of Proposition 2. If u Ern, let 6 be the unique
element of Ern satisfying n * " = ~ u. Write 7; = F(E,n). There exist is exact. We can now pass to the inductive limiL in the obvious sense. It
rational functions f,,,, g,,, on E, which are defined over 3;) and whose follows that the index of IUF(p) in I&@) is the order of the kernel of g,
divisors are given respectively by, in an obvious notation, where g = lim g, is the induced map between
-+ nvlP
H1(Fv,E)(p) and
B = 1 5 B,. To calculate the order of this kernel, we dualize the map g,.
@ general, if A is a discrete, we write A for the Pontrjagin dual of A, and
f for the dual of a homomorphism between two such groups. By Tate local
where 0 denotes the zero point of E. Since f,,, o ?ran and gz:u
have the duality, the dual of the discrete group H1(FvJE) is E(Fv), and this in turn
ind;ces a duality between (H'(F,, E ) ) ~ and
, E(F~)/X*"E(F,). Moreover,
same divisors, we can multiply f,,, by a non-zero element of 3; so that
a theorem of Cassels ([I], p. 198) shows that the dual of B, is canoni-
cally isomorphic to Further, an analysis of the proof in [I] shows
114 J. COATES INFINITE DESCENT ON ELLIPTIC CURVES 115

that : - (i). When E,.n is identified with Hom(E+ , p q n ) via the map p, It follows from (16) and (19) that there is a natural injection O F <-+ T,.(UF),
above, then 5, is precisely the map h,.. of (6); (ii). By virtue of (Id), the and the following lemma is then clear.
dual of the natural map from Bn to B,+l is the map from SF' to
Lemma 3.
' ~ ) by multiplication by n*. Hence ij = l i p Gn is none other
s ( ~induced
than the map on the right of the exact sequence (12). Thus

Eznmple 2. Take F = K = Q(G), and E the elliptic curve defined

over K by the equation y2 = x3 - 2. The prime p = 7 satisfies the
This completes the proof of Proposition 2. conditions imposed on p. A simple calculation shows t h a t #(kp(kp)) = 7.
Moreover, E ( K ) contains the point P = (3,5), which is certaidy not in
E 1 ( K ) because its coordinates are integral. We conclude from the above
To end this section, we analyse the natural exact sequences associated
index calculations that, in this case, we have
with the groups SF and BF,which are defined via passage to the projective I
limit. Putting a = x*" in (3)) and passing to the projective limit, we obtain
the exact sequence [E(f") 30 op- : E I ( F ) 80
O p e ] = 7,

Define the subgroup E l ( F ) of E ( F ) via the exactness of

2. Introduction of Iwasawa Theory

where the map on the right is the product of the reduction homomorphisms
We now introduce a certain non-cyclotomic Zp-extension of the base field
for all v dividing P. Taking the tensor product over 0 with Op., and recalling
F, which will be the basic tool used in the rest of the paper to study the
that we can identify &(k,) 30 9.with the pprimary subgroup of &,(k,), arithmetic of E over F. The most natural way to define this Zp-extension
we obtain the exact sequence is via points of finite order on E. Let Ep, = U n L 1E p , arid define Fa =
F(Epm). The action of the Galois group of 3, over F on Ell- defines a
canonical injection

It follows from (12), (16) and this last sequence that

whose image is of finite index in Z F . Thc decomposition

where the two groups on the left are regarded as subgroups of SF.Define
eF via the exact sequence
(in general, if m is an integer > 1, p, will denote the group of m-th roots
of unity) gives rise via X , to a corresponding dec~mposit~ior~ of the Galois
116 J. COATES INFINITE DESCENT ON ELLIPTIC CURVES 117

group G(3,lF) = A X I', where x,(A) C pp-1 and xoo(T')C (1 p Zp). + We omit the proof of this lemma, which is based on a slight variant
Thus the fixed field of A is a Zp-extension F, of F, whose Galois group of the criterion of Nkron-Ogg-Safarevic (see [3], or [2] for a more detailed
can be identified with I'. It is obvious that the fixed Geld of I' is the field proof).
3 = F(EP). For siniplicity, we write x (respectively, K ) for the restriction We now begin the study of the arithmetic of E over F,. Recall that
of X, to A (respectively, I'). Ep C E ( F ) if and only if EpwC E(F,). Also, as always, I' = G(F,/F).
By class field theory, there is a unique Zp-extension of the imaginary
quadratic field K , which is unramified outside y. We denote this Zp- L e m m a 6. Assume that Ep C E(F). Then, for all i > 1, we have
extension by K,. The classical theory of complex multiplication shows
immediately that the Geld F, defined in the previous paragraph is in fact
the cornpositum of F and K,. It follows that F,/F is unramified outside
the set of primes above p, and that each time above p is ramified in F,/F.
Indeed, both assertions are plainly true for K,/K, and so they remain Proof. We only sketch a proof of this well-known fact. Putting
true for the translated Zp-extension Fw/F. G, = G(F(E~,)/F), we have
We next give two important technical lemmas. If v is a prime of F,, we
define F,,, to be the union of the completions a t v of all finite extensions -+ H~(G,,E~,).
Hi(I',Epm)= lim
of Q contained in F,. Note that F,,, will not, in general, be complete.
Now Ep, is a cyclic group of order pn with p # 2, and, as G, is a group
L e m m a 4. For each prime v of F, lying above p, the group of automorphisms of Ep,, it follows that G, is also cyclic. A simple direct
calculation then shows that Hi(Gn, EP,) = 0 for all i > 1.

is finite. Proposition 7. The group E(F,), modulo its torsion subgroup, is a

free abelian group.
Proof. We assume that Ep C E ( F ) , so that F, = F(Epm), since
it certainly sufiices to prove the lemma in this case. As F,/F is a Zp- Proof. We use an entirely analogous argument to Iwasawa ([9], p. 270))
extension, the fact that v is ramified in F,/F implies that v is totally who first proved the cyclotomic analogue. Note also that the proposition
ramified in F,/F(Ep,) for all suficiently large n. IIence the residue field holds in greater generality ([Ill, p. 404). We can suppose that Ep C E(F),
k,, of v is a finite field. Recall t.hat, by our hypothesis on p, E has good since a subgroup of a free abelian group is itself free. Let r, = P",
reduction at v. Writing E, for the reduction of E modulo v, it follows that let F, be the fixed field of r,, Wn the torsion subgroup of E(F,), and
&(k,,,,) is a finite abelian group. Lht, as v does not divide p*, it is well V, = E(F,)/W,,. Define W, V, in a similar manner for F,. Taking
known that reduction moduio v induces an injection of Ep.- n E(F,,,) r,-coho~nology of the exact sequence
into &,(k,,,), thereby proving the lemma.

Remark. Lemma 4, together with the Weil pairing (13)) shows also that we obtain the exact sequence
pp- n F Z , , is a finite gronp for all v dividing 9.

L e m m a 5. Under our hypotheses on p (i.e., p splits in K , E has

good reduction at every prime above p, and p # 2), E has good reduction Now Lemmas 4 and 6 show that Idl(I',, W,) is a finite group. Since the
everywhere over the Jicld 3 = F(Ep). Mordell-Weil theorem shows that Vn is finitely generated, it follows that
118 J. COATES
I INFINITE DESCENT ON ELLIPTIC CURVES 119

~2 is a finitely generated free abelian group for all n 2 0. If m 1 n, Proof. We first remark that it suffices to prove the result when
a simple argument (see [ll],p. 403) proves that vLm/ ~ r , is " torsion free. Ep C E(F). Indeed, if this is not the case, the restriction maps induce
Hence, if m >_ n , Vr," is in fact a direct summand of vLrn , and the assertion isomorphisms
of the proposition follows easily. sl,
7 (s;)A, Skm (S$JA,

We can now repeat the definitions of the Tate-Safarevic and Selmer because the order of A is prime to p, and so, for any A-module A, H ~ ( AA)
,
groups given in $1,simply replacing F in the earlier definitions by the field has no ptorsion for all i 2 1.
F,. In particular, we now write We now assume that Ep C E ( F ) . It follows from Lemma 6, and the
usual restriction-inflation exact sequence, that the restriction map yields

In fact, unlike the situation over the base field F, it is immaterial over F, P3) H'(F, E ~ 7~H ~) ( F , , E ~ ~ ) ~ .
whether we impose local conditions a t the primes above P.
Let V be any place of Fw, and w the restriction of v to F. We plainly have
a commutative diagram
Lemma 8. We have LUFm (P) = UIkm(p), and SF- = S IF,.

Proof. It suffices to establish the first assertion, since it clearly implies

the second. We must show that (H'(F,,, ,E ) ) ~= 0 for all primes v of
F, above p. As before, let F, be the f i e d field of I',. Since, by definition,
>
F,,, is the union of the F,,, for all n 0, we see easily that where the vertical maps are the restriction maps, and where the horizon-
tal maps are the appropriate localisation homomorphisms (e.g., Xw is the
restriction homomorphism t o H1(Fw,Ep,), followed by the canonical map
where the inductive limit is taken relative to the restriction maps. By from this group to H1(F,,, E)(p)). Since
Tate local duality, (11 '(F,,, ,E ) ) ~
is dual to D, = E(F,,,)/.rr* E(F,,,),
and the restriction maps are dual to the norm maps. Now, as remarked
S: = n Ker Xw, Sam = Ker A,,,.
in the-proof of Le_mma 1, reduction modulo v defines an isomorphism w h vfl
Dn 7 E,(k,,,)/d E,(k,,,), where k,,, is the residue field of the restriction
it is clear from this diagram that i defines an injection of S$ into (SIF,)r.
of v to F,. As v is totally ramified in F,/F, for all large n, we conclude
that, for all large n, the order of D, is fmed and ,'I operates trivially on To prove that this map is surjective, it plainly suffices, in view of (23), to
D,. Hence the projective limit of the D, with respect to the norm maps is show that i , is injectivc for all places v of F- not lying above p. Let v
0, and the proof of thc lemma is complcte. be such a place of F,. Sincc we have a s u m e d Ep C E ( F ) , Lemma 5
We next show how onc can recover S k from S k = SF-. Since taking shows that E has good reduction a t v . Now the kernel of i , is equal to
cohomology comnrutcs with direct limits, we have H1(G(Fw,,/Fw), E(F,,,))(~), and this cohomology group is zero when v
docs not divide P, because F,,,/F, is unramified and E has good reduction
a t w (see Corollary 4.4 on p. 204 of [lo]). This completes the proof of the
theorem.
T h e o r e m 9. The restriction map on cohomology induces an isomor-
phwm C o r o l l a r y 10. E(F,) is a torsion group if and only if both E ( F ) is
S$ 5 (s>J. torsion and the restriction map from Ul>(p) to I U h ( p ) is injective.
 J. COATES INFINITE DESCENT ON ELLIPTIC CURVES 121
120

of the modified Tate-Safarevic groups, the standard restriction-inflation

Proof. Passing to the inductive limit, relative to the maps induced from
exact sequence, and the fact (see the proof of Theorem 9) that, for each
(a), in the standard exact sequence place v of F, not lying above p, we have H1(G(F,,,/Fv), E(F,,~)) = 0.

We now give an alternative description of the Selmer group SF_ = S I F , ,

which is often more useful. As before, let 3 , = F(Ep,), and write
we obtain the exact sequence
for the maximal abelian pextension of 3, which is unramified outside the
primes above p. Write

where, as before Dp = KP/op. The same sequence is valid if we replace F

r
by F,, and taking invariants under = G(F,/F) gives the exact sequence
Since fl, is plainly Galois over F, the Galois group G(3,lF) operates on
X, via inner automorphisms in the standard manner (i.e. if a E G(3,/F),
pick 7 E G(fl,/F) , is a, and define c . x = r x 7 - l
whose restriction to 3
for x E X,). Recall that G(3,lF) = A X r, and that x denotes the
Letting p~ denote the restriction map from III>(p) to III'F_(p), we conclude restriction of X, to A. As the order of A, which we denote by dl is prime
from Theorem 9 that we have an exact sequence to p, each Z,[A]-modulo A decomposes as a direct sum

0 -+ E ( F ) @o Dp -r (E(F,) 80~ p ) ' -+ Ker p~ -+ 0.

Now if B is any discrete pprimary I'-module, it is a standard elementary

fact that B = O if and only if B~ = 0. Hence Corollary 10 follows
From the above exact sequence, on recalling that E ( F ) D p (respectively, where A(x') denotes the subspace of A on which A acts via Finally, if A,
B are C(3,lF)-modules, we always endow IJom(A, B) with the G(3,lF)-
E(li,) BoDp) is zero if and only if E ( F ) (respectively, E(Fw)) is a torsion
structure given by ( a f ) ( a ) = a f (a-'a) for a E G(3,lF).
group.
T h e o r e m 12. There are canonical G(F,/F)-isomorphisms
We note the following description of the kernel of the restriction map

L e m m a 11. We have a canonical isomorphism Proof. Note that the second iomorphism is obvious. Also, by an en-
tirely similar reasoning to that given in the proof of Theorem 9, we can
assume that Ep C E(F), so that Ep, is rational over F,. P u t V, =
Hom(X,, Ep,). Hence both Vm and SF, are subgroups of

Proof. Since the extension F ( E p ) / F has degree prime to p, one sees

easily that it sulliccs to prove the lemma when Ep C E(F). Assuming this
We first show that V, C SL. Since Ep C E(F), Lemma 5 shows
to be the case, Lemma 5 shows that E has good reduction everywhere over
that E has good reduction everywhere over F. Let v be any finite place
I!'. The assertion of the lemma now follows i~nmediatelyfrom the definition
122 J. COATES INFINITE DESCENT ON ELLIPTIC CURVES

of F,, let H,,, denote the maximal unrarnified extension of F,,,, and let in f
l,, a simple calculation with commutators shows that
D, = G(H,,,/F,,,). Since E has good reduction a t v, we have

+
where w, = (1 T)P" - 1. On the other hand, global class field theory
(see p. 204 of [lo]). Suppose now that x E V,. By the definition of X,, for gives the following explicit description of G(fln/3,) as a G(3,lF)-module.
all places v of F, not lying above P, the restriction of x to the deconlposition If v is a place of 7, above p, let U,,, denote the local units 1 mod v of
group of v factors through the Galois group D,. It then follows easily from the completion of 3, a t v. Write &, for the group of global units of 7,)
(26) that the image of x in H1(FW,,, E ) is necessarily 0 for all v which do which are r 1 mod v for each place v of 7, above p. Let in be the diagonal
not divide P, whence x E Sjr,.
To prove that SF, C Voo,let n be an integer > 1, and take
embedding of &, in nvlP U,,,. Then Artin's reciprocity law gives the exact
sequence of G(?, IF)-modules
x, E s$,") C HO~(G(F,/F,), E , ~ ) By the definition of the Selmer
group, for each finite place v of F,, the restriction x,,, of x, to the
decomposition group of v is contained in the image of E(F,,,)/anE(Fm,,)
in 11om(G(F,,,/l;bo,,), kn). Suppose now that v does not lie above P.
Since E has good reduction a t v and (v, A) = I , Ern is mapped injectively
under reduction modulo v, whence it follows easily that, if x,,, = where A, denotes the p-primary subgroup of the ideal class group of 3,, and
I',,, modulo snE(FW,,), the extension F,,,(Q,,,)/ Fa,, is unramified, where the bar over in(&,) signifies its closure in the p-adic topology. Taking
where Q,,, satisfies .~r~&,,, = P,,,. In other words, x,,, must factor n = 0 in particular, we conclude that X,/TX, is a finitely generated
through the Galois group D, of the maximal unramified extension of Zp-modulc, whence a standard argument shows that X, is a finitely gener-
FW,,. Hence the global element x, factors through Hom(X,, Ern). Thus ated A-module. This completes the proof of the lemma.
SF,"' C Hom(X,, Ern), and, passing to the direct limit, we obtain
Remark. It follows easily from Theorem 9, Theorem 12, and Lemma 13
SF- C V,. This completes the proof of Theorem 12.
that the Pontrjagin dual of is a finitely generated Zp-module. In
fact, this is well known to be true for arbitrary abelian varieties defined
We can now use global class field theory to study the G(3,lF)-module
over a number field.
X,. Let A = Zp[[T]]be the ring of formal series in an indeterminate T
We write r, for the Zp-rank of in(&,) modulo torsion, so that
with coefficients in Zp. As usual in the theory of Zp-extensions (see [9]), it is
best to interpret the r-action on the compact Zp-module X, as a A-action,
in the following manner. Fix once and for all a topological generator 7, of
,'I and then defne T.x = (-yo- 1)x for x in X,. This extends to an action
of Zp[T] on X, in the obvious fashion, and finally to an action of A by by Dirichlet's theorem. Let 6, = [Y, : K ]- 1 - r,.
continuity.
Lemma 14. (i) X, is A-torsion if and only if 6, is bounded as
Lemma 13. The Galois group X, is a finitely generated A-module. n -+m; (ii) If 6, = 0, then X , is A-torsion.

Proof. The argument being standard (see [9], p. 255), we only give it Proof. Let p, be the Zp-rank of X,/w,X, modulo torsion. It follows
very briefly. Write 3, for the n-th layer of the Zp-extension 3,/?, and easily from the structure theory of finitely generated A-modules that X,
fl, for the maximal abelian p-extension of 3,, which is unranlified outside is A-torsion if and only if p, is bounded as n -+ oo. Assertion (i) is now
P. Since an is also plainly the maximal abclian extension of 3, contained clear, because (27) and (28) show that p, = 6, for all n >
0. It is also
 J. COATES INFINITE DESCENT ON ELIJPTIC CURVES 125
124

to show that B is a finitely generated op-module. Recall that, if N is a

plain from the structure theory that X , must be A-torsion if X,/TX, is
finite (i.e., 6, = 0). This gives (ii), and completes the proof of the lemma. discrete abelian group, &' denotes the Pontrjagin dual of N. By definition,
the Pontrjagin dual of E(F,) Dp is equal t o
In view of Lemma 14, it is natural to introduce the following weak form of
Leopoldt's conjecture on the p-adic independence of global units in number
fields.
Dualizing the canonical inclusion E(F,)@O~pr~F_,we obtain a surjection
C o n j e c t u r e A (Weak p-adic Leopoldt conjecture for 3,). 6, w
bounded as n -+ a.
Now, assuming Conjecture A, X , is A-torsion, and so Theorem 12 shows
We remark that Iwasawa [9] has unconditionally proven the analogue of that SF- is A-torsion. It ?hen follows from the structure theory for Noether-
Conjecture A for the cyclotomic Zp-extension of an arbitrary totally real ian torsion A-modules that the Zp-rank of SF, modulo torsion is finite,
number field (even though Leopoldt's conjecture remains unproven in this whence also Homo (B,op) has finite Zp-rank. Since H is a free abelian
case). While wc have not succeeded in establishing Conjecture A in general, V
group, we deduce that B itself must be a finitely generated op-module, as
we emphasize that it is valid for a fairly wide class of elliptic curves with required. This completes the proof of Theorem 16.
complex multiplication.

P r o p o s i t i o n 15. Conjecture A is valid if F ( E p ) is abelian over K.

In particular, Conjecture A holds when F(Et,,) is abelian over K , where 3. Canonical p-adic Height and its Connexion with Kummer
Et,, denotes the torsion subgroup of E(F).. Theory

For example, it follows from Proposition 15 that Conjecture A is valid The remarkable result given in this section (Theorem 17) is due to
if the base field F coincides with the field of complex multiplication K . Bernadette Perrin-Riou [12]. Using quite different methods, P. Schneider
Although we do not give details here, two proofs of Proposition 15 are has subscqucntly established an analogous result for arbitrary abelian vari-
known. Firstly, when F(Ep) is abelian over IC, the p a d i c analogue of eties and for Zp-extensioils of the base field for which the abelian variety
Baker's theorem on linear forms in the padic logarith~nsof algebraic num- is ordinary. We simply state the result, and refer the reader to [12] for a
bers can be used to prove that S, = 0 for all n 2 0. Secondly, when detailed account of the proof.
F(Eto,) is abelian over K, a generalization of the techniques of [4] can also We begin by rapidIy describing the p a d i c analogue of the Nbron-Tate
be used to prove Conjecture A (but not the stronger assertion that 6, = 0 canonical height, which is attached to E and our Zp-extension F,/F. Take
for all n> 0). a generalized Weierstrass model for E' over F
We now derive on important consequence of Conjecture A.

T h e o r e m 16. Conjecture A implies that E(F,) modulo its torsion If v is a finite place of F , we write ord, for the order valuation of v ,
subgroup is a finitely generated abelian group.
normalized so that ord,(F,Z) = Z. We always suppose that E(F,) is
endowed with its natural u-adic topology. It is a well known and elementary
Proof. Let 11 denote E(F,) modulo its torsion subgroup, and put fact, due to Nkron and Tate, that there exists a unique function
B -=I I @ o ~ p Since
. Proposition 7 shows that H is a free abelian group, one
sees easily that, in order to prove that H is finitely generated, it suffices
126 J. COATES INFINITE DESCENT ON ELLIPTIC CURVES 127

which is continuous when Q is endowed with the real topology, and which the subgroup of finite index in E ( F ) consisting of those points which belong
satisfies the following two properties :- (i) the limit as P tends to 0 of to the kernel of reduction modulo v for each place v of F dividing P. We
the expression Xv(P) - o r d , ( t ( ~ ) ) exists for every local parameter t at 0; define
(ii) for all points P # 0, Q # 0 in E(Fv) with P Q # 0, we have hp : El ( F ) -+K p
by setting hp(0) = 0, and otherwise

where A denotes the discriminant of our Weierstrass equation. Next sup-

pose that v is a place of F dividing P, and let E1,,(FV) denote the kernel
of reduction modulo v on E. As always, Kp will denote the completion of This definition does not depend on the choice of the T,, and one sees easily
K a t P, and we normalize the p a d i c logarithm by setting logp(p) = 0. It that hp is a quadratic funct.ion on EI(F),which also satisfies
is shown in [12] that there exists a unique function
(30) h p ( a P ) = (deg a)hp(P) for all a E 0.

Plainly hp has a unique extension to a quadratic function on E(F), which

which is continuous when Kg is endowed with the p-adic topology, and we also denote by hp. It is clear that (30) remains valid for this extension.
which satisfies the following two properties:- (i) for all points P # 0, Finally, it is an exercise in linear algebra to prwe that there is a unique
Q # 0 in E1,,(Fv) with P f Q # 0, we have bilinear form

such that hp(P) = (P,P)p for all P E E ( F ) , and whose behaviour with
where N F v l K denotes the local norm from Fv to Kg; (ii) for all a
P
# 0 in 0 respect to endomorphisms a E 0 is given by
and all P # 0 in E1,,(FV) with aP # 0, we have

This biliner form (, )p is our desired analogue of the Nkron-Tate height. In

where degn denotes the number of elements in the kernel of a, and where particular, it can be used to define the p-adic volume VoIp(E(J')) of E ( F )
by

w z i ,~ i ) ~
Vol F
P( ~ ( ~ =1 [E(F)
) : ~ 2or] ,
As in the case of the Ndron-Tate height, we c.an combine these local factors
to define a global height height function. Since the theory of complex where n~ denotes the &rank of E ( F ) , and 21,. . . , z,, is any maximal
multiplication shows that F necessarily contains the EIilbert class field of &linearly independent set in E ( F ) . Although we do not go into dctails here,
K, it follows that the norm from F to K of each finite place v of F is it is theoretically quite straightforward to calculate Volp(E(F)) oncc the
necessarily principal, say N F I K v = (7,)) where the generator rv is well Mordell-Weil group itself is known. The calculations made so far support
defined up to multiplication by a root of unity in K . Recall that E 1 ( F )is the conjecture that voly( E ( F ) ) is always non-zero.
128 J. COATES INFINITE DESCENT ON ELLIPTIC CURVES 129

We now explain the connection between this bilinear form ( , )p, and the Granted Theorem 17, it is an easy exercise to establish the following
theory of infinite descent, which was discovered by Bernadette Perrin-Riou result. Recall that n F dentes the 0-rank of E(F) modulo torsion.
(121. Let
iF: E1(F) 0p. c--+BF = 1im +
RF*") L e m m a 18. # 0;
(i) Coker OF is finite if and only if VO~~(E(F))
(ii) If Coker OF is finite, then
denote the natural inclusion in the exact sequence (20). On the other
hand, the composition of the natural inclusion of E ( F ) QPo Dp in Sk with
the injection of Sk into SF, given by restriction (see Theorem 9) gives
Finally, we shall also need the following important result, which is proven
a canonical injection of E ( F ) Dp in SF,. Dualizing this map, we get
in [12].
(cf. (29)) a canonical surjection
T h e o r e m 19 (B. Perrin-Riou [12]). If Conjecture A is valid, then the
map 4F of Theorem 17 i s an isomorphism.

Finally, taking the map induced between the I'-invariants, we obtain a

canonical map
4. Analogy with the Birch-Swinnerton-Dyer Conjecture

The ideas underlying this section were first exploited by Artin and Tate
which is no longer, in general, surjective. If H is an extension field of F, we [15] in their work on the geometric analogue. It seems fair to say that
write E ( H ) ( p ) for the subgroup of the torsion subgroup of E ( H ) which is they lie a t the heart of the study of the arithmetic of elliptic curves by the
annihilated by all sufficiently large powers of 7 ~ .Define mp by the formula techniques of Iwasawa theory and padic analysis. We go a little further
than previous work, in that we have avoided imposing the hypothesis that
Ill&) is finite.
If G is a p-primary abelian group, Gdiv will denote the maximal divisible
subgroup of G.
where, as always, 3 = F(Eg).
L e m m a 20. (i) N ~ ( p ) d i v= m>(p)div; (ii) #(m',(~)/m&(~)div)=
T h e o r e m 17 (R. Perrirl-Riou 1121). There exists a canonical map #(mF(Y)/~F(~)div).[UI>l(P)
: UF(P)]; (iii) If WF denotes the maximal divis-
ible subgroup of S&, we have an exact sequence

(36) 0 -+ E ( F ) @loDp -+ WF + i v 0.
m ~ ( ~ ) d+
with the following property: if we define the map
Proof. (i) and (ii) follow from the fact that LII>(y)/IIIF(~)is finite by
Proposition 2. Assertion (iii) is obtained by taking divisible subgroups in
the exact sequence (24), and nothing that E ( F ) B0,!Ip is divisible.
by OF = jF o 4.v o iF, then, for all x E E ( F ) and y E EI(F), we have
As before, let A = Zp[[T]]be the ring of formal power series in T with
coeficieritv in Zp. If A is a finitely generated A-torsion A-module, the
130 J. COATES
INFINITE DESCENT ON ELLIPTIC CURVES 131

structure theory shows that there is a A-homomorphism, with finite kernel We now study the behaviour of the characteristic power series of SFoo
and cokcrnel, from A to a A-module of the form a t the point T = 0. Recall that nF denotes the &rank of E ( F ) modulo
torsion. We define tF to be the Zp-rank of the Pontrjagin dual of IIIF(p)di,
(thus I.UF(Y)is finite if and only if tF = 0). P u t

where T is an integer > 1, and f l , . . . ,f, are non-zero elements of A. The

characteristic power series of A is defined to be fA(T) = f l . .. f,. It is
uniquely determined up to multiplication by a unit in A. We omit the It fol1ow;l from the exact sequence (36) that WF, the dual of the divisible
proof of the following elementary lemma on tlie characteristic power series subgroup of S h , is a free Zp-module of rank TF. On composing the inclusion
of A-modules. We write (A)r = AITA. of WF in S > with the restriction map from S& to SF, (see Theorem 9)) we
obtain a canonical injection of WF in SF,. Dualizing this map, we obtain
Lemma 21. (i) If 0 -t A --+ B -, C --+ 0 is an exact sequence of a canonical surjection of SF, onto wF. We then define the A-module Y by
finitely generated A-torsion A-modules, then fs(T) = fA(T).fc(T), up to the exactness of the sequence
multiplication by a unit in A; (ii) If A is a finitely generated A-torsion A-
module, then fA(0) # 0 if and ony if (A)r is finite, or equivalently Ar is
finite. M e n fA(0) # 0, we have Let f (7') denote the characteristic power series of SF,, and g(T) the charac-
teristic power series of Y. We claim that, after multiplying f ( T ) by a unit
if necessary, we have

Throughout the rest of the paper, we make the following assumption:- f (T) = TTFg(T).
(40)

Assumption. Conjecture A is valid ( i e . the weak p-adic Leopoldt con- This is clear from the exact sequence (39) and (i) of Lemma 21, because the
jecture holds for 7
, = F(Ep,)). characteristic power series of WF is plainly TTF.
Taking I'-invariants of (39)) we obtain the long exact sequence
We now collect together in a single proposition the consequences of this
assumption, which will be used later.

We can simplify this sequence in two ways. Firstly, recalling that Theorem
Proposition 22. Conjecture A implies that (i) SF, is A-torsion,
9 shows that ( 3 ~is canonically
~ ) ~ isomorphic to S>, and nothing tlie exact
(ii) SF, has no non-zero finite A-submodulea, (iii) the map qbFfiheorem
sequence
17 induces an isomorphism qbF: BF 'i
mi, have the same rank as Zp-modules.
and (iv) I.UF(p)di, and
o-+N~-+s',--+wF-+o,

whcre NF is the dual of LII~7(~)/III~-(p)div, we see that we can replace the

Proof. (i) is clear from Lemma 14 and Theorem 12. (ii) follows from two terms on the extreme right, of (41) by the finite group NF. Secondly,
Theorem 12 and R. Greenberg's [6] result that Conjecture A implies that assertion (iii) of Proposition 22 shows that we can identify SF,)^ with BF
..
X , has no non-zero finite A-submodule. (iii) was already noted in Theorem via the map qbF. Alter these modifications, we obtain the exact sequence
19, and is due to B. Perrin-Riou [12]. To prove (iv), we note that, as SF, is
A-torsion, (Sr.,)r and (SF,)r have the same rank as Zp-modules. Assertion
(iv) now follows easily from (iii), the exact sequences (20) and (24)) and
Theorem 9. This completes the proof of Proposition 22. whcre n p denotes the canonical map indicated.
132 J. COATES INFINITE DESCENT ON ELLIPTIC CURVES

L e m m a 23. The following three assertions are equivalent:-(i) g(0) # 0, wF/vF.For this reason, we define the subgroup CFof eFby
(ii) Coker aF is finite, and (iii) a~ is injective. When these hold, we have

Hence the restriction of PF to CFdefines a canonical homorphism

Proof. By lemma 20, g(0) .# 0 if and only if (Y)r is finite, and (42)
shows that the finiteness of (Y)r is equivalent to the finiteness of Coker a ~ ,
because NF is finite. Also g(0) f 0 if and only if yr is finite. I3ut yr is with Ker 7F = Ker PF. Also, we clearly have a commutative diagram
finite if and only if yr = 0 because Y has no finite non-zero I'-submodule.
The rest of Lemma 23 is clcar from Lemma 20 and the exact sequence (42).

We assume from now on that VOI~(E(F))f 0, i.e. that the determinant

of the biliner from ( , !) is not 0. Recall that we have the exact sequence
where the vertical map on the right is injective. Since Coker Or;. is finite,
it follows that CF is also of finite index in T,=(ILIF), that CokerrF is a
subgroup of finite index in Coker PF, and that
where OF is a submodule of finite index in T,.(UF). Also dualizing the
exact sequcnce (39), we obtain the exact sequence

Finally, we note that to give the canonical'map 7F is equivalent to giving

a canonical pairing
F ) o~.),SO that VF c
P u t VF = C Y ~ ( E ~ (QPO wF.
Our hypothesis that
VOI~(E(P'))f 0 implies that the map OF of Theorem 17 is injective. We
conclude both that a F is injective on E 1 ( F ) Op., and that
where {a, bIp = (7~(a))(b).We want t o stress that this canonical pairing
(, Ip,whose existence we have just proven under the two relatively mild
assumptions (which can be verified for many E and many primes p) that
IIence we have an exact sequence Conjecture A is valid and volp(E(F)) # 0, is very mysterious. At present,
n
0 -+ IIIF(p)di, -+ WF/VF -' Coker OF 0,
no example is known in which XI, and are not both 0. On the
other hand, there is little hard theoretical evidence to indicate that this is
(44) -+

always true. We attach to { , Ip the invariant

where C o k e d F is finite because V O ~ ~ ( E ( Ff) ) 0, and the order of Coker OF
is given by Lemma 18. It is plain from these remarks that a p induces a
homomorphism

i t being understood that pp(UIF) is infinite if Coker 7F is infinite. One can

immediately pose the following question. Does there exist a dual pairing
with Ker a F = Ker PF, and Coker a l p = Coker PF. There seems no reason
why the image of ,Or. is necessarily contained is the ~ u b g r o u p G & & i vof
134 J. COATES INFINITE DESCENT ON ELLIPTIC CURVES

whose restriction on the left to CF is the pairing { , )P? Of course, an This completes the proof.
affirmative answer t o this question would not only show that pp(llIF)is
finite, b u t would also prove that pp(ILIF) = 1.
We can now state our final result. Recall that yo denotes our fixed
topological generator of r. Let u = Xoo(70) be the image of yo under the Concluding Remarks
canonical character giving the action of on Ep,. We then define, for
s # 0 in Zp, 1. When t F = 0, the analogy of Theorem 24 with the refinement of
the Birch-Swinnerton-Dyer conjecture for the complex L-function L(qF, .3)
given by Gross in [8], is evident. As is explained in [8], our hypotheses t h a t
p # 2 and that p splits in K guarantees that the Tamagawa factors a t
the bad primes in the complex conjecture are prime to p, and so do not
specifically appear in the formula given in Theorem 24, which is necessarily
only valid up t o multiplication by a p a d i c unit.
The conjectural link of LI](E/F, 3) with p a d i c Lfunctions (see [4], [5]),
2. Allowing the possibility that tF > 0 (i.e. that ILIF(p) is infinite), it is
which we do not discuss in the present paper, provides the reason for
natural to ask whether tF is independent of the choice of the ordinary prime
introducing the pole term when Ep C E(F).
p. Certainly, a mystical belief in the close analogy between the behaviour
of both the co~nplexand p a d i c L-functions attached to EIF a t the points
T h e o r e m 24. Assume Conjecture A is valid, and that volp(E(F)) # 0. s = 1 suggests that tF should be independent of the ordinary prime p, and
Then (i) LI](E/F,3) has a zero at s = 1 of multiplicity 2 r~ = nF t ~ ; + that a modified form of the complex Birch-Swinnterton-Dyer conjecture
(ii) the multiplicity of this zero is exactly rF if and only if the canonical should hold even when t~ > 0. Some highly important recent work of R.
pairing { , )P given by (48) has a finite kernel on the right, or equivalently Greenberg [7] provides fragmentary theoretical evidence for this belief.
a trivial kernel on the left; (iii) If the multiplicity of this zero is exactly r F , 3. Although we do not go into details here (see [5]), a combination of
then Theorem 24 and the calculation of the values of the complex Lfunctions
L($:, k)(k = I, 2,. . .) can often be used to completely determine the arith-
metic of E relative to small ordinary primes p. We simply mention, without
proof, several examples, which it does not seem possible to treat by classical
methods. The calculations of the complex values L($:, k) used in the proof
of these results were made by N. Stephens.

where mp and pp(UIF) are given by (33) and (49), and T,~F denotes the
&ample 2. Take F = K = Q ( & f ) , and E the elliptic curve defined
Grossencharacter of 6 over F .
over K by the equation y2 = x3 - x. It is classical that nK = 0, and one
can now show that m K ( p ) = 0 for p = 5,13,17,29,37,41,53.
Proof. (i) is simply assertion (40). (ii) follows from Lemma 23 and
the fact explained above that Ker CYF= Ker y ~ and , Coker aF is finite
Example 3. Take F = K = Q(-), and E the elliptic curve defined
if and only if Coker 7 p is finite. Finally, to establish (iii), one combines
over K by the equation y2 = x3 - 22. It is classical that nK = 1, and
Proposition 2, Lemma 3, Lemma 18, Lemma 23, a.nd the formula (47), and
one can now show that ILIK(p) = 0 for p = 5,17,29. Moreover, we have
simplifies the resulting expression, noting that
V O ~ ~ ~ ( E ( --
K )1) if p = 5, and VOI~,(E(K))-- p if p = 17,29.
 J. COATES INFINITE DESCENT ON ELLIPTIC CURVES

References Received September 19, 1982

Cassels, J., Arithmetic on curves of genus 1 (VIII), Crelle 217 Professor John Coates
(1965), 180-199. Mathkmatiques, Bat. 425
Coates, J., Arithmetic on elliptic curves with complex multiplica- Universitk de Paris-Sud
tion, Hermann Weyl Lectures, I.A.S. Princeton, 1979, t o appear. 91405 Orsay, Prance
Coates, J., Wiles, A., On the conjecture of Birch and Swinnerton-
Dyer, Invent. Math., 39 (1977), 223-251.
Coates, J., Wiles, A., On padic Lfunctions and elliptic units, J.
Australian Math. Soc., 26 (1978), 1-25.
Coates, J., Goldstein, C., Some remarks on the main conjecture for
elliptic curves with complex multiplication, to appear in American
J. Math.
Greenberg, R., On the structure of certain Galoi. groups, Invent.
Math., 47 (1978), 85-99.
Greenberg, R., On the Conjecture Birch and Swinnerton-Dyer, to
appear in Invent. Math.
Gross, B., On the conjecture of Birch and Swinnerton-Dyer for
elliplic curves with complex multiplication, Progress in Math. 26
(Birkhauser), 219-236.
Iwasawa, K., On Z1-extensions of algebraic number fields, Ann. of
Math., 98 (1973), 246-326.
Mazur, B., Rational points of abelian varieties with values in towers
of number fields, Invent. Math., 18 (1972)) 183-266.
Perrin-Riou, B., Groupe de Selmer d'une courbe elliptique B mul-
tiplication complexe, Compositio Math., 43 (1981), 387-417.
Perrin-Riou, B., Descente infinie et hauteur padique sur les courbes
elliptiques B multiplication complexe, Invent. Math., 70 (1983),
369-398.
Schneider, P., Iwasawa Lfunctions of varieties over algebraic num-
ber fields. A first approach., to appear in Invent. Math.
Schneider, P., padic height pairings I, Invent. Math., 69 (1982),
401-409.
Tate, J., On the conjecture of Birch and Swinnerton-Dyer and
a geometric analogue, Skminaire Bourbaki, No. 306, Fkvrier,
1966.
 On the Ubiquity
of "Pathology" in Products
Nicholas M. Katz

Dedicated to I. R. Shafarevic on his 60th birthday

Introduction

Fix a prime number p, and an algebraically closed field k of characteristic

p. For any proper smooth k-scheme X, we denote by

its DeRham-Witt complex. One knows (c.f. [Ill, 1, 21) t h a t WOjc calculates
the crystalline cohomology of X, i.e., one has a canonical isonlorphism

which is the inverse limit of canonical isomorphisms

One also knows that H5,i,(X) is a finitely generated W-module, and that
for every n, the Wn-modles

are finitcly generated. However, there is no reason in general that for given
(i, j ) , the W-module

should be finitely generated. Indeed, in 1958 (c.f. [Se]) Serre discovered

140 N. M. KATZ "PATHOLOGY" IN PRODUCTS 141

t h a t for E a supersingular elliptic curve over k, the W-module (i.e., i = 2, is polygonally Hodge-Witt if t h c maximal Newton-Hodgc decomposition
j=O,X=ExE) (c.f. [Ka]) of ( M I F ) (i.c., using all those break-points of the Newton polygon
H 2 ( E X E, WO) which lie o n the Hodge polygon) is a decomposition into F-crystals of the
following types a) or b):
is n o t finitely generated.
Following Illusie-Raynaud and Bloch-Gabber-Kato, we say t h a t a prop-
er smooth k-scheme X is "IIodge-Witt" if, for all (i,j), the W-module a) There is a singlc Hodge slope:
I I i ( X , \YO$) is finitely generated, and we say t h a t X is "ordinary" if

lli(x,d ~ $ , ~=) O for all i, j .

One knows (c.f. [Ill-Ray]) that "ordinary" + "Hodgc-Witt."

In this note, we will show t h a t Serre's E x E example of a variety which
is not IIodge-Witt is archetypical. One precise rcsult is

T h e o r e m . Let X and Y be proper s m o o t h connected n o n - e m p t y

k-schemes, such that both HEIis(X) and Il:Ii,(Y) are torsion-free. T h e n
the product X X Y is Hodge- W i t t if and only if at least o n e of X or Y is
ordinary, and the other is Hodge- Witt. b) There are precisely two IIodge slopes, they are successive ir~tcgcrs,
Taking X = Y, we obtain say i and i + 1, and all Newton slopes lie in the open interval ]i,i -t1[,SO
t h a t the Newton polygon is strictly above the IIodge polygon except a t the
endpoints, c.g.,
C o r o l l a r y . Let X be a proper smooth connected n o n - e m p t y k - s c h e m e
with HZri,(X) torsion-free. T h e n the product X X X is IIodge- W i t t if a n d
only if X is ordinary.

As will bc clear t o the reader, the work presented here is based in a n

essential fashion upon the seminal work of Bloch, Illusie-ltaynaud, a n d
Ekedahl. I would likc to thank R. Crcw for explaining to me the "duality"
between the approach to Newton and IIodge polygons taken in [Ka], a n d
t h a t taken by I)cma.~urc[De] and by Berthclot-Ogus [B-01.

I. Reduction to a Question about Polygons L e m m a 1. Suppse that a n F-crystal ( M ,F) i s ezpressible as a direct
s u m of F-crystals,
Let ( M , F ) be a "a-F-crystal on k" (cf. [Ka]), i.c., M is a free, finitely ( M ,F ) = @(Mi, Fi).
generated W-module, and F is a a-linear endomorphism of M which in-
duces a n autornorphism of M @ Q. We say t h a t ( M , F ) is polygonally T h e n ( M , F ) is polygonally ordinary (respectively, polygonally Hodge- Witt)
ordinary if its Ncwton and Ilodgc polygons coincidc. We say t h a t ( M , F ) i f and only if each s u m m a n d (Mi, I;) is.
 "PATHOI.OGYn IN PRODUCTS 143
142 N. M. KATZ

Proof. For any subset S of Q20, let us denote by +

Applying these remarks to S = { i ) and to S =]i,i l[,we sce that the
+
conditions A and B (resp. A and x(]i, i 1[) = OVi) l~oldfor a direct sum
if and only if they hold for each summand.
Q.E.D.
the projection onto the direct sum over all X E S of all the Newton-
slope= X isotypical pieces of M @ Q . By its functoriality, rM(S)com~nutes We will make essential use of the following
with all W-linear endomorphisms of M @ Q which themselves commute
with F. Clearly (M, E') is polygonally Hodge-Witt if and ony if both of the K e y T h e o r e m (Illusie-Raynaud, Ekedahl). Let X be a proper
following conditions hold: smooth connected k-scheme, whose HZris(X) i s torsion-free. Consider the
following conditions:
>
A. For every integer i 0 the projector ~ ( { i ) )lies in Endw(M), and (EIDR) T h e IIodge =+ D e R h a m spectral sequence for X / k
Image ( ~ ( { i ) )I M), if non-zero, has all its Hodge slopes equal to i.
> +
B. For every integer i 0, the projector n(]i, i 1[) lies in Rndw(M),
and I m ( ~ ( ] i ,+ 1[) ( M), if non-zero, has each of its IIodge slopes equal to degenerates at El.
either i or i 1.+ (PO) The F-crystal is polygonally ordinary.
(PlIW) T h e F-crystal is polygonally Hodge- W i t t .
In a similar vein, (M, F ) is polygonally ordinary if and only if: T h e n we have

A holds and if x(]i, i + I[) = 0 for every integer i.

I. X is ordinary i f and only if X satisfies (IIDR) and (PO).
By functoriality, we have, for any S, 11. X is Hodge- W i t t i f and only if X satisfies (HDR) and (PHW).

To complete the reduction,we also need

L e m m a 2. Let X , Y be proper smooth non-empty k-schemes. T h e n

where we have denoted by X X Y satisfies (HDR) if and only i f both X , Y satisfy (HDR).

Proof. Clearly X satisfies (HDR) if and only if the universal inequality

the i'th projection and inclusion corresponding to the given decomposition

is an equality. The result now follows from the Kunneth formula in both
Hodge and DeRham cohomology, which gives
Therefore xM(S) lies in Endw(M) if and only if the individual nMi(S) do,
and if this is the case, then

The set of Hodge slopes (with ~nultiplicities)of a direct sum is simply the
union of the sets of IIodge slopes of the summands.
144 N. M. KATZ "PATHOLOGY" IN PRODUCTS 145

and the observation that So it suffices to show that ( M I @I M2, Fl @ Fz) is indecomposable for the
Newton-Hodge decomposition, i.e., that its Newton polygon lies strictly
above its IIodge polygon, excepb at the endpoints. (For having a t least
three distinct Hodge slopes, it will not be Hodge-Witt!)
and similarly for Y. Each of the (Mi, Pi),being indeco~nposableand not polygonaly ordinary,
Q.E.D. has its Newton polygon strictly above its Hodge polygon, except a t the
endpoints. Let us say that a non-zero F-crystal satisfis ( N > H) if its
Combining Lemma 2 with the key result of Illusie-Raynaud and Ekedahl, Newton polygon is strictly above its Hodge polygon, except a t the endpoints.
our theorem is a special case of the following theorem about F-crystals, In view of the above discussion, the polygon theorem results from the
F) and (Hrris(Y),F).
applied to (EI,',~,(X), more precise

Polygon T h e o r e m . Given two non-zero F-crystals (MI, Fl) and N> H T h e o r e m . The tensor product of two non-zero F-crystals which
(M2,1;i), their tensor product (MI @ M 2 ,Fl @I F2)is polygonally Hodge- both satisfy ( N > I-I) again satisfies ( N > H ) .
W i t t i f and oriy i f one of the (Mi, Fi) is polygonally ordinary, and the other
is polygonally Hodge- Witt.
111. Interlude: Generalities on Polygons of Hodge and Newton
Type
11. Further Reduction of the Polygon Theorem
We fix as basic data a collection of r 1 1 distinct non-negative real
In view of Lemma 2, we are immediately reduced to the case where both numbers, called "slopes,"
of the F-crystals (Mi,E;) are indecomposable as F-crystals.
If (MI,li;) is polygonally ordinary as well, thcn ( k being algebraically
closed), there is an integer j 1 0 such that
together with strictly positive integers called "multiplicities"

nl,. . .,n,.
The integer
and the assertion of the theorem is obvious in this case.
It remains to show that when both of ( M i ,I;;.) are indecomposable, as ~ = x n ~
F-crystals, and neither is polygonally ordinary, thcn their tensor product
is called the "rank." Equivalent to such data is the corresponding convex
is not Hodge-Witt. Since neither of the (Mi, F;) is polygonally ordinary,
each of them has a t least two distinct Hodge slopes. Therefore their tensor polygon in the (X,Y) plane (illustrated below with r = 3 distinct slopes).
product has a t least three dintinct Hodge slopes (for if (Mi, Fi)admits both For brevity, we will refer indifferently to either the data
a; and b; > ai as Hodge slopes, for i = 1,2, then their tensor product
+ + +
admits a1 a2, b2 az, bl b2 (as well as a1 + b2) as Hodge slopes, and by
assump tion
-+ +
a1 a2 < bl a2 < 61 62. + or to the corresponding convex polygon as "a polygon."
148 N. M. KATZ "PATIIOLOGYn IN PRODUCTS

In terms of t h a t data (A1,. . .,A,; n l , . . . ,n,), we have the explicit for- We summarize this in a diagram:
mulas
Dp(t) = C n i min(t, Xi)
i
BOp(t) = C ni max(0, t - Xi).
I

As is obvious from the picture, we have

Dp(t) + 130p(t) = N t
BOr(t) iy > 0 and monotone increasing.
As is also obvious from the picture, we have

L e m m a 3. Let P and & be two polygons. Then:

I. Q lies above (i.e., is contained in) P if and only if
BOp(t) > BOQ(t) (4) If P corresponds to d a t a (XI,. . . , X,;nl,. . . , n , ) with all Xi E Q,
consider the polynomials
for ill1 t .
11. If P and & have the same rank and the same terminal point, then f (T)= rl[(l - pX'T)"'
i
$ lies strictly above I', except a t the endpoints, if and only if we have
F(T) = D ( T -
BOr(t) > i
for all t in the interval Then for any rational number t $! {XI,. . . ,X,), we have
mi11 slope ( P ) < t < max slope (P). Dr(t) = 0rdp(qpt))
Remarks. BOr(t) = -ord,(f (pet)).
( I ) T h e second derivative of B O p ( t ) is the sum over the slopes Xi, counted
with multiplicity, of Dirac delta functions supported a t Xi: DO/: = nlSx,
when 1' is given by data (A1,. . . ,A,; n l , . . . , n,).
(2) If the polygons in question havc integer slopes, e.g., Iiodge polygons,
IV.Direct Sums and Tensor Products of Polygons
t h e dual functions Bop and D p are linear betwecri successive integers, so
one may verify iriequalitics among them just by checking a t integer values Let P and & be polygons, corresponding t o data
oft. P:(Xll...;nl...n); rankCni=N
(3) If P is the Hodge polygon of an F-crystal ( M , F ) , one verifies easily Q: ( p l , . . . , p a ; m l , . . . ,m,); rank C m j = M.
t h a t for any integer k 2 0, wc have
Then P @ Q is the polygon whose rank is N + M and whose slopcs, counted
BOp(k) = length ((FM + pk M ) l p k M) with multiplicities, are the union of those of P and Q.
= length ( FM I F M n pkM) T h e tensor product P & is the polygon of rank N M whose slopes,
counted with mulliplicities, are the numbers Xi + pj, counted with multi-
Dp(lc) = length (M/(FM +p k ~ ) ) . plicity n i m j .
150 N. M.KATZ

We rewrite this difference a s one-half of

The dual functions are given by the formulas

(where * denotes convolution, or more usefully,)

Because the (Mi, Fi) both satisfy (N > H), it follows from Lemma 3,
11, that the functions (for i = 1 , 2 )

are strictly positive in the open interval ]a;, b;[, and zero elsewhere.
For i = 1,2, let ai be the smallest Newton slope of (Mi, Fi). Then

Given non-zero F-crystals ( M I , F1)and (M2, Fz), let us denote by IIPl because (Mi, Fi) satisfies ( N > H). Therefore each of the functions (for
and 11P2 their respective Hodge po ygons, and by NPl and NP2 their i = 1,2)
respective Newton polygons. Lct us denote by HP1e2 and Bpie2 the
IIodge and Newton polygons of (M1,Ii;)@ (M2,F2), and let us denote by
IIfio2and NPlB2 the Ilodge and Newton polygons of (Ml@M2, F1@F2). has gy a sum with positive integer multiplicities of Dirac delta functions
Then we have the expected co~npatibilities supported a t various points, including the points a;,, ai, bi. Uecause the
functions f i are non-negativd, we have

where

V. Proof of the N> H Theorem

Because f i is strictly positive in ]ai, bi[, and zero outside, the first sum
after the inequality sign,
Let ( M I , Fl) and (M2, I;i) be two non-zero F-crystals, both of which
satisfy ( N > H). Let ai < b; denote the smallest and largest Hodge slopes
of (Mi, F;), for i = 1,2. Then a1 +a2 < bl +b2 are the smallest and largest
Ilodge slopes of (M1 @ M 2 ,Fl @ F 2 ) To show that Lhis tensor product +
is strictly positive in ]a1 a2, bl + bp[ except for a zero at bl +a2. Similarly,
satisfies (N > H), it is equivalent (by Lemma 3, 11) to show that the second sum

~ )O N P , ~ ~>
B O M P , ~ ~-( B ( ~0)

for a1 + a2 < t < bt + b2.

 N. M. K A T Z "PATHOLOGY" IN PRODUCTS
152

is strictly positive in ]al + a2, bl + b2[ except for a zero a t b2 + a l . So in

References
the case
bl + a 2 # b2 + a l , [B-O] I)erthelot, P., and Ogus, A., Notes on Crystalline Cohomology,
Matliematical Notes 21, Princeton University Press, 1978
we have f @ 92 + f 2 @ gl strictly positive in ]al + a2, b1 + b2[, as required. [Bl] Bloch, S., Algebraic K-theory and crystalline cohomology, Pub.
In the remaining case, Math. IIIES, 47, 1978, 187-268.
b2 +
a1 = bl a2, + [B-G-K] Bloch, S., Gabber, O., and Kato, manuscript in preparation on
IIodge-Tatc decompositions for ordinary varieities.
we have
b2 - a 2 = bl - a l . [Dem] Dernazure, M., I,ectures on p-divisible groups, Springer Lecture
Notes in Mathematics, 1972.
The final summand [Ek] Ekedahl, T., manuscript in preparation.
[Ill 11 Illusie, Id., Co~riplexedc denham-Witt, Journkes de Gkometrie
Algdbriqi!es dc Rennes, 1978, Astkrisque, 63, S.M.F., 1979.
[Ill 21 Tllusie, L., Complexe de dcRharn-Wilt et cohomologic crystalline,
is strictly positive in ]al + u2, bl + a2[.The picture of the regions of strict Ann. Scient. Ec. Norm. Sup. 4c skrie, t. 12, 1979, 501-661.
positivity of the three sums we have considered is [Ill-Ray]Illusic, I,., and Raynaud, M., Les suites spectralcs associkes au
complexe dc deILharn-Witt, to appear in Pub. Math. IHES.
[Ka] Katz, N., Slope filtrations of F-crystals, Journkes de Gkometrie
Algkbraique de Rennes, 1978, Astkrisque, No. 63, S.M.F., 1979,
113-164.
[Se] Serre, J.-P. Quelque propriktes des variktes abcliennes en car. p.,
American Journal of Mathematics, 80, 1958, 715-739.

Received October 7, 1982

Professor Nicholas M. Katz

Departrncnt of Mathematics
Princeton Univesity
Princeton, New Jersey 08540

Therefore we find that fr QP 92 + f2 QP g1 is strictly positive in

]a1 + +
az, bl b2[ in this case also.
Q.E.D.
 Conjectured Diophantine Estimates
on Elliptic Curves
Serge Lang

To I.R. Shafaravich

Let A be an elliptic curve defined over the rational numbers Q. Mordell's

theorem asserts that the group of points A(Q) is finitely generated. Say
{PI,. . . ,P,) is a basis of A(Q) modulo torsion. Explicit upper bounds for
the heights of elements in such a basis are not known. The purpose of this
note is to conjecture such bounds for a suitable basis. Indeed, R @I A(Q) is
a vector space over R with a positive definite quadratic form given by the
Nfron-'l'ate height: if A is defined by the equation

and P = (x,y) is a rational point with x = c / d written as a fraction in

lowest form, then one defines the x-height

h, = log max(lcl, Idl).

There is a unique positive definite quadratic form h such that

1
h(P) = ih.(P) + Or(1) for P E A(Q).

Then A(Q)rnod .4(Q)t,, can be viewed as a lattice in the vector space

R @ A(Q) endowed with the quadratic form h.
I assume that the reader is acquainted with the Birch-Swinnerton Dyer
conjecture. For an elegant self contained presentation in the form which I
shall use, see Tate [Ta 11. Starting with the Birch-Swinnerton Dyer formula,
and several conjectures from the analytic number theory associated with
the Gfunction of the elliptic curve, I shall give arguments showing how
.
one can get bounds for the heights of a suitably selected basis {PI,.. ,P,}
of the lattice. A classical theorem of Hermite gives bounds for an almost-
orthogonalized basis of a lattice in terms of the volume of the fundamental
156 S. LANG CONJECTURED DIOI'HANTINE ESTIMA'rES 157

domain, so the problem is to estimate tliis volurnc. Ilermite's theorem will and is the square of the volu~neof a fundamental domain of the lattice
be recalled for the convenience of the reader in the last section. 4 Q ) l A ( Q ) t o r in R 8 4 Q ) .
The conjectured estimates will also tie in naturally with a conjecture of As usual, the discriniinar~tis given by
Marshall Hall for integral points, which we discuss in $2. I shall give some
numerical examples in $3.
Manin [Ma] gave a general discussion showing how the Birch-Swinnerton The terms rpfor plA are given by the integral of Iwl over A(Qp), where
Dyer conjecture and the Taniyama-Wcil conjecture that all elliptic curves the differential form w is the one associated with the global ~riinirnummodel.
over Q are modular give effective means to find a basis for the Mordell- As Tate pointed out, ap is an integer for each p. This property is all we
Weil group A(Q) and an estimate for the Shafarevich-Tate group. The need to know about a, for our purposes. Essentially.
Taniyama-Weil conjccture will play no role in the arguments of this paper,
and I shall propose a much more precise way of estimating the regulator
and the heights in a basis as in IIermite's theorem. Manin also attributes a
"gloomy joke" to Shafarcvich in the context of his Theorem 11.1. It is there- +
except that the equation y2 = x3 + ax b is not the equation of a minimal
fore a pleasure to dedicate the conjectures and this paper t o Shafarevich. model, and one has to introduce extra cocfi?cients, cf. ['Fa I ] where the
matter is explained in detail. For simplicity of notation, I shall continue to
write the equation in standard Wcierstrass form as above.
The period a, is given by the integral
1. Rational Points

We let (P,Q ) be the bilinear symmetric form associated with the Nkron-
Tate height, namely
Finally IA(Q)t,, 1 is the order of the (finite) torsion group.
(P,Q) = h ( P + Q) - h(P) - h(Q)-
We are now interested iri estimating all the terms to get a bourid for
Then IIIIlR, which will yield a bound for 12 since ILL11 is an integer.
1
)PI = and h(P) = - 1 ~)~. Since the aP are intrgers, they work for us in getting an upper bound
2 for R.
If r is the rank of A(Q), that is the dirnensiori over R of R @ A ( & ) , then Mazur [Ma] has shown that IA(Q)lorl is bounded by 16, and hence this
Birch-Swinnerton Dyer conjecture that the L-function LA(s) = L(s) has a torsion number has a liniitecl effect OIL the ticsired estimates.
zero of order r a t s = 1, and t h a t the coefIicient of (s - 1)' in the Taylor Concerning the r-th dcrivative of the L-function. I am indebted to
expansion is given by the formula H. Montgomery and Dxvid Rohrlich for instructive discussions on the
analytic number theory of more classicid zeta functions and Lseries. In
particular, for the Ricrriann zeta function cQ(s), the Ricmarln hypothesis
implies t h a t

On the right-hand side, IIII( is a positive integer, the order of the Shafare-
vicli-Tate group.
The determinant is taken with respect to any basis { P I , .. ., P , ) . Its where t(t) tends to 0 as t + oo. More precisely, Montgomery conjec-
absolute value is called the r e g u l a t o r , tures [Mo], formula (10) that one can take

t(t) = c(1og t log log t)-

 CONJECTURED DIOPIIANTINE ESTIMATES 159

with some constant c. Without the loglog t, Titchrnarsh already observed This leaves the real period n, to be estimated from below. The Nkron
that one cannot improve the exponent (log t)-'I2, see [Ti], Theorem 8.12. differcrltial of a mini~rialmodel will not difler too much from dxly. Let US
The Rie~nannhypothesis iniplies by Theorem 14.14 that compute roughly the integral

b u t is pot known t o imply Montgomery's conjecture, or even t h e weaker

formulation with the Titchmarsh exponent without the log log t. + +
where intervals of x such t h a t x3 a x b is negative are to be disregarded
For Dirichlet Lseries over the rationals, with conductor q, Montgomery in this integral. If b2 > laI3 we change variables and let x = lb('f3u,
also conjectures that one has dx = lb1'/3du. If b2 < ~ a ( ' / ~ u .
J7or definiteness, say b2 ) laI3. Then the integral becomes
LL(.)(;)
r!
, qr(q)c;(log q)'

with a universal constant c2.

It is known in the case of elliptic curves with complex multiplication, that
Ln(s) is a Heckc L-series, which is so normalized that LI'(1) corresponds to
where -1 5 c 5 I . The integral for lul > 2 is bounded from below by a
universal constant. The integral
the r-th derivative of the Hecke L-series in its usual form a t s = 112. So to
get bounds for L$)(I), on needs bounds for the Rccke Lseries a t 3 = 112.
I once gave a general principle which allows one to prove (or conjecture)
what happens for more general 1,-series, on the basis of what happens for
tlic Rirrnann zeta function. To quote [L 21: whenever you see a t in an
estimate, with a logarithm, then replace it by dxtn, where dx = d K N f x , is also bounded from below because ( u 3 + cu + 11 is bounded from above
dK is the absolute value of t h e discriminant of a number field K, and for c, u lying in bounded intervals. A similar anslysis if b2 5 laI3 yields the
n = [K : Q]. In line with this principle, I expect lower bound

1 In,I >> 1 / ~ ' lf2 where I1 = rnax(la13, lbI2).

-L$)(I) << ~ ~ ( ~ ) c j (N)'
log
r!
Putting all this together, we arrive a t the following conjecture.
whcre N is t h e conductor of the curve, cf. [Ta 11, end of $6; and
t ( N ) tends to 0 ;is N --t co. Furthermore, c(N) may have the same shape
as in Montgomery's conjecture, namely Conjecture 1. Let II(A) = max(JaI3,lbI2). For all elliptic curves
y2 = x3 + +
a x b with a , b E Z we have
t ( N ) = c(1og N log log N)-'/ '.
Note t h a t the conductor is divisible by the same primes as the descrimin-
ant A, and ordp N 5 ordp A so N 5 lAl. For the precise order of N a t a with some universal constant c, and c(N) --t 0 as N -t 00. In fact, c ( N )
prime pJ see [l'a I], end of $ 6 , and [Ta 21 for a systeniatic way of computing may have the explicit form
N , which is an isogeny invariant. Thus the above inequalities for N in
terms of A hold when A is replaced by a minimal discrirnin;mt Amin. c(N) = cl(log N log log N ) - ' / ~
 CONJECTURED DIOPIIANTINE ESTIMATES 161

Note that if either a or b = 0, then H(A)'/'~ can be replaced by Theorem 4.2 (1) and the lower bound of Conjecture 2, which allows us to
In light of the Hall conjecture below, it should always be the case that divide by (10glA,~,l)~-' on both sides so that we end up with only one
factor of log N on the right hand side.
The above conjectures are phrased for the canonical height h. To apply
them to the ordinary height h, of the x-coordinate, we need an upper bound
In any case, H(A) is absolutely homogeneous of degree 12, in the sense 1
on Ih - i h , which will be discusscd in the next section.
that H(cA) = I c ~ ' ~ H ( A )where
, cA denotes the elliptic curve obtained On the other hand, observe that the lower bound of Conjecture 2 allows
by multiplying all the quantities a, b, x,y with the appropriate power of c, us to estimate from above. Indeed, referring to IIer~nite'sthcorem
corresponding to their weight. in $4, we have
The shape of the inequality in Conjecture 1 is analogous to the upper R =d e t ( ~= ) ~lull2.. . 1u,I2
bound for the regulator of a number field K by easy esti~natesof the residue
of the Dedckind zeta function a t s = 1, apparently first noted by Landau,
of the form
~KRK 5 c[ ~ : ~ dK)[K:Ql-l
l ~ z ~ ( ~ ~ ~
where dK is the absolute value of the discriminant. Siege1 [Si] more carefully B c:(logl~rni~ 1)'
determined precise and very good constants in this estimate. In [L 11 I made by Conjecture 2. Hcncc the upper bourid for lLUlR in Conjecture 1 yields
the following conjecture: an upper bound for lLUl of the same nature, except that we can cancel
(10gIAminlY.
. .
Conjecture 2. For all elliptic curves over the integers as above, one The above arguments do not seem t o give further insight in the ques-
has the lower bound for the canonical height: tion raised in [L 11, p. 92, concerning the ratios of successive minima, or
equivalently the lengths of elements in a n almost orthogonalized basis.
It is a good question to determine to what extent the factor cr(log N)'
for any rational point P which is not a torsion point. in Conjecture 1 must really be there. After extending Q to the field of
2-torsion points the standard proof of the Mordell-Weil theorem of course
bounds r in terms of the number of prime factors of N but also in terms
This conjecture was proved by Silverman [Si] when the j-invariant is
of the 2-rank of the class number or the p r a n k of the class numbcr of
an integer. Silverman [Sil 31 has also shown that the conjecture is best
ptorsion points, for any prime p. The question is whether this second part
possible. Taking into account the Hermite theorem, we may now give an
of the bound is, in fact, significant over the rational numbers.
upper bound for an appropriate basis of the Mordell-Weil group. I leave
Supposing r is like the number of distinct prime factors of N, one
out the integral factor lLUl on the left for simplicity.
knows that if N is a product of the first r primes, then r is asyn~ptotic
to log Nlloglog N. On the other hand, the average value of r is log log N.
Conjecture 3. There exists a basis {PI,.. .,P,) for A(Q) modulo tor-
sion, ordered by ascending height, such that:

2. Integral Points

We begin by recalling an estimate for Ih - $h,l, see Demjanenko [De],

Indeed, we have h(P) = ;[PI2,so the upper bound for h(Pl) is precisely Zimmer [Zi] or [L 11, p. 99, where I piece together the local esti~natesto
the Hermite bound in Theorem 4.2. As to the highest point, we use give a global one as follows.
162 S. LANG CONJECTURED DIOPIIANTINE ESTIMATES 163

T h e o r e m 2.1. Let y2 = z3+ a x + b be an elliptic curve with a, b E Z. because j = - 2 1 2 3 3 a 3 / ~and a, A are p-integral. Taking the sum we then
Let h be the canonical height. Then obtain a bound

The O(1) comes only from the archimedean v, there is no such constant
In particular, for v = vp with a prime number p. The theorem follows a t once. (Note: one
can get something slightly better, see the Remark in [L 11, p. 32.)

I h - -h,
I
5
1
- log H ( A )
6
+ O(1).
Theorem 2.1 suffices amply to apply the bound of Conjecture 3 to the
Proof. Let h, be the local canonical height, or in other words the Nkron ordinary height IL,. However, I want the estimate for Ih - ih,l to be
function at v, normalized as in Tate, cf. [L I] Chapter I, $7 and 58, and entirely in terms of A. For this purpose, among many others, one can use
Chapter 111, $4. Let h,,, be the local height of the x-coordinate, that is a conjecture of Marshall Hall [Hall].

C o n j e c t u r e 4. If (x, y ) is an integral point on the elliptic curve

+
y2 = x3 b with b E Z, then 1x1 << b2.
We now distiguish the two cases when v is archimedean or not.
First let v be the archimedean absolute value on Q. Then in [L 11, Actually, Stark and Trotter after some theoretical probabilistic con-
Theorem 8.4 of Chapter I, I proved siderations have suggested that the conjecture should ,be with an exponent
2 + E, or a t least some appropriate powers of loglbl occurring as a factor
of b2. I shall refer to the Hall conjecture so modified. Hall also expressed
his conjecture as an inequality

where Co is some appropriate constant. The two cases correspond to

whether j is small or large a t the archimedian absolute value. In the present whenever t , u are integers. Silverman has pointed out to me that [R-C-H-S]
case, A is an integer so loglAl _> 0. Furthermore j = -2'233a3/A, so we yields parametrizations of curves with integral points having the order of
get log(jl = logla31 - loglAl + 0(1), (To get all bounds in terms of A, see magnitude 1AI5, by polynomials g2 - f 3 with integer coefficients, having
Conjecture 4.) degree 1 + deg f . Presumably 1AI6 is the best possible power of A.
On the other hand, if v is non-archimedean, Tate has shown that Davenport [Da] has proved the IIall conjecture for polynomials in its original
form, without the E.
+
If we now use the Hall conjecture on the equation 4a3 27b2 = A.
(see the remark below), we can apply Theorem 2.1 to get

See Theorem 4.5 of Chapter 111. Recall that v(z) = -loglzlp if v = up is

the p a d i c absolute value, whence in this case
This is of special interest in that loglAl occurs on the right hand side with
+
multiplicity 1 E . As already mentioned, in some cases one might get an
even better estimate.
164 S. LANG CONJECTURED DIOPHANTINE ESTIMATES 165

Remark. In using the Ball conjecture, the "variables" a, b have fixed I shall give here his two largest solutions. I am indebted to J. Rrette for
coefficients, namely 4 and 27. Trivial changes of coordinates (and of A. using the computer to put Selmer's solutions in the Weierstrass form
by a bounded small factor) reduce such equations to the case when the
coefficients are equal to 1. For example we multiply the equation by 4233 Y2 = x3 + b, with b = - 2 4 3 3 ~ 2 .
and let u = -12a, t = 22c3b to get t2 - u3 = 4233A0.
Since we are interested in the height, it is best to clear denominators and
consider the homogeneous form
It is not completely clear how the Hall conjecture should extend to an
arbitrary equation
y2 = x3 + ax b +
to be solved in relatively prime integers (u, w, z). Then
with a, b E 2. It seems reasonable to expect:
x = u/z and y = w/z.
Conjecture 5. If (x, y) is a n integral point on this elliptic curve, then
The transformation is given by the formulas

u' =2 2 3 ~ 2 , w' = 2 2 3 2 ~-(Z),

~ a' = X + Y,
for some fixed positive number k independent of a, b E 2. One still has to divide u', z' by their g.c.d. to get x in lowest forrn. The
rank of A(Q) is 1 in each of the following two cases, and the given values
The analogous statement has been proved in the function field case for (u, w, z) yield a generator of A(Q) mod torsion. Note how the height h,
by Schmidt [Sch]. It is not clear to me what the exponent k should be. drops to about two-thirds of the height of hxlz, as it should for theoretical
Added in proofs: Stark has tentatively suggested k = 513 6, but he + reasons since X / Z has degree 3 and x has degree 2. In this special case,
has counterexamples showing that probabilistic arguments can be very one might apply Theorem 2.1 to h, and an analogue to hxlz to see this.
misleading, and that one cannot make the conjecture until it is checked The behavior is remarkably uniform for this particular algebraic family
in the function field case. For instance, he gives the example depending on D.

u = 35,208,298,044,859,842,638,816,896,575,916
which has the solution x = t6 + 2t2 and y = t9 + 3t5 + i t .
z = 94,295,149,257,506,211,891,484, 409,025
(u', z') = 956,232,024,756,251, 670
3. Some Numerical Examples

There exist some tables in the literature giving bases for the Mordell-
Weil group, mostly when the rank is 1. Selmer [Se 11, [Se 21 found some For the conductor, we use the formula in Stephens [St], p. 125, which
solutions which a t first sight appear quite large. H e works with the curve for N f *2 mod 9 gives the conductor of the Hecke character as
in Fermat form
 S. LANG CONJECTURED DIOPRANTINE ESTIMATES

and therefore where

N = d ~ ( &Nf,
3 )= 3Nfx.
b = 2212)
Since D = 2 X 173 r 4 mod 9 we find: c = 317v4 + 4u4
u = 48,869 and v = 73,265.
We have
A. -- 1.3 X lo8

Using t ( N ) = (log N log log N ) - ' / ~ , we also find:

Now we use the formula in Birch-Stephens [B-S], p. 297, giving the conduc-
tor of for y2 = x3 - Dz, namely

This puts the bound a t about 40 times h,.

If D r 3 mod 4. Here D = -317 = 3 mod 4, so N = dg(i)Nfx, and we
find:
N = 263172 = 6,431,296
u = 96, 793,912,150,542,047,971,667,215,388,941,033
log N -- 15.7
z = 195,583,944,227,823, 667,629,245,665,478,169
(u',2') = 384,647,097,245,468,469,552
H ( A ) ~ / ~ ~ N ~ ( ~718 -
) ~ o ~ N
h, -- 81
-
A. = 2 7 ( 2 4 3 3 ~ 2 ) 2 1.07 X 10''
log A. - 38.9
So h, is about one tenth of its presumed upper bound.

N = 223"912 = 3,039,948 It should be noted that most generators in Selmer's tables, for instance,
log N - 15.2 are considerably smaller than the upper bound given by the conjecture.
H ( A ) ~ I ~ ~NN ~4.9( x - ~ lo3
) ~ ~ ~ Statistically, it would seem that they follow quite a different distribution,
h, - 80.4
namely
R << loglA1.
The estimate is again roughly the same as in the preceding example.
For the exponential height, H, = exp h,, this would mean that on the
average, one has
I owe the next example to Andrew Bremner, who considers the curve H z << /Atk
for some constant k. Even the two largest Selrner points have heights
compatible with such a better bound taking k = 2; while the Bremner
This is the "other" case with complex multiplication. Bremner parametrizes point fits with k = 3 or 4. Most other points in Selmer's tables, and all the
a generalor of the Mordcll-Wcil group in the form points in Cassels' early table [Ca] or Podsypanin [Po] fit with k = 1. Within
the range of such tables, it seems impossible to distinguish a polynomial
estimate as above from an exponential estimate (for the exponential height)
168 S. LANG CONJECTUILED DIOPIIANTINE ESTIMATES

as in Conjecture 3. It would be interesting to have a more precise statistical 4. The Hermite Theorem
analysis of the possibility of a polynomial estimate.
We note that if f ( x , y) is an irreducible homogeneous polynomial of This section is essentially an appendix, t o state the Hermite theorem
degree > 3 over Z then there exists a constant k such that for any integral concerning lattices in euclidean space.
solution ( 2 ,y) of the equation
T h e o r e m 4.1. Let L be a lattice i n a vector space V of dimension
r over R, with a positive definite quadratic form. Then there exists a n
orthogonal bases { u l , . . .,u,) of V, and a basis {el,. . .,e,) of L having the
following properties.

(i) el = u1 is a vector of minimal length in L.

so the estimate is known to be polynomial in this case. This was proved (ii) e; = bi,l u l + +
.. bi,i-lu;-l ui+
by Feldman [Fe], .following Baker's method. On the other hand, for the with b;,j E R for i = 1 , . . ., r and Ibi,jl 5 112 for all i, j .
standard Weierstrass equation y2 = z3 + b, the best known result is
exponential and due to Stark [St], using similar methods, namely
(iii) 1u;I < lei[ < ( 6
~)'+~lu~l,

The theorem is proved by induction, essentially by Gram-Schmidt

As t o the constants which appear throughout in the estimates, I don't orthogonalization, and taking minimal vectors successively.
know any explicit bounds for them, although in number theory, one does A basis of the lattice as in Theorem 4.1 is called a l m o s t orthogonnl-
not expect such constarlts to be "large" (whatever that means). However, ized. One gets the following bounds for the lengths of such basis elements.
Silverman analyzed the constant in Conjecture 2, and found t h a t in the
inequality
T h e o r e m 4.2. Let {el, .. .,e,) be a n almost orthogonalized basis
h(P) > clloglAminl of L. Then:
one had c l < 5 X
hl.. lerl 5 (5)
in some examples [Si]. r(r-1)/2
(1) det(L)
The estimate of O(1) in Theorem 2.1, which comes from the archimedean
solution value, could be given explicitly, but I have not done so. Zimmer where det(L) is the volume of a fundamental domain.
[Zi] gives a good explicit form as 2 log 2 in his version.
The constants which come up in Conjecture 1 are of a much more
complicated nature, what with Ricrnarln hypothesis estimates corning up.
For IIecke L-scrics, if one is satisfied with the t(N) which is implied by Proof. The basis {el,. . .,e,) of L is obtained from the orthogonal basis
the Riemann hypothesis, then the constants could be estimated following V by a triangular matrix all of whose diagonal elements are equal to 1. The
the general approach of [L 21. Similar analytic techniques might work more volurrie of the rectangular box spanned by u l , .. .,u, is equal to the product
generally for the L-function of an elliptic curve, say assuming the functional of the lengths of the sides, so
equation and Riemann hypothesis. Cf. the remarks a t the end of [L 21. Of
course, the Montgomery conjecture (or a weaker form along Titchmarsh) det(L) = lull. .Iu, I.
seems t o lie outside the range of such techniques. These considerations The first assertion follows by using (ii) in Theorem 4.1. The second
anyhow lie a t the periphery of this kind of analytic number theory. assertion follows immediately.
 CONJECTURED DIOPHANTINE ESTIMATES 171
170 S. LANG

W. SCHMIDT, Thue's equation over function fields, J. Austra-

Note. The appearance of r ( r - 1)/2 (so essentially r 2 ) surprised me at lian Math. Soc. (A) 25 (1978) pp. 385-422.
first, but I saw no way of reducing it to the first power of r. E. SELMER, The diophantine equation ax3 + by3 + cz3, Acta
Math. (1951), pp. 203-362.
References E. SELMER, Ditto, Completion of the tables, Acta. Math. (1954),
pp. 191-197.
[I3-C-1-S] B. BIRCH, S. CHOWLA, M. HALL, A. SCHINZEL, On the C. L. SIEGEL, Abschatzung von Einheiten, Nachr. Wiss.
difference x3 - y2, Norske Vid. Selsk. Forrh. 38 (1965), Gottingen (196Y), pp. 71-86.
pp. 65-69. J. SILVERMAN, Lower bound for the canonical height on ellip-
B. BIRCH and N. STEPIIENS, The parzty of the rank of the tic curves, Duke Math. J. Vol. 48 No. 3 (1981)) pp. 633-648.
Mordell- Weil group, Topology 5 (1966) pp. 295-299. J. SILVERMAN, Integer points and the rank of Thue Elliptic
J. W. CASSELS, The rational solutions of the Diophantine equa- curves, Invent. Math. 66 (1962), pp. 395-404.
tion Y2 = X3 - D, Acta Math. 82 (1950) pp. 243- J. SILVERMAN, Heights and the Specialization map for families
273. Addenda and corrigenda to the above, Acta Math. of abelian varieties, t o appear.
84 (1951)) p. 299. H. STARK, Effective estimates of solutions of some diophantine
H. DAVENPORT, On f 3(t) - g2(t), Kon. Norsk Vid. Selsk. Euations, Acta. Arith. 24 (1973), pp. 251-259.
For. Bd. 38 Nr. 20 (1965) pp. 86-87. N. STEPHENS, The diophantine equation X3+y3= D Z 3 and
V. A. DEMJANENKO, Estimate of the remainder term in the conjectures of Birch-Swinnerton Dyer, J. reine angew.
l'ate '8 formula, Mat. Zam. 3 (1968)) pp. 271-278. Math. 231 (1968), pp. 121-162.
N. I. FETIDMAN, An effective refinement of the exponent in J. TATE, The arithmetic of elliptic curves, Invent. Math. (1974),
Liouville's theorem, Isv. Akad. Nauk 35 (1971) pp. 973- pp. 179-206.
990, AMS Transl. (1971)) pp. 985-1002. J. TATE, Algorithm for determining the Type of a Singular
M. HALL, The diophantine equation x3 - y2 = k, Computers Fiber in a n Elliptic Pencil, Modular Functions of One
in Number Theory, Academic Press (1971), pp. 173-198. Variable IV,Springer Lecture Notes 476 (Antwerp Confer-
S. LANG, Elliptic curves: diophantine analysis, Springer Verlag, ence) (1972), pp. 33-52.
1978. E. C. TITCHMARSH, The theory of the Rzemann zeta function,
S. LANG, On the zeta function of number fields, Invent. Math. Oxford Clarendon Press, 1951.
12 (1971), pp. 337-345. H. ZIMMER, On the difference of the Weil height and the Nkron-
J. MANIN, Cyclotomic fields and modular curves, Russian Tate height, Math. Z. 147 (1976), pp. 35-51.
Mathematical Surveys Vol. 26 No. 6, Published by the
London Mathematical Society, Macmillan Journals Ltd, Received August 24, 1982
1971.
H. MONTGOMERY, Extreme values of the Riemann zeta func- Professor Serge Lang
tion, Comment. Math. Helvetici 52 (1977), pp. 511-518. Department of Mathematics
V. D. PODSYPANIN, On the equation x3 = y2 + Ax6, Math. Yale University
Sbornik 24 (1949)) pp. 391-403 (See also Cassels' correc- Box 2155 Yale Station
tions in [Ca]). New Haven, Connecticut 06520
 Zet a-Functions of Varieties
Over Finite Fields at s=l
Stephen Lichtenbaum

To I.R. Shafarevich

Let k be a finite field of cardinality q = pf. Tlet be a fixed algebraic

closilre of k. Let X bc a smooth projective algebraic variety of dimension
d over k such that W = X X is connected.
k
11 is by now well-known that the zeta-function of X may be written in
the form ( ( X , s) = Z ( X , q-'), where

where P;(X, T) = det(1 - +;JT) is thc characteristic polynomial of the

endornorphism 4i,l of the ktalc cohornology groups II"(X, Q 1 )induced by
+
the Frobenius endomorphism of X. Deligne has shown [De 11, that these
polynomials P,(X, T ) have rational integral coeflicie~lts,are independent of
1, and (the ltiemann hypothesis) havc complex "reciprocal roots" of absolute
value q i / 2 .
Motivated by the analogue of the Jlirch-Swinncrton-Dyer conjecture for
Jacobians of curves ovcr funclion-fields, Tate arid Artiri [ T l ] produced a
conject,urr when d = 2 relating the behavior of lJ2(X,T) as 7' --+ q-' to
various colion~ologicnlinvariants of X. Moreover, they proved that if the
1-parl nr(X)(l) of tllc Braucr group N 2 ( X , G,) of X was finite for one
prime 1 diflercnt from p tlien the I-p:wl of the Rraucr group prime to p was
finite, and their conjecture was true [or X up to a powcr of p. Milne, in
[Ml], completed lhe work of Tate arid Artin for surfaccs by removing the
restrictions concerning the characteristic.
In this paper we propose to begin to extend the work of Tate, Artin, and
Milne to varieties of arbitxary dimension. There no longer seems to be any
reason to expect a particularly elegant formula for P2(X,T), so we are lcd
t o turn our attention to the full zcta-function Z ( X , T). If we let p l ( X ) be
thc rank of lhe Neron-Severi group of X, we now obtain thc conjcctured
 (2) (*) is equivalent to Tatc's corljecture for X of di~nension2. In
particular (*) is true if the I-part ol' I12(X,G,) is finite for any single
prime 1, by the results of Tate arid Milne.
(3) (*) is true up to a power of p for X of arty dimension, provided the
I-part of H 2 ( X , Cm)is finite for any single prime I f- p.
This formula, although remarkably simple, needs some explariation. We
mean by X(X,Go) the Euler characteristic of the sl~eafG, for the ftale (or Along the way we show that Hi(X,G,) is finite for i f 1,2, and 3, and
flat) topology on X , namely that if in addition H 2 ( X ,G,)(l) is finite for one prinle I # p, we have

(i) 1 1 2 ( x ,G,)(non-P) is finite, and

(ii) 1I3(x, C,) is the dual of a finitely-generated abelian group (up to
ptorsion).
In view of the standard fact that lli(X,C,) is isomorphic to l l i ( X , Ox),
where the cohornology groups are now taken with respect to the Zariski
topology, we see that X ( X ,Ga)= q ~ ( ~ ~ O We x ) .would like to define I would like to thank S. Chase for simplifying my original proof of
,(X, 6,) in the analogous way, but run into the obvious problcm that 1,enlma 1.2.
the H " ( x , G,) are not all finite.
IIowcver, a t lemt in good cases, it is still possible to define an Ruler
characteristic. The following seems to be a natural generalization of the 1. Basic Notations and Necessary Results
usual notion: Suppose that we have a scqucncc of cohomology groups
H" i = O,1,2,. . . with the following properties: We retain the notations and definitions of the introductiorl, and recom-
mend Milne's book [M2] as a general refcrcnce for facts mentioned here
(a) The H~ are zero for i large. without proof. If M is an abclian group, Mt,, denotes the torsion subgroup
(b) The Hi are finite for all i except i = a , i = a + 2. and Mcot0, thc quotient group of M by the maximal divisible subgroup.
fla is a finitely generated abelian group. We denote by M , the set of elemcnts in M killed by n, so M(1) = U,,,Mlm
(c)
Ila+2 is the dual of a finitely generated abclian group Da+2. is the 1-torsion subgroup of M. AX is the group of units of the ring A, and
(d) #S is the cardinality of thc finite set 8.
(e) There is a nnlurd pairing ( ,) from D ~ X+ Ha ~ -+ Q.
If U is a prcscheme ovcr X, then the sheaves G,, G, and p, associate
with U the groups r ( U , Ou), l'(U, O U ) ~and
, r ( U , O[,),X respectively. The
Then we may define the Euler characteristic in the usual way, excepta+2 that obvious sequence of sheaves on X
#!la is replaced by #(Ilk,),#llat2 is replaced by #(l[:z,) = #(Dt,, ),
and we add a regulalor term R(C.,,) = {det(hi, dj)}(-')'+l, where { I r ; ) and
{ d j } arc bases of H a modulo torsion and D ~ modulo
+ ~ torsion, respectively.
In our case, a = 1, and H 3 ( X , G,) ought t o be, up to a finite gronp, the is exact in the &ale topology if n is prime to p and exact in the flat topology
dual of the finitely generated abelian group consisting of 1-cycles modulo for any n.
numerical equivalcrlce. Let G be the Galois group of ovcr k. Then G is isomorphic t o 2
In this paper, we show that, and has the Frobenius automorphism n of Elk. a q a canonical topological
germator. If M is any topological G-module, we denote by MG (rcsp.
(*) is true for X of dimensiorl O or 1. M c ) the kernel (rcsp cokrr~lel)of the ho~nonlorphisr~~ (o - 1): M - M.
(1)
176 S. LICHTENBAUM ZETA-FUNCTIONS AT s=l

So NO(G,M ) = M G and if M is a torsion module, II1(G, M) = MG Proof. Starting with the exact sequence of sheaves on X:
arid EIi(G, M) = 0 for i > 1. It follows from ttic IIochschild-Serre spectral
sequence in flat cohornology that we have the exact sequence of finite groups

and taking cohomology, we obtain

Here II:(X, F) denoks flat cohoniology. Let H'(x, F) denote Ctale cohornol-
ogy. Recall tliat IIi(X,G,) = II)(X, Gm) and I I i ( X ,p,) = $ ( X , Pn)
Taking inverse limits with respect t o n, we obtain:
if p y n . Let 1 be any prime nunibsr, including p. l.et H>(X,Tl(p)) (resp.
I ~ ; ( ~ , T ~ (=~ ) ) B j , ( ~ , ~ (resp.
~ n ) H ) ( X , p l n ) ) . Since inverse
limit is an exact functor on the category of finite groups, we have

Starting with (1.1) and taking cohomology, we find

Taking direct limits with respect t o k, we get:

Taking projective limits, we obtain

which gives rise to:

where the notation is self-explanatory. Now assume 1 f p. As Tate ex-

plains in [TI] anti [T2], (P;(Xlq-'T)Il = (Qi(X, T)(l,where Qi(X,T) =
det(1 - uiT) and ui is the automorphism of H'(X,T~(P)) induced by
Frohenius. So 1 is a root of Qi iff q is a root of I';, which by the ltiemann By the Riemann hypothesis multiplicatiori by (ai-1) is a quasi-isomorphism
hypothesis can only occur if i = 2. It follows that if i # 2, 1 - U i on lIi(X,Ti) @ Ql/Zl for i # 2, and since I f i ( X , Tl) @ Q l / Z 1 is divisible,
is a quasi-ison~orphisrnof the finitely generated Z1-module Ili(X, ( H i ( r l TL)@ QI/ZI), =- 0 for i # 2. If i # 1, H ~ + ~ ( X ToG
, is torsion, so
i.e., ker(1 - a;) arid coker (1 - ui) are both finite. Letting ki = k q = 1 equal to Ii+'(X,l i ) ( l - t o r ~ ) ~ IP;(~-')~~'
By. definition of Pi, we have =
#ker(l - ai) and ci = C,J = #coker(l - a,), we have (Pi(%,q-')Il = kicT , # ( I I ' ( ~ , Tt) @ Q ~ / z I which
) ~ , completes the proof.
where I is the 1-adic absolute value normalized SO t h a t ( 1 11 = I - ' . If
we let biVl= #II1(X, it follows from (1.3) that b,,~= kici-1 for i We denote by NS(X) (resp. NS(W)) the group Pic(X)/ Pico(X) (resp.
different from 2 or 3. P i c ( x ) / 1 2 0 ( X ) )of divisors modulo algebraic equivalence. NS(X) modulo
torsion ( N S ( x ) ) is well-known t o be the group of divisors modulo numerical
Lemma 1.1. For all i # 1 , 2 we have: equivalence. Starting with the exact sequence
178 S. LICHTENBAUM ZETA-FUNCTIONS AT s=l 179

and taking Galois cohomology, we obtain: let T: Z i - - + ZI be a Zl-linear m a p with characteristic polynomial P ( X ) .
A s s u m e that the minimal polvnomial M ( X ) of T contains X t o at m o s t
the first power, i.e. that T is "semisimple at 0". T h e n the order of the
torsion subgroup of the cokernel of T divides P*(O).In particular, for all
where the last equality is Lang's theorem. We have the commutative but Jinitely m a n y primes 1, the cokernel o f T is torsion-free.
diagram
Proof. If T @ Q1 is invertible, the cokernel of T is torsion, and its
order is equal t o the determinant of T, up to an 6-adic unit, and so divides
the constant term of the characteristic polynomial of T. If T @ Q1 is not
invertible, look a t the diagram

and also, if K and R are the function fields of X and X,

where T' is T I Im(T). Since T @ Q1 is "semisimple a t On, T' is injective,

herice T1 @ Q1 is invertible. The snake lemma shows that
where P ( X ) denotes the group of principal divisors of X, and the bottom
row is exact by Hilbert's Theorem 90. Also, we have

is exact. Since ker(T) is torsion-free, C ~ k e r ( T ) ~ ,5

, C~kcr(T')~,,.But the
order of Coker(T1)t,, is the determinant of TI, so divides the constant term
P'(0) of the characteristic polynomial of T'.
1.3 We now need some remarks about regulators. Let M and N be two
I I 1 ( G , P ( T ) ) = 0 because of IIilbertls Theorem 90 and 1 1 2 ( ~ , %equalling
*) free finitely generated Zl-modules, and assume that (,) is a non-degenerate
0. Diagram (2) then shows P(X) = P ( X ) ~ so , diagram (3) shows Pic(X) I Zl-pairing from M X N t o Ql. This obviously extends to a non-degenerate
~ i c ( X )so~ diagram
, (1) shows t h a t NS(X) may be identified with its image
pairing from ( M @ Ql) X ( N @I Q I )to QI.
N S ( X ) ~in NS(X). If M' and N ' are two other free finitely-generated Z1-modules contained
If we let A ( X ) be the group of 1-cycles modulo numerical equivalence, in ( M @I Ql) and ( N @I Ql) respectively, we may restrict to get a pairing of
the pairing A(X) x l%s(X) + Z is nondegenerate, i.e. we claim that if M' X N ' into Q1. If we choose bases { a ; ) , { a : ) , {bj) and ( 6 ) of M, MI,
x is a 1-cycle such that (x . y) = 0 for every divisor y on X, then if z is a N arid N', then we may define Reg(M, N ) = det(a;, bj), Ileg(M1, N') =
divisor 0x1 X, (2,z ) = 0. Observe that z comes from a finite extension X' det(a:, bi.), where these regulators are defined up to a unit in Zl. We may
of X and (x, z) = (x, za) for all u implies define the index of M in MI, (MI : M ) to be (MI : M1')/(M : MI'), where
M" is any submodule of finite index in both M' and M. This is clearly
independent of the choice of MI', and we have
We conclude this section with some algebraic results.
Reg(M, N ) = (MI : M ) ( N 1 : N) Eleg(M1,N').
L e m m a 1.2. Let P ( X ) be a monic polynomial of degree r i n Z[X].
Write P ( X ) = x * P * ( x ) ,where P*(O)# 0. Let 1 be a prime number and
 ZETA-FUNCTIONS AT s=l 18 1
S. LICHTENDAUM

shows that H i ( X , PI) maps onto lZi(X, G,)[. Since the cohoniological
2. The Cohomology of Gm dimension of X at 1 is 5 2d +
1 (see [M2], Chapter VI, Corollary 1.4),
H i ( X , G,)l and hence I l i ( X , Gm)(l) = 0 for i > 2d+l. On the other hand
In this section we investigate the relationship between the behavior of we have (by [M2], Chapter VI, Remark 1.5 b) and c)) that H ; ( X , ~ ~= )0
our zeta-function Z ( X , T) as 2' + q-', and the &talecohomology of G,. +
for i > d 2. Therefore the Kumrncr sequence for flat cohomology shows
Let 1 be a variable prime different from p. We begin with the following +
t h a t Ili(X,G,)(p) = 0 for i > d 2, which is <
2d + 1 if d > I. (If
-
proposition: d = 0, ZZi(X,G,) is well-known to be zero for i > 0).
c) It follows from (1.5) that H'(x, T~(P))maps onto T ~ ( I I ~ X ,
G,,,)),
Proposition 2.1. a) H'(X, G,) is a torsion group for i # 1. and Z I ) ( X , T ~ ( ~maps) ) onto T~(IZ~(X,G,)). If i > 3, we have seen
b) Hi(X,G,) = 0 for i > 2d 1. + t h a t ffi(x, ~ ( p ) is) finite and hence ~;(Iz'(x, G,)), being torsion-free, is
c) H'(x, G,) is finite for i # 1,2,3. zero. Since fZi(X, G,)l is finite, H ~ ( xG,)(1), must be finite. Similarly,
d) Let the order of fZi(X, Gm)(l) for i 2 4 be denoted by r q , and let the fZi(X,G,)(p) is f nite.
partial Ruler characteristic .. be X ~ J .Then +
Looking now at (1.5) with i replaced by i 1, we conclude t h a t

so r;,l = bi+l,l. Since bi+l,l = ci+lki, it follows that r;,l = 1 unless

e) IZ1(X,G,) is a finitely-generated abelian group whose rank we will either lIPi(X, q-l) or there are torsion elements in either of the two finitely-
call pl. generated 21-modules Hi(W, ZI) or Hi+'(X, Zl). The first restriction only
rules out finitely many primes, and 0. Gabber has shown ([GI) t h a t so does
f) If i = 2 or 3, H i ( X l Cm)(l) = CiTr
$ ( Q l / Z l ) " i - ~ ,where Ci,l is a finite
the second. So ri,l is finite and equal to 1 for alniost all 1, which implies
group and pi,^ is a non-negative integer. Ci,{ = 0 for all but finitely many 1.
that I i ( x , G,) is finite.
d) Again we use (1.5). Since we have seen in c) that Hi(X,G,)(l) is
Proof. a) If x is a point of codimension 1 on X, let i, be the inclusion
map of z into X. Let j be the inclusion of the generic point Spec K of X
finite for i > 4, we have Tl(Hi(X, G,n)) = 0. We conclude that
into X . Then we have the exact sequence of sheaves in the ktale topology:

for i 2 4, and hence ri,l = bi+l,l for i > 4. It follows that

where the direct sum is over all points of codimension 1. We have the
Leray spectral sequence for j. : IP(X, Rq j*G,) + BP+q(K,G,). Sincc
the sheaves RQj,G, are torsion for q > 0 and since f I Q ( K ,G,) is torsion
for q > 0, we see that ZIP(X, j,G,) is torsion for p > 0. Similarly the
Leray spectral sequence for (i,),Z shows that ZIP(X, (i2),Z) is torsion for
p > 0. Then the long exact squence of cohomology coming from (2.1) shows But c4 = k41P4(X,q - l ) l ~ l , SO we are done.
that ZP(X, 6,) is torsion for p > 1. H O ( X ,G,) = k* is of course torsion.
e) is well-known.
b) For 1 prime # p, taking cohomology of the Kummer sequence
f) Let i = 2 or 3, and look a t 1Ii(X,Gm)(l). This is a countable
I-torsion abelian group with only finitely many elcnierits of order 1. By the
182 S. LICIITENBAUM

structure thcory of such groups ([K], Thcorems 4 and 9), II'(X, Gm)(l) = [ T I ] , t h a t (iv) implies (iii). So we must show that (i), (ii) and (iii) imply
( Q ~ / z ~ ) ~@( 'C;,
~ ~wherc
) p(i, 1) is a non-ncgative integer and Ci is a finite (iv). I,et C ( X ) (resp. C ( X ) ) be the free abclian group generated by the
I-group. Now look at (1.3) with i = 4. This yields one-dimensional intrgral subscl~emrs(curves) on X (resp. X). As Tate
discussed in his Woods Hole talk [T2], we have a hornornorhisrn 4 from
0 -+ lirrl
t 1 1 3 ( x ,G,,)/ln 4 H?(x, fi(/i))3 2i(H4(X1
Gm)) 4 0. C(W) t o 1 1 ~ ~ - ~ '(~fjf(,~ ) @ ( ~ - ' )which
), we will abbreviate by Il"d-(m.
n
(Tate states only that his map lands in I!2d-2(X) @ Q1, but i t is easily
Since C: -- lim II"X, Gm)/ln, the proof of c) suffices t o show t h a t if X seen t h a t in fact the map 4 factors through I12d-2(X).)
t
We see t h a t 4 induces a map dl frorn C ( X ) @ Z l to ~ I ~ ~ - ~ (I'oincarE
W ) ~ .
n
has finite torsion, C: is trivial for almost all 1, arid so C 3 = x,c: is a duality induces a map 0 frorn I12d-2(X)G t o I I O ~ ~ ( I ~ ?;(p))", ~(X, zl)
finte group. and the natural map from NS(X) to I I 2 ( x , ~ 1 ) "induces a map $ from
~ I O I ~ ( I I ~~i)",( ~ , ~ 1 t)o JIO~I(NS(X) 8 Zl, ZI). Diagrammatically, we
In order to proceed further, we must (and do) assume henceforth t h a t z
have:
for a fixed prime 1 # p, the I-component of the cohomological Brauer group
112(X,G,,) is finite. We prove in fact t h e rollowing theorem:

T h e o r e m 2.2. The following statements are equivalent:

Letting u be the composite ?I, o 0 o 41, we know that a is induced by the
intersection pairing on C ( X )X N S ( X ) by the compat,ibility of intersection
(i) 112(X,Gm)(l) is jinite. product and cup product. Since N S ( X ) is, up to torsiorl, divisors modulo
(ii) The mop h: N S ( X ) @ z ~ - + H ~ ( x , T ~ ( ~is) )LIjective
~ nu~ncricalequivalence, we see that, rriodulo torsion, u is onto and its ker-
nel is C n ( X ) @I Zl where C n ( X ) is the subgroup of 1-cycles numerically
(iii) The rank p1(X) of N S ( X ) = r k z , ( ~ ' ( X ,T~(C))" equivalent to zero. Since ?I, is bijcclive modulo torsion by assumption, we
(iv) p l ( X ) is the multiplicity of q as reciprocal root of the polynomial see that 0 is onto n ~ o d r ~ ltorsiorl.
o TTowever, by Poincark duality, and the
P2(X1T). compatibility of cup-product with the action of G, we have
(v) P 2 ( X , T) -
a x ( l - q ~ ) p l ( as
~ )T -+ q-l, with

which is obviously equal 'to r k z , ( l 1 2 ( r , l ; ) C )SO

. 0 also must be bijective
(Rl(G,) is the determinant of the natural pairing between I13((X, Gm)(l)* 3
modulo torsion, and C ( X ) @ Z1 112 d - " ~ ) G is surjective modulo torsion,
and ( N S ( X ) @ Z1), when both gro,ups are taken modulo torsion.) i.e., the Tate conjecture is true for curves on X.
(vi) Z ( X , t ) -- C x ( l - q ~ ) - p l ( ~
as) T -+ q-', where Let D ( X ) be the iriiagr of d l , Ict h be the inclusion of B ( X ) in 112"-2(x)G1
and let A ( X ) be thc group of curves modulo numerical equivalerlce. Then
we have the two exact sequences

(The reader will recognize this as a generalization of first part of Theorem

5.2 in [Tl]. We will get to the second part later).

Proof. Tate's proof in [TI] of the equivalence of (i), (ii), and (iii) works where I;i and Dl are finite groups. We see by looking a t t h e sequence (1.1)
perfectly well in this case. (See also [%I,
p.218). I t is also clear, as in and taking direct limits that we Iiave
 ZETA-FUNCTIONS AT s=l 185

(If d = 2, this reduces to (5.12)). We have already defined h, Tate's

definition of g makes sense i11 our context as well, and f is induced by the
identity on H 2 d - 2 ( z ) G . The map e is induced by the intersection pairing

and since Ig2(X,G,) is torsio11, I 1 2 ( ~ , p ( 1 ) ) H ~ ( xGm)(l).

, Now the
IIochschild-Serre spectral sequence yields: which we have scen in $1 is non-degenerate. Let {C;)(resp.{Dj)) be a
basis of A(X) (resp NS(X) modulo torsion). By Tate's lemma z.1 e is a
quasi-isomophism, with

By Poincark duality, we may identify the dual of ~ I ~ ( x , p ( l )with

)~
H~~-~(X)G.
G
is dual to We continue as in [TI] and coxlclude that g* is a quasi-isomorphism and
Again, by Poincard duality, H ~ ( X ,
p(1))

which is finite by the Riemann hypothesis. We have seen t h a t h is a quasi-isomorphism and

Summing up, we have the following diagram:

As in [Tl], the diagram is commutative, so f is a quasi-isomorphism and

z(f = z(e)z(g*)-'z(h)-l.
It also follows by Poincark duality that f is a quasi-isomorphism iff? is,
where j is the map from H2(x, ) ~Z I Z ( ~T, & ) ) ~ induced by the
~ ( ( p )to
identify on I f 2 F lT1(p)). Tatc's lemma 2.4 implies that z(f) = .z(j), as
in [ T I ] f being a quasi-isomorphism implies that pl(X) is the multiplicity
of q as reciprocal root of the polynomial P2(X,T), and that the operator
a2 - 1 on H2(X, acts "semisimply a t zero," i.e., that its minimal
polynomial P ( X ) contains X to at most the first power. This completes
wliere * denotes the Q1/Z1-dual,the row and columns are exact, El,Dl and the proof of (iv).
P1are finite groups. Recall that is in fact (H 3 ( ~ ~1 ( l ) ) ~ ) * .
We now use Tatc's Rourbaki talk [TI] as a model. We also use his talk Notice that (iv) is indcpendent of 1, so H 2 ( x 1G,)(l) finite for one 1
as a source of notations and lemmas. implies H2(X,G,)(l) finite for any 1 # p. Also, letting
We begin with the analogue of Tate's diagram (5.12):
 ZETA-FUNCTIONS AT s=1

We conclude from all this that

Theorem 2.3. Assume that X satisfies the equivalent hypotheses of

Theorem 2.2. Then
We now apply $1.2 to our situation. Let M = A(X) @ Z1, M' be the image
of (I13(x,
G,)(I))* in I I ~ ~ - ~ (@XQ1, ) ~and
,
(i) The integers p3,l of Proposition 2.lf) are all equal to pl(X). (Of
course all the p2,l = 0).
(ii) H2(X,Gm)(nonp) is a finite group.
Dcfine Rl(Gm) to be lIeg(M1,N), and observe that
Proof. (i) The exact sequence (1.5) with i = 3 becomes

so dct(Ci, Dj)Zl = (MI : M)Rl(Gm). A computation from diagram (2.2)

So P3,1 is cqual to the rank of f 1 3 ( x , ~ ~ ( p ) Now
) . looking at (1.3) with
easily yields that
i = 3, we have (since (03 - 1) induces a quasi-isornorphism on lz3(W, T~(,u)),
p3.1 = rkz, (H2(X,T ~ ( , U ) )=) ~ rkz, (H2(X,li(p)))G.But this is caught
between p l ( X ) and the multiplicity of q as reciprocal root of P2(X,T), and
since these are equal they are also equal to p3,l.
after wc remark that #F13(X,Gm)(l)cotor= #(f13(X,~ , ) ( l ) ) i ~We
~ . now (ii) Now use (1.5) with i = 2 to obtain:
see that

which in turn implies that 112(X,Gm)(l) -- H ~ ( x ,TI(^))^^^. (1.3) with

i = 3 yields
So we have completed the proof of (v). By Lemma 1.1, #El = k 4 1 ~ 3 ( q - 1 ) 1 ~ 1
SO

By the result of Gabber previously mentioned, H 3 ( X , T ~ (G ~=) 0)for ~ all

~ ~
but finitely many 1. Lemma 1.2 applies to u2 - 1 acting on I I ' ( ~T&)),,
Now P ( X , Tl),by 1.5 with i = 1, is isomorphic to T ~ ( P ~ C ~ ( XIn
) ) par-
. since we know the characteristic polyno~nialsare independent of 1, and we
ticular, it is torsion-frce, so have seen at the end of the proof of Theorem 2.2 that ( 0 2 - 1) acts "semi-
simply a t zero," (1z2(X,T1(p))Jtor is zero for all but finitely many I, which
implies our result.

We now move on to the statement and proof of our main formula.

Finally,~P~(~-')IT~ = #ZIO(X, Gm)(l). Taking everything together includ-
ing Proposition 2.11), we sce that indeed %(X,7') -
cx(1 - ~ T ) - P ' ( ~ ) Theorem 2.4. Let X be as i n Theorem 2.3. Then
 ZETA-FUNCTIONS AT s= 1
188 S. LICIITENRAUM

A' such that A injects into A' and there are isomorphisms
(i) IRl(Gm)ll = 1 for all but finitely m a n y 1.
(ii) If we define R1(Gm) to be f l , l p l ~ l ( ~ m ) ~then
;l,
h : At @ Zi -+ IIorn(B, Qi/Zi) compatible with the 41 '8.

where Cx = fpVX'(X,
Gm)-l, u is an integer, and x'(X,G,) = L e m m a 2.6. Let A be a Jnitely-generated Z [ l / p ] - m o d u l e . Let B be
a torsion abelian group S U C ~that B(p) = 0 and there exist stbrjections
dl: IIom(B, Ql/Z1) --+ A @ Zl with finite kernel3 K1 such that Kl = 0 for
all but finitely many I . T h e n there exists a finitely generated
Proof. We have seen in Proposition 2.1 that H1(X,Gm)(nonp) is a Z [ l / p ] - m o d u l eA' such that A' surjects onto A and there are isomorphisms
finite group for i > 3 and i = 0, that H 3 ( X , Gm)co~o,(nonp) is finite, 41: Hom(U, Q1/Zl) --+ A' @ Z1 compatible with the 41'3.
and that H 1 ( X ,G,n)(tor) is finite. We have seen in Theorem 2.3 that
l12(X,Gm)(nonp) is finite. Since Cx is a non-zero rational number, and Proof. Lemma 2.6 is straightforward, so we prove Lemma 2.5. We
since formula (2.4) is true after taking I-adic absolute values for each prime immediately reduce to the case when A and Ilom(R, Q1/Zl) are torsion-
1 # p, we must have IRl(Gm)ll = 1 for all but finitely many 1. The truth free. Let A = n l Z P ( A@ Zl). The hypotheses of the lemma imply that we
of (2.4) then follows formally. have an exact sequence
We still would like to interpret Rt(Gm) as an actual regulator, i.e., as
the determinant of a pairing of finitely-generated abelian groups. Since we
are ignoring p in this section, this is too much to ask for. However, we
can interpret Rt(G,) as the determinant of a pairing of finitely-generated
Z [l/p]-modules into Q. with C finite and 4 induced by the 41's. Let B* = H O ~ ( BQ/z(nonp)),
,
Referring back to diagram (2.2) in the proof of Theorem 2.2, we see that
and let A' = { x E B*: some non-zero multiple of x lies in A ) . It is
immediate that A' is a Z[l/p]-module and that we have an exact sequence
0 -+ A 4'
-+ A' 5 C, where 4' and r' are induced from 4 and r. (Use the
fact that A is pure in A.) In particular, we see that A' is finitely-generated.
We wish to show r' is surjective.
where det(Cj, Dj) was the determinant of the intersection pairing bet- By construction A/mA 1 A/mA for any integer m. Since the isornor-
ween curves and divisors on X . Since we know from Proposition 2.1 that
phism factors through A'/mAt, At/mA' maps onto it/mA, which maps
#II3(X, Gm)(l)coto,is 1 for all but finitely many 1, it follows that #El is 1
for all but finitely many 1. Thcorerri 2.4 then implies that #Dl and #El are
onto C I m C . If we let m = # C then C / m C --C and we see that T' is
surjective. It is then immediate that A' satisfies the condition of the lemma.
both equal to I for all but finitely many 1. Interpreting If 3(X, Gm)(nonp)
as the dual of a finitely-generated Z [lip]-module then follows from the next
Remark 2.7. It is now clear that the regulator formed by using
two lemmas.
the finitely-generated Z[l/p]-module dual to H 3 ( X , Gm)(nonp) agrees with
n r f n \
n (ti,).
L e m m a 2.5. Let A be a finitely-generated Z [ l / p ] - m o d u l e . Let B be
n torsion abelian group such that B ( p ) = 0 and there exist injections
Remurk 2.8. Note also that if x(X, G,) makes sense, then x l ( X , G m )
41: A @ Z1-+ IIorn(U, Q l / Z l ) with finite cokernels Cl such that C1 = 0 for differs from it by a power of p.
all but finitely m a n y 1. T h e n there exists a finitely generated Zjllp]-module
 ZETA-FUNCTIONS AT s= 1

3. The Cohomology of G,

In the prcvious paragraphs we have ignored the p p a r t of the zeta-

function, and ot)tained a formula for the prime-to-p part under thc assump-
ti011 t h a t E12(X, G,,,)(l) was finite for one prime 1 # p. In this section we We have P ( T ) = n y L 1 ( l - a i T ) , so P(q-') = @!-,(1- aiq-I). I3y the
propose a conjecture on the p-part, and show t h a t it is true for d 5 2. functional equation, this equals

Conjecture 3.1. A s s u m e that E12(X, G,)(l) is finite for some prime 1.

T h e n H 2 ( X , G,)(p) is finite, and

and since P(1) = h = #Pico(X), Cx = 0

Y-1
1-op
= 5.
On the other
hand, x ( X , Ox) = 1 - g, so x ( G a ) = ql-g. In this case, wc have

We have # I I O ( X ,G,) = q - 1, # H 1 ( x , Gm)tor = h , I 1 3 ( X , G,) = Q / Z ,

so #II"(X, G,n)cotor = I, H 2 ( X , G,) = 0. The dual of I 1 3 ( X , G,) is
Z = I I O ( X ,Z ) and the regulator pairing takes (n, D ) t o deg(nD). Its
determinant is 1, since there always exists a divisor of degree 1. So x(G,) =
h/q - 1, and we are done.

Now let d = 2. We begin with a user111lernrna on duality:

Theorem 3.2. Conjecture 3.1 is true i f d r= 0 and 1, and true under
the hypotheviv that I3r(X)(1) is finite for some 1 if d = 2. Lemma 3.3. Let X be a n algebraic surface. Assume that E12(X, Gm)(l)
i s finite for a n y 1, including the characteriutic. T h e n there is a natural
Proof. We start with d = 0. Then X = Speck, Z(X, T) = pairing from II"X,G,,~) X f14-'(x, G,) into Q / Z such that the induced
P I P ) = 0, 30 X , Q/Z) is a n isomorphism for
m a p from I I ~ ( X ,G,) to I I O ~ ( H ~ - ~ (G,),
i = 0,2,3, and 4.

Proof. The case i = 2 contained in [Ml]. The case when i = 0 will

follow front the case when i = 4. So we assume i = 3 or 4. It follows from
1 f i ( x , G,) = k if i = 0, and is 0 otherwise, SO x(G,) = Q. R'(x, G m ) = lc* Proposition 2.1 that it, is enough to prove that, for each 1, I l i ( X , Gm)(l)
if i = 0, and is 0 otherwise, so x(C,) = 9 - 1. is isomorphic t o IIOIII(II "'(x, G,), Q,/Z,). It again follows from the
Kumrncr sequence for G, that we have
Now let d = I . Then
O -+ l i ~ nH'(x, G,) @ Z/lnZ
C
n
-+ W+l(x,
T[(G,)) -+ ~ ; ( H ' + ' ( x ,(G,)) -t 0.
 ZETA-FUNCTIONSAT s=l 193

It follows from the IIochschild-Serre spectral sequence for p p that wc have On the other hand, by the main theorem of Milne's paper [MI] we have

We see from the description of the G-action on ZZi(f9, T~(G,)) in terms

of the zeta-function of X that Hi(X,T[(G,)) G
and IZi(X,% ( G , ) ) ~ are where the D, form a base for NS(X) modulo torsion, and
finite for i # 2, so lli+'(X,Tl(Gm)) is finite for i # 2,3. In particular
T~(Iz~(x, G,)) is finite, so zero, so H 4 ( X ,Gm)(l) is finite. Hut a(X)= x(X, O x )- 1 + d i m ( ~ i ~c a r ( ~ ) ) .

Since by Lemma 3.2, FZ1(X, G,) is dual to B3(X,G,), it follows that

det(Di. Di) = R(G,). Also, dim(PicVar(X)) = B / 2 . Since Pic(A1bX) =
Pic(X), P l ( X , T ) = P1(AlbX,T) = Pl(PicVarX,T). It is then well
which is dual to lim II1(X, ,Uln), which is easily seen to be EIO(X,Gm)(l). known (see [Mu], pp.180 and 206) that for an abelian variety A, Pl(A, 1) =
.C
n
#A(k). In our case this implies P l ( X , 1) = #Pico(X). Taking into ac-
Let i = 3. H 3 ( X , Gm)(l) 2: ZI~(X,,U(L))is dual to
, , # P ~ c ( X ) ~ , ,= (by Lemma 3.3)
count that # P ~ C ~ ( X ) # N S ( X ) ~=
#I13(X, G,),oto,, and that P4(q-') = (1 - q) = -#EIO(X, G,) =
-#E14(X, G,) by Lemma 3.3, we see that

Since we have assumed that Z12(X,Gm)(l) is finite, T1(H2(x,

G,)) = 0, so
H ~ ( x Tl(G,)
, --
II1(X, G,) @I Zl, and we are done. Cx = * #H1(X,
#fI"(X, C,,J#H2(X, G m ) # N 4 ( X G,)
,
=*-x(Gm)
Gm)tor#IZ3(X,G m ) c o t o r ~ ( G m ) ; ( ~ ~ O ~ ) X(G,)
'

We wish to show that our formula is equivalent to that stated by Tate

in [TI] and proved under our hypotheses by Milne in [MI]. We start with

let Pl(X,T) have deg B = B l ( X ) and reciprocal roots all . . . ,ag. By the
functional equation q 2 / a l , . . . , q 2 / a B are the reciprocal roots of P3(X, T).
Also, P l ( X , T ) has integral coefticients, so if cyi is a root so is E; = q/ai
by the Riemann hypothesis.
We have
p1(x,
q-l) = n ( 1 - q-lai) = n(1- a;') = na?'n(aj - 1)
= fq - B / 2 ~ l ( l ) .

Also, P3(X,q-') = n(l - q a ~ ' )= n(l - a j ) = Pl(1). So

 References Canonical Height Pairings via Biext ensions
Deligrie, P. La conjecture de Weil, I . Publ. Math. 1.IH.E.S. 43 (1974),
273-307. B. Mazur and J. Tate
Gabber, 0. Sur la torsion duns la cohomologie 1-adique d'une
varie'te' (to appear).
Kapiansky, I. Infinite Abelian Groups, University of Michigan Press, To Z.R.Shafarevich
Ann Arbor (1954). The object of this paper is to present t h e foundations of a theory of
Milne, J.S. O n a conjecture of A r t i n and Tate, Ann of Math. 102 padic-valued height pairings
(1975)) 517-533.
Milne, J.S. Etale Cohomology, Princeton University Press, (*I A ( K ) X A1(K) -+ Qp,
Princeton, (1980). where A is a abelian variety over a global field K , arid A' is its dual. \Vc
Mumford, D. Abelian Varieties, Oxford University Press, London say "pairings" in the plural because, in contrast to thc classical theory
(1970). of R-vducd) canoriical height, there may be many canonical p-adic valued
Tate, J. O n a conjecture of Birch and Swinnerton-Dyer and a pairings: as we explain in 5 4, up to nontrivial scalar multiple, they are in
geometric analogue. Semiriaire Bourbaki no. 306, 1065-66, W.A. one-to-orlc correspondcncc with Zp-extensions L / K whosc ramified prinws
Benjamin Inc. (1966). arc h i t c in nlin~bcrand are primes of ordinary reduction (1.1) for A.
Tate, J. Algebraic cycles and poles of zeta-functions. Arithmetic Wtwn 4 also has good reduction a t the primes of rarnificatiori for LIK,
Algebraic Geometry, IIarper and Row, New York, (1965). then a direrent method, introduced by Schneider (cf. [22] for the case of tlie
Zarchin, Yu. G. T h e Brauer group of abelian varieties over finite cyclotomic Zp-extension) enables one to associate to t / K a p-adic valued
fields, (in Russian) Izv. Akad. Nauk. USSR 46 (1980), 211-243. pairing (*). We show this to bc the same as our pairing.
Our method for the construction of the pairing is first to express the
duality between A and A' via the "canonical bicxtension" of (A, A') by G,,
Received June 30, 1982 and then to develop a theory of "canonical local splittings" of Lleute~lsions.
Partially supported by N.S.F. grants Our pairings arc thcn defined in a manncr analogous to Blocti's defir~ilion
of the classical R-valued pairing. \Vhrreas for Bloch it suffices t o split
Professor Stephen Lichtenbaum certain l o r d cxten:;ions, to obtain uniqueness we must ask for an ezpccially
Dcpartrnent of Mathematics cohcrcwt Snnii:y of splittings of the local extensions, i.e., a splitting of the
Cornell University local bicxtcnsion.
Ithaca, Ncw York 14853 We treat si~nul~aneously the R-valucd and p-atlic valucd theories, and
express our results in a 51niSoi-rn" In:mncr in ternis of the notion of a Y -
valued rar~otiicl~l pairing for n general value group Y satisfying some axioms.
The connection between biextcnsions and hrights is, to be sure, not
surprising.
Firstly, Zarhin [24] pointed out that arbitrary (not necessarily canonical)
splittings of tlie canonical birutension are equivalent to Nkron type pairings
between 7cro cycles and divisors.
S c c o ~ d l ybicxtensions
, have been used to define theta (and sigma-) func-
tions, as is explained of 13rcen [5] and in a manuscript i r ~prtparation by
196 MAZUR AND TATE HEIGHT PAIRINGS VIA BIEXTENSIONS

Norman [19]. Both of these authors point out that, although the concept of 1. Local Splittings
biextensions is riot explicitly mentioned in the theories of p a d i c theta func-
tions of Mumford, and of Barsotti [2] (see also Cristante [6]) it is directly Let IC be a field cornplete with respect t o a place v which is either
related t o these theories (via the theorem of the cube). One might also try archimedean or discrete. If v is discrctc, let 0 = OK denote the ring of
to relate Nkron's approach to p-adic theta functions [16], [17], [la] directly v-integers in K , T a prime element in 0 and k = O/KO the residue field.
to biextensions. If AIK is an abclian variety over K , anti v is discrete, wc denote by A
Thirdly, the theory of p a d i c heights for elliptic curves of complex mul- (or sometimes: AIO) the N h o n model of A/K over O. If v is archirnedean,
tiplication (and p a prime of ordinary reduction) has been developed by we let A denote A / K .
Perrin-Riou [20] and Bernardi [4], using a p a d i c version of t h e sigma- In the non-archimedean case, A0 (or denotcs the connected com-
function. Here the padic sigma-function plays a role analogous t o that of ponent of zero in A, i.e., the open subgroupscheme of whose closed
the classical sigma-function in Nkron's theory [15] for archimedean primes. fiber A; is connected. If U10 is any scheme over 0, its closed fiber TJ X Ok
Nkron has also developed a theory of p-adic valued height pairings us- is denoted Uo.
ing his padic theta functions [18]. In the case of elliptic curves of com-
plex ~nultiplicatiorl,an explicit connection between Nkron's definition and
(1.1) O r d i n a r y a b e l i a n varieties.
8ernardi's has not yet been made (to our knowledge). What is the relation
A and Alfc are c;~llctlordinary if v is discrete, the characteristic of k
(if any) between Nkron's p a d i c height and ours?
is # 0, and the special Gber of A satisfies the following equivalent condi-
Since the explicit expression for the local terms of our canonical p
tions.
adic pairing involves the canonical p a d i c theta functions (of Mumford
and Barsotti), we would find it useful to have a practical algorithm for
computing these functions. In a subsequent paper we will discuss this issue (i) The formal completion A{ of A. a t the origin is of multiplicative
in the case of elliptic curves. type, i.e., is isomorphic to a product of copies of 6& over the algebraic
In this connection, one should also note that the beginnings of a (mod p)- closure of k.
valued theory of height for general elliptic curves (with ordinary reduction (ii) For p = char k, the connected component of the kernel of the
at p ) can be found in ['Ll]. Iiomornorphisrn p: A; --+ A: is the dual of an &ale group scheme over k.
Our construction of padic valued canonical heights requires ordinary (iii) A$ is an extension of an ordinary abelian variety by a torus TA.
reduction at the primes of ramificatiori of the chosen Zp-extension. Can
one find a generalization or replacement of our construction valid for all If L / K is a finite Geld extension arid A/K is ordinary, so is AII, and
Zp-extensions? For elliptic curves with complex multiplication, R. Gross has formation of A0 comrnutcs with the basc change of rings OK -+ 01,.
some ideas on this; see J. Oesterlk, Construction de hauteurs archimkdiennes If AIK is ordinary, then A/K has good reduction over k (equivalently:
et p-adiques suivant la mkthode de Bloch, p. 175-192, in Skminaire de AID is an abelian scliemt~)if an only if TA = 0.
Theorie des Nornbres (SCminaire Delange-Pisot-Poitou), Paris, 1980-81;
Progress in Math., Vol. 22, Birkhsuser Boston, Basel, Stuttgart, 1982. (1.2) E x p o n e n t s .
We are grateful t o M. Artin, L. Breen, J. Coates, L. Moret-Bailly, By the exponent of a finite abelian group G we mean the sniallest integer
P. Norman and B. I'errin-Riou for pleasant arid informative conversa- m > 0 such that mG = 0.
tions concerning algebraic spaces, biextensions, theta-functions, and p-adic In this paragraph, suppose that v is discrete. Let mA = m * / ~denote
heights. the exponent of Ao(k)/Ag(k). Now suppose that k is finite. Let TA denote
We also thank 0 . Gabber for providing us with significant help in work- the "rnaxirnal torus" in A0 which exists by, e.g., [SGA 31 exposk X N , Tlirn.
ing out $5. 1.1. Let nn = n * , ~dcnote the exponcnt of /l;(k)/TA(k).
We refcr to the numbers mA and nA as the exponents of A.
198 MAZUR AND TATE

(1.4) p-splittings
The exponents are sensitive to isogenous change of A.
As for their dependence on the basc field K , m n 1 ~ 1adrnits a finite Let I / , V, W, arid Y be abelian groups, X a biextentiori of (U, V) by W
upper bound for all finite unrarnified extensions K 1 / K , while n ~ l is~ t and p: W-+ Y a hornornorphisrn. A p-splitting of X is a map
independent of K' provided A is ordinary and K1/K is a finite totally
ramified extension. If we drop the assumption that A be ordinary, then
such that
nnlKt adrnits a finite upper bound for all finite totally ramified extensions
K1/K.
(i) $(w + +
x) = y(w) $(x) for w E W, x E X.
(1.3) B i e x t e n s i o n s and p a i r e d a b e l i a n varieties. (ii) For each u E U (resp. v E V) the restriction of to ,X ( r c s p X,)
For an introductiori to the concept of bicxte~ision,we suggcst reading 52 is a group honioniorphism.
and $3 of [12]. For a fullcr treatment of this notion, see expos& VII and
VIII of [SCA 711. A useful and pleasantly written introduction to this fuller (IIcre ,X (resp. X,) denotes the part of X above {u) X V (resp.
treatrrirnt may be found in the first 5 4 pages of $1 of [5]. U X {v)) and is a group extension of V (resp. U )by CV.) Note that we are
Lct A>K denote the dual of AjK arid EfK the canonical biextcnsion of expressing the action of W on X additively. We will continue to do so even
(AIK,AllK) by GmIK expressing the duality ([SCA 7 I] Exposk VII, 2.9). wheu W = Gm.
If v is archirriedean, let En dcnote t h r canonical biextensiori EfK. If v is
discrete, let EA (or ETo) dcnote the canonical biextension of (AO,A') by (1.5) C a n o n i c a l p-splittings.
Ilet Y be an abclian group and p: K * -+ Y a homomorphism.
GmlO, i.e., the unique such biexterision whose general fiber is EfK (whose
existence and uniqueness follow from [SGA 7 I] ICxposk VII, 7.lb).
If BIK is any abelian varicty, to give a bicxtension Elrc of (AIK,BIK) Theorem. There exists a canonical p-aplittingl
by 6,/1c is rquivalent to giving a K-honlomorphisrn A: BIK - - (and
EIK is the pullback of EfK by ( I , A)), or to giving a K-hornornorphism
in the following three cases:
A': -+ AIK (the dual of A). (1.5.1) v is archimedean and p(c) = 0 for c such that lclv = 1.
Again, if v is arcliirnedean, let E denote ElK, while if v is discrele, E (1.5.2) v is discrete, p is unramified (i.e., p ( ~ *= ) O), and Y is uniquely
(or E/o) will dcnote the pullback of E;b, viewed as biextension of (A0, B) divisible by mA.
by Gm/0. (1.5.3) v is discrete, Ic is finite. A i.9 ordinary, and Y is uniquely divisible
The abelim varieties AIK, BIK will be said to be paired, if a bicx- by mAmgnAn8.
tension EIK of (AIK, BIK) by GmIK (equivalently: a K-homornorphisrn If both (1.5.2) arid (15.3) Iiold, they yield the same $ J ~ .
A: BIK -+ is fixed. The: image of $, satisfies the following inclusion relations:
In what follows, we suppose AIK, BIK are paired abelian varieties over
I<, with EIK the bicxtension expressing the pairing. in case (1.5.1)
In all cases, archimedean or non-archimedean, the set E ( K ) of points
of E with coordinates in K is a set theoretical biextension of the groups in case (1.5.2)
(A(K), B ( K ) ) by G m ( K ) = K*. Our aim is to iritroduce canonical split-
tings of biextcnsions obtained from this one via various types of homomor- + 1
mAmenAne P(o*), in case (1.5.3).
phisms p: K * -4 Y.

'The propcrtics eharaclerizing this p-splitling ur~iquelyarc explained in (1.9) below.

200 MAZUR AND TATE IIEIGIIT PAIRINGS VIA BIEXTISNSIONS 20 1

Before beginning the proof of the theorem, we give some lemmas on On the other hand, it is easy to check that this formula defines a p-splitting
set-theoretical biextensions. Let U , V, W and X be as in (1.4), i.e., X a 1C, on all of X, if +,
is a p-splitting of XO.
biexter~sionof (U, V) by W. For integers rn and n we define a map
Another case of unique extendibility of p-splittings is given by

which for each ( u ,v) E U X V takes the fiber ,Xu of X over (u, v) to the (1.7) Lemma. Suppose W' is a subgroup of W and XI a subset of
fiber ,,Xnv. The map (m, 1) is defined as multiplication by m in the group X such that X' is a biextension of (U, V) by W'. Let p: W---+ Y be a
X, for each v E V, the map (1, n ) is defined as multiplication by n in the homomorphism and let = p I W'. A p'-splitting +' of W' extends uniquely
group ,X for each u E U, and finally, to a p-splitting 1C, of W.

defn
(m, n) = (m, 1) 0 (1, n ) = (1, n) 0 (m, 1); Let W = Ui(wi + W') be the expression of W as disjoint union of cosets
of W'. Then X = U(wi + X') is a disjoint union because this is true on
the commutativity of (1, n ) and (m, 1) results from the compatibility axiom each fiber over U X V. If x = w, +
x' E wi X' and + +
extends +', we
for the two laws of composition in a biextcnsion. We have the rules must have
$(XI = P( wi) + +'(x').
and On the other hand, it is easy to check that this formula defines a p-splitting
(m, n)(x + w) = (m, n)x + mn w for w E W. 11, of X, if 4' is a p'-splitting of XI.

If p: W+ Y is a homomorphism and + is a p-splitting, then (1.8) Proposition. Suppose 1J and V 'are compact topological abelian
groups and X a topological biextension of (U, V) by W, such that the projec-
tion pr: X + U X V has local sections. Then X has a unique continzious
In particular, if Y is uniquely divisible by m and n, then we have splitting 1C,: X ---t 08.

Since the space of points of a projective variety over a locally compact

field is compact, an immediate consequence of 1.8 is:
This leads to
(1.8.1) Corollary. Define v: K *-+ R by v(x) = - loglxl. If K is
(1.6) Lemma. Suppose U0 and V0 are .subgroups of U and V, and that locally compact, then E ( K ) has n unique v-splitting which is continuous
m and n are integers > 0 such that mU C U0 and nV C VO. Let X0 be from the v-topology in E ( K ) to the usual topology in 08.
the part of X lying over IJO X V0 and let p: W---+ Y be a homomorphism
of W into a group Y which is uniquely divisible by mn. Then a p-splitting
+,: X 0 -+ Y extends uniquely to a p-splitting 1C, of X. Proof of 1.8. Since the projection X -, U X V has continuous local
sections and its fiber is real affine 1-space, we can use a partition of unity
Indeed (m, n ) X C XO.Hence, if + extends 4, we must have of U X V to obtain a continuous global section 8: U X V +X.
Unicity: If and q2are two splittings, then (+1-$2)0s is a continuous
biadditive map of U X V into W. The image of such a map is bounded and
closed under multiplication by 2, so is 0.
204 MAZUR AND TATE IIEIGIIT PAIRINGS VIA BIEXTENSIONS

canonical p-splitting $,. For x E E(K) there is an integer v such that Let p, p' : K * -+ Y be homornorphismv and c, c' E End Y. Then
(mA,1)2 E ?r" + +
E(o); writing (mA,1)x = nu y, we have

( n ~~ , B ~ B E) Et(6)
Y

and then $,,(x) is given explicitly by the formula

(1.10.2) Change of field.
Suppose u: K --+ L is a continuous homomorphisms of local fields. Let

be a homomorphism. Then
If we arc simultaneously in case (1.5.2) and (1.5.3), then p o $0 = 0, hence
the $, we have just defined in case (1.5.3) does indeed coincide with that
defined for (1.5.2). This concludes the proof of 1.5, or rather its reduction
to our result (5.11, 5.12) on the existence of unique splittings of certain i.e., the diagram
formal group scheme biextensions. E(K) 5 E(L)

R e m a r k . At the cost of increasing n~ and n g a bit, one can avoid some

of the technical complications of 5.11.1 by replacing the "toric completions" is commutative. Indeed, the right side has the characterizing properties of
At, B t , Et by the formal completions A f , Bf, Ef of A, B and E a t the zero the left, in each of the three cases.
points of their special fibers. Then Ef is a biextension of ( A f ,B f ) by 6f, in
the category of formal groups over 6, and the existence of a unique splitting (1.10.3) Change of abelian variety.
for it is given by a proposition of Mumford; see (5.11.5). But then instead First, let A1, TII be abelian varieties over K and f : A1 -+A, g: B1 ---+ B
of nA, n~ one must take integers NA, N g such that K-homomorphisms. Let El be the biextension of (A1,B1) by GmjK ob-
tained from E by pullback via ( f , g), so that we have a commutative
diagram
El - , E
9
(1.10) Functorial properties of qbp. 1 1
In this paragraph, without stating it explicitly, we suppose that, in each fXg
A,XBI A X B
situation considered, the canonical p-splittings which occur in formulae are
defined, i.e., that we are in one of the three cases (1.5.1-3). If no indication Let p: K * --+ Y be a homomorphism and $, (resp. $1,) the canonical
of a proof is given, it is because the stated formula follows immediately p-splitting of E (resp. of El). Then
from the unique characterization of canonical p-splittings given in (1.9).

(1.10.1) Change of value group, and linearity i n p.

Let p: K * --+ Y and c: Y -+Y' be homomorphisms. Then (1.lO.4) S y m m e t r y .
If E is the biextension pairing A and B, then its "mirror-image" 'E (cf.
[SGA 7 I], exp. VII, 2.7) is a biextension, pairing U and A. Morcover, there
 HEIGHT PAIRINGS VIA I~IEXTBNSIONS 207
206 MAZUIt AND TATE

his pairing, Schneidcr makes use of "local spIittingsJ1obtainetl by considera-

is a car~onicalidentificatiorl of sets 'I;;(K) = E ( K ) . Any p-splitting of li: is
tion of local universal norrns. In this paragraph we invest,igatc the corincc-
a p-splitting of ' E .
tion between the canonical biextension splittings of (1.5) and Schneider's
splittings.
(1.10.5) Trace.
Suppose L > K is a finite extension, which we assume Galois for
(1.11.1) Local Zp-extensions. Let K be a complete local Geld with finitc
siniplicity. Let G = Gal(I,/K) and suppose p: L* +Y is a honiomorpl~isrn.
residue field. Let p: K * 4 Q, be a non-trivial conlinuous homomorphism.
Let E(L, K ) be the part of B whicli lies over A(L) X B(IC), and define a
Such a hornornorphisn~cxterids uniquely to the profinite completion of K t
niap Tr: B(L, K ) - l i ( l ( ) as the map which is the trace Bb(L) -t E b ( K )
on the Iibcrs of LC (:B for b E B ( K ) . ' h e n for x E E(L, K ) ,

and by local class field theory, j determines a Zp-extension L/# whose vth
layer K,/K is the unique cyclic extension of degree p" such t h a l

(sum in E~(L)).

(1.11.2) Universal p-norms.

By (1.10.2) and (2.10.1) this equals Fix a p as in (1.11.1). If GIK is any commutative group scheme, let
the subgroup of universal p-norms, G ( K ) C G ( K ) , be defined as the
intersection of the images

for all v.
In particular, if p~ = p for all T E G I then
Examples. If G = G,, then G , ( K ) is the kernel of the homomorphism
p: G,(K) = K *-+ Q,. Tlic group of universal p-norms A ( K ) for our
And if p = 0 0 NLIK, where 0: K *--+ Y , then this becomes abelian variety A is of finite index irl A ( K ) if either AID is ordinary of good
reduction or L / K is unramificd, ( [ l l ] 4.39; and if L I K is unrarnified, 4.2
and 4.3).

whence, if Y has 110 [IJ : K]-torsion, (1.11.3) Biextensions of universal p-norms.

Itecall t h a t E(K,, K ) C lS(K,) is the set of points which project t o
A ( K , ) X I j ( K ) . Define the subset &(K) C E ( K ) t o be the intersection of
the images of
(1.11) L o c a l u n i v e r s a l n o r m s . T r K y l K :E(KV,K ) +E ( K )
In p2] Schneider defined a p a d i c height pairing using an approach
for all u,(cf. 1.10.5).
modelled on I3loch's definition of tlic arc hi me dear^ height p n i r i r ~ gTo define
208 MAZUR A N D TATE IIEIGfIT PAIRINGS VIA BIEXTENSIONS 209

(1.11.4) Lemma. I ~ A ( K )is o f f i n i t e index in A(K), t h e n B(K) inherits and +,

= +p,, the canonical p,-splitting of E(K,). We have by (1.10.5.3)
the structure of a biextension of (A(K),R(K)) by 6,(K) via the natural a commutative diagram:
inclusion
Grn(~) c 6m(K)
1 1
@(K) C E(K)
1 1
A(K) x B(K) C A(K) X B(K). where += $1 is the canonical p-splitting of E ( K ) .

Proof. For each b E R ( K ) we must show that (1.11.7) Claim.

There i.9 a n integer c E Z, such that, for all u,

is exact. This is (essentially) lemma 3 of $2 of [22].

Our proposition follows from (1.11.5) because
Schneider's nletl~odfor constructing his padic analytic height may be
phrased in terms of biextensions as t'ollows:

(1.11.5) Lemma. Let A([<) be of finite index i n A(K1. T h e n there is and consequently by (1.11.6.1), +E(K) is contained in pU-' p(K*) for all
a unique p-splitting of the bieztensions E ( K ) which takes E ( K ) t o zero. Y, hence is zero.
To prove the claim, let 0, = o(K,) be the ring of integers in K,.
Proof. An easy application of (1.G) and (1.7). We refer to the above p- Note that the NCron model of AIKY over OY is, under assumption (a) or
splitting (which exists when i ( is~ ) index in A(K)) as Schneider's
of finite (b), the base-change to 0, of A. Let m u = ~ A I K , , n u = ~ A I K , and
p-splitting. m', = mH/KUln', = ~ B I K , be the exponents of AIKY and UIKY.
By the theorem in (1.5),
(1 . l l . 6 ) Proposition. Suppose that either

(a) I,/K is unramified,

or if L / K is unramified, and
(b) Alo and U10 have good reduction, and are ordinary.

T h e n i ( K ) is of finite index i n A ( K ) and the canonical p-splitting of

E ( K ) is equal t o Schneider's p-splittiny.
if AIK and RIK have good, ordinary reduction.
Proof. By (1.11.2), A(K) is indeed of finite index in A ( K ) . Note that, in case (a), m u is bounded independent of u (by the number
For each u 2 1, let of components of A - . If we are not in case (a), then m u = m', = 1 and
1 ~ )
nu,(resp. n:) is bounded independent of u (by the number of points of A
(resp. B) in the residue field of L). Our claim follows.
210 MAZUR AND TATE IIEIGIIT PAIRINGS VIA BIEXTENSIONS

5 2. Interpretation in Terms of Zero-Cycles and Divisors; where

Relation with Nhron's Canonical Quasi-Functions

(2.1) The s y m b o l [R, D , c]. (2.1.6) For each D and each a0 E (A - Supp D)(K) there is a K-morphism
In tlie next two paragraphs, K can be any field; AIK is an abelian variety
over K , A)K its dual, and LC = BA, cf. (1.3).
Consider the set T of triples (a, D , c) consisting of:
(i) A zero cycle of degree zero, n = zi
ni(ai), a; E A(lC), on AK, all such that
of whose components are points ui of A rational over K. g(a) = [(a)- (ao), D, 11-
(ii) A divisor D algebraically equivalent to zero on AIK, whose support
is disjoint from 8. The properties (2.1.1)-(2.1 .G), for K arid its algebraic extensions, charac-
(iii) An element c E K *.
terize the symbol [a, D , c]. Indeed, suppose [ and [ Iz are two such
symbols with those properties. Let their difference be
A triple (8, Dl c) E I' deterniines a point [a, D , c] E C (K ) in a well-known
mariner, and every point of E ( K ) is of this form. The symbol [a, D, c] obeys
the following rules.
This makes sense by (2.1.1). By (2.1.2), 6(a, D , c) is independent of c, and
(2.1.1) [a, D, c] lies over the point (a,b) E A ( K ) X B ( K ) , where by (2.1.5), it depends only on the class of 1). Choosing divisors D j in this
class such that none of them passes through zero, and whose supports have
defn
a = S(R) = C n;a; an empty intersection, we find using (2.1.6) t h a t there are morphisms
i
and hi: ( A - Supp Di)--+ 6,
defn
b = Cl(D) = Thc point in A1(K) representing
the class of the divisor D . such that
hi(.) = b((a) - (O), Di, 1).

(2.1.2) [n,D,c] -+ c [n,D, 11 for c E K * . These fit together to make a morphism

(2.1.3) Addition in , E ( K ) for a = S(a) is given by

This h is constanl since A is complete, and h = 0 because h(0) = 0. Since

the elements (a) - (0) generate the group of zero cycles of degree 0, this
(2.1.4) Addition in Eh(lC) for b = C1 D is given by shows that b(8, D , c) = 0 for all (8, D , c ) and proves unicity.
A further property of the syrnbol is invariance under translation

(2.1.5) If f is a rational function on AK whose support is disjoint from a

we have if a, and D, denote the images of a and D under translation by a. Indeed,
[a, ( / ) , I ] = [a,0, f (411 the left side has the properties characterizing the right.
 HEIGHT PAIItINGS VIA BIEXTENSIONS 213
212 MAZUR AND TATE

its proof, his symbol is characterized by those three conditions togcther

(2.2) Interpretatio~lsof splittings as pairings between disjoint
with his condition (iv), which states that the function a H [(a) - (a,), Dlp
z e r o cycles and divisors. is bounded on bounded subsets of (A - Supp D ) ( K ) . We have (cf. 2.16)
In this paragraph we recall the connection between p-splitlings and
NCron type pairings; cf. Zarhin 1241, where also conditions are given for
p t o admit some splitting.
Let P be the set of all pairs (A,D ) such that (a, D, 1) E T. Suppose now
A morphism like gao,o takes bounded sets to bounded sets. It sufices
p: K *-- Yiis a homomorphism, and +:
E ( K ) 4 Y is a psplitting of E ( K ) .
therefore to show that our canonical splitting $, is bounded on boundcd
Put
subsets of E ( K ) . In the archimedean case (1.5.1) this is true because
bounded sets are compact and & is continuous. In case (1.5.2)) it is a
Then it is easy t o check that the symbol [A, I?]+, which is defined for consequence of the following lemma, whose proof we leave to the reader.
(A, d) E P, and takes values in Y, satisfies the rules
(2.3.2) Lemma. A bounded subset T of E ( K ) which lies in
(2.2.1) [a, D] is biadditive. EO(#)= U V E Z ( ~+" E ( R ) ) is contained in a finite union of the sets
T" + E(R).
(2.2.3) [a,, Da] = [a, Dl. In cases (1.5.1) and (1.5.2) the canonical p-splitting dp is determined
by +, became p(x) depends only on v(x). Hence, those two cases of the
Conversely, given a symbol satisfying (2.2.1)) (2.2.2), and (2.2.3), we theorem of 1.5 are simply the expression in terms of bicxtensions of I3loch's
+
obtain a p-splitting of E ( K ) by putting interpretation in terms of extensions of N d r o n ' ~theory of canonical quasi-
(2.2.4) $([A, 11,c]) = P(C)+ [a, Dl. functions and his pairing (X,a),. Thus these two cases are not really new;
nor are biextensions essential for them, since the characterizing properties
This is clear, once one proves t h a t +(x) is well defined, i.e., t h a t (2.2.4) is of +, can be expressed in terms of single extensions.
independent of the representation of x as a triple x = [a, D l c]. We leave
the details of this t o the reader. Thus, a p-splitting of E ( K ) is the same a8
a pairing [A, D] satisjying the above three properties.
5 3. Global Fields
(2.3) Canonical pairings.
Now suppose we are in the situation of the theorem of (1.5). In particular, In this section K denotes a "global1' field, by which we mean a t first
we suppose t h a t K is local ay in $1. Define the canonical p-pairing (on P, only this, that there is given a set U of places of K such t h a t each v E U'
with values in Y) by is either archimedean or discrete, and such that, for each c E K C ,we have
[a, DIP = +,?([a, Dl 11) Ic), = 1 for all but a finite number of ,u E U. In particular, the set 21, of
archirr~cdeanplaces in V is finite.
where +p is the canonical p-splitting of (1.5).
For each v E U, let I<, denote the completion of K a t v, and for v U,,
let 0, be the ring of integers in K,, .rr, a uniformizer and k(v) = o,/T,o,.
(2.3.1) Proposition. Suppose v(x) = l o g l x l . Then the canonical
Let AIK and BIIc be abelian varictics and F i E / K a biextcnsion of
v-pairing [a, Dl, coincides with Nkon'a symbol (D, A),, defined in $9 01(151.
( A I ~B, / K ) by 6,/K, i.e., A/K and BIK are paired abelian varieties. For
each v of V, we consider the local theory for AIKu discussed in $1, and let
Indeed, our conditions (2.2.1). (2.2.2), (2.2.3) are Nbron's conditions (i), the symbols A,, E,, for all v E U; A ; , r n ~ * , for v @ U;, T A , , , ~ , + "for v
(ii) and (iii'), and according to NCron's Theorem 3 and his remark (d) after
214 MAZUK AND TATE IIRIGIIT PAIRINGS VIA BIEXTENSIONS

such t h a t k ( u ) is finite havc t11c mcanings explained in thc beginning of 51; For each c E K * we liavc then
and similarly for the abelian varicty BlK.
Let Y be a commutative group, and suppose that we arc givcn a family
p = ( P v ) v E Vof honiomorphisrns p,: K: -+ Y , such t h a t pv(O:) = 0 for all
but a finitc number of v's, and such t h a t the "sum formula" CUE,, p,(c) = 0
+
bccause $,(xu 4- c ) = $,(xu) p,,(c) and Cup,(c) = 0. Thus the right side
of (3.1.1) is indcpcndrnt of the choice of x E E ( K ) ahovc ( a , b). Moreover,
holds for all c E K *. the map E ( K ) -+ A ( K )x 13(K) is surjective, because there arc local scctions
Dcfinc the topological ring AK as the restricted product n',,,
K , where ( E is a linc buridle on A X B, minus its 0-section). IIer~ce(3.1.1) defines a
a vcctor x = ( z v ) v E Vis in AK if x u E 0, for all but a firiitc number of map A ( K ) X B ( K ) -+ Y. The map so defined is biadditivc, because the &,
v E V - V,. Wc have a canonical homomorphism K -t AK . It is convenient are splittings.
to view a family p = (p,) as above, as a homomorphism
(3.2) Definition. If, in the situation of (3.1)) the p,-splitting $, is the
canonical p,-splitting of Theorem 1.3, for each v, thcn the pairing (3.1.1)
which annihilates the image of 0: for all but a finite number of v E V - v, is called the canonical p-pairing, and is denoted by ( , ) p .
as well a s the image of K *.
Suppose t h a t we are givcn for cach v E V a p,-splitting, $,, of E(K,), (3.3) Thus, thc conditions for thc canonical p-pairing
and t h a t $,(Ic(o,)) = 0 for all but a finitc number of v's:

(3.1) Lemma. L e t p = (P,) a n d let (4,) be as just described. T h e r e t o be defined for a family p = (p,) of homomorphisms p,: K: --,Y are
i s a unique pairing (3.3.1) For cach v E V,, we have p , ( c ) = 0 if lcl, = 1.
(3.3.2) Thcre is a finite subset S C V - V , (possibly empty) such that
pv(o;) = 0 for v @ S U V , and such t h a t A, is ordinary and Ic(v) finite for
s u c h t h a t if x E E ( K ) lies above (a, b) E A ( K ) X R ( K ) , t h e n v E S.
(3.3.3) S u m formula CvEV pv(x) = 0 for x E K*.
(3.3.4) Y is uniquely divisible by M N , whcre

where xu E E(K,) i s the i m a g e of x u n d e r the i n c l u s i o n K C K,

Proof. We first note t h a t (3.1.1) is a finitc sum. Indced, our abelian

varieties A, B and the biextcnsion E come from abclian schcnles AIR, BIR
and a birxterlsion E of (Altz, I l l R ) by Gmln for somc finitely gcnerated A homomorphism p: A k -+ Y dcfincd by such a family (p,) whcre (3.3.1-
subring R C K. 4) are satisfied is callcd admissible. This notion dcperids upon thc "globid"
This R is contained in 0, for almost all v by the fact t h a t lclv = 1 for field K with its v, AIK, nlK and Y. If Y is a uniquely divisible group,
almost all v, for c E K*. Wc have A ( K ) = U A ( K ) , the union over all such then the notion docs not depend on 1jIK.
R1s, and for x E E(I1) we have x u E E(o,) for all but a finitc number of v. The values of thc canonical p-pairing are in the following subgroup of Y:
We can therefore define a map $: E ( K ) -4Y by
216 MAZUR AND Tn'l'E HEIGHT PAIRINGS VIA BIEXTENSIONS 217

(3.4) F u n c t o r i a l P r o p e r t i e s . An important corollary of this rule is the following. Suppose A and R

In this paragraph, wc let A , I1 be paired abclian varieties over a. "global1' are abelian varieties over K , A' and B' their duals, and EA(resp ,YB) the
field K , and without staling it explicitly we suppose that, in each situation biextension of ( A ,A') (reup. (13, U')) by Gm expressing the duality. Suppose
cor~sidercd,the canonical p-pairings which occur in formulac are defined, f : A--t B is a homomorphism and j': B'--+ A' its dual. This means that
i.e., that the p's which occur arc all admissiblc. Each subparagraph follows the pullbacks (f X l ~ t ) * aridE ~( I A X f')*EAare canonically isomorphic
directly from its local counterpart in (1.10). biextensions of (A, B') by 6,. IIence for a E A ( K ) and b' E B1(K) we have

(3.4.1) Change of value group, and linearity in p.

Let p: A 2 -+ Y and c: Y -+ Y' be homomorphisms. Then
where the canonical p-pairing on the left is relative to EB and that on the
right is relative t o En.
for a E A(K), b E B(K).
The group of homoniorphisms (3.4.4) Symmetry.
Let 'E be the "mirror-image" of the biextension E (cf. [SGA 7 I] exp.
VII 2.7). Thus ' E is a. bicxtension of (BIK,AIK) by GmIK. Then
which are admissible for A, I? form an End(Y)-module, and we have:

for c, c' E End(Y), (I E A(K), b E B(K). for a E A(K), b E A(K), where the lcft-hand side refers to the canonical
ppairing A ( K ) X B ( K ) -r Y (coming from E) and the right-hand side
(3.4.2) Change of field refers to the canonical p-pairing B ( K ) X A(K) -+ Y (coming from 'E).
I,et a : K -+ I, be a homon~orphisrn of "global" ficlds such that I, is If A is symmetrically paired with itself (by a biextension E of (AlK, AIK)
of finiLc degree over K , and the choscn set of places Vr, for L is the full by GmlK such that there is an isomorphism of biextension E Z 'E) then
"inverse iinage" of the chosrn set of placcs VK for K. the canonical p-pairing
Dcnotr by the same lettcr u the induced rriappi~igA ;( -+ A; and also
the mappings iriduccd on groups of rational points.
Then for an admissiblc p: A; -+ Y we have

is a symmetric bilinear form.

for n~ A(K), b E B(K).
Remark. This is notably the case when A is the jacobian of a smooth
(3.4.3) Chavge of abelian variety. projective curve X I K and the biextension E is the one determined by the
First, let A l l B1 be abclian varieties over K and f : A1 - + A , g: B1-+ B canonical &divisor (cf. [13], $2, 53).
K-homomorph~sms.IJet h:l be the biextension of (AlIK, by GmlK
obtained from E via pullback via ( f , g). Then (3.4.5) Trace.
Let K C L be a finite Galois extension of "global" fields. Assume again
that VL is the full inverse image of VK. We have a natural norm homomor-
phism NLIK: +A ;( compatible with local norms NLWIKv:1,; -* K :
218 MAZUIt AND TATE HEIGHT PAIRINGS VIA RIEXTENSIONS

NLIK: L* -
for v E VK and w E Vr, "lying above" v, and with the global norm
K * . Assume that Y has no [L : K]-torsion. Then for a E
A(L) arid b E B ( K )
Then the canonical p-pairing

is defined. Note that if K is a number field, V the set of all places, and lclv
the normed absolute value a t v E V, then "dividing" the exact sequence
(3.5) E x a m p l e s of c a n o n i c a l p-pairings.
(3.5.1) The Ne'ron pairing: Suppose our places v have absolute values 0 + Re, -+ Wo --* (ideal group of K ) -+ 0
c ++ lclv satisfying the product formula vEVW

by the sequence

units of K ideals
0 -+ -+ K* +
Taking Y = IW and pv(c) = - loglcl, for each v we obtain a canonical roots of 1 roots of 1 of K
pairing
A ( K ) X B ( K ) + R. we obtain an exact sequence

This is Nkron's canonical pairing, as follows immediately from 2.3.1. @v&

0 -* + Y -+ (ideal class group of #) -+ 0.
image of units of K
(3.5.2) A slight refinement of (3.5.1) which uses essentially the same local
splittngs and therefore is in some sense an old story, hut which does not Thus Y is (non-canonically) a product of a finite group, a rcal line, and
seem to have been considered much and whose value we are unable to (card V, - 1) circles.
estimate, is obtained as follows. Suppose that each m, = 1. Let
(3.5.3) Remarks. 1) Y maps canonically t o the ideal class group, so we
get a canonical pairing

be a real vector space with basis elements e, in one-one corresponde~lce A ( K ) X B(K) --* (ideal class group of K).
with the places v E V. Let
This pairing is not new; see [10], [23] for a function Geld version, and for
something in number fields, see Duncan Buell's paper Elliptic curves and
class groups oJ quadratic fields, J. London Math. Soc. (2), 15, (1977), 19-25.
and let Z denote the subgroup of clements of W of the form 2) Let us specialize (3.5.2) Lo the case of K a real quadratic field of class

x
vEV
loglcl,ev, c E K*.
number one, arid AIQ, BIQ paired abelian varieties over Q.
Then Gal(K/Q) operates on A(K), B ( K ) arid Y. Let the superscript
+ or - refer t o the maximal subgroup on which the nontrivial element of

Gal(K/Q) acts as multiplication by +1 or -1. Thus B ( K ) + = B(Q), and

the canonical p-pairing induces a pairing
and for each v E V , put

What is the meaning of this pairing?

 IIEIGMT PAIRINGS VIA BIEXTENSIONS 221
MAZUR AND TATE

Since Q, is totally disconnected and Ztorsion free, pv = 0 for ar-

8 4. p-adic Height Pairings
chimedean v. For nonarchimedean v, p, is unrarnified unless p = char (k(v)),
because otherwise 0: has no infinite pro-pgroup as quotient.
Fix paired abelian varieties AIK, BllC and suppose that K is a global
The homomorphism p (and PGal) is admissible for A, if and only if A is
field in the strict sense, i.e., is either a finite extension of Q or of I;'(T), ordinary a t v if p, is ramified (and if, in the function field case, such places
F a finite field. Let V be the set of all places of K.

(4.1) Let p: A
;( - Qp be a continuous admissible ho~nomorphisn~,
A, K ) so that the canonical pairing
(for
are finite in number).
If PGal is nontrivial, then p ~ cuts
the condition that p ~ factors:
~ l
~ outl a Zp-extension L / K defined by

Such a Zp-exterlsion is called admissible for A if peal is admissible. Thus

is defined. The space of such homomorphisms forms a Qp-vector space a Zp-extension L/K is admissible for A if and only if it is ramified a t only
Vp = Vp(A,K). If p # char . K , then Vp is finite-dimensional. a finite set S of places of K , arid A, is ordinary for each v E S. The
The canonical pairing ( , )p induces a trilinear functional admissible Zp-extensions arc in (1 : 1)-correspondence with one-dimensional
sut)spaces of V,(A, K).

(4.3) T h e d e t e r m i n a n t f o r m .
For this paragraph and the next, let U = A', and E = E ~ .
For a E A(K), b E B ( K ) we have:
By the Mordell-Weil theore~n,the groups A ( # ) and A f ( K ) arc finitely
(4.1.1) generated. They are of the same rank since A and A' are isogcnous.
(a, b),, E P-"- P(AK) C QP
Let the common rank be r and let (P,)l<il,, (resp. (Qj)lljSr) be a
where v is the maximum of the two numbers basis for A(K) (rcsp. A1(K)) mod torsion.
Then for admissible p the determinant
(4.1.2) Ma? (ordPmAe) and Max ( o r d p m ~ , m 4 n
, ~,ne,) defn
vfinrte vES
6(A1P) = detl<i,j<r(Pi, Qj)p
where S is as in (3.3.2). is defined, and is, up to sign, an invariant of A and p. Since ( , )p is
linear in p, the determinant 6(A) is a honiogencous form of degree r on
(4.2) Admissible Zp-extensions. the Qp-vector space V,,. A line in this space, and also the corresponding
Let Zp-extension is called singular (/0r A ) if b(A,p) = 0 for p in the line.
PGal: G K Qp
(4.4) Schneider's p-adic a n a l y t i c height.
be a continuous homomorphisms of the Galois group G K of an algebraic
Supposc, in this paragraph, that K is a number field and that A has
closure of K into the additive group of p a d i c numbers. Then by the
good, ordinary reduction a t all places v of K , of residual characteristic p.
reciprocity law, p ~ , l induces a continuous homomorphism
Thcr, Schneider [22] has defined a height pairing

We say that peal is admissible (for A, K ) if p is.

222 MAZUIt AND TATE IIEIGIIT PAIRINGS VIA BILCX'J'ENSIONS 223

using "local splittirlgs" defined by universal norms in the cyclotomic (4.5.8) P r o p o s i t i o n : Let w E Wo = Wr. These conditions are eguiv-
Zp-extension. Ilcfine the hornornorphisnl p,: GK -+ Q, cutting out the alent.
cyclotomic Zp-extensions by p,(a) = logp(u) where u = u(u) E Z i is defined
by f = <" for all ppower roots of unity I. r.
Elere logp is the p a d i c ( 1 ) The element w is in the image of the natural projection
logarithm.
From (1.11.6) and the definitio11 of Sclmeider's height pairing ( , ), t W7n + Wo.
lirn
we inirnediatcly have: m

(1') There is a family of elements wm E W,, such that N,,,,,w, = Wn

P r o p o s i t i o n . Schneider3 height pairing is equal to our canonical for m >_ n 2 0, and wo = w .
p, -pairing. (2) For each m 2 0 there is un element w , E W, such that N,w, = w.
Schneider conjectures that this height pairing is nondegenerate.
Proof. (1) and (1') are clearly equivalent, and (1') =+ (2) trivially. A
Remark. Schneidrr's construction of a p-adic height pairing generalizes standard compactness argument gives that (2) =+ (1').
kind of y ~ , , : G K -+ Q,; for K any global
immediately to the followi~~g
field.
Call an elernent w E Wo satisfying the above conditions a universal norm.
Let K be a global field.
14.4.1 ) f i e homomorphism PGaZ i s r a m i f i e d onZy a t a f i n i t e s e t If p ~ ~GK l :--+ Q p is a nontrivial continuous homomorphism cutting out
o f vaZuations v of k , and for each such v , A has good ordin- the Zp-extension L / K set r = Gal(L/K) arid let K,/K dcnote the nth
ary reduction. layer, so t h a t I', = Gitl(I>/lCn).
Suppose that All< is an abelian variety; set
Then (L.11.6) again insures t h a t Schrieider's p-pairing coincides with the
canonical p-pairing.

(4.5) G l o b a l u n i v e r s a l n o r m s . The univcrsal norms (for the r-module W) in Wo = A ( K j @ Zp form a

Let r be a topological group isomorphic to Z,, and written multiplica- hp-submodule:
tively. 1,et rn = r p " . Let W be a Zp-module admitting an action of 1' such U P ( ' ~ K ) 4 K )@ a, c
that if W,, = Writ (the fixed subrnodulc under the action of r7&) then W ,
which we rcfer to as the module of universal p-norms. Fix BIK an abelian
is a Zp-module of finite type for each n 2 0, and W = Un Wn. variety and IclK a biextension of ( A , B ) by 6,.
Let
(4.5.2) P r o p o s i t i o n . Let p s a t i s f y ( 4 . 4 . 1 ) . The univeruul p-norms
are degenerate {or the canonical p-pairing:
for integers m 2 n 2 0. Set N , = N,,o
Form the projective limit linl W m ,compiled via the mappings
t
m for all u E U , ( A I ~ ) and /3 E B ( K )@ Zp.

Proof. I3y the functorial property (3.4.3) we may assume that B is the
dual of A and E = E*. For each n 2 0, let a n E A(Kn) @ Zp be such t h a t
224 MAZUR AND TATE IIEIGIIT PAIRINGS VIA BIEXTENSIONS

Consider these two special cases.

CASE 1. r a n k A ( K ) = 1.
wherc p, = p o N K , / ~ . Then the anti-cyclotomic canonical pairing ( , ) p a , K is totally degener-
By (4.l.L), (a,, P),, E & - p n ( A ; ( n ) ,where Y, is given in terms of the ate. What are the universal norms? If (P,P),= # 0 it follows that the
exponents for AIK,, , 1jIK, als in (4.1.2). By the same reasoning a8 in (1.11.7) anti-cyclotomic Zp-extension is the only Zp-extension of K, singular for A.
v,, admits a n upper bound (say N) of n , giving:
CASE 2. r a n k A ( K = 2.
Let & E A ( K ) be such t h a t P, Q generate A(K) mod torsion, and
for all n. U& = -Q for u # 1 E Gal(K/Q).

Writing p = UP, + vPa,K E vp

for (u, V) E Q, X Qp, the determinant,
Remarks. 1. Let K be a quadratic imaginary field, and p a prime 6(Alp), defined in (4.3), is easily seen to be this quadratic form:
number. The anti-cyclotomic Zp-extension of K is the unique Zp-extension
L / K such that L / Q is Galois with nori-abelian (necessarily "dihedral")
Galois group. By considering Birch-IIeegner points on factors of jacobians
of modular curves one may produce (for any quadratic imaginary field K)
examples of prime p, abclian varieties AIQ and finite field extensions K / K
such that if p: GK ---+ Qp cuts out the anti-cyclotornic Zp-extension L/I< which represents zero if (P,&)Pa,K = 0 or if (P, P)pc- (Q, is a square
(where I, = L.K) the Zp-module Up(AIK)of universal p-norms is of positive in Q,. It would be interesting to have some cases wherc ( P ,
rank. ([7], [9]). #0
and (P,P)p, (Q, Q),,, is a nonzero square in Q , . ~In such a case one would
We have no examples where rankzpUp(AIK)is positive, and p cuts out have precisely two Zp-extensions of K which are singular for A. It would
a Zp-extension diITerent from (a base-change of) an anti-cyclotornic Zp- then be especially interesting to investigate the arithmetic of A over the
extension. various finite layers of these singular Zp-extensions.
2. Let Dp(AIK) denote the left-kcr~ielof the canonical p-pairing, where
p is admissible. By (4.5.2)

Although a t present, we have no examples wherc I3 = A', E = EA and

5 5.Riextensions of Formal Group Schemes
where this is a strict inequality, we expect that such examples exist.
3. An interesting sprcial case to investigate is that in which A is an (5.1) R e v i e w OF formal schemes.
Our references for this paragraph are [EGA I] $10 and Knutson's [8],
elliptic curve over Q, with rank A(Q) = 1. Let P E A(Q) be a generator
C11. V, $1. All our rings are assumed to be noetherian (and our schemes,
mod torsion. Let p be a prime of ordinary reduction for A.
formal schemes, ctc., will be locally noetherian).
Let K be a qu~drilticimaginary field and p , , ~ :GK ---t Q p a continuous
ho~nomorphismwhich cuts out the anti-cyclotomic Zp-extension of K.
A (noetherian) topological ring A is called adic if A has an ideal I (called
an ideal of definition) such that the topology on A is the I-adic topology,
Note that Vp(A, K ) is 2-dimensional, generated by p, (cf. 4.4) and p , , ~ .
and A is separated and complete for its topology. Since A is noetherian,
Using the behavior of our canonical pairing under the action of Cal(K/Q)
(cf. (3.4.2)) one sees that
( P I P),.,, = 0.
2 ~ nthis regard, see forthcoming publications of G u d r u n Brattutrh.
226 MAZUR AND TATE HEIGIIT PAIRINGS VIA BIEXTENSIONS 227

A has a largest ideal of definition I (which is the radical of any ideal of If GIs
is a (formally) smooth formal group sche~neover the formal scheme
definition). S such that the structural rnorphisms 9 -+ S induces an isomorphism on
If A is adic, the topological space Sprc:(A/I) (for I any ideal of definitior~) underlying topological spaces, we say that 9 is a formal group over S .
is independent of I, and is called the formal spectrum of A.
The formal spectrunl of A may be endowed with a canonical local adic- (5.3) Examples.
ringed space structure ( 1 , ox), denoted Spf(A) such t h a t there is a canonical Most of our examples of formal schemes can be obtained by completing
isomorphism l ' ( X , o X ) A and such that t h e functor A I-+ Spf(A) is a a scheme X along a closed subschcme Y.
fully faithful functor from the category of adic rings to the category of (5.3.1) Completion along the closed fiber.
local adic-ringed spaces. By definition, an a f i n e formal scheme is a local Let XIS be an S-scheme locally of finite type and set Xn = X x s Sn.
adic-ringed space isomorphic to Spf(A) for some adic A. Let Y = Xols, be the closed fiber.
The topological ring A is called the a f i n e coordinate ring of the afine Denote by 2 the completion of X along Y. Then 2 is a formal 3-
formal scheme Spf(A). schenle. In the notation of ([ICCA I] $10) it is given as an adic (Sn)-system
The category of formal schemes is the (fully faithful) subcategory of local of schemes (XnlsJn.
(adic) ringed spaces "modelled on" affine formal schemes. (Cf. [8], [EGA I] If X), is another S-scheme locally of finite type, we have a bijection
$10.)
Recall that o is a complete discrete valuation ring with uniformizer n .
:,et S = Spec 0, and S, =, Spec o/nn+lo for n > 0. In particular,
So = Spec(o/n~)= Speck where k is the residue field of 0. (5.3.2) Formal completion at zero.
We let 6 dcnote the adic ring o with no as ideal of dcfinition, and
Let XIs be a smooth cornrnutative group scheme of dimension d. Let
3 = Spf(6). eo denote the identity clement of the closed fiber Xols,. Form X f , the
Let XIS be an S-scheme and set Xn = X X s S,, viewed as Sn-scheme.
Let Y C Xo be a closed subscheme. completion of X aL the point eo. Then Xf is a formal group over the
formal scheme 3. Moreover, Xf is an affine forrnal scheme whose affine
If XY denotes the completion of X along Y, then XY is a formal 3-
coordinate ring is isomorphic t o a power series ring in d variables over 6
scheme. Let R be a local 0-algebra separated and complete for the n-adic
with the maximal ideal as ideal of dcfinition.
topology.
Let [l be R viewed as adic local &algebra with ideal of definition n h .
Then Spf h is an afine formal scherne and it makes sense t o consider (5.3.3) Formul completion along subtori.
x'(R), the R-valued points of the formal $-scheme XY. This set may be Let XIS be a smooth (commutative) group scherne, and To C Xo a torus
viewed in a natural way as a subset of X(R), the set of R-valued points over k, contained in the closed fiber. Let Xt denote the formal group scheme
of the S-scheme X. Which subset? Since R is local, one can reduce over 3 obtained by completil~gX along To. Sirice To is an ifline k-group
this question to the case where X is afiinc, where it is easily resolved. sche~ne,a straightforward application of Serre's criterion (also compare [8]
Specifically, if x E X ( H ) , denote by x X s So E Xo(Ro) its specialization to V thm. 2.5) shows that X t is an afline forrnal group scheme over 3.
the closed fiber. Then There are natural homomorphisms of formal group schemes

(5.2) Formal group schemes and formal groups. (5.4) Tori.

If S is a formal sche~ne,a formal group scheme over S is a group-object By definition, a torus (of dimension d ) over any base scheme S is a
in Lhe category of forrnal schemes over S. From now on, formal group group scheme over S which, locally for the 6talc topology, is isomorphic to
scheme Incans commutative formal group scheme. a product of d copies of 6,.
228 MAZUR A N D TATE HEIGHT PAIRINGS VIA BIEXTENSIONS 229

Let ks be a separable algebraic closure of k, and G = Gal(ks/k). Let where ( X n ) t is the formal completion of Xn along To and Yf, is the formal
R be a local 'artinian ring' with k as residue field. If T is a torus over R, completion a t zero in Xn/Tn.
let Tolk denote its closed fiber. The category of tori over R is equivalent Since (Xn)t is canonically (Xt),, passage to the limit gives a sequence
to the category of tori over k, the equivalence being given by passage t o of affine formal group schemes over S:
closed fiber T H To; the latter category is anti-equivalent to the category
of free abelian groups of finite rank endowed with continuous G-module
structures, the anti-equivalence being given by &,,,-duality (cf. [SGA 31,
exp. VIII).
Consequently, given Tolk, for each n >
0, there is a torus Tnlsn given
The sequence is exact in the sense that the rnorphism X t Y f is
formally srnooLh, with kernel 5?. The affine coordinate ring of the 3-formal
-
up t o canonical isomorphisni, such that T, Xs,, So = To. group Y f is a power series ring over 0 in u variables where u is the
The system of tori ( T ~ ~ ~-, )has
, , ~tlle property that T, X sn S, T, codimension of To in Xo.
for n 2 rn, and determines a formal group scheme f which we refer to
(5.5.4) Split tori.
as the formal torus over 3 determined by Tolk. If To is a split torus over k, then the affine coordinate ring of Xt is
isomorphic to an 0-algebra of the form:
(5.5) T h e g e o m e t r y of t o r i c completions.
Let XIS be a commutative smooth group scheme over S, Tolk a torus,
and
PO: To--+Xo
where p is the dimension of To. Since To is split Pic(To) = 0, and hence
a closed immersion of klgroup schemes. Let ?',j. = ( T , , ~ , , ) , ~be
~ the Pic(Xt) = 0 ([SGA 21 exp. XI prop. 1.1, plus the fact that To is afine).
formal torus over 3 determined by TOlk as discussed in (5.4).
For each n 2 0,there is a unique closed immersion (5.6) Riextensions of f o r m a l g r o u p s c h e m e s over 3.
Let A, B , C be formal group schemes over 9 and E a biextension of (A, 8 )
by C over 3. Thus S is a formal 3-scheme which is a C-torsor over A X j. 8,
of &-group schemes extending po (SGA 3 exp. VIII thm. 5.1). locally trivial for the Zariski topology and, moreover, E is endowed with
two group-extensions structures
Invoking [SGA 31 exp VI, we may form the quotient group scheme Xn/Tn.
The sequence

is exact in the sense that p, is faithfully flat (it is, in fact, smooth) with where Ae = A X i B viewed as formal group scheme over 8, and E ( s )
kernel T,. denotes E as 8-formal scheme via its natural projection to B, etc.
The sequence (5.5.1) induces a sequence: The group-extensions ( A ) , (tB) are required to be co~npaLiblewith the
C-torsor structure of h', and with each other (cf. [SGA 7 I] exp. VII).

4 ~ h notation
e is not totally misleading. If X O / G is an abelian variety, there is, in fact,
3 ~ h notation
e is not misleading. This ? is, in fact, the completion along the closed fiber an abelian variety over S whose formal completion at the zero part of the closed fiber
of a torus T over S (determined uniquely up to canonical isornorphism), but this is
is yf. See ([SGA7 I] exp. IX 57). Howcver, we make n o use of this fact.
irrelevant to us.
 HISIGHT PAIRINGS VIA MEXTENSIONS 23 1
230 MAZUR AND TATE

In the specific casc that the C-torsor E admits a global section

(5.9) Biextensions by 6n11S
o : A x g B --+ li' (in the cstrgory of formal 3-schema) the biextcnrion Recall that bm13is thc formal completion of Bmls along its closed fiber
structure on E determines, and is determined hy two 2-cocycles Gmlk. It is determined by the system (Gmls,)nyo.
If E is a biextension of (A, 8) by 6,) thc 6,-torsor l3 rnay be taken to
p a : (A X A)B-+ CB be the corriplen~cntof the zero-scction in a line bundle 1 over A X S 8.
$Jo: (B X B)A+ C A In particular, if Pic(A X 8 ) = 0, as is the casc if A and B are toric com-
pletions At, B t along split tori, any such I3 admits a section a: A X B -+ E ,
and hence can be describcd by the 2-cocyles p a , $Ja as in (5.6).
where pa is a morphism in the category of formal 8-schemes (a 2-cocycle)
determinirig the group-law in tlie usual way on the Ca-torsor
(5.10) Canonical Reductions.
Let U , V be two smooth group schemes over S with connected fibers. Let
T(U)o C Uo, ?'(V)o C Vo denote k-tori contained in the special fibers. Let
Ut, Vt be the completions of U, V along ?'(TJ)o,and T(V)o respectively.
Let
and $Ja is similar. For more details, sce 52 of [12]. The 2-cocyclcs pa,$,
satisfy the relations (a), (b), (c) of $2 of [12].

(5.7) The category of biextensions.

Let 13115XT3(A, 8; C) denotc the category of hiextcnsions (of formal S- denotc the exact sequences of formal group schemes over S as provided by
schcmrs) of (A, 8 ) by C. See ((51, [SGA 7 I] exp. VII) for a discussion of this (5.5.3).
category.
Let 13iexti(A, 8 ; C) denote the group of iso~norphisrnsclasses of bicxtcn-
(5.10.3) Propositions. T h e functor "inverse image of biextensions"
sions of (A, 8) by C , and let Uiext;(A, 8 ; C ) be the group of autorrlorphisms induces a n equivalence of categories
of thc trivial biextension.

(5.7.1) In general, wc have that iliext$(A, 8; C) is the group of 3-nior-

phisrns h: A X B -4 C which arc bilincar in thc sense that tlie induccd Note that this proposition is very close to the statement of Corollary 3.5
mappings of ([SGA 7 I] exp. VIII). Indeed, it is identical except that we are dealing
ha: A B --+ CB; h ~ BA
: --+ CA with formal group schemes rather than group schemes. We shall adapt the
proof of that corollary to our situation.
are hormomorphisms (of formal 8-groups, arid A-groups, respectively). Elere we rely on some of the results and notions explained in [8], (especially
chapter V). In particular, we shall consider the global etale topology of for-
(5.8) Canonical trivializations. mal 3-schenies (as in loc. cit. V $1) and its induced topos, which we denote
These statements are equiva.lent: T . We also shall consider formal algebraic spaces over ?, (loc. cit. I1 $2) and
(5.8.1) 1311CXT3(A, 8 ; C ) is equivalent to the punctual category. group objects in that category (formal group algebraic spaces over 5).
(5.8.2) I3iext$(A, 8 ; C) = 0, for i = 0 , l . We may also consider bicxtensioris for the topos T ([SGA 7 I] exp. VII
52). For exarnpie, 131EXTT(Ut,Vt; 6,) means the catcgory of biextensions
(5.8.3) Every biextensions I3 of ( A , B ) by C in the catc:gory of Q-schemes
of (Ut, Vt) b y 6, iu the topos T , i.e., a bicxtensioris E in this category is
admits a uriique trivialisation (i.e., a lC-splitting $J: I<-+ C; cf. (1.4).
232 MAZUR AND TATE HEIGIIT PAIRINGS VIA BIEXTENSIONS 233

a sheaf in T (and conscquently a formal algebraic space over S ) endowed It then follows from the general fact ([SGA 7 I] exp. VIII 1.5.2) that
with the structure of biextension.
A key step in the proof of (5.10.3) is:

(5.10.4) P r o p o s i t i o n . B I E X T ~ ( ? ( U ) ,Vt; 6,) i s equivalent t o the punc- To show that Biext\(?(U), Vt; 6,) vanishes, we note that (5.10.2) repre-
tual category. sents an exact sequence of sheaves in the topos T and hence our problem
is reduced to the following two vanishing statements:
We prepare for its proof.
(5.10.5) P r o p o s i t i o n . A n y formal group algebraic space over 9 is a
formal group scheme over S.

Proof. Let B be a formal group algebraic space over 5, and, keeping t o But (5.10.7) follows immediately from (5.10.6) and the argument of
the notation of loc. cit., let B1 denote its first truncation (loc. cit. V llefn. proposition 3.4 of. [SGA 7 I], ext. VIII. To see (5.10.8), consider an exact
2.4). Then B1 is an algebraic space over So and it inherits an algebraic sequence of formal group schemes over S.
space-group sLructure from that of 8 . By Murre's theorem ([I], Theorem
4.1, [14]), B1 is a group scheme over So. By (loc. cit. V Theorem 2.5) B is
a formal group scheme.
Since Z f ( k ) = 0, and & -+ Zf is formally ktale, there is a unique lifting
(5.10.1) Corollary. Let G I , C2 be formal group schemes over 9 (com- 7: Zf --+ & sending the k-valued point of Z f to zero in &(k). It is immediate
mutative, as always). T h e n a n y element e i n E X ~ \ ( G ~ , G
is ~representable
) t h a t 7 provides a splitting of (5.10.9).
by a n extension 0 -+ G 2 -+ & -+ G1 -t 0 in the category of formal group Since the "Cartier dual," M = I I ~ ~ ~ ( ? ( u6,) ) , is ktale, one easily
schemes over 5. ) , 6,) = 0. The proposition then follows from
sees that ~ i c x t $ ( ? ( ~ Vt;
Now form (5.8).

(5.10.10)Conclusion of the proof of the proposition (5.10.3).

By ([SGA 7 I] V113.7.6) and the exact sequence (5.10.1), one sees that the
where the underline means as sheaves for the ktale topology over 5. One category U I E X T ~ ( YVt;
~ , 6,) may be identified up t o equivalence with the
easily sees that M,(the "Cartier dual" of ?'(u)) is representable by an adic category whose objects are S-biextensions E of ( U t , V t ) by 6, supplied
(Sn)-system of locally constant group schemes (torsion-free of finite rank). with trivializations of the induced biextensions &' of (T(U), Vt) by 6,.
But N = 0. To see this, we may suppose that k is scparably algebraically Morphisms are defined evidently. By (5.10.4) we then have an equivalence
closed, and (using (5.10.6)) we are reduced to proving that if of categories

is an exact sequence of formal 5-group schemes, then (e) splits. But, for and a symmetrical reduction of Vt t o Z f establishes our proposition.
each n >0, (en) = (e) X S Sn may be seen to be an exact sequence
of group schemes over Sn,and splits by ([SGA 7 I]exp. VIII, Prop. 3.3.1). (5.11) C a n o n i c a l trivializations.
Moreover, a splitting of (e,) determines uniquely a compatible splitting Let U, V, T(U),-,, T(V)o be as in (5.10). Thus Uls, Vls are smooth group
of (en) for a11 n. Thus fi = 0. schemes with co~inectedfibers.
234 M A Z U R AND TATE HEICIIT PAIRINGS VIA IIIEXTENSIONS 235

(5.11.1) Proposition. Suppose t h a t ~f (equivalently: Y J ) i s of rnul- M =h k ( 6,,) ~ ' takes

which , the unique k-valued point of 2f X Zf to
tiplicative type. T h e n HIEXTT(ut, V t ;6,) i s equivalent t o the punctual zero. Sirice M is a projective limit of' locally constant group schernrs, cp,
category. is trivial as well and consequently E is the trivial biextension.

Proof. Using (5.10.3) it sufices to show (5.11.5) A m i l d weakening.

If one is willing to assume that the formal group 21 is of "finite height"
(5.11.2) Lemma. 1f Y f i s ordinary, t h e n B I B X T ~
(Y 2f ;6,) i s punc- which is indeed all that occurs in any s ~ r i o u sapplication lhat we envision,
tual. then (5.11.2) also follows immediately from ([12], 55, Prop. 4).

This is a simple exercise which can be seen in a variety of ways. For

example:

(5.11.3) Proof of (5.11.2): In analogy with the proof of (5.10.3), set

References

PI M. Artin, Algebraization of formal moduli: I, Global A n a l y s i s .

Papers in honor of K. Kodaira. (1970) 21-71. Princeton Univ.
Then M is easily seen to be a projective limit of locally constant finite Press.
(formal) group schemes over 3.
Moreover, 3 is seen to vanish by the prdof.. of (SGA 7 I exp. VIII, Prop.
PI I. Barsotti, Considerations o n t h e t a - f u n c t i o n s , Symposia Malhe-
matica (1970) p. 247.
3.3.1). ~ f, ; 6,) 7ICxt1(2f, M ) and the latter group is seen
Thus ~ i e x t ' ( Z PI S. Illoch, A note o n height pairings, T a m a g a w a n u m b e r s a n d t h e
to be zero by the argument demonstrating (5.10.8). B i r c h a n d S u h n e r t o n - D y e r cozjecture, Inv. Math. 58 (1980),
65-76.
A

Finally, since IIoms(Zf, M) = 0, we see that l4] D. Bernardi, l l a u t e u r p-adique s u r les courbes elliptiques.
Sdminaire de Thhorie des Nornbres, l'aris 1979-80, SCrriiriaire
~ i e x t : ( ~ Z~f,; 6,) = 0 Delange-Pisot-Poitou. Progrcss in Math. series vol. 12. Birkhauscr,
I3ostori-Bascl-Stuttgart. (1981) 1-14.
as well and (5.11.1) follows from (5.8). L. Breeri, h n c t i o n s thc^ta ct the'or2me d u cube, preprint of the
[51
Laboratoire associC au C.N.R.S. No 305, Universitd. de Itcnnes I,
(5.11.4) Paraphrase of part of the above proof, using cocycles. 1982.
From the fact that M is a projective lirnil of locally constant finite [61 V. Cristaritc, T h e t a functions a n d Barsotti-?'ate groups, Anriali
(formal) group schemes over 3, and that N = 0, the reader may verify della Scuola Norniale Superiore di Pisa, Serie IV, 7 (1980) 181-

a: Y f X Z f -
that if E is a biextension of ( y f , Z f ) by 6,) then there is a section
E such that the 2-cocyclc (5.6.2) 1~1
215.
M. Harris, S y s t e m a t i c growth of Mordell- W e i l groups of abelian
varieties in t o w e r s of n u m b e r fields, Inv. Math. 51 (1979) 123-
141.
is trivial. From the relations satisfied by &, cp, (cf. [12], 52 (a), (b), (c))
it follows that p, may be viewed a s an 3-morphisms from Z f X to zf
236 MAZUR AND TATE HEIGHT PAIRINGS VIA BIEXTRNSIONS 237

D. Knutson, Algebraic Spaces, Lecture Notes in Math. 203, [EGA I] A. Grothendicck, dle'ments de ge'ome'trie alge'brigue. Publications
Springer Berlin-Heidelberg-New York, 1971. Mathkmatiques, I. H. E. S. 4 Paris (1961).
P . F. Kurchanov, Elliptic curves of infinite rank over I'-extensions, [SGA 21 A. Grothendieck, Cohomologie locale des faisceaux cohe'rents et
Math. USSR Sbornik, 19, 320-324 (1973). Thebre'mes de LeJachetz locaux et globaux, North-Holland Publish-
J. Manin, The Tate height of points on an abelian variety. Its ing Company - Amsterdam (1968).
variants and applications. Izv. Akad. Nauk SSSIl Ser. Mat. 28 [SGA 31 M. Demazure and A. Grothendieck et al, Shminaire de Gkomktrie
(1964), 1363-1390. (AMS Translations 59 (1966) 82-110). Algkbrique du I3ois Marie 1962/64. Sche'mas en Groupes 11. Lecture
B. Mazur, Rational points of abelian varieties with values in towers Notes in Mathematics. 152, springer, I3erlin-Heidelb&g-~ewYork
of number fields, Inv. Math. 18 (1972) 183-266. (1970).
D. Mumford, Biextensions of formal groups, in the Proceedings of [SGA 7 I]A ~;othendieck et all Skrninaire de Gkometrie Algkbrique du
the Bombay Colloquium on Algebraic Geometry, Tata Institute of Bois Marie. 1967169. Groupes de Monodromie en Ge'ome'trie
Fundamental Research Studies in Mathematics 4, London, Oxford Alge'brique. Lecture Notes in Mathematics. 288, Springer, Berlin-
University Press 1968. Heidelberg-New York (1972).
D. Mumford, A remark on Mordell's conjecture, Amer. J. Math.
87 (1965) 1007-1016.
J. P. Murre, On contravariant functors from the category of pre- Received October 13, 1982
schemes over a field into the category of abelian groups, Pub.
Math. I. 11. E. S. no 23, 1964. Supported by NSF
A. Nkron, Quasi-fonctions et hauteurs sur les varikte's abe'liennes,
Annals of Math. 82, no 2, (1965) 249-331. Professor Barry Mazur
A. Nkron, Hauteurs et fonctions thtta. Rend. Sci. Mat. Milano Department of Mathematics
46 (1976) 111-135. Harvard University
A. Nkron, Fonctions thtta p-adiques, Symposia Mathematica, 24 Cambridge, Massachusetts 02138
(1981) 315-345.
A. Nkron, Fonctions th2ta p-adiques et hauteurs p-adiques, pp. Professor John Tate
149 -174, in Skminaire de theorie des nombres (Skminaire Delange- Department of Mathematics
Pisot-Poitou), Paris 1980-81, Progress in Math., Vol22, Birkhauser Harvard University
Boston-Basel-Stuttgart (1982). Cambridge, Massachusetts 02138
P . Norman, Theta Functions. Handwritten manuscript.
B. Pcrrin-Riou, Descente infinie et hauteur p-adique sur les courbes
elliptiques c i multiplication complexe. To appear in Inv. Math.
M. ltoscnblum, in collaboration with S. Abramov, The Birch-
Swinnerton-Dyer conjecture mod p. Handwritten ma~iuscript.
P. Schneider, p-adic height pairings I, to appear in Inv. Math.
J. Tate, Variation of the canonical height of a point depending on
a parameter. To appear in Amer. J. Math.
Ju. G. Zarhin, Ne'ron Pairing and Quasicharacters, Isv. Akad.
Nauk. SSSR Ser. Mat. Tom 36 (1972) No. 3, 497-509. (Math.
USSR Izvestija, Vol. 6 (1972) No. 3, 401-503.)
 The Action of an Automorphism of C
On a Shinlura Variety
and its Special Points
J. S. Milne

In [8, pp. 222-2231 1,anglands made a very precise conjecture describing

how an automorpllisrn of C acts on a Shirr~urav;trieLy and its special points.
Thc results of Milnc-Sliih [Is], whcri c o ~ i ~ b i n ewith
d the result of Dclignc
[5], give a proof of the cor~jecturc(including its supplcrrient) for id1 Shirnura
varieties of abclian t y p (this class cxclutlcs only those varieties associated
with groups having factors of cxccptional type and most types D). Here
the proof is extentied to cover all Shirnura varieties. As a comxquencc,
onc obtains a complete proof of Shirnura's conjecture oil the existence of
cnr~onicalmodels. The main new ingrrtlicmts in ttir proof are the results of
Kaahdan [7] arid the rnethods o l norovoi [2].
In the preprint [ 7 ] , T<aahdan shows t h a t the conjugate (by an a u t o ~ n o r -
piiisrri of C ) of the quotient of :I Ilcrrnitiar~syn~rr~ctric dorriain by an arith-
~ u c ~ tgroup
ic is a variety o f the same I'orrii. (For a precise staternerlt of what
we u w frorn [7]. sec (3.2).) In s c c ~ i o n s2 and 3 we apply this result to prove
ttic, following weak form of I,anglantls's conjrcturc:
(0.1) let M 0 ( C , X + ) bc thc coriuectcd Shirnura variety dcGr~cdby a
simply-corinected semi-sirriplc idgrbraic group G arid IIermitian symmetric
donlain X+;111cr1, for any auLornorpt~lsrn r of C , there is a connected
Shimlira variety IMO(C', XI+) for which t h e exist compatible isornor-
pllisms
p: TMO(G,X+)-+ MO(G', XI+)
$: CA, -+ G a l .GIA'.

111 121 13orovoi shows Lhat the analogue of (0.1) for non-corinect,cd Shimura
varieties irnplies the existence of c:inoriical models for all Shirnura varieties.
Wc adapt his r r ~ r t l ~ o dtos sl~ow,in scc1,ion 4, 5, and 6, that (0.1) implics
Langlands's conjccturr for all conncctcd Shirnura varieties.
 TIIE ACTION OF AN AUTOMOltPIIISM OF C 241

In the final section we review the main consequences of this result: pair ( + G I+Xi-)also satisfying (C), a special point 'h E +X+,and an
1,anglxntls's conjecture i n its original form; the existence of canonical models iso~norphisnl$, = (g H +g): G ( A f ) - - t r C ( A f ) . (See also [l4]; in general,
in the sense of Delignc; l,nnglandsls conjccture describing the action of we shall use the definitions of [l4] and [15] rather than [8].)
complex corljugatior~ on a Shirnura variety with a real canonical model;
the existence of canonical models in the sense of Shimura. An cxpository T h e o r e m 1.1. Assume that G is simply-connected; then, with the
account of this, and related n~aterial,can be found in [ l l ] . above notationa, there ezists an isomorphism
I am indebted to P. Deligne for several valuable conversations and,
especially, for suggestions that led to the elirniriatiorl of a hypot,hesis on the
congruence nature of a.rit,flmetic subgroups in the statement of t . 1 main ~
theorem. such that

Notations. The notatiorls are the same as [4]. In particular, a reduc-

tive grollp C: is connected with centre Z(G). A superscript -t denotes a
topological connected co~nponent,and C ( Q )+ drnotes the inverse image of
Gad@)+ under G(&) -t Gad(R). The symbol S denotes I L c s ~ / ~ and, G ~ ,
Remark 1.2. This is a weak form of part (a) of Conjecture C o
for any Ilornornorphism h: S -t C , ph denotes the restriction of ho to the
[15, p. 3401. In G we shall see that it leads to a proof of the full conjecture.
first factor in S c = G m X G,.
If (G, X ) and ( G t , X t )arc pairs defining Shimura varieties, then a map
( G , X ) + ((:',Xt) is a hon~ornorphisnlG + G' carrying X into X'. A n Remark 1.3. The real approximation theorem 13, 0.4) shows t h a t
inclusion (TIh ) -+ (GIX) will always mean that T is a maximal torus of G(Q)+ is dense in G(R)+. Therefore, for any x E X+, G ( Q ) + z is (real)
G. dense in X + , and its image in 1'\X+is ~ a r i s k dense.
i It follows that there
If C: is a group over Q , thew C1 = GQ, .By a hornomorphisrrl GAt + GtAf is a t most one map p, satisfying the conditions of the theorem.
we mean a family of homornorphisnls of algebraic groups G1 -+ G{ whose
product rriaps G ( A ~ into) C t ( A f ) . (As usual, Af = ( Z / m Z ) @ Q.) Remark 1.4. Let G be a semi-simple group over Q, and let i: T + G
The closure of G(&) in (;(A/) is denoted by G(Q)-. be the inclusion of a ~naxinlaltorus. Then Aut(G,i) = T % T/Z(G).
We say that a group G satisfies the llasse principle for H i if the map of Fix a finite Galois extension L / Q and consider triples (G1,i', 4) where
it
Galois cohornology groups II'(&, G) -+ rIlIIi(Q1, G) is injective, where 1 (G', T -+ G') is isomorphic to (G, i) over L, and .II, is an isomorphism
runs through all primes of Q including 1 = oo. (GI i)Af 5 (C', (i.e., an isomorphism G A j + G i / carrying iAf into
If V is an algebraic variety over a field k, and T: k + K is an inclusion iaf). Given such a triple, choose a11 a: (G, i ) ~ (GI, if)^ and define
of Gelds, then T V denotes V @,+,, K = V Xsp,, k spec K. /3 = P(G', if,$) E T(A[) by the equation $ o p = a. Let j7 be the image
The 111airldrlinitiorls concerrlirlg connected Shirnura varieties are reviewed of ,8 in T(A[)/T(L). Then (G', i', 4 ) I-+ p(G', i', 4) defines a one-to-one
in an Appendix. correspondence

{isornorphisrn classes of triples (G', i', 4)) H (T(A[/F(L))Gal(L/Q)

$1. Statement of the First Theorem
Consider now (TIh) (GIXt). Using the element B(T,
p h ) ex-
To a pair (C, X + ) satisfying (C) (see the Apperldix), a special point plicitly defined in [14, 3.181, one obtains from this correspondence a pair
h E X + , and an autornorphisnl T of C , 1,;~uglands[8] associates another ('CIri: Tc-r 'G) together with an isomorphism 4,: CAf -+'GAJ carrying
 THE ACTION O F AN AUTOMORI'HISM O F C 243

i into ' i . These are the objects in (1.1). The map 'h: S -t T% 'G is that Proposition 2.1. Let (GIX+) and (G', XI+) satisfy (C), and let
4:
whose associated cocharacter is TP,,, and " X + is the ' ~ ~ ~ ( ~ ) ~ - c o ~ ~ j u g a c ~ GA, -+ G'AI be a n isomorphism such that $(G(Q)+) = Cf(Q)+. For

class containing 'h . a n y finite e'tale morphism p : Mo(G,X+)-t Mo(G', XI+)compatible with
One other fact we shall need concerns the class 7 of ('GI ' i ) in H1(Q, T) 4, there exists a n element g E G(Q)+ and a n isomorphism 4,: G 3 G'
(equal to the image of P(T,p h ) under such that p = M0(+,) o g. I n particular, p is a n isomorphism.

Proof. Choose an h E X+ and write p([h]) = g'[h']l some h' E X'+,

g' E GI(&)+ (see the Appendix). Let g-' = +-'(y'); then

The existence of 4 shows that the image of 7 in H ~ ( Q , , T )is zero, for all
finite 1; its image in H 1 ( R , T ) is represented by ~ ~ ( - - 1 ) / ~ ( - - l[14,
) 3.141.
Remark 1.5. Theorem 1.1 (in fact, Conjecture C O ) is proved in
((51, [Is]) for Shimura varieties of abelian type. Since we shall need to make
use of this result for groups of type A, we outline the main steps in its proof. satisfy the same conditions ay p and 4,and p o y[h] = [h']. It therefore
For pairs ( G , X + ) with G the symplectic group and X+ the Siegel upper sulfices to prove the following proposition.
half-space, (1.1) is shown in [15, 7.171 to be a consequence of a statement
about abelian varieties of CM-type. This statement is proved in [5]. Let Proposition 2.2. In addition t o the hypotheses of (.2.l), suppose there
G be of type A, and suppose that G is almost simple over Q. Then G can +
exist h E X+ and h' E X'+ such that p([h]) = [h']. T h e n is defined over
be embedded into a symplcctic group ([4, 2.3.10]), and the following easy Q and cp = Mo($).
lemma can be applied.
Proof. There exists a unique isomorphism @: X + q X ' + lifting all cpr,,r:
Lemma 1.6. Let ( G , X + ) satisfy (C), and let be a reductive sub-
r \ X+-+ I" \ X'+ and sending h to h'. Let @, be the map
group of Gad. Suppose that some h X+ factors through pRI and let a H @ o a o @ - I : Aut(X+) + Aut(Xt+). For any
X& be the ~ " ~ ( ~ ) + - c o n j class u ~ a containing
c~ the composite h' of h with
-
H -+ zad. A s s u m e that pad satisfies (C3) and let U be the s i m p l y 'cotr-
nected covering group of g a d . T h e n (11,~ &satisfiesj (C), and there is a n @,(a)induces an automorphisn~of Mo(G',X'+); in particular, it lies in
embedding M"(H, x;) L* M O ( G ,X + ) compatible with I I ( A f ) c-+ G(Af) the cornrnensur;~,bilitygroup of any I" E x ( C ' ) and therefore belongs to
under which [h'] I-+ [h]. If h is special, so also is h', and i f (1.1) holds for GfU"(Q)(see [I, Thrn. 21). Consider a q t G(Q)+, and write q, and q~
( G , X t ) and h , t h e n it does also for (H,x&)and h'. for its images iri G O ~ ( Q ) and
+ G(Q)T. Then q, and qf define the same
automorphisnl of M o ( G , X + ) , and so @,(q,) and +(qf) define the same
$2. Morphisms of Shimura Varieties automorphisnl of Mo(G', XI+). They therefore have the same image in
Gfad(Q)+A(relGI) = Gr(Q);+ Gfnd(Q)+. Therefore +(qf) E G f ( Q + )
A morphism cp: Mo(G,X+)--+ MO(G',X'+) will be said to be finite and
@(a)+
(and @,(q,) = +(qf) in G " ~ ( Q ) + ) . As G(Q)+ is Zariski dense in G, we
ktale if, for any r' E C ( G f ) , there exists a I' E C ( G ) such that co~icludethat 4 is defined over Q . Write 4, for 4 regarded as a Q-rational
map, and consider

is finite and ktale.

 Tl1E ACTION 01' AN AUTOMOItPMISM OF C 245
244 J. S. MILNE

It remains to show that this map is the identity. We know t h a t it is

C o r o l l a r y 2.6. Let ( C , X ) satisfy 14, 2.1.1.1-2.1.1.3]; t h e n the canoni-
cal m a p ( G ( A ~ ) / Z ( Q ) - ) , ( ~ , G ~ ~ ( -+Q )A u t ( ~ ( C , x ) )identifies
finite and &tale, maps [h] t o [4y1 o h'], and commutes with the action of
G(Q);. Wc have Lhcrcfore to prove thc proposition in the case t h a t G = G'
and 4 is the identity map. The equality noted parenthetically in the last
paragraph shows that in this case + , ( q ) = q for q E G(Q)+/%(&). The
real approximation theorem [3, 0.41 statcs that G ( Q ) + is dense in G(R)+,
and so @, is the identity map on G a d ( n ) + . As @, = & @ this means t h a t
@ centralizes Gad ( R ) + , which implies that @ = i d [18, I1 2.61. Proof. Let p E A u ~ ~ ( ~ , ) ( M ( G , x ) )Then
. there exists a g E G ( A ~ )
such that g 0 p maps [h, l] to [h', 1] for some h E X+ and h' E XI+.
C o r o l l a r y 2.3. The map Then g o p maps MO(Gde',X + ) into M0(Gde7,X + ) and we can therefore
apply 2.2.

C o r o l l a r y 2.7. A s s u m e , i n (..6), that the centre Z of G satisfies the

Hasae principle for 111for finite primes. T h e n
identifies h u t G i Q ) +(MO(G,x+)) with { g * o I &g = a-I in Gad(Af)).

Proof. An automorphisrn of M O ( G ,X+)commuting with the action of

G(Q)+ commutcs (by continuity) with the action of G(&)T. It can therefore
be written g o M O ( o )for some g E G(Q)T and o E Aut(G). In order for this Proof. This follows from (2.6) as (2.4) follows from (2.3).
map to comnlute with the action of G(Q)+, a-l and & g must be equal.
In a must be an inner automorphism, a Gad(Q), and so the
map is t h a t defined by g * a.
$3. Proof of a Weak Form of (1.1)
C o r o l l a r y 2.4. If Z = Z(G) satisfies the Hasse principle for H1,
then A U ~ ~ ( ~ ) + ( M O ( G , X
=+ Z) )( A f ) n C(Q);/Z(Q); for example, if G This section is devoted to proving the following result.
is a n adjoint group, A U ~ ~ (MO(G,X+))
( ~ ) + = 1.
P r o p o s i t i o n 3.1. Let ( G , X + ) satisfy (C), and a.wume that G is simply
connected; then, for a n y automorphisrn T of C, there is a pair (G1,X'+)
Proof. The h y p o t l w h implies that if an element of Gad(Q)+ lifts t o satiujying (C) for which there exist compatible isomorphisms
an clement of G ( A f ) , then it lifts to an element of G(Q).

Example 2.5. The last corollary applies to the Shirnura varieties defined
by simply connected groups without factors of type A,, n 2 8 (see (3.8)
below). For these groups G(Q); = G ( A f ) and so
We begin by recalling a theorem of Kazhdan. Let X + be a Hermitian
symmetric domain, so that the identity component G of A u t ( X + ) is a
product of connected non-compact simple real Lie groups. For I' an arith-
metic subgroup of G, I' \ X+ carries a unique structure of an algebraic
The propositions can also be used to conipute the automorphism groups variety, and so T(I' \ X + ) is defined, T E Aut(C).
of non-connected Shirnura varieties.
246 J. S. MILNE TtlE ACTION O F AN AUTOMORPHISM OF C 247

T h e o r e m 3.2 ( K a z h d a n ) . (a) The universal covering space XI+ of Choose X: + r(Tlg \ X + ) 9 3 its to make thc- following diagram commute
T(I'\ X + ) is a Hermitian symmetric domain. with t h e upper arrows, and choose Lj to rrlake it commute with the lower
(b) Let G' be the identity component of Aut(X1+), and identify the arrows:
fundamental group I" of r ( r \ X+) with a wbgroup of G'; then 1" is a XT
id
+
5 x:
lattice in G'.

Proof. The assurnption that I' is an arithmetic subgroup of G means

t h a t there exists a group G I over Q and a surjective homomorphism
f : G1(R)+ -+ G with compact kernel carrying an arithmetic subgroup of The double coset r1LjI''C Aut(X;) is well-defined, and we let
G I into a group commensurable with .'I If G I is the symplectic group, tlwn
the theorem follows from the theory of moduli varieties of abelian varieties.
If G1 has no Q-simple factor Go such that GOR is of type E6 or & or has
factors of both types D R and DH, then the Q-simple factors of C I can Then ro is a subgroup of A u t ( X f ) and is independent of the choice of K.
be embedded into syrnplectic groups, and this case follows from the last (In [7] it is denoted by G".)
case. When I' \ X f is compact (so that C1 has Q-rank zero), the theorem There is a map 7 ++ 7 f : ro-+ G(AJ)/z(Q) that can bc characterized
is proved in [6] (it also follows from Yau's t.heorem [21] on the existence of as follows: for all I' = K n C(Q) (as above), and all g E G ( Q ) n 7f K ,
Einstein metrics). The remaining cases are treated in 171. t h e diagram (3.3.1) cornrnutes with 5 replaced by 7 . We have thercforc a
canonical embedding
Remark 3.3. If I' is irreducible (for example, if G is Q-simple) then I" is
also irreducible because otherwise r ( r \ X + ) , and hence l ' \ X + , would have a
finite ktale coverirlg t h a t was a product. Consequently, when rankR G' > 1,
Margulis's theorem [lo, Thm. 11 shows that I" is arilhrnetic. Both 7, and 7~ act on T M " ( G , X + ) , the, first through its action on Xt
and the second through its action on M O ( G ,X-t). These actions are equal.
Let (G, X + ) be as in the statement of (3.1). 11, proving the proposition,
we can assume that C is almost simple over Q and is not of t y p e A (because Lemma 3.4. Regard I'o as a subgroup o f G ( A ~ ) / z ( Q ) .
when G is of type A we know much more - see (1.5)). This last assumption
implies that G I is not compact for any 1. (a) ro i.9 dense i n G(Af)/%(Q).
Choose a compact open subgroup K of G ( A f ) cont;lining X ( Q ) , and let (b) I', n Z ( A f ) / z ( Q ) = 1.
I' = G(Q) n K be the corresponding congruence subgroup. Then (c) For a n y compact open subgroup K of G ( A f ) containing Z(Q),

On applying (3.2)) onc obtains a Hermitian symmetric domain x:,

a real
Lie group G' such t h a t G' = Aut(X;)+, and an irreducible lattice r' in
Proof. (a) it is clear from the definition of 7 H 7f that, for any K,
G' such that T M;{(C, X + ) = I" \ x:. ro + G ( A f ) / Z ( Q ) K is surjective.
For any g E G(Q), let ,'I = I' n g-lI'g. There are two obvious maps
1, g: I', \ X+ I$ I' \ X+, namely the projection map and the projection (b) Suppose 7 E ro is such t h a t YJ E %(Qf)/z(Q). Then the remark
preceding the statement of the len~rnashows that 7, centralizes roin C'.
map preceded by lelt multiplication by g. On applying r , we obtain maps
AS I'o has finite covolume i n C', this i~nplicsthat 7, is in l h c centre of G'
[17, 5.4, 5.281, and so 7, = 1.
248 J . S . MILNE
THE ACTION OF AN AUTOMORPEIISM OF C 249

(c) We can assume that K is the group used in the construction of r o . Proof. Margulis's theorem gives us a pair (Gl, 4s) satisfying (a) and
such that 4s is surjective with a compact kernel. We can suppose that I's
It is then clear that 1'' = l'o n K / % ( Q ) . 'L'hc second q u a l i t y follows from
the Grst and (a). is irreducible. Then GI is almost sirnple over Q and so cannot be of type
A. Therefore Gl1 does not have any compact factors and $1 111ust be an
isomorphism.
Later we s l d l show that T'" is contained in thc identity component G'
Let A be a subgroup of ro,s of finite index, and let AA =
of A u t ( ~ : ) , but for the prcsrut wc define I': = I'on GI.
We now fix an integral structure for C arid define, for any firiile set S of ((fi) E I J r E s r ( G ~OG,)
, I
fi(X) E Q1 fi(X) = f11(1)1 all 111' E S1X E A).
finite primes, G(A$) = n,,, G ( Q )x n,,, G(Zl). Let Toss = I.~~C(A!);
Thcn, for all sufficientlysmall A, AA is independent of A and Spec& = GI.

then %'o,snI': can he regarded as a subgroup of G b ZG'X G1. nleS This shows the uniqueness.
When we enlarge S, to S' say, then G1 does not change and
$ J ~ ,= ~ 4s.
G ~ We~therefore get a map
L e m m a 3.5. The group I'& is a n irreducible lattice i n G',.

It follows from ( 3 . 4 ~ )that I ' O is a discrete subgroup of

Proof.
G' X c ( A ~ )and
n,,, G(Zl) is compact, the projectiorl C', X nl,:C(Zl)
crele groups to discrete groups [20, p. 41, and so
-
, therefore I'lsis a discrete subgroup of G' X c ( A ~ ) As .
G i takes dis-
k discrete in G',.
such that for all finite S , $f(Gl,s) is commensurable with I'o,s, where
G l , s = Gl(Q) n GI(A!).
factors of ClR, and consider
Let GelornPbe the product of the anisotropic

Let U be a compact open subgroup of nlEs

G1, and let
r = (u x G(Z[)) n r:.
ws
Then U r o , s \ G> = I" \ GI, which carries an invariant finite measure. It
follows that is of finite covolume in G',.
For any finite set S , fo,s dr 3-l(r0.s)is commensurable with G l ( Q ) s .
Our assurnptior~that G is almost simple over Q implies that ro,sis
Ultimately we shall show that I'o = G l ( Q ) / Z ( Q ) , but we begin with a
irreducible (cf. 3.3).
weaker result.
Now assume that S is nonempty and sufficiently large t h a t rank(G$) 2 2.
Then Margulis's theorem [lo, Thrrl. 71 shows thaL I:'s is arithmetic. More
precisely, there is the following result. L e m m a 3.7. The group r0 C $ ( G ~ ( Q ) ) z. ( A ~ ) .

4 s : Gls
properties:
-
L e m m a 3.6. There exists a n algebraic group G1 over Q and a m a p
cb, where Gls = G 1 X~ nlES
Gl,lr having the following
and
Proof. Let 7 E I'o, and let 7 E A u t ( ~ f X
as an element of h u t ( x f ) X G ( A f ) / % ( ~ ) For
maps Fo,s into ro,s.
.
) G l ( A f ) map to 7 regarded
large enough S, 7 E I ' o , ~ ,

Choose an irreducible representation T: Gyd c=+ GL, of Gyd, and

consider the representation &;i. o T of n G1(Q). The Zariski closure of
rs i n G l ( Q ) such that GS(rS)
(a) there ezists a n S-arithmetic subgroup
is cornmen.surable with I':,s;
(b) write $.JS =
kernel, and each
X HlES
$1; then 4,
is a n isomorphism.
i s surjective with a compact
( & ~ O I ' ) ( ~ ' ~ , ~ ~ Gin~ GL,
t o show that there is a morphism a: G1
is &y. Lift
-
( Q ) )is T ( G ~ ~and
) , so 110, Thm. 81 can be applied
G7d whose restriction to ?o,s
to an isomorphism a: G1 -+ G I . Thcn a and &? agree on
a subgroup of F o , s of finite index (therefore also on a subgroup of GI(Q)s
Moreover, (el,$s) i~ uniquely determined b y the conditions (a) and (6).
of firkite index). As ro,s is dense in nlEs G(Q1)/Z(Q), this shows that
250 J. S. MILNE THE ACTION OF AN AUTOMOItPIIISM OF C

a = on nG(Ql)/Z(Q). In particular, a! is an inner automorphism, zg E Z(Af)/Z(Q). The map g ++ zg iy a homomorphism

i.e., a E G7d(Q). Moreover, a has the property that it lifts t o G1(Ql) for
all 1 E S, and hence for all 1 because we can extend S. The next lemma
shows that this implies that a lifts to an element a1 E G1(Q). This element
a1 has the same image as 7 in AU~(X:) X Gad(Af), which completes the If we knew, rrs is conjectured, that GI(&)/%(&) is simple, then this homo-
proof. ~norpliismwould have to be zero, and we would have achieved our i ~ n -
mediate goal of showing that ro 3 GI(Q)/Z(Q). Instead, we argue as
Let G be a simply connected semi-simple group over a follows.
Lemma 3.8.
number field k, and let Z = Z(G). Then H1(k, G) -t nu
finite I l l ( k v , G )
Lct F be a totally real Galois extension of Q with Galois group A; let
G , = ResFIQ GF and let X $ be such that there is an embedding
is injective, provided G has no factors of type A,, n 2 4.

Proof. We can assume G is absolutely almost simple. Then

Z(z) = 2 / 2 2 X 2 / 2 2 or Z / n Z , n 5 4. If Z(x) = Z(k), then the result
Then A acts on G, and X$, and G is the unique subgroup of G, such t h a t
is obvious from class field theory. In any case, z(R) = Z(L) for I, a Galois ,- .
extension of k with Galois group S3 or Z/nZ, n 5 3. The exact sequence n s a G~ lbz-' (G*)F is an isomorphism. The group A as continues to
act when we make the above convtructiorls for (G,,X:). In particular, A
acts on GI, and there is an inclusion GI -* GI, that identifies GI, with
RespIQ GI. We can conclude:
n
shows that is suffrcesto prove t h a t ~ l ( L / k Z, ) -+ H1(Lv/kv, Z) is injec-
tive. In fact it suffices to do this with k replaced by the fixed field of a
L e m m a 3.9. For any F as above, the diagram
Sylow subgroup of Gal(L/k). But then the Galois group is cyclic, and the
result is obvious.

Lemma 3.7 implies t h a t ro and G1(Q) have the same image in

G " ~ ( A=
~ )G : d ( ~ f ) . If we form the quotient of M O ( G , X + )by the action
of Z(Af), we get Z ( A f ) \ M O ( G , X + )= M O ( G a d , X + ) .Therefore,
commutes, where X, = ResFlQ 2.

Let 7 E GI(&); we shall show that 7 (mod Z(Q)) E ro.

We have proved (3.1) with G replaced by Gad. An argument of Borovoi L e m m a 3.10. There exist fundamental maximal tori Ti C GI,
(see 5.2a below; it is not necessary to assume there that G is simply i = 1 , . . .,k, and elements 7i
= Ti(Q) such that 7 = 7 1 . . . 7k.
connected) shows that the map GY,
-+ G$ defined by $J identifies G i d
with an inner form of Gad. Therefore $ J f : GA, --+ GIAt has the same Proof. Let U be the set of g E G l ( R ) such that the centralizer of
property, and so +f JZ(G) is defined over Q, i.e., $Jf identifies Z 5 Z(G) g is a compact maximal torus. Then, the usual argument using the Lie
with Z1 % Z(G1). algebra, shows that U is open in Gl(R). Moreover, U generates Gl(R). Let
From (3.4) we know that I'o n Z ( A f ) / Z ( Q ) = 1. Therefore any ele- 7 = 7;. . . ytk with 7: E U . According to the real approximation theorem,
ment g of GI(&)/%(&) can be written uniquely as g = 7 zg, 7 E ro, the ~ e Gt l ( Q ) n U is dense in U. We can thcrcfore choose 7i E G1(Q) n [J,
 TIIE ACTION OF AN AUTOMORPFIISM OF C
252 J. S. MILNE

= 7(72 . . .7k)-1 also lies in U . As

df is an isornorpism; this is what we had to prove.
i =2 , 3 , . . . , k, so close to that
71 E G I (Q),the elements 71, . . . , 7k fulfill the requirements of the Icmma.
C o r o l l a r y 3.11. Suppose i n (3.2) that T is a congruence group,
Thus we can assume 7 E T ( Q ) C G ( Q ) ,wlicrc T is a maximal torus such
i e . , that there exists a simply-connected semi-simple group GI over Q and
that T(R) is conlpacl. This last condition on 1' irnplies that T splits over a
a surjective homomorphism f : GI(R)--+ G with compact kernel carrying a
CM-field L, which can be chosen to be Galois over Q . Let F be the maximal
totally real subfirld of L and 1ct G, = G. The construction of congruence subgroup of G(Q) into a 8ubgroup of r with finite index. T h e n
r ( r \X+) is the quotient of a Hermitian symmetric domain by a congruence
Borovoi recalled bclow in $4, gives a reductive group H , of type A1 such
group.
that I: C Ha C Gl*. After possibly extending 17 we can assume no (Zl,),
is anisotropic. Consider

$4. Embedding Forms of SL2 into G

In this section, we recall some results of norovoi [2]. Let ( G , X + ) satisfy

(C), and let (T, h) C ( G , X ) . Throughout the section, we assume the
Note that 7 E Ila(Q)/Z*(Q)nH,(Q). A theorem or Platonov arid RapinEuk following conditions hold:
[l6] shows that U,(Q) has no lion-central normal subgroup. Thercbre, the
lower niap is zero on II,(Q)/%,(Q) n H,(Q) and so 7 niaps to zero. It
(4.la) G = ResFIQ C', where F is totally real and G' is absolutely
therefore lies in I'o.
almost simple;
We have shown To 3 G I ( Q ) / Z ( Q ) . Therefore, for any compact open
(4.1 b) the maximal torus T' C G' such that ResFlq 2'' = T splits over
K C G ( A f ) containing Z ( Q ) , we have a finite &ale map
a quadratic, totally imaginary extension L of F.

Remark 4.2. The condition (b) irnplies that TI, and any subtorus, is a
We therefore have a finite ktale map product of one-dimensional tori, and therefore satisfies the Hasse principle
for H1.
As T i is split, we can write
which is G(Af)-cquivariant when we identify Gal with GIAr by means of
4f. When we apply r-' to this map, and repeat the whole of the above Lie G i = Lie T i @ (Lie Gi),
construction for ( G I ,x:)
and T-l, we obtain maps aER

where It = R ( G b , T b ) . An a E R will be said to be totally compact if it is

a compact root of (G' @s.,o R)c for all embeddings a: E' 4 R. Let Rntc
The composite map satisfies the liypothescs of (2.1) and tlierefore is an denote the set of roots that are not totally compact. Then, for a E Rntc,
isomorphism. This implies that the first map

is an isomorphis~rland, on applying r-', we get that is defined over F , and we let H', be the corresponding connected subgroup
of GI. Write 11, = RcsFIQ H', and Z, = Z ( H , ) .
254 J. S. MILNE TIIE ACTION OF AN AUTOMOILPlIISM O F C 255

P r o p o s i t i o n 4.3 (Borovoi, OniYEic). Z(G) = na Za, (0 E Rntc). maps )I 3 t h and )1 ;--t = p t $ p- induce a C-linear, equivariant
isornorphism p+ 3 ttl,(sec [4, 1.2.141).
Proof. From the formula [p,p] = k, valid whenver a = k @ jl is the (b) It follows from [4, 1.2.71 that X'+= X+. Sincc
Cartan decomposition of a non-compact real Lie algebra, it follows t h a t
~nt+ c ~ n t >
c R . Thus

h and h' define thc same IIodgc filtration on Lie(G), but this implies that
they are equal [4, p. 2541.
and the proposition follows by applying ResFIQ,
P r o p o s i t i o n 5.2 (Borovoi (21). Let (G, X + ) , (G', X''), and T , p,

C o r o l l a r y 4.4. Let T = T I Z and Z, = Z a / Z where Z = Z(G). and be as in (3.1). Consider i: ( T , h ) c,(G, X + ) .

Regard Z , ( A ~ ) / Z ~ ( Q )as a s ~ b g r o u pof ??(A/ )/T(Q). Then
na
z,(Af)/z,(Q) = 1, (a E nntc).
(a) There exists an inclusion it: T ~ - G'r and a g E G ( A f ) such, that
$ o ad(g) o iA, = ~ L I .
(b) There exists an h' E X' factoring through it such that ph, = ~ p h
Proof. This follows easily from the fact t h a t r\Za = 1. (as maps into T).
- (c)
The pair (G', j ' ) is a form of (G, j); its class in l['(Q, T ) ,
T %T/z(G), has image zero in H L ( Q l , T )for all 1 f oo and is represented
55. Completion of the Proof of (1.1) ) 1 = 00.
b y ~ p ~ ( - l ) / p ~ ( - lfor

In this section, we use the rricthods of norovoi [2] to deducc (1.1) from Proof. (a) Let p ( ~ [ h ]= ) gt[h'], h' E XI+,g' E C ( A f ) (see the Appendix).
(3.1). For any t E ~ ( T ( Q ) ) ,t[h] = [h], and so $(t)g'[ht] = g'[hl], i.e.,
Let ( G , X + ) satisfy (C),
. . and consider ( T I h) C ( G , X + ) . For any
(gtP1$(t)gt ) [h'] = [h'].
q E T(Q), q[h] = [h], and so there is a representation ph of T on the tangent
space th t o M O ( G , X + )a t [h]. This implies that 9'-'$(t)gt E Gt(Q). Let g = $J-'(~'-'); then
$ o c&g o i: TA/ + G ~ maps I T ( Q ) into C t ( Q ) . As T ( Q ) is Zariski dense
Lemma 5.1 (cf. [2, 3.2, 3.31). (a) ph -
@ a , a E R(Gc, To), in T , this means lhat $ o a d g o i is defined over Q ; we denole it by i'.
(a,P h ) = 1. (b) Let h' be as in (a); then phr = 7ptL bccause (cp o g ) ( ~ [ h ]= ) [h'] and
(b) Suppose that ( C , X t + ) also satisfies (C), and that p o g is l'(Q)-equivxriant. Thercfore (b) follows from (5.lb).
(l",h') c (G, XI+); (c) As $ o & g is an isomorphisni ( G ,i)*, 2 (GI, iP)Af, it is clear that

if ph - pht, then h = h' (and Xt+ = X').

t h e class of ( G I i) in I l l ( Q l , T ) is trivial for all finite 1. The clement
adp(-1) is a Cartan involution on G; therefore the class of p(-1) in
H 1 ( R , T ) corresponds to the (unique) compact form of G c . Similarly,
Proof. (a) According t o (GI) thew is a decomposition of 9 = Lie(G) the class of ~ p ( - 1 ) correspo~ldsto the compact form of Cb = G c . The
conclusion is now clear.

such that Adp(z) acts trivially on a z on P+, and as z-I on P-. Remark 5.3. If in (5.2) we replace cp by cp o g and $ by 4 o u d ( g ) , then
Thus h,C jl' if and only if a(p(z)) = z , i.e., ( a l p ) = 1. The canonical p and $ are still compatible, b u t now p ( ~ [ h ]=
) [h'] itrlrl $ o i = it.
256 J. S. MILNE THE ACTION OF AN AUTOMORPHISM OF C

i
C o r o l l a r y 5.4. Consider ( T ,h) -+ ( G , X+),and assume that G is and taking ~ [ h , ]t o [ + h a ]The
. map
simply-connected and that 'I' 2'/Z(G) satisfies the ZIasse principle for
II'. Let ( T , 'h) i l , 'x+)
( ' G , be as in $1. Then there exist isomorphisms

f i e s [h,] and is compatible with & t : Il',(Af)--tIZ',(Af). Therefore

(2.2) shows t h a t & t , regarded as elenrent of ~ i z ~ ( A flies) , in Hzd(Q),
i.e., t E % , ( A f ) / z , ( ~ ) . Now (4.4) shows t h a t t E T ( Q ) , and so p 0 t and
a- d ( t ) 0 $ fulfill the requirements of the theorem.
such that

(a) P(+]) -- ['hl;

$6. Compatibility of the Maps p, for Different Special Points
(b) p and 4 are compatible;
(c) $ o i A f = ' i A f .
Let ( G I X + ) salisfy (C) a n d let (1',h) C (G, X + ) . Then Langlarlds's
Proof. I,rt (G1,X'+) be as in (3.1), and let i': T' '--+ G' and h' E X" constructions lead t o the definition of a map
be as in (5.2;~)and (5.2b). Choose p and ?C, as i n (5.3). Then ( 5 . 2 ~ )shows
t h a t t h w c exists an isornorphisnl (GI, i f ) 5 ('C,'i) (cf. (1.4))) and (5.2b)
shows t h a t the isoniorphisms carries h' into ' h . T h e corollary is now clear.
t h a t is compatible with the maps of the same name, G ( A ~ )-t 'G(Af),
G a d ( A f )-t 'Gad(Af) (see [l4, 581).
We now prove (1.1). Let ( G , X + ) , h, and T be as in the statement
of (1 1 ) We can assume t h a t G is almost simple over Q. Then
P r o p o ~ i t ~ i o6.1.
n Assume that G is simply connected, and let p, be
G = R c s f i / g G I for sonic totally real field TIarid absolutrly almost-simple
a n isomorphism r M O ( G , X + )-+ M O ( + G , ' X + ) with the properties (a) a n d
group G , . 1,ct (T, h) c ( C , X+)and let 7'1 be the ~naxinlaltorus in G I
(6) of (1.1). Then (1.16) holds for a l l y E G a d ( ~ ) + A ( r e l G ) .
such t h a t ResF,/Q 7; = T. As lk is anisotropic, T I splits over a CM-
field L 3 F'. Let F' be t h e maximal totally real subfield of I, and let
Proof. Let G , = ItespIQ(Gp) for some totally real field F, and
G , = I t ~ s ] , . GI
/ ~ where G' = G I F . Let h, be the composite of h with t h e
canonical inclusion G c-,G,. Then the G,(R)+-conjugacy class X , con- h, be the composite of h with G c - 4 G,. Then there is a commutative
taining h,, togrther with G,, satisfy (C), and ( I .G) shows t h a t it sufices t o diagram
prove (1.1) for ( G , , X , ) and h,. This allows us to assume t h a t the original r M O ( G ,X + ) 2 MO('G, ' X + )
objects, ( T ,h) C-t ( G IX + ) , satisfy (4.1).
Let p and cC, b r isomorphisms as in (5.4). As both 4 and $, (see $1)
are isornorphis~ns(G, i)Af 5 ('c, tilere exists a t E T(A/) such t h a t
$, = & ( t ) o 4. To any non totally compact root a , there corresponds a (('G, ' X + ) a n d ('G,,'X:) defined using h and h,). This shows t h a t p,
subgroup IT, of G containing T (see 54). Let Ilb, = I I ~ ' and , let X: (for G) is compatible with the action of g for all g E G , ( A f ) (any F ) a.nd all
h
be the Ilgd(R)+-conjugacy class containing h, % (S -+ Ha -+ 112). g E W,(Q)+ (any Ha c G,; see $4), but these elenients generate a dense
Then p maps r M o ( I I L , X:) into MO('II,,' X:) because it maps r(g[h,]) subgroup of G a d ( ~ ) + l \ ( r eG).
l
t o &t-l($T(g))[Tha] for all g E Hb,(Af) and l l ' , ( ~ f[h,] ) . is dense in
MO(li',, X ; ) . Sirice Ilb, is of type A l , we know (1.1) for it: there exists an Remark 6.2. Theorem 1.1 and (6.1) prove part (a) of Conjecture C 0
isomorphism p,: rM0(fl',, X:)-t MO(+II',,'X,f), compatible with +, (sec [14, p. 3401) for simply connected groups.
 THE ACTION OF AN AUTOMOltPIIISM OF C

We shall now need to consider two special points h and h' for a given Consider the diagram
( G , X + ) . To distinguish the objects 'GI 'X+,p,, . . . constrl~ctedrelative
t o h from the similar objects constructed relative t o h', we shall write the
former 'lhG, 'phX+, p,,h,. . . .
Gal(L/Q)
Let P ( r l h) and P(r,h') be t h e elements of ( G a d ( ~ L ) / C a " ~ ) )
corresponding
-
to h and h' respectively (see (1.4)). The image of P(T, h) and
P(r, h') in H 1 ( Q , Cad) are trivial a t the finite primes and are equal a t the
infinite primes (see (15, p. 315-3161). As Gad satisfies the Ilasse principal
for 111([9, VII. 61) this shows t h a t P(T,h) and c ( r , h') have the same image
in H1(Q, G ~ and ~ therefore
) B ~ F ( Th1)P(r,
, h)-' lies in G " ~ ( A J ) / G " ~ ( Q ) .
Moreover, there is an isomorphism f : 'yhG 4 ' P ~ ' Gsuch t h a t
The upper square commutes because of (6.1). The outside of the diagram
commutes becausc both maps send [h] to ['h'] = [f o 'h] and the action of
T ( ~ to
) thal of Thus the lower triangle commutes.
'lhtg.

Before proving the general case, we need some lemmas.

Define
L e m m a 6.4. Consider a n inclusion i: (G, X ) c=+ (GI, XI). Theorem
(6.3) holds for h and h' (as elements of X) if and only if it holds for i o h
and i o h'.
to be ('*"B),-' o M O ( f ) . It carries the action of Tfhg,y E G(Af), into the
action of + g. (See [15, p. 312-3181.)
T ~ Proof. Since i induces an embedding M O ( G , X + ) % MO(G',XI+) and
t h e maps 4 r , i o h , 4r,iohl, and 4(7; i o h', i o h) restrict to 4 ~ 7 4T,hll
, ~ ~ and
T h e o r e m 6.3. Assume that G is simply connected; then for any q5(r; h', h) on M O ( G X , + ) , the sufficiency is clear. The necessity follows
special h, h' E X+, the diagram from the facts that G1(Q)+[i 0 h] is dense in MO(G',XI+) and q5,,ioh, and
a ( r ; i o h', i o h) o q5T,ioh have the same behaviour with respect to the IIecke
operators.

L e m m a 6.5. If (6.3) is true for the pairs (h, h') and (h', h"), then it is
also true for the pair (h, h").

Proof. This is immediate, since d(r; h", h') 0 4(r; h', h) = q5(r; h", h).

L e m m a 6.6 (Borovoi). Let G be a simple noncompact group over R

and let X be a G(R)-conjugacy class of homornorphisms S -+ G satisfying
commutes. (C1) and ((72). Suppose h, h' E X factor through the same maximal torus
T c G. Then h' = h of h' = h-'.
Proof. We first prove this under the assumption that
h' = &q o h, q E Gad (Q)+. In this case B = q (see [15, 9.11) so t h a t Proof. Let K , be the centralizer of h(S). Then [4, 1.2.71, K , is a
M O ( f )= ( T * ~ o' ~ ) h', h) = ~ O ( Thl,
4O(r; ; h) o (+lhq). maximal connected compact subgroup of G. The pair (C, T) dcter~ninesthe
260 3. S. MILNE TI115 ACTION O F AN AIJTOMORPIIISM OF C 261

set of compact roots in R(Gc, Tc), which determines K,. The centre Z of T = U' sornc r , where U is the unique one-dimensional non-split F-torus
IC, is 1-dirne~lsional(loc. tit,.), and h dcfines an isomorphism S / G , 5 Z; split by L, and therefore I '(I,/F, 1') x ( F X/ N L X )'. T h e penultimate
it is thcrcfore determined up to sign. assertion shows that c is represented by a family (ct, . . . ,c,) of totally
positive elements of F X . After adjoining JcT to P', i = 1 , . . . , r , we can
We now complete the proof of (6.3). As usual, we can assume that G assume c = 1. Thcn c, = t-l . a t some t E T(L) and so, after replacing /3
is almost simple over Q and therefore that G = ItesFlg G I , where F is with Pt-', we can assume it lies in G l ( F ) . Regard P as an element of G(Q).
totally real and G1 is absolutely almost simple. Let ( T , h ) C (G, X + ) and Lemma (6.5) and the first part of the proof show that we need only prove
(TI, h') c ( G , X + ) , and lct TI, T i C G I be such t h a t RcsFIQ TI = T and t h e theorem for &P o h and h'. This means t h a t we can assume t h a t h and
ResFlq T i = T'. There exists a CM-field L splitting both I; and T:. After h' factor through the same torus. But then they must be equal because,
replacing G with G, = ResFtlQ where F' is the maxinial totally in the context of (6.6), h-' does not lie in the same connected component
real subfield of L, and using (6.4), we can assume that L is a quadratic as h.
extension of F. As and T:,L are split, there exists a p E G1(L) such Finally, suppose l(wvo) # 0, say wvo = sawlo with l ( w ~ , ) < l(wvo)
t h a t PT1P-' = Ti. For each real prime v: F C-+ R of F, choose an and s, thc reflection corresponding to the root a. If a is compact a t v,
extension (also denoted by v) of v to L and write Zf, for B g F , , R , any then s, lifts to 7, E N ( F v ) (see [15, p. 3081) arid we can replace yVowith
yuo7, . Then w , , ~is replaced with w',~ and we can apply the induction
-1
F-group 11. As TI,, arid Ti,, are compact, there cxisls a 7, E G,(R) such
that yv2;,v7;1 = Ti,,. Let c, = p-' . up, where o generates Gal(L/F). hypothesis. Supposc therefore that a is not compact a t v and define
Then c, E N(L), where N is the normalizer of T, and so it defines a class H L C G I , 11: > TI, Zf; of type A,, as in $4. Let ZZ, be the derived
c E H1(Z,/F, N). AS 7;' . v(P) E N ( C ) and group of Ilb, and let T, = 11, n 1'. Then T, -- U ,which implies t h a t
N'(L/F, T,) + $, II1(Lv/Fv,T,) is surjective. Choose c, E I l l ( L / F , T,)
mapping to (c,) where c, = 1 for v f v, and cvo = 6(s,) where S is the
boundary map
where L denotes complex conjugation, we see that c maps to 1 in
l l ( l , v / ~ wN,) . Let w, be the image of 7,' v(P) in W(L,), wherc
W = N I T . The image w of c, in W(L) is L because v(w) = w;' . LW, and N, = Norm(%). Then c, maps t o 1 in $N'(L,/F,,N,). T h e Hasse
L acts trivially on W(Z,,) (see [15, p. 3071). Thus c, E T(L) and principle therefore shows that c, splits in II1(L/F, H a ) : (c,), = g-' . og,
g E H,(L). L i f t s, to n, E N,(L); then v,((c,),) = n z l . m , and so
gn;' E H,(R). Since we know the theorem for ltesFlq H a , (6.4) allows
us to replace (T, h) with (&g o T, &(gn;') o h). This replaces ,8 with
The following diagram is useful: /3g-', yvO with rvon,g-l, 7, with v # v,, wvo with gs,'wvog-l =
g ~ ' , ~ g - ' , and w, wilh gwvy-l, v # v,. Thus C l ( w , ) is di~niriished,and
we can apply the induction liypothesis.

$7. Conclusions

We now prove (6.3) by induction on 1 = ~ , l ( w , ) , where l(w,) is the We have shown that Conjecture C0 of [15, p. 340-3411 is true for ( G , X + )
length of w, as an element of W(C). Suppose first t h a t 1 = 0. Then whenever G is simply connected. As is remarked in [15, 9.61, this implies
'
7,' . v(P) E 7'(Lv) and so c maps to 1 in ZI (Lv/Fv, T ) for all v. Note that the general case.
 THE ACTION OF AN AUTOMOItPIIISM OF C 263
262 J. S. MILNE

(C1) for all h E X + , the IIodge structure on Lie(GR) defined by h is of

Theorem 7.1. The conjecture of Langlands [8, p. 232-2331 (see also
[15, p. Sll]) is true for all Shimura varieties. type { ( - I 1 1 1 1 (01 011 ( 1 1 -111;
( C z ) d h ( i ) is a Cartan irivolution on Gkd;
Proof. In [15, 9.41 it is shown that this conjecture is equivalent to (C3) Gad has no non-trivial factors defined over Q that are anisotropic
Conjecture C O . over R.

Theorem 7.2. Canonical models (in the sense of (4, 2.2.51) exist for Such a (G, X+) definrs a connected Shimura variety M O ( G , X + ) . The
all Shimura varieties. topology r(G) on gad(^) is that for which Ihc images of the congruence
sulr)groups of C(Q) form a fundamental system of neighbourhoods of the
identity, and
Proof. This is a consequence of (7.1) (see (15, $71).
MO(G,x+) = lip r \ X+

Theorem 7.3. The conjecture of Langlands describing the action of where the limit is over the set C ( G ) of torsiori-free arithmetic subgroups of
complex conjugation o n a Shimzira variety having a real canonical model G a d ( Q ) that are open relative to the topology T(G). For h E E X +[h] , and
[8, p. 2341 is true. [hIr denote the iniages of h in M '(G, X + ) and r. M o ( G , X + ) I' \ X+.
,4ny LY E G U d ( Q ) +acts on M O ( G ,X + )by transport of structure: a[hjr =
Proof. This again follows from (7.1) [a o I L ] , ( ~ ) . Any g E GI&); acts aa follows: Let r E C(G)and let K
be a compact open subgroup of G ( A f ) such that r contains the image of
Theorem 7.4. The m a i n theorems of [IS], viz. (4.6) and (4.9), are K G ( Q ) + ; then g E q K some q E G ( Q ) + , and g[h]r df = [adq0 hlBrQ-l.
true for all Shimura varieties. These actions combine to give an action of G(Q)T X Gad(&)+ (semi-
direct product for the obvious action of C a d ( Q ) + on G(Q)T). The map
Proof. They are proved i11 [13] under the assumption that G is classical q H (q,uJq-l) identifies G(Q)+ with a normal subgroup of the product,
and thc, canonical model exists, but the first assunlption is only used to and the quotient
simplify the p o o f of [13, 1.31, and we can instead deduce this theorem
from Proposition 2.1 above.
continues to act on M o ( G , X + ) . The completion of Gad(Q)+ for the topol-
Remark 7.5. Theorem 7.4 gives a definitive answer to the question of
ogy r(G), Gad(Q)+A(relG), is equal to C(Q);,G(g)+C""Q)+, and this
Slhnura 119, p. 3471.
ideritificat,ion is compatible with the actions of the groups on Mo(G,X + )
(see [4, 2.1.6.21).
Remark 7.6. For Shiniura varieties of Abelian type, (7.1), (7.2), and
Any x E M O ( G , X + )can be written x = g[h] for some g E G ( Q ) I and
(7.3) were first proved in ([5], [15]), [4], arid [12] respectively.
h t X + . For suppose xr = [h]r.;then, for any 1'1 c ,'I Xr, =; 71.1 [ h ]some
-yrl E F; let 7 = limrl,l y r l , and let 7 = g * a;then x = ? [ h ] = g[a(h)].
If G is simply connected, then G(Q)+ = G(Af); moreover
Appendix

We say that (G, X+) satisfies (C) if G is a semi-simple group over Q and
X+ is a G(R)+-conjugacy class of maps S -4 G$ for which the following
if K is a compact open subgroup of G ( A f ) containing Z(Q), and r is the
image of K n C ( Q ) in Gad(Q)+.
hold:
 TI1E ACTION OF AN AUTOMORPIIISM OF C 265

References 1191 G. Sllirnurn, O n aritlinietic automorphic functions. Actes, C o n g r b

Intern. Math. (1970) Tom 2, 343-348.
A. Uorel, Density a n d maxirrlality of arithmetic subgroups. J. Rcine [20] G. Shirnura, Arithmetic Theory of Automorphic Functions. Publ.
Angew. Math. 224 (196(i), 78 89. Math. Soc. Japan 11, Princeton University Press, Princeton, 1971.
M. Borovoi, Canorlical models of Shimura varieties. IIarldwritten [21] S-T. Yau, On t h e Ricci curvature of a compact Kahler manifold
notes datcd 26/5/81. a n d t h e complex M o n g c - A m p h equation, I. Comm. P u r e Appl.
P . Delignr, Travaux de Shiniura. SCm. Bourbaki F6vrier 71, Expos6 Math., 31 (1978), 339-411.
389, Lecture Notes in Math., 244, Springcr, Uerlin, 1971.
P. Deligric, VariCt6s de Shirriurn: intcrprc~t;ttiorrniodulairc, r t tcch-
niques d c rorrstruction d c 111od6lescarioniques. I'roc. Symp. P u r e Received April 29, 1982
Math., A.M.S., 33 (1979), p;trt 2, 217- 290.
P . Drlignr, Motifs r t groupes de 'l'aniyama. Lecture Notes in This work was completed while the author was at The Institute for Advanced
Math., 900, Springer, 13crlin, 1982, pp. 261-279. Study and was supported in part by NSF grant MCS 8103365.
D. Kazhdan, O n arithrrrctic varictirs. Lie Croups and Thcir Repre-
sentations. Ilndapest, 1971, pp. 151 -217. Professor J. S. Milne
I). Kazhcian, O n aritlrrnctic varieties, 11. Prcprint. Department of Mathematics
R. Langlarlds, Automorphic rrpresentatiorls, Shimura varieties, a n d University of Michigan
nlotives. Eiri M&-chen. I'roc. Symp. Pure Math., A.M.S. 33 (1979)) Ann Arbor, Michigan 48109
p a r t 2, 205 246.
R . Langlands, 1,es dCbuts d'urie formule dcs traces stables. Publ.
Math. Univ. Paris VII. (To appear).
G. Margulis, Discrete groups of motions of manifolds of nonpositive
curvature. Arncr. Math. Soc. 'l'ransl. 109 (1977), 33-45.
J. Milne, T h c arithmetic of :~utornorphirfunctions. In preparation.
J. Milne and K-y. Shih, T h r aclion of corrlplex conjugation on a
Shimura variety. Annals of Math., 113 (1'381), 569-599.
J. Milne and K-y. Shih, Autornorphisrri groups of Shirnura varieties
a n d reciprocity laws. Amer. J. Math., 103 (1981), 911-935.
J. Milne and K-y. Shill, Langli~rlds'sco~lstructionof t h e Taniyama
group. Lecture Notes in Math., 900, Springer, Berlin, 1982, 229-
260.
J. Milne and K-y. Shih, Conjugates of Shimura varieties. Lecture
Notes in Math., 900, Spririgcr, Bcrlin, 1982, pp. 280 356.
V. P l a t o ~ ~ oa vn d A. Itapirltuk, O n t h c group of rational points of
three-dirnrnsional groups. Soviet Math. Dokl. 20 (1979), 693-697.
M. Ragunathan, Discrete Subgroups of Lie Croups. Erg. Math. 68,
Springer, Ilerlin, 1972.
I. Satake, Algebraic Structures of Symmetric Domains. Publ. Math.
Soc. Japan 14, Princeton IJriivrrsity t'rrss, Princeton, 1980.
 The Torelli Theorem
for Ordinary K3 Surfaces
over Finite Fields
Niels 0. Nygaard

To I.R. Shafarevich

Introduction

Shafarevich's and Piatetski-Stlapiro's proof of the Torelli theorem for K3

surfaces over C [13] is one of the most beautiful proofs in complex algebraic
geometry.
In one of its several fornlulations the theorem states the following: Let
(X, L) and (XI, L') be polarized surfaces over C and assume that
4: H2(X',Z)---+112(X,Z) is an isomorphism compatible with the cup-
product pairings and the IIodge structures. Assume further that 4 maps
the cohomology class of L' to that of L then ( X , L ) 2: (X', L') and in fact
4 is induccd by an isomorphism between the polarized surfaces.
011e can ask whether an analogous theorem holds in positive cllaracteris-
tics, i.e., whether there are linear algebra data associated to a polarized
K3 surface defined over a field of characteristic p > 0 such that these
data uniquely determine the isomorphism class of the K3 surface. Let
me explain why there seems to be hope for such a theorem: Combining
the Torelli theorem and Kulikov's result, t h a t the period map is proper,
one shows that the moduli space of polarized K3 surfaces of some fixed
degree, say d, (here the term "K3 surface" has to be broadened slightly
to include certain singular surfaces) is isomorphic to the period space.
The period space is a disjoint union of Shimura varieties S h K ( G ,p ) where
G = 0 (2,19). h s u m e that we can construct a moduli space over Spec z[;],
n a suitable integer, for polarized K3 surfaces of degree d. Then we have
in fact constructed a model for the period space and one can show that
this model is canonical in the sense of Shimura varieties. Let p be a prime
not dividing n and consider the reduction mod p of the moduli space. We
268 N. 0. NYGAARD THE TOIIISIJLI TIiTCOItEM FOR K3 SUILE'ACRS 269

then get a variety over Fp which is on the one hand a moduli space for T h e o r e m 1.2. Let X o / S p e c k be a n ordinary K3 surface. T h e m a p
polarized K3 surfaces of dcgree d in ch;~ractcristicp and on t h r other hand
the reduction of the canonical rrlodrl of a Shimura variety. Langlands has (130. classes of liftings X/ Spec A )
conjectured a purely group theoretical (or linear algebraic) description of -+ {Iso. classes of liftings G / Spec A )
the set of Fppoints of the reduction of a Shimura variety [9]. In the present
case it then makes sense t o use the term "periods" for the elements of this defined by
set and t o view Langlands' conjecture as a conjectural Torclli theorem for X / Spec A I-, $x/ Spec A
polarized K3 surfaces over Fp. One instance t h a t indicates t h a t Lhis makes
is a functorial isomorphism.
sense is the suprrsingular case. Here Langlands' conjecture predicts t h a t a
supersingular K3 surface is uniquely determined by its crystalline cohomol-
ogy. This was conjectured by Ogus in [ll]arid recently proved by him [12] Proof. [lo] theorem 1.3.
as a consequcncr of Rudakov's and Shafarcvich's results on degenerations
of K3 surfaces in characteristic p [14]. Since hcight one groups arc rigid, there is precisely one lifting C; of
B T $ ~to Spcc A. Si~~iilarly
dtale groups arc rigid so there is precisely one
lifting G? of $goto Spec A. It follows t h a t if G is any lifting of $xo t o
Spec A we have an extension
1. Review of the Canonical Lifting and the Kuga-Satake-
Deligne Abelian Variety

In this section we recall some of the results of our previous paper [lo] liflirlg the extension
and of Deligne's paper [4].
Let k be a perfect Geld of characteristic p.

Definition 1.1. A K 3 surface X o / k is said t o be ordinary if the In articular we can consider the lifting of $x0 defining the trivial
following equivalent conditions are satisfied. extension G = G i X (:$. 13y the theorem there exists a unique lifting
XCan,,/SpecA such that $xc=,,,= G X G gI. Now let A = Wn and put
Xn = X,,,,w, then we get a proper flat formal schcrne {Xn)/ Spf W.
(i) The formal Brauer group B T $ ~has height 1.
(ii) The frobenius F: H2(OxO)4 1I2(Ox0) is bijective.
T h e o r e m 1.3. The formal scheme {Xn)/Spf W i~ algebraizable and
(iii) The EIoclge polygon and the Newton polygon of H&,s(XO/W) coin-
dejines a I(3 surface XcUn/SpccW . Furthermore Xcanhas the property
cide.
that any line bundle o n Xo lifts uniquely to X,,,.

Let A be a n artinian local ring with residue field k and let X/ SpecA be
Proof. [lo] proposition 1.6.
a lifting of the K3 surface Xo/ Speck.
By Artin-Mazur[2] the enlarged formal Brauer group $xAdefines a
pdivisible group on Spec A lifting $xo / Spec k. Next recall from [4] the following facts: Let (X, L ) / Spcc V be a polarized
In [lo] we proved the followirig theorem.
K3 surface where V is a discrete valuation ring with residue field k a finite
extension of Fp and fraction field K a finite extension o f Q p
Choose an isomorphism K rr C and base extend X t o a K3 surface
x, / Spcc C .
270 N. 0. NYGAARD THE TORELLI TI3EORb;M FOR K3 SURFACES

We let p2(Xc,Z(1)) denote the primitive part of the cohomology, Le., It follows from theorem 1.6 t h a t we have isomorphisms

W)Y-

T h e o r e m 1.4. (Kuga-Satake-Deligne). There exists a finite eztension

K t / K and a n abelian variety A/ Spec K t with the following properties.

(i) Ahas complex multiplications by the even Clifford algebra where a) is an isomorphism of rational Ilodge rtructures and b) is an
C = c + P ~ (Z(1))
x ~ ,of the bilinear form o n p2(xc,~ ( 1 ) ) . isomorphism of Galois modules.
The canonical lifLing of A0 is characterized by the fact that the frobcrlius
(ii) H1(Ac, Z) is isomorphic to c + P ~ (~x ( 1~ , a right C-module,
) )as
lifts to A,,,. It follows that we get an isogeny U A t ~End(Acun,c)
~ ~ @Q
and left multipliction defines a n isomorphism of algebras
and hence by c) above an isogeny U A E End(&) @ Q. Since the frobenius
on A. cornmutes with the action of C it follows that U A c 13ndc(Ac) @ Q.
The isogeny oa induces an vutornorphisrrl of H 1 ( A o , Q ) compatil>le with
the IIodge structure and the right C-niodule structure hence an elerrlent
(iii) p is a n isomorphism of f o d g - structures. U A E ~ n d c ( H ' ( ~ Q)).
c,
We define an algebra autoniorphisrn oxo of C+l'2 (XCan,c,Q(1)) by
(iv) Tensoring with Zl and identifying Zl-cohomology with l-adic e'tale
cohomoloyy p induce.? an isomorphism of e'tale sheaves o n Spec K t

In [lo] it is proved that ax induces an autornorphism

where f : XKI-+ Spec K t and y: A--+ Spec K t are the structure maps.
(v) A has good reduction. compaLible with the cup-product and the Hodge structure and such that
Fx0 is identified with the geometric frobcnius urtder the canonical isolnor-
Assume now that (Xo, Lo) is a polarized ordinary K3 surface over Fq and phism
consider the canonical 1ifLirig Xcan/Spec W with the unique polarization f p2(Xcun,c, Q(1)) @ QI P:t(x~,
Q((1))-
lifting Lo. Applying the above Ltieorern to (Xcan,L) we get an abelian
variety A over a finite extension of Qp.

P r o p o s i t i o n 1.5. T h e reduction A. of A (which is a n abelian variety 2. T h e Torelli Theorem

by (v)) i s ordinary.
We fix ay before an isomorphism Qp 2: C.
Proof. [10] Proposition 2.5. Let (Xu, L 0) be a polarized ordinary K 3 surface over Fq,q = pn. Define
1
M(Xo) = I J 2 ( X c a n , c , Z(1)) and let v t M(X0) be the cohomology class of
the lifting of Lo. We extend Fxo to M ( X o ) @ Q by putting Fx,,(u) = v .
The key result of [lo] is the following:

no, i n T h e o r e m 2.1. Let (Xo, Lo) and (Xb, Lb) be polarized ordinary K 3
T h e o r e m 1.6. A is isogeneous t o the canonical lifting A,,, of
surfaces over Fq. Assume that 4: M(Xh)--1 M(AYo)is a n isomorphism
the sense of ordinary abelian varieties.
272 N . 0. NYGAARD THIS 'I'OREL1,I TtIrCOREM FOR K 3 SUILl'ACES 273

compatible with the cup-product pairings and such that ~ ( v ' =

) v. Assume It is clear from the assumptions that axo. 4 = d, . ax;, so we get, for
further that there is an integer m such that $.If'?; = FTU. 4 then (Xo, f 0 ) Y E c'P2(Xcan,c, Q(l))1
-
and (Xb, Lb) are isomorphic over Fq.

Proof. Enlarging Fq if I ~ C C C S S ~we

~ ~may
, assume

We shall prove that the map

This shows that unl.$(oA,) is in the ccrrter of C + P ~ ( ( ~ Y , , , , ~Q(I)),
, which is
4: H ~ ( x : , ~ , ~~,( 1 ) ) + can,^, ~ ( 1 ) ) a ccntral simple nlgc+ra over Q , hence 02' . $(a = ,X) E Q and q5(aA,)=
X . a,,.
is compatible with the Ilodgc structures. Since 4(v1) = v, it is clear t h a t Assume that x E U1(&, C ) is an cigeavector for frobenius correspond-
we get an isornorphisrn ing t o an eigcnvaluc p E C , i.e., we have a n , . x = p . x then
4: Y 2 ( ~ L a n , c~, ( 1 ) ) p 2 ( x c a n , c , Z(1))
-+

and hence an isorrlorphisn~of algcbras

It follows t h a t 4(x) is an eigcnvector for frobenius on H 1 ( A c , C ) cor-
4: C + P ~ ( X : ~ ,~, ~( 1, )-)+ C+P~(X~,,,C,
%(I)). responding to the cigcnvalue X-' . p.
We havc isoniorphisms of t h o d u l e s Assurne that the reductions of A and A' are defined over F , r . Then
all the eigenvalucs of frobenius have absolute value pT/2 by the Itieniann
) 11 y&, Z)
c+P ~ ( x ~ ~ , ~~, C( 1, ) = hypothesis. In particular Ip) = pT/2 and ] X - - ' p l = p7/2 so ]XI = I. Since
X E Q we have X = f l , so 4(u%,) = a;, and hcnce passing to a qr~adratic
c+p2(xc,,,c, Z(1)) I I 1 ( A c l Z),
5
extension of the grour~dfieldwe can assume $(on,) = UA. It follows t h a t
hcnce we have an isomorphism the isomorphism
4: H1(A:,, Q)--+ f l l ( A c l Q)
4: l l l ( A & , Z) 7 I I ' ( A ~ , Z).
is compatible with the frobenius maps.
The frobenil~smaps on Ell(A&,Q ) and 1l1(Ac, Q ) arc Iiliips of right The reduction map
modules, hence are given by Iclt multiplication by elenlc~lts
IIom(A,,,, A',,,,) -, tIorn(Ao, A:)
a*! E C I
?+ ,2
(x:,,,,, ~ ( 1 ) ) '
is a bijection. 13y theorem 1.6 we have
and

The frobcnius rnaps on C+P~(X:,,,~,

Q(1)) and C + P 2 ( l y ~ ~ , , ,~c(, 1 ) )
induced by Ilk; and Fx0 are given by Hence the reduction map

is bijective.
 TIIE TORELIJ TIIISOREM FOR K 3 SURFACES

Let I ~ o r n , ( l ~ ' ( A &Q),

, I l l ( A c , Q)) denote the set of liomornorphisms
4:zI'(A&,Q ) -4H '(Ac, Q) satisfying ant = an $J.
It is clear that the obvious map
and a commutative diagram

maps into ~ ~ o r r l , ( H ~ ( AQ),

& , N1(Ac, Q)),since this is true after tensoring
with Q1 and identifying
SO 4: ~~(x',,~,~, ~ ( 1 -+) ) ~ ( 1 ) is) compatible with the IIodge
structures. Since v' and v have IIodgc type (0,O) it follows that also

By a theorem of Tate [15] the obvious map

is compatible with the Hodge structures.
Now we apply the Torelli theorem over C to deduce that there is an
isomorphism
$J: ( X c a n , ~ L)
, (XLan,cl
L').
is a bijection. It follows from this and the previous remarks that
This ivomorphim is defined over a finite extension K/Qp, which we can
choose large enough to contain W(Fq). Let R be the integral closure
of W(Fq) in K . Then (X,,,, L)R and (Xf,,, L)R are polarized surfaces
is a bijection. over Spec R whose generic fibers are isomorphic as polarized surfaces. By
By the above 4 E 1 1 o r n , ( H ~ ( ~ &Q),
, I13(Ac, Q)), hcnce is induced by an the Mumford-Matsusaka theorem [16] this implies that (X,,,, L)R and
element of IIorn(A, A')@Q, and so, being induced by a geometric morphisni, (X',,,, L')R are isomorphic. The residue field of R is a finite extension
4 is compatible with the IIodge structures. of Fq and so it follows that (Xo, Lo) and (Xb, Lb) become isomorphic over
Now the commutative diagram a finite extension of Fq.

References

[I] Artin, M., Supersingular K3 surfaces. Ann. Sc. kc. Norm. Sup. Lie
and the fact that p is an iso~norphismof IIodge structures implies that skrie, t. 7, 543-568, (1974).
[2] Artin, M., Mazur, B., Formal groups arising from algebraic
varieties. Ann. Sc. EC. Norm. Sup. 4e &riel t. 10, 87-132, (1977).
[3] Berthelot, P., Cohomologie cristalline des schemas des caracte'r-
is compatible with the 1Iodgc structures. It follows from [lo] lemma 3.2 istique p > 0. Lecture Notes in Math. no. 407, Springer Verlag.
that we have injective maps of Hodge structures Berlin, Heidelberg, New York, (1974).
[4] Deligne, P., La conjecture de W e d pour les surfaces Kt?. Inv. Math.
15, 206-226, (1972).
 I~cligne,P., Vurie'te's abe'liennes ordinaires sur u n corps fini. Inv. Real Points on Shirnura Curves
Math. 8, 238 -213, (1968).
Uclignc, P., Illusie, I,., Gristaux ordinaires et coordone'es canoniques.
In Surfaces Algcbrique, Serninaire Orsay, 1976-78. Lecture Notes A. P. Ogg
in Math. no. 868, Springer Verlag. Berlin, iIeidclberg, New York,
(1981).
Katz,. N.,. Serre- Tate local m o d u l i of ordinary abelian varieties.
In Surfaces Algcbrique, Serninaire Orsay, 1976-78. 1,ccture Notes
in Math. no. 868, Springer Verlag. Berlin, IIcidclberg, New York, Let 13 be a quaterriion algebra over Q , i.e., a central simple alrgbra of
(1981). dimension 4 over Q. We assume t h a t FI is indefinite, aqd fix an identification
Kuga, M., Satake, I., Abelian varieties associated t o polarized K3 of I?, = B @ R with iM2(R). T h e discriminant L) of I1 is thcn the product
surfaces. Math. Ann. 169, 239-242, (1967). of a n w e n number of distinct primes. T h e completion Elp = B Jj Q,, is a
Langlands, R., S o m e c o n t e m p o r a r y problems w i t h origins in the skew field if p I I) and is isomorphic to M2(Q,) if p y D , and D = 1 if and
J u g e n d t r a u m (FIilbert 's problem 12). In MatliernaLical dcvelopntents only if 13 is isomorphic to M 2 ( Q ) , i.t'., B is not a skew field.
arising from IIiltwrt problems, S y ~ n p in . Pure Math. vol. XXVIII, T,et cr H at denot,e the canonical i n v o l u t i o n of B; ttlus (ap)' = /3'a1, and
part 2. American Math. Soc. Providence, Rhode Island, (1976). +
the trace s ( a ) = a at and the n o r m n ( a ) = a d are in Q for any a E B.
Nygaard, N., T h e Tate conjecture for ordinary K3 surfaces over If R oi onc of its corr~plclionsis isomorphic t o a rriatrix algebra M 2 ( K ) ,
fin it^ fields. 'I'o appear in Inv. Math.
Ogus, A, Stipersingular K3 crystals. Journeks dc GEometrid Algd-
then (::)I = (yc a n d the trace and norm correspond t o the trace
and dctcrrninant of a two-by-two matrix. An e1cmrr:t a of E) is an integer
brique. Rennes 1978, Asthrisque no. 64, 3-86, (1978).
if s ( a ) w~rdn(n)arc in %. An order in B is a subring 0 of B, consistJirig of
Ogus, A., A crystalline l'orelli t h e o r e m for supersingular K3 sur-
integrrs and of rank 4 over Z . T h e corresponding local orders are written
faces. This volume.
Op -- O @ Z p , a n d we have 0 = I1 np Op.
Sh;tfarevich, I., Piatetski-Shapiro, I. A l'orelli t h e o r e m for surfaces
Givcn an integer 11' 2 1 which is relatively prime to I), an Eichler order
of type K3. Math. U.S.S.R. Izvestija 5 , 547-588, (1971).
of levd I' is an order 0 such t h a t the local order Op is maximal if p /Y F,
Shafarevich, I., Iludakov, A., Degeneration of K3 surfaces over
and if p / F ,there is an isomorphism of TIp onto Mz(Op) carrying Op onto
fields of finite characteristic. Prcprint.
Tate, J., E n d o m o r p h i s m s of abelian varieties over finite fields. Inv. {(:, :) : a, b, c , d E Zp . Thus an Biehler order of level 1 is a m a x i ~ n a l
Math. no. 2, 134-144, (1966). order. All ordcrs in th'is paper will be Eichlcr orders. Eichler has shown
Murnford, D., Matsusaka, T w o f u n d a m e n t a l t h e o r e m s o n d e f o r m a -
tiona of polarized algebraic varieties. Amer. Jour. of Malh. 86, [3,41:
(1964).
1) There is only one Eichler order of a given level F, u p to conjugation,
i.c., tile general one is a0a-', where a E B X .
2) 0 contains a unit of norm -1.
Reccivcd August 10, 1982
T h e group O ( ' ) of units of norm 1 in 0 is thus of index 2 in 0 X . It is a
Professor Nicls 0 . Nygaard
discrete subgroup of SL(2,R ) a n d so s c l s on the upper half-plane g). Since
Department of Mathematics
O ( ' ) depends only on D and F, up to conjugation, we have a well-tieli~~cd
University of Chicago
Rientann surface 0 ( ' ) \ 2). If D > 1, this Riernann surface is compact
Chicago, lllirwiu 60637
278 A. P. OGG REAL POINTS ON SIIIMUItA CUItVES 279

and is t h e Shimura curve S = If D = 1, then S denotes the usual

phism of L will both be denoted by ru r-, a'. Consider an e~nhedding
compactificxtion by adding cusps, i.e., S is the ordinary modular curve p: L-+R. If D > 1, i.e., if l? is a skew field, the11 1, is a quadratic field,
X o ( F ) .According to Shimura, S has a canonical model defined over Q. and the local algebra Lp = I, @ Qp is necessarily a field when p divides D.
If D > 1, then S ( Q ) is empty since S ( R ) is empty, as shown by Shimura Thus the primes p wliich ramify in I? cannot split in L; this is a necessary
and Shih; the proof will be recalled below. If D = 1, then S ( Q ) is known and sufficient condition for L to be crr~beddablein B. If L = Q ( a ) , and
by the work by Barry Masur. The next problem along these lines is to +
s = a a', n = a d , then to give an embedding cp is to give an element
determine, in the case D > 1 , the rational points on a quotient S/(w) of S of B with trace s and norm n . We assume that x2 - sx + n = 0 has two
by an Atkirt-Lehner involution; these involutions will be described in detail distinct roots, i.e., if 1, is not a field (possible orlly if D = I), then this
in $2 below. A rational point on such a quotient corresponds t o a pair of polynornial has two tlislinct rational roots.
conjugate points on S over some (imaginary) quadratic number Geld, b u t of Let 0 be a fixed Eichler order of level F in U . Given an embedding
course not all quadratic points on S need arise in this manner. For example, p : L- B, its order is R = p-'(0)) an order in L; in this situation one
if S is hyperelliptic over Q, or more generally, if S/(w) is isomorphic over Q often says that p is an optimal embedding of R into 0. If 7 E 0 X , then
t o the projective line, then there are infinitely many such quadratic points. we have an equivalent embedding p,: L -t U , by cp,(a) = 7cp(a)7-1; it has
In $5, we complete the program of determining the values of ( D , F ) for the same order l ? as cp. If R = Z[a], and 9 = a t a', n = a d , then to give
which S is hyperelliptic, whicli was done in [12] for D = 1 and in [lo] for an optimal embedding of R into 0 is to give a primitive element p of 0
I;' = 1, and determine for wliicli of these values S is hyperelliptic over R with trace s and norm n; p is impriinitive if p - a E p 0 for some a E Z and
or Q. some prime p. If 7 E 0 the11 the conjugate element pl = 7P7-' (often
The main part of the paper is devoted to t h e study of the real points
on S/(w); it is hoped that this will be of use later on in the study of the
written pl - p) defines an equivalent embedding. llet v(ll, 0 ) denote the
number of inequivalcrit optimal rrnbeddings of R irito 0.
rational poiuts. If w = w(m)is associated to an exact divisor m of D F Sirnilaxly, in the local situation, we havc optimal embeddings of llp into
(see $2 for the notation), then real points on dm) = ~ / ( w ( m ) )will in Op, and two of them are cquivalcnt if conjugate by an elcrr~cntof 0:. Let
general arise from certairl ernbeddings of fi into 0, and t h e number v(m) vp(R, 0 ) denote the number of incquivalent optirrlal cmbeddings of Itp into
of classes of such embeddings is given by Eichler's results on embeddings, Op. Thcn vp(R, 0 ) = 1 if p Y D F . (We can take Op = M2(Zp), and (: 8')
which we review in $1. Then s ( ~ ) ( R )consists of v(m)/2 disjoint circles, if is certainly a primitive element with given norm X I ~trace. Conversely,
there are no cusps or elliptic fixed points of order 2; otherwise Lhis formula
must be suitably modified.
let(: :) = a be such an element. We change bases in Z: by e2 = (:)

The basis results on the arithmetic of quaternions which are used in this +
and el = a r e n , so a e l = ne2 and el ae2 = se2, as desired. This works
paper can be found in the papers of Eichler or in the recent lecture notes (
if 7 = (el e2) = a&--by ) z
y-cz y E 0 2 , i.e., if we can find x, y E Zp with
[la] of VignCras. 111 Z or Zp we write a I 6 if a divides 6, a 11 6 if a divides 6 +
cx2 (d - a)xy - by2 E Z F . This is possible unless p divides c, d - a , and
exactly, i.e., if a divides 6 and a is relatively prime t o b/a, and a - b if a 6, in which case a is not primitive.)
and 6 are associates.
It is a pleasure to thank Y. Ihara for several helpful conversations. Theorem 1 (Eichler). Let h(ll) be the cla.93 number of R . Then

1. Eichler's Embedding Theorem.

More ~recisely,suppoee we are given for each p I D P an equivalence class of

Wet B be a quaternion algebra over Q and let L be a 2-dimensional al-
optim.al embeddings of Rp into Op. Then there are exactly h ( R ) inequivalent
gebra over Q. The canonical involution of 13 and the non-trivial automor-
optimal embeddings of I2 into 0 which are in the given local clawes.
280 A. P. OGG REAL POINTS O N SIIIMUILA CURVES 28 1

(t)
i
Remark. 1) Here we use the Fact that the class number h ( 0 ) is 1; in 2 (if = 1)
general, when 11 is not ncccssarily indefinite, thc left sidc would be a sum
over h ( 0 ) classes. The result is not stated in exactly this form in [5], but b) If p f 11 I" (say f - pk), then vp = pk (if (f) = 0).
the proof is the same.
2 ) Let Ill be the principal order in L, i.e., the order consisting of all
O (if ( t )= -1)
integers, and let Rf = Z -k f K l be thc ortlcr of conductor f . Their class

-- pZk
numbers arc connectctl by Ucdckiritl's formula
(if r,. )
d) If ppF f I then up - (if j' p2k+')'

Proof (sketch). Parts i) and ii) are very fa~niliar,so let us assume that
p2 I F and Op '= : a , b, c , d E Zp . Then Tip = Zp[a], wlicre we
where (k) (4+
is the Legendre symbol, if I, is a field, and oll~crwise = 1.
assume t h a t 3 =. s ( a ) and n = n ( a ) satisfy
In that last case, say I, = Q + Q = {(: :) : a , d E Q), so Rl = z z
has four units k l , *c, where 1 = ( o1 0'), r = (A
and h ( R I ) = 1; then n=O and f 11 3 1
i f f 2 3 n - p 2 a n p
iff 2' < n - f 2 and f 1 3 - 1.
3) The generalized Legendre symbol for the order R = Rf is
Thus we take n as small as possible. If up 2 1, the11 we can write
Rp = Zp[P] with
( ' )
=
( 1, so Fln(p)ln. Conversely, if F I n , then
is primitive in 0, with the desired norm and trace. Thus up 2 1 if
and only if F I n, so up = 0 exactly in t h e cases listed in the statement of
The forrriulas for the local f;xtors up(R, 0 ) are then: the theorem, and we may assume that F ( n frorn now on.

Theorem 2. Let f be the conductor of R and let F be the level of 0.

Note t h a t conjugating
b (mod F ) unchanged.
(icfi 1by an clement of 0 ; leaves a(rnod F) and

Then vp -=vp(R, 0 ) is given below, according to various cases; divisibilities

are to be understood as in Z,, and $p is the multiplicative function with
g p ( p k )= p k ( l + I l P ) and *,(n) = 1 if P I n .
Suppose that p
p )/a, and (i :) and
f , so we are in case a), and L
(: :) are two solutions, so up
( p 12= 1, n = 0. Then
2. Conversely, let
a = (:c fi) have trace 3 and norm 0. Then either

i) If p / D, then up = 1 - (f).
ii) 11 F, then up = 1 +
If p
iii) Suppose p2 I F. and ae2 = 0 = a l e l , so a -
(;F, :), or p )/d and a - (i:). Thus up = 2
in this case, in agreement with a).

a) If (pf)' I F, then up =
I
Suppose p f from now on. Then p 3, so if a = (ic is primitiveI :)
(otherwise). in 0, with norm n and trace Y, then p divides both a and d and so p
282 A. P. OGG REAL POINTS ON SIIIMURA CURVES 283

cannot divide both b and c. If p [b, let do be a standard representative The genus y(m) of s(~
=)~ / ( w ( m ) )is related to the genus g of S by
of d modulo P. Then 7 =
-
y)
(d2do E 0: and 7-'a7 = Fco dol ) -
--
tdO -
a(do) a. Conversely, any do with do(3 - do) 0 (rnodF) leads t o a
solution rw(do), since P I n. Similarly, if p I b but p j'c, we find that a - where e(m) is the number of fixed points of w(m) on S. Let P E S be
a fixed point. Suppose that P is represented by z E B. (Fixed cusps
(7 .".).
-
o(ao) = with bo divisible by p and a0 a standard representative occur only when m = 4; cf. [12, Prop. 31.) Then p(z) = 7(z), where
of a modulo k'. Thus up is the number of solutions x (mod F ) of ~ ( -3x) r 7 E o('). We replace p by 7 p and assume that p(z) = z; we assume also
O(mod F) plus the nunlber of solutions y (mod F ) of y(y - 3) that s(p) > 0, replacing p by -p if necessary. Then Q(p) is a11 imaginary
n(modpF).
-
-
In case a), for example, where n = 0 and f (1 3, and (pf)' I F, say f quadratic number-field. Since p' generates the same ideal 1 = l', we have
pk, the first equation has as solutions z r 0 , s (mod 1~'/pk),which is 2pk p' = cp, where t E o(') n Q(p). In general, t = -1 so p 2 = -m; the
solutions (mod P'), and the second has as solutions y 0, 3 (mod Flpk-l), + +
other possibilities are E = 1 i = 1 ~ 4 if , m = 2, and E = 1 - (3, if
or 2pk-' solution (mod F). The other cases are similar.

-
m = 3; here and in the following cn denotes a primitive n-th root of 1.
The order R = Q(p) n O contains Z[p]; if it is larger, ttien we claim t h a t
m +
3 (mod 4) and R = Z[(1 p)/2]. (Assume that p 2 = -m and let
2. The Atkin-Lehner Group.
As always, let 0 be an Eichler order of level F in the quaternion algebra
I? of discrirninant D. As shown by Eichler, the group of (nori-zero fractional
+
--
a = ( a bp)/2 E 0, where a, b, E Q . Then a = s(n) E Z, and -bm =
+
s ( a p ) E s(l(m)) C m . Z, by (I), so b E Z. Since n ( a ) = (a2 mb2)/4 E Z,
we have a h (mod 2) if m = 3 (mod 4), and otherwise a and b are even,
using the fact that p is primitive in 0, since it generates the ideal I(m), in
two-sided) ideals of O is as follows. We have the obvious ideals xO for x E the case m 0 (mod 4)).
0 X . In addition, if m 11 D F , we have an ideal I = I(m), non-obvious if Thus, t o z E @ representing a fixed point p E 5 of w(nz), we have
m # 1, with I2 = mO. To define it, it is enough to describe the local associated two optimal ernbeddings of R into 0 , corresponding to p and p'
conipletion Ip at any prime p. We take Ip = Op if p m. If p I m (with trace 2 0). Consider an equivalent embedding, defincd by 7p7-',
and p I Dl then I , is the ideal of all non-units in 0,. If pk 11 m arid F, where 7 E O X . If n(7) = 1, then 7/~7-' fixes 7(z), which also represents P.
and we identify Op with

($ i)Op - 4($A).
{(p:c

Note that
I
t ) : a, b, c, d t Zp , then we identify 1, with
Suppose then that n(7) = -1. Then 7p7-' futes 7 ( ~E) @, which represents
the complex conjugate point P E S, according to the real structure of S
defined in [17], and discussed more fully in the next section. If P = P,
i.e., (changing the notation) if Z = 7(z), where 7 E O with n(7) = -1,
then z and hence z is fixed by 7p7-', so TI"T-' = p'. (We do not have
(1) s(l(m)) c m z . rp7-' = p, for then 7 would be a unit in R arid have norm I .) Cor~versely,
The quotient group of all ideals modulo the obvious ideals is generated
suppose that p' = 7p7-', where 7 E O X . Then p' Fies 7(z), SO 7(z) = z
by the I ( m ) and YO is isomorphic to C i , where r is the number of prime
or Z; actually 7(z) = Z, i.e., P = P, since if 7(z) = z , then 7 E O(p)
factors of Dli' and Cz is the group of order 2.
We can write I ( m ) = pO = 011, where p E 0 has norm m, using
arid 7p7-I = p and not p'. Thus I' = P if and only if /I - p', and the
number of fixed points of w(m) on S is the nurnber of inequivalent optimal
Eiclder's results that any ideal (left, right, or two-sided) of 0 is principal
embeddings of R into 0, such that p 0 = I(m). Now any two such p are
and O has a unit of norm -1. Then O = pO Xp-ll since 0 = iiOp-l,
locally equivalent at any prime p dividing m (by Theorem 2 and its proof).
and also o(') = pO(l)p-'. Hence p defines an autornorphism w(m) of
By Theorem 1, then, we have (expect in the case D = 1 and m = 4)
S with w(m)' = 1, called the Atkin-Cchner involution a..sociated to m;
w(m) # 1 if m f 1. These involutions from the Atkin-Lehner group

PI W = {w(m) : m 11 D F ) -- C;.
284 A. 1'. OGG REAL POINTS ON SIIIMURA CUIlVES 285

where R ranges over Z [ & % ] , and also Z[(1 +

@)/2] when Proof. If n(7) = 1, then Zpl = $Zp) has the same image as Zp in S
m E 3 (mod 4), and %[i]when m = 2. and hence in s(,). Suppose that n(7) = - - I . Tlien cr = r p defines w(m),
Thc elerncnts of W arc rational autoniorphisnis of S, i.e., they arc defined so Zp, = a ( Z p ) has the same image as Zp in s ( ~ ) .
over Q, and so they act on S ( K ) for any field K.

3. Real Points, Especially when D > 1.

m -
Let p E U(m), m # 1. T h r corresponding order R = Q(p)n O is either
Z[@] or Z[(1 + P)/2], the second possibility occurring only if
1 (mod 4). The proof of this is the same as that of the correspond-
ing fact in $2, replacing -m by m. Thus our sets Xp are indexed by the
equivalence classes of certain optimal embetldirigs. Clearly Xp = XPp,so
Let p: F~P)--,S be the natural map. According to Shimura, the real the number of sets is v(m)/2, whcre v(m) is the number of inequivalent
structure of S, i.e., the action of co~nplexconjugation on S, satisfies elements of U(m), provided that ,O and -/3 arc never conjugate under O X .
+
In the galera1 case, write v(m) = vl(m) 2v2(m), whcre we have vl in-
equivalent P U(m) with P - -P, arid uz inequivalent pairs +p with
p f -/3. The number of sets X p is thus
whcre t is any unit of norm -1 in 0 .
Given m 11 D F , let Y = p ( z ) satisfy = w(m)P, i.e., &) = ~ c ( z ) ,
where 7 E o('),n(C1)= m, arid p 0 = I(m), in the riotation of 52. Then
= p( z), where p = ~ - ' - - y - ' ~ has norm -m and PO = I(m). Writing The nuniber v(m) is given by Eichler's crnbedding thcorcm; as above, the
= (: :) t M2(R), we have s = (az + b)/(cz + d) or az + 6 = c/zI2+ dz; local factor in the formula is 1 a t a prime dividing m, since thc ideal and
hence the local embedding is futcd, so we get:
comparing imagiriary parts, we have s(P) = a + d = 0, so P2 = -pP' = m.
This is not possible if m is a square and D > 1, so B is a skew field; in
particular, taking m = 1, this prows [I71 that S ( R ) is empty when D > 1.
In any case, let p,: @- s(,) =. S/(w(m)) be the natural map, and
put
P r o p o s i t i o n 1. u(m) = C Rh(R)npfm
R ranges over Z [ f i ] , and also Z[(1++)I21 i f m -
u,,(R, 0), for m f 1, where
l(mod 4). (Notation:
Even i f m is a square, a: = fiis an element of B satisfying s ( a ) = 0 and
a2 = m.)
(6) U(m) = {p E 0 : P O = l(m), s(P) = 0, P2 = m),
The number v(m) is easily computed, so at least we know in all cases
and whether s(")(R) is empty or not. Now S ( ~ ' ) ( Ris) a real 1-manifold, i.e.,
a union of disjoint circles. If D > 1 and if 0(') has no elliptic fixed points
1 , Xp is a circle and P 1-/3, for all
of order 2 with real image in ~ ( ~ then
P E U(m), as wc shall show, so the numtwr of coniponents of s(,)(R) is
for /3 E U(m). Then X p = p,(Zp) is contained in s(,)(R), and these sets
cover s(,)(R) if I) > 1; for D = I , we must add in S(R)/(w(m)). Note
that Zp : c(x2 + y2)- 2ax = b is a geodesic arc in B; if c = 0 it is a vertical
linc: and if c # 0 it is the circle (cx - a)2+ ( ~ y = +
) ~a 2 bc = rn. If D = 1 i11 this case. This formula will have to be modified in the exceptional cases.
(so we take a, b, c, d E Z) and m is a square in Z , then ;5p begins and ends Suppose that P E U(m) satisfies -P = -yp7-', where 7 t O X . Then
a t a cusp; otherwise, X p is closed. P = r2pr-2, so r 2 cornmutes with 3!, and 7 (which generate 13) and so
lies in the center Q of B. Thus -y2 = +1, and 72 = -1 if R is a skew
field. In either case 0 contains the Z-module (1, P, 7, p7)Z generated by
Lemma. If pl = 7pr-1, where 7 t 0 then Xp, = X p .
286 A. P. OGG REAL POINTS ON SMlMUIlA CURVES 287

anticommuting elements ,f3,7 with P2 = m and 72 = i l l so the (reduced) i) If -p = 7P7-', wh.ere 72 = -1, t h e n one e n d o f X p = pm(Zp)is
discriminant D F of 0 divides the discriminant 4m of this module; thus a real E - p o i n t , and every real C-point arises i n this way.
m 11 D F I 4m. Suppose that 72 = -1. Then 7 is an elliptic element of ii) If -0 = 7p7-', where 72 = 1 (necessarily D = I), t h e n one end
o('), of order 4, with a unique fixed point zo in B. Since B7 = -7P, we see of X p is a fixed point of w(m), and every real fixed point of w(m) arises i n
that P(zo) is also fixed by 7, and in the lower half-plane (since n(P) < O), this way.
so p(zo) = z0. Thus zo E Zp and p,,(zo) is a real image of an elliptic fixed iii) T h e two ends of Xp are of opposite type (as i n i) and ii)) if and
point of order 2, or a real "E-point" for short. Note that pl = is also only if R = 0 n Q(p) contains a unit of n o r m -1.
in U(m), and zo E Zp,. The sets Zp and Zp, meet at a right angle at zo
+
(say because pl = (1 + -y)-'P(l + ?), and 1 7 acts as a rotation through Proof of iii). If -p = 7/3r-1 = 71~r11, then 71 = 7p, where
90), so Xp and Xp, meet a t Po = cpm(zo)a t an angle of 180, as they
should. It may happen that P -
PI, in which case X p is a circle, i.e., a p E 0 and p commutes with p, i.e., p E R x , and conversely. The type
changes only if n(p) = -1.
component of s ( ~ ) ( R ) ;we discuss this possibility later. Conversely, let zo
be an elliptic fixed point of order 2 mapping to a real point Po on ~ ( " 1 .
Now suppose that Xp, meets Xp, in a non-cusp, where Pi E U(m) and
Then zo E zp, where P E U(m), and 20 is fmed by 7 E 0, where 72 = -1.
P2 is not equivalent to +PI; we can suppose that there is a z E Jf) with
Then fi7(zo) = f o ,so PI = /3-y has trace 0 (as shown early in this section),
i.e., /3 and 7 anticommute and we are in the situation just discussed. Thus z E ;5p, f l Zp,. Then /32 = PIT, where 7 E O X and n(-y) = I , and
7(z) = z. Thus z is an elliptic fixed point of order e = 2 or 3, and we can
-0 = 7P7-1, with 72 = -1, if arid only if an end of Xp is an E-point. suppose that 7 is of order 2e = 4 or 6. We have p2 = p l y = 7',81,since
The number of elliptic fixed points of order 2 on S is v(Z[i],0). If this
= -Pi If e = 3, then 7/317-1 = = --Ply = -P2, contrary to
number is # 0, t,hen 4 1
F, and the number is either 2' (if 2 D F ) or 1
27-1 (if 2 (1 DF), where r is the number of prime factors of D F . In either the assumption. Thus e = 2, so we are in the situation i) of Propositiori 2:
case, the E-points form a single orbit under W, with (w(2)) as stabilizer
P r o p o s i t i o n 3. If the set8 Xp, a n d Xp, meet i n a non-cusp P, d
when 2 11 D F . In particular, they all look the same, on S or on ~ ( " 1 .
6

Consider now the second case; /37 = -7P, with 72 = 1, possible only
PI I f ~ t P 2 ,t h e n P is a real E - p o i n t .
if D = 1. Then n(7) = -1 (otherwise 7 = r'), so 7' = -7. Then
a = P 7 defines w(m) and a2 = -m, so a has a fmed point z E B. Since a
We have seen that real E-points can occur only if m I D F I 2m, and of
aniiconlrnutes with /3 and with 7, it fixed P(z) and 7(z), which are therefore course ~ ( m > ) 0 and v(Z[i], 0) > 0. I,et us show next that these cases
z, so z E Zp n 2,. Thus X p and X7 = pm(Z7) meet at a real fixed point do occur, a t least if m is not a square (automatic if D > 1). Suppose
first that D F is odd; then rn = Dl!' and m r 1 (1nod4), since D is the
p,(z) of iu(m). Conversely, let P = p(z) E S be a real fixed point of w ( m ) ,
product of an even number of primes, each r 3 (mod 4), and p r 1 (mod 4)
not a cusp; we can assume that a(z) = z and P(z) = z, where a defines
if p I F (we are assurriing that i E 0 ) . P u t Rl = Z[(l + f i ) / 2 ] and
w(m) and p E U(m). We assume that m > 3 and hence a2 = -m. (If
m = 2 or 3, then F divides 8 or 12 and S = Xo(F)has genus 0, and the R2 = Z [ f i ] . We realize Ll in standard form as
real locus of any quotient of S is a single circle.) Write a = p 7 , where
7 E 0 and n(7) = -1; then 7(z) = 2 and s(7) = 0, so -a! = a' =
7'P' = 7P = -P7, and we are back in the situation just discussed. This
proves the first two parts of: ( A prime p ramifies in R if and only if p I m and (f) = -1, i.e., p I Dl
P r o p o s i t i o n 2. Let PE U(m). The possibilities for P - -P are as
so this is the quaternion algebra of discriminant D.) Put /3 =
( -+)
0
follows. and 7 = A). Then 0 = {(,; :,) :a, b E RI is an order in l3 of
288 A. P. OGG ItEAI, POINTS ON SIIIMURA CURVES 289

discriminant D F , and a n Eict~lcrorder of level P'. (This needs to be verified

only a t primes p dividing F . Tlien p --
1 (rnod4), so i E Zp. We have
an iso~nor~hism
of Op onto the standard order

with 7 c (i o,) and /3 rt (" A). ) Thus we d o have real 6-points in this
n0
case. Furthermore, Q(/'?) = 111, while PI
Q ( p l ) n 0 -- R 2 , so [j and arc c c r t a i ~ ~ lnyo t equivalent. Finally, we we have w h a t we want. Note t h a t in either case we have
claim t h a t if 1) > 1 anti we arc in this s i t u a t i o ~ ~thcn
, t h e component of
s("')(R) contairling these sets is
Ilet us treat first t h e case m = 2t = D F . Here we take /3 = (1 + p)7 =
7(1 - p), so P2 = m , p2 = -1, and P p = -pP. Thus there is a real
E-point a t a n end of X p , a n d we need to know whether p p is equivalent
+
t o /3 or not. Now (p l)-l/'?(p + +
1) = 7(p 1) = y ( l - p)p = Pp, so the
general solution a E U X of spa-' = P p is (p + 1 ) a = x + yP E Q(P).
+
Here a E 0 if and only if x y P is a11 e l e ~ n e n of t 12 = 0 n Q ( P ) = Z [PI
of norm 2, i.e., p - p p if a n d only if x 2 - m y 2 = &2 is solvable with
i.e., there are only two E-points on t h e componrnt. This is truc because x, y E Z. If this is not solvable, a n d D > 1, thcn we have #(m) = u(m)/2
t h e stability gronp of the con~poncrltin ~ / ( w ( r n )is) an elcrr~enti~ry abclian - ~ X p with
as in thc previous case. It i t is solvable, thcn we h a ~ e . 2 ~circles
2-group iicting freely a n d transitively on the IS-;)oirlts on the component,
a n d so i t is of order 2. Thus t h e nurnlwr of c o r n p r ~ c r l t sor S ( ~ ) ( ~isI )again
/3- -P, plus circles Xp with P ,)'-P, so #(m) = (v(rn) -t 27-2)/2, with
u(m) == h ( 4 m ) = h ( Z [ f i ] ) .
#(m) = v(m)/2, as in (9), although the reason I t x i changed; iri this case Suppose now t h a t m = t a n d take P = 7 ; as noted above, we have
we have zTp2 special components as just discussed, plus circles Xp where R = 0 n &(P) = Z[P]. We have only t o decide whether /3 and p p are
P+ - P.
Thus we arc reduced t o t h e case D F = 2 t , where t is odd, and
m = t or 2t. If p I t , then p = 3 (mod 4) if p I D and p 1 (1r1od4) if
p I F. We assume t h a t t is n o t a square (automatic if 1) > 1). 'l'hen
- +
equivalent or not. Again (P l)--'p(p
+
+ I ) = Pp, so we ask again whether
therc is a n a E 0 with (p 1 ) a = x -typ, i.e., whether x 2 - m y 2 = f 2
is solvable with x, y E Z or not.
Thus we have proved:
we a realize B a ( ) :6 , ( ) (The odd ramified prirnes
Theorem 3. Let D > 1 a n d m 11 D F , where m i.s not a .square
a r e those dividing t and r 3 (mod 4), i.e., those dividing D.) T h e order ( s ( ~ ) ( R ) is empty if m is a q u r e ) . Let u ( m ) be the number deJined in
--
0 = I(;,:) : a , 6 E ~ [ ( f i ] } has d i s c r i r n i n a ~ i~tt a n d the desired comple- Proposition 1. Then the number # ( m ) of components of s ( ~ ' ) ( R ) is v(m)/2,
unless u(m) > 0 , i E 0 , D F = 2t with t odd, m = t or 2t, a n d
tion 5, = 0, a t any odd prime p; t h e proof is the same as above. Write x2 - m y 2 = +2 is solvable with x, y E Z , in which case #(m) =
Ji 0 0 1
+
(u(m) 2T-2)/2, where r is the number of distinct prime factors of DF.
our Eichler ordcr will contain a
with index 2. If 2 1 D, so 02 is to
be the maximal order, then we j r ~ s tadd in (1 + p)(l + ~ ) / 2 ,which has

-
Example. Consider t h e case F = 1, D = 2p, m = p (a prime). T h e n
trace I and norrr~( 1 - t)/2 E Z. This also works - when 2 1 F, as follows. S(P)(R)is empty if and only if fi 6 O, i.e., p --
1(1nod8). If p 5(n1od a),
W r i k t = b2 + c Z , which is possible, and map 0 2 into the standnrtl order then #(p) = u(p)/2 = h(p). If p 3(mocl /I), then i E 0 a n d x 2 -py2 = ~ t 2
290 A. P. OGG REAL POINTS ON SHIMURA CURVES 29 1

is solvable, so #(p) = (1 + h(4p))/2. The fixed points of w(2) on S are as the cusp 0, so YOo consists of these two arcs and contains just two cusps.
follows: Thus S(R) has 2r-1 components, a single orbit under W, with stabilizer
group (w(F)), when F is odd (arid > 1). O n each component, w(F) is a
i)
ii)
two points in Q(i), if p
two points in Q(-),
--
if p -
3 (mod4);
5 , 7 nod 8).
"reflection," with two hxed points.
Let us assume from now on t h a t the genus of S is > 0; otherwise S ( R )
consists of one circle.
Let P be even (and > 2). We have connected 0 to oo to 112, and hence
(Cf. Shimura [10] for the rationality.) The images in S(P) are then also 0 to 2 / F , applying w(F); we express this by writing ( F / 2 , 1 , F, 2), the
rational ovcr Q, so we see t h a t s(P)(Q) is non-empty (with at least two cusp l / y being listed by its denominator y. If 2 11 F, then w(2) sends F
elements when p z 7 (mod 8)) if s(P)(R) is. to F / 2 and 1 t o 2 (cf. [12]), so we have connected up 2 and F / 2 , and
Y w contains four cusps. Thus we have in this case 2r-2 components, each
containing four cusps, and forming a single orbit under W. The stabilizer
4. Real Points when D = 1. of a component is .(w(2), w(F)), with w(2) a rotation and w(F) and w(F/2)
reflections. Suppose now that 4 1 F. Then the matrix a = (: 'i2),
In this section, S = X o ( F ) is the ordinary modular curve of level F ;
nornlalizes 0(') and so defines an involution of S, defined over Q, fming
B = M 2 ( Q ) and 0 = . Let p: g- S be t h e the cusp m ( y = F') and interchanging the cusps y = 1 and y = 2. If 8 1 F,
then the cusp y = F / 2 is fixed by a, so F / 2 and 2 are connected and YW
natural map, where denotes t h e upper half-plane pj together with cusps. cor~tainsfour cusps. Thus we have 2'-' components, a single orbit under
We recall the description (cf.[ll]) of the cusps of S. For each positive divisor W, with (w(F)) as stabilizer, when 8 1 F. Suppose finally that 4 11 F, say
y of F we have p(t) cusps where t = (y, F l y ) and x is taken modulo li' = 4F', where F' is odd (and > I). Then a sends the cusp y = 2F' t o
t (and prime t o t). These cusps are conjugate ovcr Q, and their field of the cusp y = F', so we have (2F1, 1, F , 2, F'); w(4) sends (1, F ) to (4, F'), so
rationality is Q(ct), ct = e2*ilt, so they are real (or rational) only if t = 1 F' is connected to 4, and w(F) sends (2, F') to (2F1, 4), closing the circle.
or 2. Thus we have rational cusps with t = 1, forming a single orbit Thus we have 2r-2 components, a single orbit under W with stabilizer
under W ; cf. [12] for the action of W on the cusps. If 4 ( F, then t h e (w(4), w(F)), with each component containing six cusps, when 4 11 F. In
rational cusps with t = 2 also form a single orbit under W, consisting of this last case, w(F) and w(4) act as reflections, the fixed points of w(4)
2' elements if 8 1 F, and of 2r-1 elements (stabilized by w(2)) if 4 11 F. being cusps (with t = 2), while w(F1) is a rotation. Thus:
,.
The set of real points of S is the union of the sets YE = p(ZE), where ZE
is the geodesic arc E(Z) = f in and 6 E 0 satisfies S(E)= 0 and n(6) = Theorem 4. Let S = Xo(F), F > 4. Then the components of S ( R )
-1, or E E U(1). Thus YE is an arc in S(R), beginning and ending a t a cusp form a single orbit under W, and are described by the followiny table:
and passing through no other cusps. We do not care whether p: ZE-4 YE is
one-one or not, although t h a t is the case when F > 1 for the "fundamental Case #Components
arc" (cf. [9])Z(l 0 ) = {iy : y
0 -1
>
01, which lies in a fundamental domain
2%F
27-1
#Cusps per component
2
stabilizer in W
(w(F))
for r o ( F ) = o('). 2 j/ F
2'-2 4 (w(F), w(2))
Let us determine the component Y w of S ( R ) containing the cusp 41 P 27-2 6
connecting m t o O = w(F)(oo), so it is fixed (w(F),w(4))
oo = 1/F. It contains Y 27-1 4
(0-1)) 81F (wp'))
by w(F). Taking E = (i 2'))
the set ZEis the line Re(z) = -112, connect-
In all cases, w ( F ) acts on a component as a reflection, the two fixed points
ing m t o -112. If F is odd (and > 1, say), then -112 is the same =
being non-cusps. I f 2 11 F, then w(F/2) is cr reflection (with non-cusps as
292 A. P. OGG ILKAL POINTS O N SIIIMURA CURVES 293

fixed p i n t s ) and w(2) is a rotation, and if 4 11 F, then w(F/4) is a rotation We suppose t h a t m # 2 , 4 from now on.
a n d w(4) is a reflection with cusps (with t = 2) as fixed points. Let us suppose next t h a t m is rrot ;t square, in which case the situation
is sorncwhat similar to t h a t when D > 1. Accordirrg to Tllcorern I, w(m)
has real fixed points only if m = F or F / 2 , in whirl1 cases ~ ( ~ ) / ( w ( m ) )
Now let m I/ P, m > 1, and let pm: $)- s ( ~ be )t h e natural map. cor~sist~s
of 2'-' or 2'-2 half-circles.
Then s ( ~ ) ( R )consists of the old part ~ ( ~ ) / ( w ( r nplus
) ) t h e new part: 'raking now m = E' or F / 2 , let us d c t e r m i ~ l rt h e componerlt X w
of S('")(R) containing t h r cusp m. It contains X,,going from co t o
a fixcd point of w(rn), where it meets Xp; here E U(l), P E fJ(m),
anti cp = -PC. Specifically, if m = I;' we can take c = (h
p .= ( mO O') J

where X, = pm(Z,) = ~ , / ( w ( m ) )is known, by Theorem 4. We continue

t o assume t h a t S has genus > 0, t o avoid trivial special rases. Propositions
and if 2m = F we can take r = (: ) = ( (
R = 0 f l &(P) = Z[p] -- Z [ f i ] . T h e general solution of c1/3 = -/3c1,
2 and 3 are still available, as we made !lo a s s u ~ n p t i o ~about
l D being > 1 1 E 0 X , is ~1 -- cp, where p E R X . If p = 7', 7 E R, then
in t h a t part of $3; recall t h a t these proposilions can apply only if F I 4m.
We have notcd t h a t Zp begins and ends a t a cusp if and only if m = n 2
is a square. This can also be seen by the rule in [12]: let I' be the cusp
of S represented by x l y , where a s usual y I F and x is taken modulo Thus we are interested in p only modulo squares and take p = a + bp to
t = ( Y , F l y ) . Write y = yry" and t = t't", where y' and t' divide m and y" be a base unit in R. Suppose -1 = n(p) = a2 - mb2, so the other end
and t" divide k ' l w ~ .Then PO =. w(m)(P) is represented by x O / ~ Owhere , of Xp is a n E-point, by Proposition 2. The next arc is -Yp,, wlicre Dl =
yo = (mly') . yl', to = t , and x0 -- -z(mod t'), xO E x(mod i"). On the pel = PFPE U(m), going from thc by-point t o a futcd point of w(m), where
other hand, the complex conjugate is represented by -x/y (cf.[ll]), so it nicets X,, ,
-
P = Po if a n d only if yI2 = rn and z r -x (mod t"), i.e., t" =. 1 or 2.
Let us take care of the cnscs m = 2,4. If m = 2, then S ( R ) consists
of 2+-' circles on which w(2) acts as a rotation, so S ( R ) / ( W ( ~ ) )is also
2'-' circles. The situation of I'ropositior~s 2 and 3 does not apply (for If m = It', we have 2 = , fixing the cusp b/(l - a). T h e
F I 4m = 8 implies g = 0, which was excluded), so the new p a r t consists denominator of this cusp is-divisible by r l o odd prime divisor of rrL (since
of ~ ( 2 ) / 2= v ( z [ ~ ]0,) / 2 circles Xp. Thus: a 2 - mb2 = -I), so it has y = 1 or 2 and is connected to co on S ( R \ .
\ ,I
by 'I'heorem 4, m d wc are back or1 t h c half-circle with which we began.
-
For F = 2m, we havc 2 = a2+bm(l-a) az-ab(m+l)/2+b(m-l)/2
2bm(a-1) -a2-bm(1-a) fixing ),
Suppose m = 4, which is the only case in which w(m) has fixcd cusps; +
the cusp (1 - a b)/2(a - l ) , again connected to cm on S ( R ) . Thus, if
. ,
here s(R)/(w(~)) consists of 2'-2 half-circles, with end points a t the cusps n(p) = -1, we have
with t = 2. T h e discussion above shows t h a t the sets X p , P E U(4),
old
will end a t cusps with t = 2, so we havc a t most 2'-2 components.
Actually, it is 2'-'; this is seen from the action of W, or directly: taking
(
2(1+F1) 2+F1
/3 = - r r , 2(1+1.) , where P. = IF1, the arc Zp runs from -112 to
(2 + P1)/(2F1) and so X p connects the cusps with y = 2 a n d y = 2F1, and
and tho formula #(m) = v(m)/2 holds, as we have 2'-' (if 2 )(F) or 27-2
closes off the compofient containing the cusp co. Thus:
(if 2 11 F) such components, a single orbit under W, plus ordinary corn-
poricnts X p will1 p )( -p. Suppose now t h a t n(p) = 1, so the olhcr end
 REAL POINTS ON SIIIMURA CUItVES 295
294 A. P . OGG

of half as many circles as docs S ( R ) , and we also have v(m)/2 circles X p ,

of Xo is also a fixed point of w(m), where X p meets X,,; 61 = cp aa be- p ,)'-P. Thus we have:
fore. If m = F, then c l = f i e s the cusps x1 = - b ( l a) and +
x2 = b/(l -a), which may or may not be connrcted t o oo or1 S ( R ) . (E.g., if T h e o r e m 5. Let S = X o ( F ) Imve genus > 0; let m 11 F, with 7n not
m = 119, then p = 120 + l l p and 2 2 = -11/119 -
oo, while if m = 21, of the form n 2 , n 2 3. T h e n the number #(m) of components of s ( ~ ) ( R )
then p = 55 + 12P and x l = -3114 -- 117 is on the other component is given by formulas (11) to (Id), if m = 2,4, F, F / 2 , F / 4 , a n d otherwise
of S ( R ) . ) By Theorem 4, the test is as follows. Let y1 resp. y2 be t h e #(m) = (#(I)-I- v(m))/2.
greatest common divisor of m and the deriomirlator of x1 resp. x2. Then
xi is connected to oo on S ( R ) if and only if y1 or y2 is a power of 2. If Finally, we have the case m = n 2 , n 2 3; then each Zp begins a n d ends
),
F = 2m, then t l = a-bm a-b(m+l)/2 fuing t h e cusps (1- a+ b)/2(a- 1), a t a cusp, so each X p is an arc with a cusp a t least one end. T h e general
(b - a - 1)/2(a + I), which again may or may not be connected t o oo on element of U(m) is p =
"
nz, -nz,
), where F 1 n2c, i.e., d = F1.m divides c,
S ( R ) . If the fixed cusps of t l are connected to co on S ( R ) , then we have +
a n d bc = 1 - a 2 n 2 ;thctn fixes the cusps ( f l an)/nc. According t o [ll],
one such cornponrnt of s ( ~ ) ( R )correspondirlg t o each component of S ( R ) , for cxarnple, the standard form for the cusp represented by a given rational
with only one class in U(m) associated to such components. Iri the contrary number is as follows: write the rational number in lowest terms as r l y s ,
case, X m contairls two or more half-circles fro111 S(R)/(w(m)); actually t h e where y I F , y > 0; let t = (y, Fly). We now take r a r ~ ds modulo t and
number is exactly two, since the elementary ;~belian2-group W is operating act or1 the I r k by (i S!L), getting ?.sly. In our case we have t = n or 272;
transitively. In this case, we have two clases from U(m) per component, since x/y and -x/y are interchanged by w(m) arid by complex conjugalion,
and so no rrlodilication here of the expected forrnrila #(m) = v(m)/2. we have for each such y exactly (p(n)/2 real cusps x / ~on s ( ~with ',
As t o the I(-points, they have bcrn taken into account if n(p) = -1, x E G = ( Z / n Z ) X / & 1. We call thrse the new cusps on s('")(R); the
so assume t h a t n(p) = I and i E 0 . By the argurrlrnt of 53 (where we old cusps are on S(R)/(w(m)). Changing the notation, the set of all new
assumed that m is not a square, b u t not thnt 11 > I), w ( F ) acts ils complex cusps on s ( ~ ) ( R )is the set of all
conjugation on the elliptic fwcd points of order 2 or1 S ; if 2 11 F, then so
does w(F/2), since ~ ( 2 leavrs ) them fixed. Thus the E-points are real if
m = I;' or P / 2 . The argurnerlt preceding Theorem 3 (deciding whether t h e
where x E G a n d y > 0 is a divisor of d = F / m w i ~ h(y, d/y) = 1 or 2. T h e
c o r ~ p o n e n t scontair~ing13-pc;ints consist of one or two arcs) is still valid, so
field of rationality of l'(z, y) is the real subfield of ~ ( e ~ " ~ 1 "In) . general,
we get, for m = F or F / 2 :
P ( z , y) is a n end of two arcs Xp and sc has t,wo neighbors, for which we

{ [$!{I
(if a2 - mb2 = -1) need a formula; using the action of W, we may assume t h a t y = 1 or 2
(13) #(m) = A + 8)/2 (if a 2 - mb2 = I), with no rsseritial loss of generality.
Given an elen~entof G, we represent it by x E Z, coprime t o 2n. We can
where A is t h r riurr~berof cornporierits of S(R) if the fixed cusps of c r (above) solve cx = -1 + a n , with c, a E Z and d I c, getting an clement 0 of U(m),
are conr~ectctlto oo there, and otherwise A = 0, arid B = 2r-2 if 2 / F , with b = -41 + an). One end is (-1 + a n ) / n c = x / n = P ( x , I ) and the
i E 0, a n d s2- m y 2 == f2 is solvable, and otherwise B = 0. +
other is (1 an)/nc. If n is ever1 (this is the easirst case), then this fraction
The remaining case where w(m) stabilizes a component of S ( R ) is is already in lowest terms, so (1 + a n ) / n c = (1 + a n ) / ( n . d . c l d ) = P(zl, d),
m = F / 4 . In this case S(R)/(w(m)) consists of 2r-2 circles, by Theorem where x' = c/d = l / d x E G. Thus P ( l / d x , d ) is the only neighbor of
4, and there are no real E-poirits, so we have P ( x , I), and more generally P ( l / d x , d l y ) is the only neighbor of P ( x , y),
when n is even. If d > 1 (and 2 1 n), t h r n each new component contains
exactly two cusps and we arc done, so suppose t h a t d = 1; write now
I'(x) = ll(x, 1). NOWari exception to the general rule occurs only when
In the rernainirlg cases ( m f 2,4, F , F / 2 , F / 4 ) , S(R)/(w(m)) consists
296 A. P. OGG REAL POINTS ON SflIMURA CURVES

x = x-' in G; then the corresponding co~nponentmay be a loop X p , component, with no cusp equal t o one of its neighbors, looks like
with only one cusp, or it may be conncctcd t o the old part. P'or exaniplc,
p= )A:( 1 0
fixes P ( 1 ) and anticon~mutcswith t = ( o -,), which f i e s oo, (16). . . ,P ( x , y), I'(l/xd, dly), P ( x d / l , y), P ( 4 / x d , dly), P ( x d l l 6 , y), ...
so P(1) is on t h e component Xm containing m. Also, PI = (n2c;n) I - 9 and its length is 2k, i.r., it contains 2k new cusps, where Ic is t h e order of
1-na/2 -l+n/2 4 i r i G. 15xccptions can occur only if d = I; then P(x) is cqual to one of
l i e s P(1) and anticon~mutcswith r l = (7t12+ +n1,2), which fixes
its neighbors if x 2 = 1 or 4x2 = 1 in G. If so, then P(x) is connected t o
1 / ( 2 + n ) , i.c., the cusp 112, which is connected t o m on the old part. Thus,
a n E-point or t o a Gxed point of *w(rn). Now I'(1) is connected to oo, by
in either case, P ( x ) is the only IWW cusp on its cornponcnt when x 2 = 1 in
t h e same proof as above (this proof did not nrcd n t o be even), and oo is
G , so we have:
connectcd t o P(1/2), since P = 2 -(=)/'
( )
~ixcs(I + n ) / 2 n = p ( 1 / 2 )
Theorem 6.1. Let I' = dm, where m = n 2 , with n even, n 2 4 a n d anticornrnutes with t = (A I:), which fixes co. Thus the component
(and m 11 F'). Xw contains

i) If d > I , then each new component of s("')(R) contains exactly two

cusps, P ( x , y) and P ( 1 lxtl, d l y ) , a n d
Similarly, whenever f x satisfirs s2 = 1 (rnodn), then an old cusp lies

where s is the number of prime factors of d.

ii) If d = 1, then each component of s(~")(R)contains either two new
cusps, P ( s ) a n d I'(x-I), or one new CUSP P(x), if x 2 = I in G . Thus
-
prime divisor of n is
x2
-
between P ( x ) a n d 1'(x/2). Since there are 2'-'

1 (mod4), then we have 2'-'

old cusps and 2.-'
solrrtions f x, t h e corrcspondencc is one-one. Similarly, if i E 0 , ix., each

-1 (rnodn), and an E-point lies between P(x) and P(x/2). To see

such

solutions f x of

this, write n = c2 +e:, where (c, e l ) = 1 and (say) c is odd, c l is even; then
1 = a c l + a l c for s u i h b l e a , a1 E Z. P u t x = a c - a l c l . Then xc+cl = a n 2
+
and - x c l + c = a l n 2 . P u t b = a1 - a x , bl = a a l x . Then bc = 1 - a 2 n 2

where A is the number of solutions of x2 = 1 i n G.

and b ~ c l= I - uqn2, so D = c n ~
1 and Dl =
( -clna
-bl
-al,,a
1 are

Remark. Then
in U(m). They anticommute, since t = DPlnv2 -
- (d
-n-a('+z2)
is in
-z )
U(-1). Thus Zo meets Zp, a t t h e fixed point of t, so X p meets Xp, in an
23-point, and we claim that t h e cusps a t t h e other ends are P ( x ) arid P(x/2).
(if 8 1 n )
(if 2 11 n a n d x 2 z -1 (mod n ) is not solvable)
+
To see this, write n = (u2 v2)/2, c = UV, cl = (u2 - v2)/2, whcre u a n d
+ + +
v arc coprime odd integers. T h r n 1 a n = c a l a ( n cl) = a l u v a u 2 , +

-
(otherwise).
+ +
Suppose n odd, n >_ 3, from now on; we retain the notation of the
beginning of t h e paragraph preceding Theorem 6.1. A neighbor of P ( x , 1)
is (1 + a n ) / n c . This fraction must be p u t into lowest terms; note t h a t this is P(x). Similarly,
v(alv + a u ) -
a n d (1 a n ) / c = ( a l v au)/v in lowest terms; since

-alcl + ac x (mod n),

+
(nc, 1 a n ) is a power of 2. If d is odd (i.e., if F is odd), we can choose
c t o be either odd or even; taking c odd we have (1 3- a r r ) / ( n . d . c l d ) =
P ( l / x d , d), a n d taking c even we have P(1/4xd, d). Thus (applying w(y)),
t h e two neighbors of l'(x, y) are P(l l x d , d/y) and P ( l / l z d , d l y ) . A general
298 A. P. OGG REAL POINTS ON SI-IIMURA CURVES 299

+ +
in lowest terms, and ((u -- v)/2)(a(u - v)/2 a l ( u v)/2) 3 x/2 (mod n), This is in general of length 4k, where k is the order of 4 in G . Exceptions
so we have P(x/2). can occur (with a cusp equal to one of its neighbors) only if d = 2, and
Thus, whenever x2 = 1 in G, an old cusp or an E-point lies between 4x2 = 1 in G, so P ( x , 1) resp. P(x, 2) is equal to its neighbor P(1/4x, 1)
P ( z ) and P(x/2). Now let 1 be the order of 2 in G. If 1 is odd, then resp. P(1/4x,2). Note that P ( x , 1) and P(x, 2) are interchanged by w(2).

-
the component of P ( x ) has length 1 (contains 1 new cusps), where x2 = 1 By Theorem 4, a component of S(R)/(w(m)) is a half-circle whose ends are
in G. In fact, it contains exactly the new cusps P ( x . 2i), while ordinary fued by w(m) and interchanged by w(2), and any E-point (present only
components have length 21 = 2k, as wr have seen. If 1 = 2k is even, then if each prime dividing n is 1 (mod4)) is fixed by w(2). Thus, given
ordinary components have order 2k (k is the order of 4 iri G); so do the 4x2 = 1 in G, and d = 2, so P ( x , 1) is connected to a Fixed point of w(m)
special components, since P(x/2) is on the component i f x2 = 1 in G, and or an E-point, this fixed point or E-point is in turn connected to the image
again our component contains exactly the new cusps P ( x . 2i). Thus: P(x, 2) of P ( x , 1) under w(2). The component in question thus contains all
P ( x .4', 1) and all P(2x . 4', 2), and is fixed by w(2), so it contains all cusps
Theorem 6.2. Let F = dm be odd, where m = n 2 , n > 3. Let k P ( x .2', 1) and P ( x . 2i, 2), and no other cusps. Thus a special component
resp. 1 be the order of 4 resp. 2 in G = ( Z / n Z ) X /(+l). has length 21, where 1 is the order of 2 in G. If 1 = 2k is even, this is the
same as the length 4k of ordinary components, while if 1 = k is odd, then
If d > 1, then each new component of s ( ~ ) ( R )has length 2k, so special components have length 2k while ordinary components have length
i)
4k. Thus:

where 3 is the number of prime factors of d.

Theorem 6.3. Let F = dm, where 2 11 d and m = n 2 ( n > 3). Let k
reps. 1 be the order of 4 reps. 2 in G.
ii) If d = 1 and 1 = 2k is even, then s ( ~ ) ( R )has p(n)/4k components,
each o j length 2k.
iii) If d = 1 and 1 = k is odd, and A is the number of solutions of i) If d > 2, then each new component of s ( ~ ) ( R )has length 4k, so
x2 -- 1 in G, then s ( ~ ' ) ( R ) has A special components of length k, plus
ordinary component3 of length 2k. Thus
2'-l(if i y i n 0 ) where s is the number of odd prime factors of d.
#(m) = A/2 + p(n)/4k, and A =
2' (if i E 0 ) ' ii) If d = 2 and 1 = 2k is even, then all components have length 4k,
and
We suppose from now on that d = F / m is even. Suppose first that, 2 11 d.

- -
#(m) = p(n)/4k.
We have cx = -1 +arb as above, and can take c E 0 or 2 (mod 4) as we
iii) I j d = 2 and 1 = k is odd, and A is the number of solutions of
wish. 'l'akirig c 2 (mod I), we find that (1 + an)/nc = 1'(1/2xd, d/2), and
taking c O(mod 4), we get I'(1/4zd, (1). Thus the neighbors of P ( x , 1) arc x2 = 1 in G, then s ( ~ ) ( R ) has A special components of length 2k, plus
P(1/2xd, d/2) and 1'(1/4xd, d). Applying w(y) and ~ ( 2 ~where ) , y 11 d, y ordinary components of length 4k, and
odd, we get in general
. . .,P(1/4xdl d/y), P ( x , y), P(1/2xd, d/2y), . . .
. . . ,Y(l/4xd, d/2y), P ( x , 2y), P(1/2xd, d/y), . . . Suppose next that 4 11 d. By the usual argument, we find that the neigh-
bors of P ( z , 1) are P(1/2zd, d/2) and P(1/4xd, d) and more generally P ( x , y)
and hence
is betwen P(1/22d, d/2y) and P(1/4xd, dly), for an odd exact divisor y of d.
Applying w(4), we have P(x, 4y) between P(1/2xd, d/2y) and 1'(1/4xd, d / 4 ~ ) ,
 A. P . OGG REAL POIN'I'S ON SI-IIMURA CURVES 301

so a new component is obtained by dividing S by ttic hyperelliptic involution v has points raional
over R or Q or not. If D = 1, then the first pitrt was done in [12]-actually,
this is the hardrst case, and the second questior~does 1101arise, since S has
rational cusps; wc assume that D > 1 frorn now on.
Thc methods for determining whether S is hyperelliptic or not are well
in general of length 6, hut of length 3 if d = 4 and 16x2 = 1 in G , and known (cf. [12] for the case 11 = 1 and [IOj I'or the case F = I), so we give
only then. As t o the old part, since S ( R ) consists of 2r-2 circles, with w(m) here only an outline of the proof and a statcmerit of the results. It turns
acting bs a rotation on each cornponcrit if d = 4 and carrying one circle out that the hyperclliptic involution v is always in W, so whether S/(v) has
onto another if d > 4, we have 2'-"circles if d = 4 and 27-3 if d > 4. real poirit,s or not is contained in the results of $3.
Thus: If S is hyprrclliptic, and F' I F ,then Sn,,,*ris also L'subhypcrclliptic," i.e.,
either hypc~rcllipticor of genus < 1 (since it has a function of degree 2 if S
Theorem 6.4. Let F = dm, where 4 11 d and m = n 2 ( n 2 3). does). A hyperclliptic curve does not admit Ci iis a group of autornorphisms
(cf. [I'L], Proposition I), so in our case a t most three primes divide DP',

i) If d > 4, then # ( m )= 2'-3 + 2s-2p(n), where s is the number of

iz., r < 3. l'hus, if S is hyperelliptic, then D is the producl of orily two
primes, and F is 1 or a prime power; then hyperelliptic involution v is in
odd prime factors of d.
W i f F > I.
ii) If d =- 4, then #(m) = 27-2 + cp(n)/4 + 4 2 , where A is the number
of solutions of x2 = 1 in G. - divide DF. Then S lias good reduction
Let 1 bc any primr which doc,s not
modulo I , a non-singular curve S dcfinecl over F(1), of the same genus g.
By the work of Ihara (cf.[7], csp. p. 293), the "supersingular points" on 3
Finally, we have the case 8 1 d. By Theorem 4, S(R)/(W(WL))consists are all rational over F(12), and their number is
of 2lP2 circles. Wc find that the neighbors of P(x, 1) are P(1/2xd, d/2)
and P(l/4xd, d); applying ,w(d), we find P(x, d) between P(1/2xd, 2) and
P(1/4zd, 1). Similarly, solving (c/2) . x = -1 + a n , with d I c, we see
t h a t P(x, 2) lies belwcen P(1/2xd, d) and P ( l / x d , d/2) and P(x, d/2) lies where ij is the genus of S = Su,rl,.. On the other hand, if S is hyperelliptic
between P(1/2xd, 1) and P(l/xd, 2). Thus our component looks like then so is 3, and so 3 (being a double cover of a curve of genus 0) has a t
+
most 2(12 1) points over F(12). Thus
P(x, I), P(1/2xd, d/2), P(2x, 2), P(1/4xd, d), P(x, 1)

and lias length 4 as do all riew components (using the action of W). Thus:
if S is hyperelliptic. The formula for the genus is
Theorem 6.5. Let F = dm, where 8 1 d and rn = n 2 ( n 2 3). Then
#(m) = 2'-2 +
28-2cp(n), where s is the number of odd prime factors of d.

where $(F) = Fn,,,(l +lip), ek is the number of (inequivalent) fured

points of order k and o is the number of cusps; e2 = u(Z[i], 0), e3 =
5. The Hyperelliptic Problem. Y(Z[S.?],O ) , and for us u = 0 (since D > I). In going from g Lo ij we
+
multiply $(F) by (1 I), and ek by 0, I, or 2, depending on cases, so we
In this section our aim is to deter~ninefor which values of ( I 1 , F ) the have
curve S = S D , is ~ hyperclliptic, and for each such value, whether S is
hyperelliptic over R or Q or not, i.e., whether the curve S / ( v ) of genus 0
302 A. P. OGG REAL POINTS ON SHIMURA CURVES 303

In particular, if S is hypcrelliptic we get (by (21) and (23)) noted in the example a t the cnd of 53, w(2) has four fixed points, two over
Q(i)and two over Q ( G ) , so it is not possible that Ci acts as a group
of rational autornorphisms; this elimirlatcs D = 142. Another argu~rlentis
found in Michon [10], p. 224, and a proof that the cases D = 133, 177 do
In particular, if F = 1 we can take 1 = 2 or 3 (since S has genus 0 if not arise; another way to eliminate these lwo values is to find a few more
D = 6), so D = pq ( p and q prime, p < q ) with ( p - l)(q - 1) <120. Then points on the curve besides the (g - 1) = 8 supersingl~larpoints, so there
are more t h a t ten points over F(4). At any rate, we have the first part of:
certainly D < 300 and we can use Table 5 (pp. 135-141) in 121, as follows.
In this table, for D = pq(p < q), the last column gives t h e dimensions of
S + + , S '-, S -+, in order, where S is the space of new (A la Atkin- Theorem 7. There are exactly twenty-four values of D for which
Lehner) cusp forms of weight 2 for h ( D ) and, e.g., S+- is the subspace S = So,, is hyperelliptic; i n each case the hyperelliptic involution v is
on which w(p) = 1 and w(q) = -1. Now it has been known for a long in W . These values, together with the genus g of S and the m for which
time (since Eichler's work on the trace formula and its applications), t h a t v = w ( m , ), are:
for any D the space S is essentially the same as the space of holomorphic
differentials on S, i.e., the two spaces have the same dimension and the
actions or the respective IIecke algebras are equivalent. This led Kazhdan
and Maaur to surmise that the Jacobian J ( S ) must be Q-isogenous to the
new part of the Jacobian Jo(D) of X O ( D )which
, was proved immediately by
Ribet [13]. The actions of the two Alkin-1,ehner groups on S are "opposite,"
i.e., if the Dirichlet series correspondi~~g
to a nor~nalizednewform is

For three of these values, namely D = 57,82,93, the curve S is not hyper-
then the operator w ( ~ on) new forms for r o ( D ) has cigenvalue -ap = T I ,
elliptic over R; for the other twenty-one values, S is hyperelliptic over Q.
while the operator w ( ~on) J ( S ) has eigenvalue a p ;cf. Atkin-Lehner [I] and
Eicl~lcr[GI, respectively. Thus, for us, the last column of Table 5 should
be read backwards; for exan~ple,when D = pq, the genus of S / W is the Proof of the last part. Since O is a maximal order, there are real poi~lts
dinlension of S--. Checking through the table, we find twenty-four values on s(") if and only if fi E R. This is always the case if m = D and we
D = 26,. . .,206 for which S is hypcrclliptic, with hypcrelliptic involution find that fi B exactly in the three cases listed; for example
@ B
v E W; these valucs are listed in Theorern 7 below. For example, for when L) = 57 since the prime 3 splits in Q(m) but ramifies n B.
D = 206 = 2 . 103, the last column in Table 5 reads 0 , 5 , 4 , 0 , so S has For the remaining twmty-one valurs, we can show that S is hyperelliptic

-
genus 9 and v = w(206), while for D = 82, it is 1,0,2,0, so g = 3 and over Q by finding in cach case suitable "points of complex multiplication"
v = w(41). Instead of using the table, we could of course compute the which map to rational points on ~ ( ~ For 1 . any value of D ,the construction
genera in each case by formulas (22), (3), (4). To show t h a t there are no is as follows. Let 1 be 1, 2, or a prime 3 (mod 4), so I, = Q(Q) has
further values of D for which S is hyperelliptic, i.e., to show that S is discriminant -4, -8, or -1 and odd cl:lss number h . We assume that each
hyperelliptic but v f i n W is not possible, we need check only those S for prime p dividing D does not split in L, so L B. To be specific, we can
which cach w(m), m # 1, has a t most four fixed points, i.e., g ( m ) = 19/21 or realize Ll as
+
[(g 1)/2], by a result of Schoenebcrg (cr. [12], p.452). Checking th-ough
'l'able 5, only D = 133,142, 177 need a closer look. For D = 142, as we
304 A. I-'. OGG REAL POINTS ON SFIIMURA CURVES

and take 0 to be a maximal ordcr containing the ring of all such clcrnents
with u , b intcgral in ~ ( f i ) . (Clrarly B is indefinite, so i t is enough to
check that an odd prime ra~nifiesin 13 if and only if it divides D). Then 0
contains anticommuting elcrnents p = (0,A) and p = ( fi 0
E U(m),
so the point P E S defined by the fixed point of p in @ maps to a real
point of s ( ~by) the
, results of $3. Shimura has shown (cf. [16], p 5 8 ) that
the exact field of rationality of P is thc class field C ( L ) of the imaginary
quadratic numbcr-field L. Thus P is of degree h ovcr I,, and its image
on s(")is of degree h over Q, so we have a rational point on s(') if we The curve S is not hyperelliptic over R for exactly six o j these values,
can embed L in 11 for one of thc nine values 1 = 1,2,3,7,11,19, 43, G7,163 namely (Dl F ) = (6,17), (10, 13), (14,3), (15,4),(21,2),(26, 3).
for which h = 1 . This can always be done if 11 = pq < 2227, ay one
checks; on particular, the twenty hyperclliptic curves S with v = w(D) are This leaves thirteen curves S which are hyperelliptic over R. For the
hyperclliptic over Q. nine with m = DF, we find rational points on s ( ~ coming
) from complex
The only other case is D = 58. In this case S is of genus 2 and hence multiplication by fl,where 1 is a prime with h(-1) = 1, 1 r 3 (mod4),
autonmtically hyprrelliptic over its field of definition Q, since t h r r r is a
positive cmonical divisor (of degree 2) defined over Q. Also, we have
and );( is 1 if p I F and is not 1 if p I D. We can take 1 as follows:
already noted in the example a t thc end of $3 that S(P) has rational points
if it has real points in the case L) =. 2p.

Finally, let S -- S D p be hyperelliptic, where D > t and F > 1. Then

the hyperclliptic involution v must be in W -- C l = Ci, so there are no
dilficult cases. The inequality (24) reduces the problem to checking out
a finite number of cases; this nurnbcr is reduced further by renlembering
Then B = {( a
-lbl a'
) discriminant D.
b, : o,b E ~ ( 6 ) has We have

that must by subhyperelliptic, and so D is among the values just

detrrrnincd (plus the values for which g 5 1). For example, if D = 6 then
anticom~nutingelements p = ((O, t ) and ,b' =
with p2 = -1 and

$(F) 5 78 if 5 )/ 17 and $(F) 5 100 if 7 1/ F, and we find exactly six

P" a
m. There is an obvious order containing p and 0.We construct an
hyperelliptic curves S6,P',nanlcly F = 11,19,29,31,37 with v = w(6F)
a
Eichler order 0 containing by taking Op to be a maximal order containing
if p 1/ F; if p I F , we map = OP onto the standard order by
and F = 17 with v = w(34). The result is:
P I-+ :( A)
and p I-+ (y=i_;3
; note that POp is the required two-sided
Theorem 8. There are exactly nineteen values of ( D l F ) , with D > 1 ideal. Thus 0 is an Eichler order of lcvel F , containing ,i3 E U ( m ) and p
and F > 1, such that S = Sl),F is hyperelliptic. In each case the hypercl- with p2 = -1, with pP = - p p . Thc futed points of p are in Q ( G ) and
liptic involution t: is in W . These values, together with the genua g and the Thus S is hyperclliptic over Q
map to real, i.e. rational, points on s("').
m jor which v = w(m), are: in these nine cases.
306 A. P. OGG REAL POINTS ON SHIMURA CUItVES 307

This method will not work in two of the remaining four cases. The A. Ogg, Rational points on certain elliptic modular curves. Proc.
method depcrids on finding anticomrnuting p and P in 0 with p2 = -1 and Symposia Pure Math. 24 (1973), 221-231.
,B E U ( m ) . In the case D = 10, F = 19, m = 38, the discriminant of O A. Ogg, IIyperclliptic modular curves. Bull. Soc. Math. France
would then divide 4.1.38, so 1 -- 5, which is not allowed (sincc h(-1) = 1). 102 (1974)) 449-462.
Thus we cannot prove that Slo,lgand similarly S14,5is hyperelliplic over K. Ribet, Sur les variktks abkliennes A multiplications rCelles. C.
Q by this method and must leave the problem until another occasion. For R. Acad. Sci. Paris 291 (1980), Skrie A, 121-123.
s15,,
we use 1 = 7, B = {(-:, i) : a , b ) ,( T - k )
~ ( f i )= G. Shimura, On the zeta-functions of the algebraic curves unifor-
mized by certain automorphic functions. J. Math. Soc. Japan 13
and p = (-,0 o).
1
At the prime 2, we can map p onto (141 1 and p onto (1961), 275-331.
G. Shimura, On the theory of automorphic functions. Ann. of
p ( 1 2 (note that flE Z2). Thus SI5,,is hyperelliptic over Q, as Math. 70 (1959), 101-144.
is S39,2,by the same proof.
Except for two unresolved cases, then, the curves S which are hyperel- G. Shirnura, Construction of class fields and zeta functions of al-
liptic over R are hyperelliptic over Q. gebraic curves. Ann. of Math. 85 (1967), 58-159.
G. ~ h i n i u r a ,On the real points of an arithmetic quotient of a
bounded symmetric domain. Math. Ann. 215 (1975), 135-164.
M. F. Vignkras, Arithrnktique des Algkbres de Quaternions. Lecture
References Notes in Matllernatics 800. Berlin, Heidelberg, New York: Springer,
1980.
A. 0 . L. Atkin and J. Lchner, IIecke operators on ro(m). Math.
Ann. 185 (1970), 134-160.
B. Birch, W. Kuyk (ed.), Modular functins of one variable IV. Received February 26, 1982
Leclure Notes in Mathematics 476. Berlin, IIeidelberg, New York:
Springer, 1975. Professor Andrew P. Ogg
M. F~ichler,~ b e die
r Einheiten der Divisionsalgebren. Math. Ann. Department of Mathematics
114 (1937), 635-654. University of California
M. Eiclllcr, ~ b e rdie Idealklassenzahl hyperkomplexer Systeme. Berkeley, California 04720
Math. %. 43 (1938), 481-494.
M. Eichlcr, Zur Zahlcntheorie der Quaternionen-Algebren. J. Reine
Angew. Math. 195 (1955)) 127-151.
M. Eichler, ~ b e die
r Darstellbarkcit von Modulformen durch Theta-
reihen. J. lteine Angew. Math. 195 (1955)) 156-171.
Y. Ihara, Congruence relations and Shimura curves. Proc. Symposia
Pure Math. 33(1979), part 2, 291-311.
13. Mazur, Rational isogenies of prime degree. Invent. Math. 44
(1978), 129-162.
B. Mazur and 11. P. F. Swinnerton-Dyer, Arithmetic of Weil curves.
Invent. Math. 25 (1974), 1-61.
J. -F. Michon, Courbes de Shimura hyperelliptiques. Bull. SOC.
Math. France 109 (1981), 217-225.
 Special Automorphic Forms on PGSp4
I. I. Piatetski-Shapiro

To I.R. Shafnreuich

In a classical situation special autornorphic forms were studied by Maass.

Let us recall their definition. Denote by H t h e Siegel half plane of genus
2. Consider Sicgel's modular forms of a given weight with respect to the
Siegel full modular group. It is lcnown t h a t they have the following Fourier
decomposition:
f ( 2 )= aT exp 2xi tr(T%),

where T runs over the matrices of the form (r;2 ; ) !T n,rl rn t Z . p o t

dT = 4nm - r 2 , eT = ( n ,r , m). The Maass space (following Zagier) is the
space of those I(%) such t h a t the cocnicients aT depend only on dT and
e T . The forms which lie in the Maass spacc do not satisfy the Ramanujan
conjecture. That was one of the reasons why Maass studied these forms.
The aim of this paper is t o study a similar space for the group P G S p 4
over an arbitrary global field k. Our special forms do not satisfy the
Ramanujan conjectures either. We prove that all special forms are lifts
from z2. We also prove the local analog of this result.
Our results give an intrinsic characterization of the Weil lihing frorn
SL2 to PGSpa = SO5 in terms of Fourier coeficicnts. In $1 we introduce
a loc;tl analog of the Maass spacc which are representations over a local
field k having the U-property. A main result of $1 is that any preunitary
represeritatior~with the [/-property is a lilt from S t 2 .
In $2 we introduce a noti011 of special autornorphic forms (a global analog
of Maass forms), and we prove that special cuspidal auto~norphicforrr~sare
z2.
lilt frorn Our proof is conlpletcly diITerent from that of Maass-Zagier
and is based on representation theory and on R. IIowe's general theory of
dual reductive pairs. Howe's theory of the rank of representations of Spz,
is also used here. In the appendix we give a simple proof of one corollary
of Howc's theory.
I am very grateful t o my student D. Soudry for his help in the preparation
of this manuscript. I thank N. Wallach for his valuable rrrn:~rks a r ~ d
 I. I. PIATETSKI-SIIAPIRO SPECIAL AUTOMORPIIIC FORMS ON PGSpq 31 1

linguistic corrections. -
2) To any nontrivial character ?C, of k p we can attach the Weil repre-

Notations and Basic Dejinitions

sentation w+ of Sp2,(kp) which acts on S(k;), the space of Schwarz-Uruhat
functions on an n-dimensional vector space over k,. We will view Sp2,,(kp)
-
Let J = (-:2 :). as the collection of pairs ( g , E ) where g E G a n ( k p )and E = *1. Then we
have that the Weil representation w+ is the unique representation satisfying
GSp4 = { s E GL4 I t g ~ =g X(g)J, X ( g ) E k * )
G = PGSp4 = G S p 4 / { X l , tI X E k * )
(i) 0 ta-1 )>( ) P ( x )

Denote by & the parabolic subgroup of G which preserves a nontrivial

vector. Let & = F'N where 11' is the Levi subgroup and N the unipotent
radical. Let Z be the center of N , and let I) be the centralizer of Z in Q.
Let k be a global Geld with A its ring of adeles. Any character of
where ~ ( tis) the well-known Weil constant. The similar formulm will be
Sk \ Sn is obtained as $ ~ ( s =) 4 ( t r T S ) , where 1C, is a fixed nontrivial
true for the adelic case.
character of k \ A and T is a symmetric matrix with entries in k . We call
qhT nondege~lerateif det 7' # 0 . Denote by OT the stabilizer of 4~ in M, 3) We now introduce the notion of a dual reductive pair due t o R. IIowe
and by 0;. its connected component. [8] in a ~lightllymodified farm. Let k be'any field. Let G I ,G 2 be two
Let p ( g )be an automorphic function on G . Introduce thc Fourier coelfi- reductive subgroups of Sp2,(k). We will call the pair ( G I ,G 2 ) a dual
cient ,- reductive pair if

(a) s l g z = ~ G J I ; Vgi E Gi,Vg2 E G z .

It is easy to see that (Hence GlG2 is a subgroup of Sp2,(k).)
p,(SGJ) = c p , ( g ) , V6 E O d k ) . (0.2) (b) The full centralizer of G ; in Sp2,(k) lies in C I G z ( i = 1,2).
We call p ( g ) special if it satisfies (0.2) for any S E O k ( A ) and for any
The typical example of a dual reductive pair will be the following:
nondegencrate T.
G I = SO,, G 2 = S p 2 [ . In our paper we consider the case m = 5 and
1=1.
Notations Related to the Weil Representation
4) Let k be a global field and let ( G I ,G2) be a dual reductive pair lying in
Sp2,(k). Denote by Z the 2n dimerisional space in which Spz,(k) acts. P u t
The theory of the Weil representation is well known. We recall that the
Z = Z1 @ Zz where Z1 and Z2 are Lagrangian (isotropic) subspaces. We
main facts are:
know that the Weil representation of Fp2,(n) acts on the space S ( Z ~ ( A ) )
-
1) there exists a metaplcctic group Sp2,(kP),where k p is a local field.
- the space of Schwarz-Bruhat functions on Z1(A).We define
This group is a double covering of SP2,(kp).Similarly there exists a group
of G2,(~) where A is the adele ring of any global field k.
 SPECIAL AUTOMORPIIIC 1'OILMS O N PGSp4 313

It is known that this function has moderate growth and hence the integral We will tlcr~ote( , ) the symmetric form on X which is preserved by
J'GSp4. We will denote by ( , ) the skew syrnmctric form on a two-
dimensional space Y preserved by SL2. This way we define a dual reductive
pair G, SL2. In fact, it defines a liltirig from to G. z2
Assume thal the automorphic form p(g) is a lift, i.e., it car1 be written
in the form ,-
makes sense, where f 2 is any cuspidal automorphic form.
Assume thal f 2 E ~ 2 where
, 7r2 is a cuspidal autornorphic representation;
then the set of functions of the form (0.5) is an automorphic representation
PI, [lo]. where n(h) is a cuspidal autornorphic function on x 2 ( A ) . Then [3]
Now we describe a dual reductive pair, which is important for our
purposes.
Consider the space
where W(h,$p, a ) = SK\a
$ ( - ~ ~ ) u ( ( A f ) hda
) and ZT = (y2, -T) is the
X I
= {T E M4(k) T J is skew sym~nctricarid tr(T) = 0). special point in %I which depends only on T. N is the unipotent subgroup
of SL2, which preserves yz. Since any 6 E O;.(A) preserves z ~ we , have
X is a 5-dimensional space over k. We get a representation of GSp4 on X
by
g: T--t g-ITg. This shows that any form which comes as a lifting is special.

It is easy to check that this action preserves X. We put a symmetric non-

degenerate from ( , ) on X by defining
$1 R e p r e s e n t a t i o n s over a L o c a l Field with U - P r o p e r t y

Let (T, V) be an irreducible srnooth representation of Gk where k is

P u t G =. PGSp4. Then the above action of G on X imbeds G as the a local field. Let $, be a riondegenerate character of Sk. We say that
(T, V) has the I/-property (uniqueness) if, whonvcr a linear functior~alLT
connect-ed component of the orthogonal group O5 preserving this form.
s E
satisfies l T ( ~ ( s ) v )= ~ / ~ ~ ( s ) l( ~ ( vS)k , v E V, T nondegcrirrate), then
Denote it by 0:.
There exists a basis of X such that the symmetric form is represented ) any 6 E 0%.Assume that ( T , V ) is a preunitary
l ( ~ ( 6 ) v )= l ~ ( v for
irreducible smooth adnlissible reprcsentslion with the U-property. We will
by
prove that rig is irrrducible wl~cnwe consider .rr as n unitary reprcscntition.
01000
We also prove the followi~igrigidity property. If (xi, Vl) and (n2,V2) are
two such representations and nllq rr then T I --7r2.1 It is easy to
deduce from [3] that the Weil lifting from x2 lias the U-property. In this
paper we prove the converse, that is, that any representation satisfying the
U-property comes as a lift from x2.
For this realization of G, P is a parabolic subgroup preserving an isotropic
line and Q a parabolic subgroup preserving a two-dimensional isotropic
subspace. l ~ h i rigidity
s property is false without the U-property, even if nilq is irreducible.
 SPIXIAI, AUTOMORPHIC FORMS ON P G S p i

L e m m a 1.1. Let (n, V) have the [J-property; the space of the linear pk (mo) = lo. Now define
functionals o n V satisfying

is at most one-dimensional. where so is a. sufficiently large open compact subgroup of S. This vo satisfies
the properties of the lemma.
Proof. Consider the group R = O;.S. Let /I be any character 0 $ . Let
1 be a linear functional satisfying Remark. If vo E V satisfies (1.4), thcn for any T with det T = 0 we
have lT(v0) = 0. In order to prove this, if sufFices to verify that a set of T
such that l ~ ( v 0 #
) 0 co~isistsof T of the form m ~ o m - l ,m E sup X(m).
It is known that the space of linear functionals satifying (1.1) is at most This last set is closed and does not contain any T such that det T = 0.
one dimensional 1.11. III our case p = 1. Let ly- be a linear functional such that (T-nondegenerate)
We describe a generalized Kirillov model. Let 1' be a non-degenerate
symrnetric matrix of order two. Ilct (x, V) be an irreducible smooth repre-
sentatiorl of G' (k a local ~lonarchimedc:mGeld). We say that the integral We denote by 1 y (the unique up to a scalar multiple) functional IT such
that
lT(x(6)v) = lT(v), Vv E V, V s E O$. (1.7)
Let 1'"': L;--+ C be any functional such that linv(n(6)x) = linv(x),
exists if it stabilizes for large open con~pactsubgroups of 5'. Let 7;,. . . , T2 Vx E Li, V6 E 0;;. We denote by 7r the natural action of OTi on I,,. Denote
be representatives of classes M-equivitler~tsy~nmctricmatrices. P u t I
by Ly = (1 L; linV(l)= 0 ) .

Denote Li = V / K and
functions
V - L; the natural projection. Introduce the
will be an invariant functional for some T. It is easy t o check that any
invariant functional has the form (1.8).
d ( m ) = cpi(n(m)v). (1.2)
L e m m a 1.3. If ( n , V ) is an irreducible smooth representaion, which
does not satisfy the U-property, then there exists a vector vo E V, vo f 0

L e m m a 1.2 Let X(m): M -

Lio be a smooth function with a compact
support. Then there exists vo E V such that
). 0, VT.
with l ~ ( v o=

Proof. If (n, V) does not possess thc U-property then there exists io
rp$(m) = X(m) and rp',o(m) = 0, i # io. ( l a 4 )
such that L!,, is nontrivial. Using Lemma 1.1 we can construct vo # O such
that p,o(m) E Ly, Vm, Vi. Thus vo satisfies the assumptions of our lemma.

If ( T , V) has the U-property, then according to Lemma 1, the spaces Li

Proof. It sulGces to prove the lemnla for X(m) = lo .e,(m) where eo(m)
are one-dimensional and the group OFr, acts on Li trivially. Hence we can
is a cl~;tracteristicfunction of a sufllcierltly small open compact neighbor-
view pi(m) x i functions with values in C.
hood of a point mo E iLI and lo L;,. First we construct vl V such that
 SPECIAL AUTOMORPHIC FOILMS O N PGSpq 317

Assume now t h a t (T, V) is an irreducible preunitary representation satis- Lernma 1.5. Assume that P = MS and Q = FN contain the same
fying the U-property. Uorel su6group B of G. Then for any nondegenerate symmetric T except
for a subset of a smaller dimension,
Lemma 1.4. Under the appropriate choice of the measure d m on
OSi \ M, we have
where BI = B n M .

Proof. It is well known that B l \ M is a complete algebraic manifold

of dirnensior~1. O& is a conriccted algebraic group acting on this manifold.
Moreover, - Hence either its image is one point or dense in the manifold. It is easy to
V @L'(o+~
f~: \ M), determine when its image is a point. In this case 0%C B1.Such T lies in
a subset of positive codirnension.
where the summation is with respect to those i such that pi is nontrivial.

Proof. Consider the completion V of V . According to the result of the Theorem 1.1.
(1) If (n, V) i.9 a preunitary ~ m o o t hirreducible representation with the
appendix, we have the following spectral decomposition with respect to the
U-property, then for any Q-invariant subspace Vl, we have
unitary operators ~ ( s ) s, E S

(2) If (TI, VI), (nz, Vz) are two such reprcsentations, satisfying T ~ J ~ ~ T

The integration is with respect to all nondegenerate symmetric matrices. as unitary representations, then z ~ 2 .
(Recall that any character of S has the form $ ~ ~ ( =
s ) $(tr T S ) where 1' is
a symmetric matrix of order 2.) The meaning of (1.11) is t h a t there exists Proof. Let Vl be a Q-invariant subspace of V. Denote by
a map that attaches to each v E a measurable function f v ( T ) E VT such
that
( ' " 1 ~ )= / I ~ ( T ) \ ~ ~ T J
First we show that Fi is a n M invariant set. Indeed, let mo E Ti artd
where d T = dp,(T) is a spectral measure, corresponding to r / S . m l E M. We show that mom1 E El. There exists GOsuch that p$(fio) #
Let v E V be a srnooth vector. Then there exists a n open compact 0, arid there exists G l , close to m l , such t h a t f i l = r72; 16 k o b J b E 131,
subgroup Mo C M which stabilizes v. f v ( T ) is preserved by Mo ;w well. 6 E O$,. (Imnrna 1.5). We gct that G o G l E E;, which implics mom1 E Ri.
Let 7'1 be a nondegencrate symmetric matrix of order 2. Then its orbit l h i s shows that E; = 0;; \ M. Using the same argument as in the proof
r ,

under Mo is an open neighborhood of TI. This irnplies that fv(T) is of Lernma 1.2, one can show t h a t for any mo E Ei,,3vo E Vl such that
locally constant on the set of the nondegenerate T's. According to our r z ( m ) = 1, m E U(mo) and p:O(m) = 0, if i # io or m 6 U(mo). Since
h', = 0;: \ M , Vi, such that cp' is nontrivial, we have (see 1.10) that the
construction, f,(T) defines a map V + VT which commutes with the
natural action of S. (On VT, S acts according to the character +r.) Since
T has the U-property, all the spaces VT should be one-dimensional, and 0$
acts trivially on O$.. Thus we can identify f v ( T ) with one of the function
set of such functions is dense in V .
Now we prove the second statement of our theorem. Let :I V1
be a Q-isomorphism, which exists according to our assumptiorl. Put B1 =
- V2
M n Q. It is clear that if v E Vl is invariant with respect to a conlpact
 SPECIAL AUTOMORPIIIC FORMS O N PGSp,l 319

open subgroup U of B1, then p t ( m ) is locally cor~stant.We show that if This implies (see Theorem 1.1) t h a t (n, V), viewed as a representation of Q,
p t ( m l ) = 0, then p k V ( m l ) = 0 (for v as above). Indeed, p t ( m l ) is a linear can he written as i n d g r , where T is an irreducible unitary representation
functional which is uniquely defimd by its transfornlatio~llaw with respect of D ,on which Z acts according to a nontrivial character 4. IIence, ?- is
to S. of the form o @ w+, where Q is a unitary representation of x2. Consider
It is obvious that is nontrivial for x l iff the corresponding map for x2 the Weil lifting 0(o,$). 0 ( u , 4 ) has the U-property, and its restriction t o
is nontrivial. Denote by S'O the subset Q is isomorphic to o @ w+ [3]. The rigidity theorem (Theorem 1.1) can be
proved in the case when we assume a priori t h a t only one representation is
{ v E V1 Ip>(m) is locally constant with compact support and unitary. This implies that n = O(u,+).
p;(m) = 0 for i f i o ) . We consider now the archimedean case. The idea of the proof is similar
to the nonarchimedean case. It is easy to see that if k = R, then there
Similarly we introduce S:. It is clear that a. sends S y t o SF. Since a are two equivalence classes of nondegenerate symmetric matrices of order
is a linear map preserving zero and commutir~gwith right translations by 2. The first class consists of definite matrices and the second consists of
elernents of B1, a. is a multiplication by a constant on S?. Ilcnce, on Sy , indefinite niatriceu. In the case k = C, there is only one equivalence class
a commutes with the action of M . Since @Si is dense in and since of nondegenerate symmetric matrices.
M and Q generate G, X commutes with the action of G. This proves our Let (T, V ) be a smooth, preunitary, irreducible admissible infinite dimen-
theorem. sional representation of G. Let us recall that this means that there exists
in V an invariant herrnitiari form. We let I1 be the completion of V with
Lemma 1.6. If Z C G i s a n y unipotent subgroup a n d ( a , V ) is an respect to this form. Then V consists of all the vectors in I1 lying in
irreducible infinite dimensional unitary representation of G, t h e n i t has n o the domain of definition for all the operators of the enveloping algebra of
vector i n V which is invariant with respzct t o Z . G. It is easy to see that V is a nuclear space. Denote by V' the space
of continuous linear functionals on V. Then we have a rigged (equipped)
Proof. It follows from the Howc-Moore theorem that any matrix co- space:
efficient of an infinite dimensional unitary irreducible representation tends v C H C V'. (1.12)
to zero [7]. For the definition of a rigged space see [6].
P u t Q = FN where F is the reductive part and N the unipotcnt radical Let T be a symmetric matrix of order 2. Denote by LT the space of all
of Q. Denote by i! the center of N. Let D be the centralizer of Z in Q . It continuous linear furictionals on V satisfying
is well known t h a t any representation T of L) on which Z acts according to
a given nontrivial character 4 is of the form o @ w+, where w+ is the \Veil
representation of = E2. N which corresponds to the character and +
o is an irreducible unitary representation of n2.
r is irreducible (unitary) and denote by V$ the subspace { v E V I lT(v) = O,VIT E LT).
ifl o is irreducible (unitary). Let pT be the natural projection from V onto VT = V$ \ V. P u t

Theorem 1.2. A n y irreducible smooth preunitary representation

(n,V) with the U property can be obtained as a Weil lifting from some ir-
We now consider the equivalence classes of allsymmetric matrices. Accord-
reducible unitary representation of E2.( W e assume here that ( x , V ) is
ing to the spectral theorem which is proved in 161, we have
n o t a trivial representation.)

Proof. Let (n, V) be as above. It has to be infinite-din~ensional.Accord-

ing to Lemma 1.6, it has no vector which is invariant with respect to Z.
 SPECIAL AUTOMORPIIIC FORMS ON PGSp4

where M iis the stabilizer of 2;. Take a finite place p l # po. Thcn take the non-zcro vector u l E npl
In t11e appendix it is proved t h a t except for the case k = C arid n the which was defined in Lemma 1.2 (see the remark after the lem~na). Let
Weil representation - the contribution of the degenerate terms is zero. u = @upE n be a nor)-zero vector sucll t h a t u p , = U O , up, = U I . Lel J(g)
We now explain how t o prove a lcmma similar to Lemma 1.2. In order be the automorphic function in n which corresponds to u. We will show
t o prove it, wc 11avc to use the same argument as in Lemma 1.2 and also t h a t under our assumptions
the fact2 t h a t the Lie operators act locally on the runctioris pl(m).
This
implies that any smooth function with compact support on M , \ M defines
some vector iri V . From the definition of u l , it follows t h a t
Now theorems 1.1, 1.2 can be proved exactly as in the nonarchimedean Vp E PA VT, d e t T = 0.
) 0
f ~ ( p=
case.
If det T f 0, then since f(y) is a special automorphic form, fT(P) defines

-
an invariant linear furlctional on n. Using the property of uo wc obtain
thai, JT(p) = 0 for p E PA and invertible T. I t is well known t h a t Pk \ PA
$2 Special Automorphic Forms is a dense subset of G k\ GA. Hence, from (2.1) it follows that: f 0, a
contradiction.
In this section we will prove t h a t any cuspidal special automorphic form
on (2 is a lifting from x2 ('r11~0r~m 2.2). It seenis t h a t thc assurnption C o r o l l a r y . Let 7i be a preunitary irreducible special automorphic rep-
tliat the autorr~orphicform should bc cuspidal is not necessary. Wc need it resentation. Then xlQA is irreducible as a unitary representation.
sincc the lifting of noricuspidd autornorphic forms from z2 was not yet
investigaled. Proof. This follows from Theorem 2.1 and Theorem 1.1.
Lct r be a preunitary irreducible spccial autornorphic representation.
T h e o r e m 2.1. Let a = @ r p be an irreducible automorphic special Let f E n. P u t
r
representation; then each component np has the U-property.

Proof. Let J(y) E n. It is easy t o see t h a t

Then !$(by) = f$(B), VS E Dk.EIencc the set of f+lu, is an automorphic
represenlation of DA. Consider the map J --t J$luA. Denote it by g.
It defincs a homomorphism n --t I n dQD i r . We know 1h;tt the image of n
is nontrivial. From the above corollary we know t h a t nlQ, is irreducible
where
as a unitary representation. This iniplies t h a t nlQ, = ~ n d z : ~
as unitary
representations and r is irrcducible.
Assume t h a t np, does not have the U-property. Then, according t o
lemma 1.3, there exists a non-zero vector uo E nposuch t h a t for any linear T h e o r e m 2.2. Let n be a n irreducible cuspidal special automorphic
functional of type lp,
we have representation of G. Then there exists an irreducible cuspidal automor-
phic representation a of =2(A) with a missing character such that
1j;!"(uo) = 0. (2.3) n = O ( 0 , $J).

Proof. We have already proved tliat n J Q A= 1nd:' where r is the

2 ~ N.
. Berristein told me a very sirnplc proof of this statement based on an cnlbedding
into a representaLior~which is induced from the parabolic subgroup. unitary irreducible autornorphic reprcscntntior~of D A . It is known t h a t any
 SPECIAL AUTOMORI'HIC FORMS O N PG'Sp4

irreducible representation of DA with a given restriction to the group ZA Proof. We first prove the following lemma. Let k # C.
( 2 is the center of D) can be uniquely written in the form a @ w+, where o
is an irreducible representation of x 2 ( A ) and w+ is the Weil representation Lemma 1. Let (T,H)be a n irreducible unitary representation of D
of DA which corresponds to the character 4. It is easy to see that if T is s ~ i c hthat Z acts nontrivially. Consider the spectral decomposition of I f with
autornorphic, then so is a . It remains to prove that a is cuspidal with a respect to S . T h e n the spectral measure of the aet of degenerate characters
missing character +. We first see that of S i.9 zero.

f ( z )d z = 0 V f E n, (2.5) Proof. It is known that any irreducible unitary representation of D such

IZ, \z* that Z :tcts according to a nontrivial character 4 is of the form a@w+ where
W+ is the Weil representation corresponding to $J and a is an irreducible
since otherwise n possesses a Whittaker model 131. . . However, any local
component of n cannot have a Whittaker model because it represents a
unitary representation of x2and is hence infinite dimerlsional. In the case
k = C , it is not true because w+ is a reprcsentatiori of D itself and hence
lifting front x 2 [ 3 ] . Since the unipotent subgroup of z2
C D and Z are we can take a to be any representation of z 2 ( C ) ; for instance, the trivial
conjugate in G, we get that the restriction of any autornorphic function representation. Write S = X Y where X is the unipotent subgroup of SL2
cp E T ton2, is a cuspidal function, and hence, using the results of [3] we
- of N.
+
have that is a rriissirig character of u.
We now prove t h a t K = O(a,+). We know that n and O(a,+) are special Taking S = {(I2
0
S)
and Y is the maxirnal abelian subgroup

12
-
I S-symmetric , X corresponds to
automorphic representations and we also know that their restrictions to
Q A are of the form ~ n d g ; , where T = a @ W+ [3]. Using Theorern 1.1 we
obtain that n g O(a, +). Since the multiplicity one theorem is true for such
and Y corresponds to ( (ii)).
",
Consider the spectral decomposition of
wd with respect to S. The spectral measure is concentrated on matrices
representations, we obtain t h a t n = B(a, +) [3].
of' the form T = ( $1 and we can take as a spectral measure just
the staridard ~ ~ e b e s g umeasure
e dp. For the representation a the spectral
Appendix measure is concentrated on matrices of t b forni 1. = (: 0) where a is in
a union of cosets module ( K * ) 2 ,and we can take the spectral measure to be
The aim of this appendix is t o prove the following theorem, which is due d a . Hence the spectral decomposition of T with respect to S is concentrated
t o R. Howe. on matrices of the form (O0 a 0 1 p
) +(p p) =
, .
(k p1i; a) wit11 measure
Theorem. Let (n, I ) be a n infinite dimensional irreducible unitary dpd,. The degenerate matrices are those with a = 0. Clearly,
..
the spectral
representation of G = PGSp4(k), where k i s a n y local field. I n the case measure of the set of such matrices equals zero.
k = C we exclude the W e i l representation. Consider the spectral decom-
position with respect t o S : Lemma 2. Let k # C and ( 7 , H ) be a unitary representation of D o n
which Z acts according t o a nontrivial character; then the spectral measure
of the set of degenerate characters of S is zero.

Proof. This follows from the fact that H is a direct sum (maybe con-
where T is the symmetric 2 X 2 m a t r i z which defines the character
tinuous) of irreducible unitary representations satisfying the assumptions
&(S) = +(tr TS). T h e n the spectral measure of the set of all degenerate
of Idernma 1. Wc now prove our theorem for the case k # C . Consider the
matrices equals zero.
324 I. I. PIATETSKI-SEIAPIRO SPECIAL AUTOMORPIJIC FORMS ON PGSpq

spectral decomposition of II with respect to Z. We have References

D. Zagier, Sur la conjecture de Saito-Kurokaws (D'apres 11. Maass),

Serninaire dc Thkorie des Nombres, Paris 1979-80. Sdminaire
Ddlange-Pisot-Poitou. Birkhauser.
where 7 corresponds to the character $(yz) = + 7 . The spectral measure 11. Maass, ~ b e cine
r Spezialschar von Modulforrnen zweiten Grades.
of (7 = 0) equals zero, since otherwise there exists a vector v E 11 which is Invent. Math. 52 (1979), 95--104.
invariant wiLh respect to ~ ( 2which
) ~ contradicts a well-know11 theorem of I. I. Piatetski-Shapiro, On the Saito-Kurokawa lifting. Preprint.
IIowe and Moore about the vanishing of matrix coeficients at infinity [7]. February 1981.
For any 7 # 0, Z acts according to a nontrivial character on N ,and hence M. E. Novodvorsky and I. I. Piatetski-Shapiro, Generalized Bessel
we can apply Lemma 2. This proves the theorem in the case k # C . models for the syniplectic group of rank 2, Mat. Sb. 90(2) (1973),
We now consider the case k = C . 246-256.
We consider the spectral decomposition (2). For the same reasons the I. I. Piatetski-Shapiro, 1,-functions for G S P ( ~ )Preprint.
.
spectral measure of ( 7 = 0) equals zero. Each of the spaces I i , can have I. M. Gelfand and N. Ya. Vilenkin, Generalized functions, Vol. 4,
a subspacc II: which is a multiple of w*,. II: is the sub-space of 117 such Academic Press, 1964.
that the spectral measure for the spectral decosnpositiorl of 117 with respect R. Uowe and C. Moore, Asymptotic properties of unitary repre-
t o S is concentrated on degenerate characters of S . The reason for this is sentations. J. Fun. Anal. 32 (1979), 72-96.
the following: any irreducible unitary representation of Dl such that Z acts It. Howe, 0 series arid invariant theory, Proc. Symp. Pure Math.,
according to a nontrivial character, has the form a @ w+,, where a is an XXXIII, Part I, 275-286.
irreducible unitary representation of SL2. Ilence a can be either infinite R. Howe, The notion of rank for representations. Preprint.
dimensional or trivial. In the first case Lhe spectral measure concentrated R. IIowe and I. I. Piatetski-Shapiro, Some examples of automorphic
on degenerate characters is zero. The proof of this fact is the same as that forms on GSp4. Preprint.
of Lemma I. In the second case the spectral measure is concentrated on
degenerate characters.
Now consider the space Received June 22, 1082

Professor I. I. E'iatetski-Shapiro
Department of Mathematics
Yale University
By construction H 0 is invariant with respect t o Q. We have t h a t H 0 is
Ihx 2155 Yale Station
invariant with respect to the Lcvi subgroup M of P, since M transforms
New IIaven, Connecticut 06520
degenerate characters of S to degenerate characters of S. Thus 15' is
G-invariant and hence H = IIO. It. I-Iowe proved that if the spectrum of an
infinite dimensional irreducible unitary representation of G is degenerate,
then H is the Weil representation [9].
 Sous-vari6t6s d'une vari6t6 ab6lienne
et points de torsion
M. Raynaud

To Z.R. Shafarevich

Soicrit A une variktk abkliennc dkfinie sur le corps des nornbres complexes,
T le sous-groupe de torsion de A et X un sous-schkma fer~nkintkgre de A.

ThdorBme. S i T n X est dense dans X pour la topologie de Zariski,

alors X eat le trandat6, par u n point de torsion, d'une sous-varie'tk abi-
licnne de A.

Cr rksultat g6nkralise d'ur~cpart [9], qui traite du cas oh X est une

co~irbect ci'autre part [2] oh T cst rcmplack par sa cornposante n primaire
pour n enticr > 1.
Indiquons le priricipc clc la dkmonstratiori dans le cas oil X ne contient
pas dc triinsl;Lt6 d'une sous-variktb ab6lirrinc non nulle de A. Pour ktablir
la finitude tic X n T, on choisit, uri bon no~nbrcpremier p et on ddmontrc
d'une part que 13 torsion p-primnire contc~iuedans X + n cst finic et born&
ind6peridan1ment de a E A, ce qui sc h i t par un raffir~ernent facile du
rbsultnt de I3ogornolov; d'autre part, on prouve la finitude de la torsion
prcmihc B p, par une rnhthode p-adique.
Pour ccla on se ranlknc au cas oh A est un R-sclidma abdlien, o& 12
est uu anneau de valuation discrkte cornplct, de corps dcs fractions K de
caract6ristique 0, dc corps rksiduel k algkbriquemerlt clos de caractkristique
p. On suppose tic plus que R eet non ramifik, done isonlorphe j. l'anneau
W ( k )drs vccteurs dc Witt B cocfficicnts darks k . Le schkrna X e ~ main-
t
tenant un sous-schkma de la fibrc gdnkrique AK de A.
Examino~lsd'alord le cas oil A rst un relkvement canoniquc (au scns
de Serrc-Tate), de sa fibre spkcialc A,. D a m ce cas, lc Frobenius absolu
F de A, se relhve en u11 endomorphisme 4 de A, an-linbaire, oh OR est
le Frobenius de R = W(lc). Alors 4 laisse Fixe les points de torsion de
A ( R ) d'ordre premier A p, ct done aussi X si on supposc: que ces derniers
sont dcnses dans X, pour la topologie dc: Zariski. Une ktude analytique
328 M. RAYNAUD Points de torsion 329

locale (cf. 5.3.1) permet alors d'cn dkduire que X est le translatk d'une 1.2. Soit R un anneau de valuation discrkte complet, de corps des
sous-variktk abklienne de A. fractions K de caracthristique sbro el dc corps rksiducl k algkbriquement
Dans le cas gknkral, le Frobcnius F de A, ne se relkve pas en un en- clos de caractkristique p > 0. O n suppose quc pR est I'iddal maximal de R
dornorphisme de A, et on doit remplacer A par son extension vectorielle (i.e., 11 cst abolunicnt non ramifik) et pour n >0, on pose R, = Ii/p7'+'R.
universelle forrnelle &. Cette modification crke quelques complications plus Pour tout n > 0, soit An un 12,-schbma abdlien tel que
techniques que thkoriques. Tout d'abord f n'est plus un schkma en groupes
rnais un schkrnn formel en groupes et sa fibre gknkrique EK est un groupe
analytique rigide. Ensuite I'cndomorphisme 4 agissant sur l'algkbre de Lie
de CK fait intervcmir plusieurs pcntes; on doit skparer la contribution de et soit A = lim A, le R-schkma abklicn formel corrcspondant. Si A: est le
-+
chacune d'elles ct traiter spkcialement la pente 0. R,-schkma abklien dual tle A,, nous dirons que A' = lirn A: est le schkma
Comme application, on montre que la torsion situke sur une courbe non --+
abklien formel dual dc A. Par exemple, dans le cas algkbrisable, A et A'
elliptique de A, est finie, bornke uniformkment aprks translation, du rnoins provicn~ientpar conlplktion le long de leurs fibres fermkcs d'un Il-schkrna
si X n'est pas contenue dans une surface abklienne de A. abklien A et de son dual A'.
Pour tout n 2 0, l'analogue de (1) fournit une suite exacte de
R,-sclibmas cn groupcs lisses:
1. Extensions vectorielles universelles.

Ides rksultats 6noncks dans ce numdro sont extraits de [7].

qui fait de E(A,) une extension vectorielle universelle dc A,. Passant 8 la
limitc inductive sur n , on obtient une suite exacte de R-schkmas formcls en
groupes lisses:
1 .l. Soient S un schkma, A un S-schkma abklien, A' le S-schkma abklien
dual de A. I1 existe une extension universellc de S-schkmas en groupes:

que nous appcllerons l'extension vcctoricllc formelle univcrselle de A. Donc,

par construction, si I'on part d'un R-schkma abdlien A, de cornplktion
de A par un S-groupe vectoriel V(A). Vu son caractEre universel, l'extension formelle A, (4) se dkduit dc (1) par cornplitiori formellc le long des fibres
(1) est unique B isomorphisnic unique prhs. De plus, on montre que sa fernlkes et rcdonne (3) par rkduction mod, p n + l .
formation commute aux changernents de base S' -+ S et que, si on suppose
S affine pour sirnplificr, on a un isonlorphisrne canoniquc de suites cxactes
de I I O ( S ,Os)-modules localement libres: 1.3. I1 nous faut maintenant rappcllcr le caractkrc rristallin de (4).
Notons d'abord que R n'ktant pas rarnilik, I'idbal niaxirrial pR, tie R, a une
structure naturclle d'idkal B puisvanccs diviskes. Par passage la limitc sur
n, on en dkduit urn isornorphisme canoniquc de R-modules:

oh la premikrc ligne est la suite exacte d'algkbres de Lie dkdnite de (1) et

la seconde est la suite exacte courte qui dkcrit la cohornologie de De Rham oh le second rnernbre dksigne la cohorriologie cristallinc dc A, 8 coelficicnts
relative de A' sur S , en degrk 1. dans W ( l c ) -- R, artlicau des vectcurs dc Witt de longueur irifinie B coclfi-
330 M. RAYNAUD Points de torsion 33 1

cients d a m k. L'analogue de (2) fournit un isornorphisme: Consid6rons sur Ic k-schhrna abdlien A, le rnorphisme de Froheniuu ab-
s o h E': A,--+& et le ~norphismcde IG-obenius rclatif Flk: A. - / i ~ l ) , de
sorte que 1'011 a F = uk 0 Flk. Par lc cllangemc~ltde base OR, on dkduit
cle A (rcsp. &(A)) dcs R-schhmas forrncls en groupcs A(') et &(A)(')et par
d'oh l'on dkduit un isornorphisme canoniquc: transport de struct,urc par I'autornorphisme U R , on a &(A)(')= &(A(')).
D'nprks 1.3, au k-niorphisrne Flk: A, 4 /iL1), corrcspcnd un rnorphisme
de R-schkmas forrnels en groupcs &(l;jk): &(A)---+ & ( A ) ( ' ) ,que nous notcrons
Soit ~riairttcnaritB = linl Bn un second R - s c h h a nbdicn formel, B' +lR. l3n composarit 4/Ii avec l'isornorphisrne absolu gn: &(A)(')+ &(A),
---t
n on ot)ticnt un endomorphisme absolu, un-linkaire 4: &(A)-+ &(A). Au
son dual el &(a) son extension vcrtoriellc forrnellc univcrselle. niorphisme de Frobenius relatif F'lk: A, -+ AL'),
correspond par dualitk
Soit u,: A, -+ 13, un k-morpltismc de sch6ntiw abkliens. Alors, d'aprks le rnorphisrne du Vrrschiebung V/k: A~')--+A',. I1 en r6sultc que si on
([7] chap. 11. $1) pour tolit n 2 0, il existc un rnorphisrric dc Rn-sch6mas identifie Lie &(A) B IIE,,,(A',) par T, Lie 4 s'identifie au Verschiebung V,
en groupes: 0;'-lir16aire, de la cohomologie cristnlline.
En(uo): E(&) -+ E(Bn),
caractkris6 par Ics propriktbs suivantes:
1.5. Dans la suite, Ic schkma abdlier~ formcl A 6tant Furk, on pose
E,(uo)rst Ic k-rnorphisrnc canonique provcnant du caractkre univer- & = &(A), 2, = V ( A ) et on note K : l + A l'application dc paxiage au
i) quotient qui intervient dans (4). On a donc une suite exacte dc I{-sch6rnas
sel des extcnsioris E(A,) et E(B,).
forrnels cn groupes:
ii) 13n+l(uo) se rGduit en lL(u,) modulo pn+', et donc les En(uo)
dL:fiuissent url rr~orphisrne&(u,): &(A)-+ & ( B ) dc R-schdmas forrnels en
groupcs.
iii) Le diagramnlc suivnnt est cornmutatif:
qui par r6duction modulo k, fournit une suite exactc de k-groupcs algk-
lie f (u,) briques
Lie&(&) -----4 LieE(B)

qui fait dc Eo l'extcnsion vectorielle uuivcrselle dc A,. Aux suites exactcs

(9) e t (10) correspo~idcritdes suites cxnctes d'alg6bres de Lie:
oh 11:: B', 4 A: est 1e ~riorphisnledual de uo c t Iliris(u:) est I'application
d6duite de ub par passage i la cohomologie cristalline.

1.4. Notons uk: Spcc(k)- Spec(k), I'isomorpl~isrnedkfini par le Frobe- (lo1) 0 -+ LIpL -t M/pM + A/pA -t 0.
nius de k qui envoie a E k sur U P . On note UR I'autorrtorphisme de
Spec(R) = Spec(MT(k))qui relbve uk ct O K la restriction de nil B Spec(K).
Pour tout R-schdnia (ou sch6ma formel) X on note x(')1c 12-schkrna (OU 1 1.6. 11 rksulte de 1.3 que & ne ddpcnd pas du choix du R-schkma abdlien
sclihma formel) ddduit de X par le clti~ngerncrltde base O J ~ . 011a donc un forrncl A qui relkve /Io. Plus prdcisi.rncnt, si est ,111 autre R-sch6ma
isornorphisrnc i~bsolux(')-+X not6 encore OH. ahdicn formel qui relkve A,, on a unc suite exact,? dc J2-schkm:1 forlnelu en
 M. RAYNAUD Points de torsion 333

groupes, analogue de (9): 2.3. Iteprenons leu notations du numkro prkckdent. A la suite exacte (9)
de R - s c h h a s formels en groupes lisses correspond, par passage B la fibre
gknkrique, une suite exacte de K-groupes rigides arialytiques lisses:

qui s'identifie A (10) par rkduction modulo p. Le groupe formel vectoriel (12) 0 -+ VK -+ &K -+ AK -+ 0.
ti slidentifie par l'application logarithme B un sous-R-module de La suite exacte d'algkbres de IAe, associke A (12) se ddduit de (9') par
M = Lie&. tensorisation avec K. En particulicr, nous noterons H I'algkbre de Lie de
De plas, la correspondarlce: &K identifike A HETi,(Ab) @ R K par T A K, de sorte que l'on a la suite
exacte:

ktablit une bijection entre: L'endornorphisnie absolu, on-linkaire 4 de & induit un endomorphisme
- d'une part les R-rnodulcs E , facteur direct de M, tels que = L/pL arc-lindaire de l'espace rigide &K e t not6 encore 4; Lie (4) induit le Verschie-
(cornme sous-k-vectorirls de M/pM) bung V sur 11.
- d'autrc part, les R-schdrnas abkliens formels 3 qui rclbvent A,.

Remarque 2.9.1. Supposons que le R-schknia abklien formel A soit

la complCtion formelle d'un R-schkma abklieri A e t soit AK la K-variktk
2. Gkomktrie formelle et gkomCtrie rigide. abklienne fibre gCnCrique de A. Par restriction de ( I ) 8. la fibre gi:ndrique,
on obtient un K-groupe algkbrique EK extension vectorielle universelle de
AK. Si I'on applique A EK le foncteur "Gaga" analy tique rigide, on obtient
2.1. Soit X un R-schdma forrnel de type fini. Rappelons [8]qu'il lui
un groupe analytique rigide ( E K )(qui ~ ~n'est
~ pas rkunion d'un nombre
correspond c:ar~oriiquemcnt u n K-espace analytique rigide X K , la "fibre
fini d'ouverts affinoYdes si A # 0) e t EK est un sous-groupe ouvert rigide
gdnkrique" de X. Ddcrivons cette correspondance dans le cas 06 X est
de ( E K ) ~ ; ~ prkcisdrnent, on dkduit ( E K ) de
(plus ~ ~lK~ A partir de (12), en
afine. Soit A l'algkbre de ddfinition de X. Donc A est une R-algkbre,
ktendarit VK en un K .vectoriel). Mais en gkndral, I'endomorphisme 4 de EK
cornplkte pour la topologie p-adique, topologiquement de type Gni. Alors la
ne s'ktend pas A (EK),,, et, a fortiori, ne provient pas d'un endomorphisme
fibre gknkrique XK de X cst l'espace rigide analytique, afinoi'dc, associk A
algkbrique, UK-lindaire de EK. C'est pour cette raison que nous allons
l'algkbre de Tate AK = A BR K .
devoir travailler avec les structures formelles e t rigides analytiques.

2.2. SoitY un sous-schkma formel fermi: de X, alors YK est un sous-

espace analytique rigidc fermk de XK . Avec les notations prbckdcntes si Y 3. Eldments indkfiniment p-divisibles.
mR
est dkfini par I'idkal I de A, Y K est ddfini par I'itlkal IK= I K de AK.
Rkciproqucment, si YK est un sous-cspace anlytique rigide fermd de XKJ 3.1. ConsidCrons un groupe comrnutatif G, un nombre premier p e t le
Y
il cxistc un plus petit sous-schkma forrnel fern16 de X , de fibre gknkrique systkme projectif:
Y
Y K . De plus, y est R-plat: si Y K est dkfini par I'idkal J de A K J est -+ G -I: G 3 G ,
dkfini par le plus grand idkal I de A tel que I @ A K = J. (13)
La correspondance Y H Y K ktablit une bijection entre sous-schkmas dans lequel I'application G 4 G est la multiplication par p. Soit lim G
4-
formels fermds R-plats de X e t sous-espaces analytiques rigides fermks de
XK. Par analogie avec l'opkration d'adhkrence schkniatique dans le cas la limite projective de ce systkme. Un klkment de lim G s'identifie A une
Y
algdbrique, n o w dirons que est l'adhkre~iceforrnelle dc YK dans 1.
t
334 M. RAYNAUD Points dc torsion 335

suitc (z,), n >0, d'klCrrlcrlts dc G tels clue px,+l = x,. Soit pnCle noyau Notons que la ~riultiplicationpar pn dans X est Ctalc sur la fibre gknkrique
de la multiplication par p" d a m G et soit ]im t
,,,G la. lirnite projective XK, en particulicr pnX(R) est lini et donc on a la suite exacte:
"p,n"
du s y s t h e projectif:

Nous dirons que les Clkr~ientsde n pnX(R) sont les e'le'ments inde'/iniment
n
Par passage & la limite sur lcs suites exactes: p-divisible3 de X(R).
Notons Tpm X(11) (rcsp. Tp- X(k)) la torsion p-primaire pdivisible de
X(R) (resp. X(k)).

on obtient la suite exacte: Lemme 3.2.2. L'application n pnX(R) -+ X(k) a un noyau fini
n
(resp. est injective) si et seulement si l'application qm
X(R) -+ Tpm X(k)
est surjective (resp. bijective).

En cffet, lc diagramme conimutatif B lignes exactes:

oh a,((z,)) = x,. Si de plus le systkme projectif (14) satisfait ii la condition
de Mittag-Lemer, a, est surjectif.

1 12 1
O-t pnX(k)-+l@X(k)-r X(k)
3.2. Soit R comme dans 1.2 et soit ,T un R-schCma formel en groupes "p,n" '(P"
conimutatif e t lisse, B fibre ferrnke X, = XIi @ k connexe
niontre que 1e noyau de l'application n pnX(R) -+ X(k) est isomorphe
n
Lemme 3.2.1. Les applications: au conoyau de la flkclie lirn ,nX(R) --+ lim ,nX(k). Cette dernikre
e t
"p,n" "p,n"
lim
t X(R) -) lim
c I(&) --+ Iim
t X(k) application se d6duit tle I'applicatiori TpmX(R) -+ Tpm X(k) par application
I'P" " P" ''~" du fo~icteurI I o I I ~ ( Q ~ / .),
~ !d'oh
~ , le le~nme.

dkduites dea applications canoniques X(R) -+ X(Rn) -+ X(k), sont bijec- Ezemples 3.2.3. Supposorls que X soit un schCma abklien forrricl. Alors
tive~. X,(k) cst pdivisible, donc n pnX(k) = X(k).
n
Notons Xpoo lc R-groupe p-divisiblc construit sur les noyaux dcs niulti-
En effet, on a X(R) = lim
t-
X(Rn) et doric, par associativitk des li~nites plications par pn dam X. On a une suite cxactc de R-groupcs p-divisibles:
n
projcctives, l i i X (R) = l@ am
c I(&). Par suite, il s u l t de montrer
"P" "P"
que pour tout entier n 2 0, l'application lim
e X(Rn+,) 4 lim
c X(Rn) oh (Xp-)inf cst connexe et (Xpm)et est ktale isomorphe B ( Q ~ / Z , ) ~oh
, h
"P" "P"
cst le p-rang de X,. La suite exacte (16) est scindCe en rCducLion modulo p
est bijective. Or comme X est lisse, l1a.pplication X(IZn+l) -+ X ( R ~ est
) et il rCsulte de 3.2.2. que llnpplic,ztion n pnX(R) --r X(k) cst bijective si
surjective; par ailleurs son noyau est anrrulC par p, ktant n
 Points de t,orsion 337
336 M.RAYNAUD

I1 sufiit de montrer quc ) induit I'identitk sur lim &(R)et d'aprks 3.2.1. il
e t seulcrrier~tsi (16) est scindde. Ce sera lc cas si h = 0, ou bien si X, est t

ordinaire et si X est le rcl&vement canonique de Serre-Tate. "P"

sufit devoir que 4, induit I'idcntitk sur lim&,(k), oc 4, est I'endomorphisme
4-

3.3. Reprenons les notations de 1.5. absolu de & mn P

k = E, induit par 4. Par construction, 4, provient, par
fonctorialitk de l'extcnsion vectoriclle univcrselle, du Frobenius absolu J' de
Lemme 3.3.1. - &(R)4 liIn
L'application Jim
7
e A(R), dkduite de x (9), A,, de sorte quc l'on a un diagramme cornmutatif:
"P" "P"
ert bijective; e n particulier, tout e'le'ment inde'finiment p-divisible de A(R)
se relkvc en u n e'le'ment inde'finiment p-divisible de &(R).

D'aprks 3.2.1, il sufit de voir que l'application lim &(k)4 lim ff (k) est
bijective, ce qui rksulte de la suite exacte:
'7, '7,
Or F induit l'identitk sur A,(k) et on a dkj& remarquk que I'application
lim E,(k) -t linl A,(k) ktait bijective.
e C
('P" <'P"
dkduik dc (10) et du fait que V ( k ) est annul6 par p.

Remarques 3.3.2. i) Pour la dkrnonstration du t l r k o r h e annonck clans 3.5. Soicnt R comme dans 1.2 et A un R-schkma abklien de fibre
l'introduction, lcs seuls points indkfiniment p-divisiblcs de A(R) que nous gknkrique Ak.
auroris & considdrcr sont c e w de torsion.
ii) On peuL rnontrrr que la torsion p-primuire ir:dffinirnent p-divisible T h k o r e m e 3.5.1. Soit XK un sous-sche'ma fermk intkgre de AK con-
de & ( [ I ? ) w t nulle, de sorte q ~ l'application
c t- & ( H ) -r
a,: linl &(R) est tenant u n ensemble Zariski dense de points inde'jiniment p-divisible3 de
injective. "P" A ( K ) = 411). O n suppose que X n'est stable par aucune translation n o n
nulle de Arc. Alors il existe une partie E l de x,
Zariski dense duns X K ,
3.4. Considkro~lsl'tmdornorphisme absolu 4 dc & et rappelons que dont l'image dans A(k) est u n seul point.
4 = on o (din) (1.4). Notons que ) ieduit 11n endo~norphisn~e du groupe
& ( R ) . Rn r b t , soit .c t & ( R ) ,vu conime une section de & nu-dessus de Corollaire 3.5.2. Soit X K u n sous-sche'ma ferme' intkgre de AK con-
S p e c ( 1 l ) Cornposant x avcc c$/* on obtient une section de &I1) au-dcssus tenant u n ensemble Zariski dense xde points indljiniment p-divisibles de
de Spec(ll), puis on transporte cettc section en une section dr & nu-dessus A(k). O n suppose de plus que l'une des conditions suivantes est rkaliske:
de Spec(ll) par le el~angenrentdc basc (oa)-l. On a donc 4 0 s = #(x)oon.
i) eat forme' de points de torsion.
ii) L'application n pnA(R)--, A ( K ) a u n noyau jini (cf. 3.2.2).
n

Alors X K est le translate; par u n point indkjiniment p-divisible, d'une

sous-varikte' abe'lienne de A K .

Les klkments indkfiniment p-divbihles de & ( R ) La dkrnonst.ration du thkorkme fera l'objet des trois numkros suivants.
P r o p o s i t i o n 3.4.1.
Montror~scomrncnt le t h k o r h c entraine le corollaire.
sont jixe's par 4.
 Points de torsion 339

de y ne commute pas A la restriction A un sous-schkma forn~clouvert de

Soit AK Ic sous-schbma en grouprs de AK form6 drs translations de AK E.
qui laisscnt stable XK. Soit A lc sow-scll6rna en groupcs de A , adh6rcrlce
scl~kmatiqeed r BK d;ws A et soit C Ic R-scl~d~na abilien quotient AID. ii) L'espace rigide YK est rkduit (c'est-B-ciirc ses anneaux locaux sont
rhduits) cornme il rksulte de I'rxistence d'urr cspace rigide r&iuit, sous-jacent
Notons El1(resp. X G ) l'image ddc C (resp. X K ) dans C K . Alors XI'
h un espace rigidc donnk (conskquence dc I'cxcellence dcs anneaux de la
t s t Xariski dcnsr clans X'f( et X'f( n'cst stable par aucurle translation nor^
gkornhtrie rigide (11 (3.3)).
nulle de C K . I1 rrsttc ivoir que X> cst r6duit h un point ral,ionrlcl. D'aprks
le thborkn~e3.5.1, 11 rxiste une partie Cy de XI1, Znriski dcnse dans X';(
dont l'irnage dans C(k) cst rkduite i un point. Lcs lrypothkscs i) c t ii) Nous clirons clue y (resp. YK) a t l'ndhhence furmelle (resp. rigide) de
se consrrvcnt qua11c1 011 rernplace C et A par C1I ct C (dans le cas ii) A dans & (rcsp. &K).

cela rdsulte par excrnplc de 3.2.2) ct assurent quc Its fibres de I'application
X" C ( b ) sont finies, donc X'f( cst h i et par suitc rkduit B un point
ration nel. 4.2. Soit Y l'adhkrence rigide dans d'une famille A de poirlts indk-
finirncnt p-divisibles de & ( R ) .Les points de Asont fix& par q5 (3.4.1)) donc Y
est stablc par 4. On supposc dc plus clue Y t,st un espacc rigide irrkductible.
4. Etude rigide analytique. Soit d sa dirncnsion. Commc Y est rkduit et K dc caract6ristique 0,
Y est alors lisse sur K en dchors d'un fermk rigide Z f Y ;I'ouvert
cornpldrncntsirc U est le plus grand ouverl de Y oii le faisceau des formes
D;ms les nurn6ros 4 ct 5 on reprend Ics hypothkses et notations de 1.5 difKrentielles de degrk 1, fly, est localement libre de rang d.
et 1.6. Rappelons (2.3) que I'on a not6 11I'algkbre de l i e de EK et que q5 induit
4.1. Soit A unc partie de & ( R =
) tK[K). sur H lc Verschiebung V issu de la cohornologie cristallirle de A:. Soit

Lemme 4.1.1. I1 existe un plus petit sous-sche'ma formel f e r m t Y de

& te1 que A C & ( ] I ) . 1)e plus Y est 12-plot et so Jbre gCnCrique Y K (2.1) la dCconlposition de II suivant les pentes de V. Par translation, on ktend
est le plus petit sous-espace unalytique rigide f e r m t de LK contenant A. la dkcornposition (17) en une d6co1nposition du fibrk tangent 1' A tK:

1311 effet, soit U, un reconvremcnt ouvert a E n c du It-schdma formel &

e t soit A, I'anrleau dc U,. ConsidCrons la famillc I], j E J , dcs faisccaux Soit y E A. L'espace tangent dc Zariski B Y en y, Ty(y) est un sous-K-
cohkrents d'iddaux dc O x qui s'annulcnt sur A. Cette farnille est non vide, vectoriel de la fibre T(y) de 1' en y. Comrne y est un point fixc souu 4,
car clle contierit 0; clle est filtrante croissante; enfin cornme les anneaux l'application tangente B q5 en y, T+(y) induit une application rrK-linkaire
A, sont noethCrirns, cllc conticnt I I I ~ddmcnt maxirr~alI. Alors I d6finit T(y) -r T(y) qui coincide avcc V: I[--+ 11 q u a d on identifie II A T(y)
I t plus petit sous-srhbina lorrncl fern16 Y dc & Lcl que A C Y ( R ) . Si par translation. Conirnc Y cut stable par q5, TU(y)est un sous-K-espace
ddsignc le s o u s - s e l l h a formel ferrnd dc Y, ddduit dc Y en divisarlt Oy par vectoriel de T(y) stable par T+(y) et par suite:
son faivccau d'itlksux de R-torsion, alors g(12) contierit aussi A, donc y = y
e t par suite y est 12-plat. La dcrnikre assertion rksulte dc 2.2.

Remarques 4.1.2. Proposition 4.2.1. Sur l'ouvert de lissite' U de Y, le fibre' tangent Tu

a' U est localement libre et I'on a la de'composition:
i) La dCfinition clc Y et de Y K est globale; en particulier la lormation
 Points de torsion
340 M. RAYNAUD

i) P est stable par v

06 les faisceaux Tu n Tx I U s o n t localement librev de rang constant.
l'application exponcntielk est ddfinie sur p et rkalise un isomor-
ii)
phisme d11 groupe additif rigide sus-jaccnL 2 P sur us sous-groupe ouvert
Nous allons travailler avec le faisceau des formes diK6rcntielles sur Y, fly, rigide U de tK.
plutBt qu'avec le libri tangent To qui n'existe que sur U. Soit R le faisceau Par excmple, on peut prendre P = pTIrie(f), avec = 1 si p # 2 e t
dcs formes diffi.rcnticlles sur t K .A la dkco~rlposition(18) correspond une r = 2 si p = 2, Le sous-groupe ouvcrt rigide U de lK ,st nkcessairement
dCco~nposiLiond u d e stable par 4 et IJexponentiellctransporte le Verschiehung V sur M en 4 I U.
n = @xqnx. Dksormais or1 identifie ( P , V) avec (U, 4).
O n a une surjection canonique u: R I Y 4 R y dont la fibre en un point Pour X t Q, posons Px = Hx n P. Quitte alors 2 rempIacer P par la
y E Y ( K ) s t la surjcction R(y) + Ilr(y), duale de I'injection sur les cspacn somme dirccte dc PA, on peut supposer que P vkrifie de plus:
tangents T Y ( ~ ) L T(TJ).Pour touL 1 E Q,consid6rons le faisceau colrkrent iii) P = $xEsPx, avec PA = I & n P.
Nx sur Y , conoyau dc l'applicatioll cornposke:
Alors, en restriction U , la dkco~npositionT = $ f i du fibrk tangent
provient de la dkcompositiorl de P en produit des groupes additifs PA.
Dans la suite de ce numkro, rlous allons ktudier les points ration~lelsde
Soit N = $xEeNA ct soit v: fly -+ N . l'application sornme des surjections
certain3 sous-espaces rigides dc P. Rappelons en particulier que si Z est un
canoniques nu -, N*.Alors v est encore surjective e l sa fibre en y E Y ( K )
espace analytique rigide lissr, partout dc dirr~ensionn, % ( K )a une structure
est bijective si et seulement si la relation (19) est vkrilikc.
naturclle de variktk K-analytique, partout de dimerrsiorl n au sens de ([3],
Soit Y' le sous-espace rigide analytique fermi. de Y forn~kdes points
0; le rang de N a L > d (cf. [GI, Ienlme 3.6). 11 rfsulte des considfrations 5).
prPcCdcntes e t de (19) que Y' contient A, ct par suiLe Y' = Y. En particulier
N I U est partout dc rang? d. Conlmr U est rkduit e t flu localernrrlt libre
de rang dl la llkchc surjective v I U : nu -+ N 1 U est nfcessairen~entun 5.2. Soit Y comrne d a m 4.2. Alors P n Y est un sous-espace rigide ouvert
isomorphisme. Par suitc Nx I U est localement librc, de rang conslant sur de Y e t un sous-espace rigide fcrmk de P. 1,es PAn Y sonL des sous-espaces
U ,p~iisqueY esl irrfductible. l'ar clualitk cntre Tu et f l u , on en dCduit rigides fermks de P n Y; P n Y e t les PA n Y sont stables par V .
4.2.1.
Proposition 5.2.1. Supposons que l'origine 0 soit u n point lisae de
Y . Alors pour A E Q) PA n Y est lisse e n 0) d7espace tangent PA n Ty(0).
5. Etude analytique "molle."
Soit Pi = P,, et notons nx: -+ P P i lit projectiorl de noyau PA.
I Alors f i n Y est la fibre au-dessus dc 0 dc xx I P n Y . D'aprhs 4.2.1le rang
5.1. Reprenons Ies dkcornpositions (17) e t (18): de l'application linkaire tarigente ,i xx I P n Y est constarit sur l'ouvert
de lissitd de P n Y. Cornme K est de caractkristique 0, nx I P n Y est
une subinlmersion au voisinage de 0 ([3], 5.10.6) et er1 particulicr s a fibre
La dfcornpositio~ldc T ne provirnt pas d'une dfcomposition en produit au-dessus dc 0, f i n Y est lisse en 0. L'assertion sur I'espace tangent est
I du groupe rigide E K 1 rnais elle en provient localenlent au sens andytique par ailleurs immkdiate et indbpendallte des questions dc lissitk.
LLmou'' .
D r facon prfcisr appelons rkseau P de H tout sow-12-module libre P de
II, trl quc P Qn K = I I . Soit P un rfseau de 11 contenu dans Lie(&) tel

I1 que:
342 M. RAYNAUD Points de torsion

5.3. soit encore:

P r o p o s i t i o n 5.3.1 Soit P u n R-module libre de rang fini, m u n i d'un
endomorphisme v o g l - l i n k a i r e , avant une seule pente X > 0. Soit Z u n
sous-espace rigide analytique de P, ontenant 0 , lisse en 0 de dimension T I
et stable par V. Alors il existe u n entier 3 2 0 tel que p s P n Z ( K ) soit u n D'oh plalaaf, = paai,,Va. Cornme a > 0 e t a;,, = 0 pour la/ 5 1, on a
sous-It-module de I-', de rang n. ai,, = 0, donc hi .= 0 . C'cst dire, qu'avcc les rkductions d6jB faites, Z ( K )
cdincide avcc &, d'oh la proposition.
Soit C l'espacc tangent B Z cn 0. Alors B est un sous-K-vectoriel de
P @ IZK, stat)lc par V, de dimension n. La catdgorie des K-vectoriels de Remarques 5.3.2.
dimension h i e , rnurris d'un Verschiebung V Ctant semi-simple ([dl, chap.
IV), on pcut trouver un suppldmer~taireE' dc E dans P @n K , stable i) Lcs calculs prdcddents restent valables lorsque X = 0, rnais ils per-
par V. La pente X de V est dgale B a / b , oh a et b sont dcs rnticrs > 0 , mettent simplenient de conclure que les sdries C , ai,,ga sont & coefficients
premiers entre eux. Cornrnc le corps rksiduel k dc R est algdbriquement dans Q, (autrernent (lit que, Z se desccnd sur Qp, au voisinage de zkro), ce
clos, il rksulte de [ I ] chap.
, IV, que l'on pcut trouvcr une base e l , . . . , e , qui n'implique pas une lindarisation de Z .
dr E e t une base f l , . . . , f,,, de E', telles que vb(e,) = pae;, i = I, . . . ,n ii) La proposition 5.3.1 ne s'ktend pas au cas oil V agissant sur P
et v b ( f J ) = pa j3,j = 1 , . . . , m. Quitte B multiplier les ei et les fi par adrnet plusicurs pentcs > 0. Par exernplc, supposons que P ait une base
une puissance de p, on pcut supposcr quc e; et f3 sont dans P . Soient & le (e.f.g) tclle que:
It-module de base e l , . . . , e, e t S le R-modulc de base f l , . . . , f,. Quitte V(e) = pe (pente I),
enfin B rernplacer Y par un sous-rkseau, on peut supposcr que P = & @ S.
Notons n : P 4 & la projection sur le facteur &. Vu le choix de & , n Z I (pente 112)
est &talcen 0. Quittc alors B rernplacer P par p s P , & par ps&, pour s 2 0 V ( s )= p f
co~~venablc, on pcut supposer quc x I Z est un isomorphisme Z 7 Q . Soient
X I , .. . ,x, leu coordonndcs de & relatives 11 la base e l , . . . , en ct yl,. . . ,,y e t soient x, y, z les coordonnkcs associkes B la base el f , g. Alors la sous-
les coordonn6es de S relatives B la base f l , . . . , f,,. Alors i! adrnet une variktd de P d'kquation x -- yz cst lisse, passe par l'originc, est stable par
reprksentation pararnktrique de la forrne: V e t ndaiimoins n'est pas linkarisable prcs de 0 .

yi = hi(xl,. . . ,x,), i = 1,. . . ,m

5.4. llcvenons au sons-espace rigide Y de t K . 1,orsque 0 est un point
lisse dc Y, lcs cspaccs rigides PA nY sont lisscs en 0 (5.2.1) e t pour X > 0 , on
oh h i = C, ai,,ga est une skrie A coefficients dans R, tendant vers 0 quand
peut lcur appliquer 5.3.1. Si maintenant y E A, est un point lisse quelconque
la1 -+ oo et sans terrne de dcgrk 5 1.
de Y , on se rarn&rie au cas prkcCdcnt par translation par -y. En rdsurnk
Le fait que % soil stable par V b se traduit par leu identitks:
des N o 4 ct 5, on ddduit le corollaire suivant:

C o r o l l a i r e 5.4.1. Soit Y l'adhe'rence rigide dans &K d'un ensemble

A de points &(R)indkfiniment p-diviaibles. O n suppose Y irre'ductible et
o n note 1J l'ouvert de lissitk de Y . Soit X une pente > 0 de V (17) et
supposona que dans la de'composition (20) du jibre' tangent ?;, n Tx I U
soit localement libre de rang n x . Alors, pour tout y E An IJ(K), il existe
u n sous-ll-module l l y , x de Ill n L i e ( & ) , de rung n x , stable par V, contenu
344 M. RAYNAUD Points de torsion 345

duns le domaine de dkfinition de l'application exponentielle relative d tK dans A K .

+
et tel que y C X ~ (soit ~ Icontenu ) Y avec pour evpace tangent e n
~ , ~ duns Puisque nous allons travailler avec les espaces rigides, nous changeons de
n
Y (TU I;, I UY). notations e t dksignons par XaIgl'adhbrence de Zariski tic
dam A K .
dans AK et
par X l'cspace rigide associd, adhkrcnce rigide de
On suppose dksormais que est formk dc points indkfiniment
pdivisibles, que Xulgest intkgre et n'cst stable par aucune tranlation non
6. Fin de la ddmonstration du thdorkme 3.5.1. nulle de AK.
ConsidCrons la suite exacte de R-schkmas formels en groupes introduite
6.1. Commencons par un lernme "algkbrique": dans (1.5):
L e m m e 6.1.1. Soit X u n sous-sche'ma fermk intkgre d'une variktk
abklienne A , X et A ktant de'/inis sur u n corps alge'briquement clos c. On
suppose que X contient une famille Zariski dense Xx, X E A, de sous- et la suite cxacte des K-groupes rigides associke:
+
sche'rnas de la forme Bx ax, oir Bx est une sous-varie'tk abklienne n o n
nulle de A et ax E A(c). A1078 X est stable par les translations d'une
sous-varie'te' abklienne n o n nulle B de A. D'aprks 3.3.1, on peut trouver un ensemble h dc points indkfiniment
pdivisiblcs de & ( R ) tel que -/r(A) = C . Soit Y 1';tdhi.rence rigide de A
dans lK. Comme lK cst rduniori d'un nombre h i d'ouverts afino'ides, Y
(La dkmonstration qui suit a ktk modifike sur dpreuve). Pour chaque
n'a qu'un nombre h i de composantes irrbductibles Y,, i E I . Alors Y,est
+
X E A, choisissons un point ax E Xx(c). On a Xx = ax Bx. Quitte B
I'adhkrence rigide de A i = AnY,(R). Soit Ci = r(A,) C X ( K ) = Xalg(K).
agrandir Xi, ce qui ne change pas la proprikti. de densitk, on peut supposcr
+
que BA est une sous-variktk abklienne maxin~alede A tclle que ax Bx soit
Conirne Xals est irrhductible, il existe i, E I tel que CIosoit Zariski dense
dans XUIg(K).Pour dkrnor~trer3.5.1, on peut, quitte B remplacer par
contenu d a m X . Alors, d'aprks Dogomolov (Math. USSR Izvestija vol. 17
(1981) N o 1, Theorem 1 p. 58) les variktks abLliennes Bk possibles sont en
CIo et Y par Y,o,supposcr Y irrkductible.
Soit A' la partie de A contenue dans l'ouvert de lissitk U de Y. Alors
nombre fini. Conirne X est irrkductible, on se r a m h e , quitte A restrcindre
A, au cas oh Bx est une variktk abklienne B indkpendante de A. Mais alors Y cst aussi l'adhkrence rigide de A' tlans tK,donc X est aussi l'adhi.rericc
rigide de C' = ?r(A1)d a m AK et par suite XuIgest I'adhkrencc de Zariski
B laisse stable les Xi par translation, doc laisse stable X.
de C'. Ilernplacant C par C', on peut supposer A C U(K).

6.2. Nous en venons A la dkmonstration dc 3.5.1.

6.3.
Soient donc A un R-schkma abklien et A le schkma abdlien formel com-
L e m m e 6.3.1. Soil
plbtion de A le long de sa fibre ferrnbe A,.
Soit C un ensemble de points de A(R) = A(I1). Comnie A est propre
sur R, I'adhkrence formelle (4.1) X de C d a m A est algkbrisable, c'est-8-
dire est la complktion for~nelled'un sous-schkrna fermb X de A ( [ 5 ]5.1.8). la suite exacte d'algibres de Lie associke 6 (1.2) et soit 11' u n s o u s - K -
Nkcessairement X est I'adhkrence de Zariski de dans A. Alors la fibre vectoriel de I, g RK , stable par le Verschiebung V de 11. Alors 11' est
$ contenu dans la composante Io de pcnte 0 de II.
gknkrique XK de X est I'adhdrence de Zariski de C dans la fibre gknkrique
AK de A, tandis que I'espace rigide XK, fibre gkndrique de X est I'adhkrence
rigidc de dans AK. Ainsi le foncteur "Gaga rigide analytique" transforme Reprenons la suite exacte
I'adhdrence de Zariski XK dc C dans AK en l'adhdrcnce rigide X K de
346 M. RAYNAUD Points de torsion

associke 11 (9). I'ar tensorisation avec K on obtient (12') et par rdduction 6.4. Reprenons la suibe exacte d'alghbres de Lie:
modulo p, on obtient

associke h (9) et soit M = M, @ M a la dCcomposition de M en sa com-

qui d'aprks (1.1) s'identifie h la suite exacte: posante M, oh V est bijectif et sa composante M a oh V est topologique-
ment nilpotent. Alors M, BR K est la composante H, de pente 0 de II,
tandis qu'en rkduction modulo p, M,/phd0 est annulk par F, donc contenu
dans LIpL. On peut donc trouver un R-module z,
facteur direct de M , tel
z
que > M, et E/pX = LIpL dans M I P M D'aprds 1.6, correspond B
Les R-modules M et Bi,fl(A',) sont munis des opdrateurs F et V tels que un certain relkverneut formel 3 de A,, quotient de &, de sorte que l'on a
F V = VI;' = p, et l'ideritification prdcddcrite est compatible avec ces une suite exacte de R-schdmns formels en groupes:
opkrateurs. Conlme k' ar~nuleR i b , F annule L/pL.
Notons M' 1e 12-rriodule Illn M C If = M B R K . Par hypothkse II' est
stable par V, donc par F , ct par suite M' rst stable par 1'. D'autre part,
M' cst contenu dans L puisquc 11' est contcnu dans L BRI< = Lie VK. (les rclkvements 3 que l'on obtient de cette fason sont c e w pour lesquels
Par construction M / M ' est sans R-torsion, et il en est dc rrii;.rne de L/M1. la suite exacte (16) de groupes pdivisibles est scind6e).
L'application M t / p M ' -t L / p L est alors injective, donc F est nu1 sur Considkrons alors le morphisme rigide f K -+ AK, fibre gknkrique de
M1/pM'. C'cst dire que I'opkrateur F sur M ' est divisible par p, donc
V opkrant sur M' est inversible et HI = MI @, K est contenu dans la
a. Par construction de z, on a z QDn K 2 11, et il rksulte alors de
6.3.2. que la restriction de a B l'ouvert de lissitk U de Y a une application
cornposante tle pcnte 0. linkaire tangente nulle. Comrne K est de caractkristique zkro, que Y est
irrdductible, rkduit c t contient des points rationnels, l'image de Y par
L e m m e 6.3.2. Dan3 la de'composition (20) du fibre' tangent d U: 7-i est un point z de A(K) = A(R). Soit z, la spkcialisation de f dans
A,(k) = A(k). Le diagramme commutatif:

ne figure que la pente X = 0.

En elkt, supposons qu'il existc X > 0, tel quc Tu n TA I I / , soit non nul,
donc localcn~rncntlibre dc rang nx > 0. Il'aprks 5.4.1, pour tout y E A,
il cxiste uri sous-12-module librc lly,X de Hi, de rang n x , stablc par V, tcl
montre xY(R) est une partie de A(&) qui se spkcialise en z, E A,(k), en
+
que y exp(FIy,x) soit contenu dans Y. Conlrne X > 0, il rdsulte de 6.3.1
particulier = ~ ( hse)spkcialise en z,, d'oh le thkor&xne3.5.1.
que Hy.x n'est pas contenu dans Lie VK, donc Ny,x = (I,ie.rr)(Iiy,x)est
un sous-R-module non nu1 de L i e n K . I1 en rdsulte que ~ X ~ ( N engendre
,,~)
+
une sous-varidtk abdlienne non nulle By dam A K . Commc x(y) e ~ p ( N ~ , ~ )
est contenu dans X ( K ) = Xalg(K), +
x(y) By est contcriu dans Xal,. 7. Complkment sur la ddmonstration de Bogomolov.
L'rnsemble des ~ ( y )y, E A cst Zariski dense dans Xalget doric (6.1.1),
XaIgest stat)le par les trarislat,ions d'une sous-variktk abdlienne non nulle 0, A une
7.1. Soient c un corps algkbriquernent clos de caractkristiq~~e
B de AK, en contradiction avec lcs hypothkscs faites sur Xaly variktk abblienrle d6Gnie sur c, X u n sous-schdma fermk inlhgrc de A qui
348 M. RAYNAUD Points de torsion 349

n'est stable par aucunc translation non nullc de A. Notons Z le fcrmk de sous-variktk abflicnnc de A (cf. (21 p. 703). Vu la dklinition de U, y: est
+
X adhkrcnce des sous-schkrr~asde X de la forrne a U , oh a E A(c) e t B de dimension 0, done Y ct Ua,b sont de dimension 0.
est unc sous-vari6td abklicnnc non nulle de A. D'aprbs 6.1.1, on a Z # X. Prouvons maintenant 7.1.1. Soil L un sous-corps de c, de type fini sur
Soit U = X - 2. Q, tels que A, X et Z soient dkfinis sur L. Notons la cl6turc algkbrique
Illknonck suivant, prkcise le rksultat dc Bogomolov e t nous allona le dc L dans c et C le groupe de Galois de sur L. D'aprks ([2], cor 2 p. 703))
dkmontrer par les m h e s techniques: il existe un nonibre entier 1 > 1, e t g E G, tel que si x E A@) cst d'ordre
une puissance dc p, alors g x = 12. On ktend g en un Lautoniorphisme
P r o p o s i t i o n 7.1.1. Soit p un nombre premier. Alors il existe un entier not6 encore g dc c.
N > 0 tel que, Va E A(c), la torsion p-primaire contenue dana U + a est Si alors a E A(c) et si u + a est un point d'ordre une puissance de p de
finie, de cardinal 5 N. + + +
U a , on a g(u + a ) = 1(u + a ) e t d'antre part g(u a ) = gu gu E fJ g a , +
donc 11 E Ua,pa et on applique 7.1.2.
Soit 1 un enticr > 1 e t notons 1~ la multiplication par 1 dans A. Pour
tout couple ( a ,b) .de points de A(c), nous allons const,ruire par rkcurrence
un certain sous-schfma fcrmf UF~ de U. I'renons U$ = U et supposons
-
8. Points de torsion d'une sous-varidtd.
avoir d6fini UL?', sous-schk~nafermP de U. Soit u$) lladhkrencc de u$)
dans X ct d6linissons UF;+')par la formule: 8.1. Dans ce numkro, nous dfmontrons le rksultat knonck dans I1introduc-
tion. Soicnt donc A une variktk abklicnne dkfinic sur un corps algkbrique-
ment clos c dc caractkristiquc 0, T le sous-groupc dc torsion dc A(c) e t X
un sous-schkma fcrrnk intkgrc de A tel que T n X soit Zariski dense dans
X. I1 nous faut rnontrer que X est le translatk d'urlc sous-variktk abklienne
de A.
L n );(I: forn~entune suite dfcroissante de sous-schkmas fermfs de U, donc
stationnaire, dc valeur UaVbpour m > 0. On a lA(Ua,,
+
(m+l) a ) C + a+
b
et donc lA(Ua,b a ) C va,b+ b et clairernent, Ua,b est le plus grand sous-
Quittc & passer au quotient par un sous-groupe de A, on peut supposer
que X n'est stable par aucune translation non nulle dc A, ce qui nous permet
d'introduire le fermk strict % dc X (cf. 7.1) et l'ouvert complkmentaire
s c h h a fermd dc U posskdant cette propriktk. U = X - %. Nous devons alors nlontrer que X cst un point de A.
Par dcs rkductions klkmcntaircs dktaillkes dans ([!I], 10.2 et 10.3), on se
L e m m e 7.1.2. Le schima Ua,b est fini et de'pend de faqon constructible rarnkne au cas oh A est un El-schkma abklien (Mini sur un anncau R d u
de (a, b ) . E n particulier, le cardinal de Ua,b est borne' par un entier N type d h i t dans 1.2, de caractfristiquc rksiduellc p > 2 e t oh X et U sont
indipendant de (a, b). des sous-schkmas de la fibre gknkrique AK de A. Soient R une cl6ture
8 algkbriquc dc K ct G = Gal ( R I K ) .
Le caractkre constructible de Ua,b est irnrnkdiat sur la dkfinition, une fois Prockdant cornme dans ([!I], li), on choisit unc dkcomposition
notk que si (a, b) varie d a m un fcrmk intkgre Z de A X A, alors pour rn
-
fixk, la formation de u:?) commute au passage aux fibres au-dessus d'un
ouvert non vide de 2. 11 rcstc i voir quc fJaTb est fini. Soit x E A(c) tel que du G-module T(R), oG T 1 ( R ) = T1(IC) est la torsion rationnelle, pdivisible
+ +
(1 - 1)x = a - b et soit Y = Ua,b a x. Alors lA(Y) c y. Si Y1 est (somme directe dc la torsion d'ordrc premier i p ct de la torsion p p r i ~ n a i r e
la rkunion dcs composantes irrkductibles de Y dc dimension maximum, on pdivisible, ratio~lnellle sur K ) et oh T" est un supplkn~entaire(non
) F. I1 cxiste alors une cornposante irrkductible Y: de Y'
a aussi l ~ ( Y l C canonique) de TI dans T. Ritppclons ([!I], 5.2.2) que I'action de G sur T"
e t un enticr n > 0, tcls quc l i ( Y i ) = F:. Donc Y:eut le translatk d'une
--I
cst forte dans le scns suivant: VM cntier > 0, 3n entier > 0, tcl que si
E
350 M. RAYNAUU Points de torsion 35 1

x E ll"(K)est d'ordrc p' avcc r > n, alors l'orbitc de x sous G a un l e r cas: il existe une courbr: clliptique E de A qui laisso S stable par
cardinal > M. txanslation, auquel cas S est une surface dliptiqi~cde fibre I;, de base
Par aillcurs, d'aprcs 7.1.1, la torsi011 p-primaire conknue d a m U a, + une courbe non elliptique C de U = AlE. Cornme C ne contient qu'un
a E A ( K ) est finie, borrlde par iln entier N indkpendant dc a. nombre fini de poinls de torsion, Ics poi~llsde torsion de S sont situks sur
un nornbrc fini de courbes de genre 1, E;, i E 1. Pour x variable dans X ( c ) ,
les interscctions Ei n ( X - x) ont un cardinal bornk, d'oh le rksultat.
8.2. Soit z un point de torsion de U(R). On a x = x' -t x", avec t

x' E T 1 ( K )et x" E T1'(1~).Alors x" cst un point de t,orsiorl p-primaire de 2 k m e cas: Les translations qui laissent S stable sont en nombre fini.
U - x'. Comme U e t x' sont ddfirlis sur I<, toute l'orbite de x" sous G cst Alors S ne conticnt qu'un nonlbre fini de courbes de genre 1, E,, i E I
contenue tians U - 2'. Puisquc cettc orbite a au plus N blCrnents, il existe (6.1.1) et si Z est la rkunion des Ei, I'ouvert U = S -. Z ne cor~tientqu'un
n, indkpcndant de x" tcl que l'ordre de XI' divise pn. nombre fini de points de torsion et on conclut cornme dans le ler cas.
Soit X' I'image de x dans A par la multiplication par pn. Alors, tout
point dc T n U(K)a urle image dans X' qui est contenue dans T' n X 1 ( K ) . R e m a r q u e 9.2. . La proposition 9.1 laisse ouvert le cas oil S est une sur-
Cornme T n U btait Zariski dense dans U , T 1 n X ' est Zariski dense dans XI. I face abklienne. Rappclo~istoutcfois que pour prcsque tout nontbre premier
Mais T' est constituk dc points de A ( K ) = A(IZ) indbfinirnent p-divisibles. p, la torsion prcrnikre A p, contcnue dans X + a , est borr~keindkpcndamrnent
I1 rbsulte alors dc 3.5.2 que XI est le translat6 ci'une sous-variktk abdienne t de a cornme il rdsulte de ((91, 6.3.1).
de A; par suite il en cst de rni.rne de X ct donc X est rbduit A un point.

9. Application aux courbes.

Soil A ilne variktk ahklicnne dbfinie sur un corps c algbbriquement clos

de caractkristique 0 et soit X uric courbc fermbe, intkgrc, non elliptique
de A.

Proposition 9.1. O n suppose que In surface S de A engendre'e par les

diffkrences (x-x') entre points de X n'est pas tine surface abklienne. A l o r s
il existe u n entier N tel que, V, E A(c), la courbe X +
a c o n t i e n n e au plus
N points de t o r s i o n .

Soit a E A ( c ) tel que X + a contienne un point de torsion t. Alors la

+
courbe X a - t contierit 0 ct contient le n&me nombre de points de torsion
que X + a . Pour ktablir 9.1, on pcut donc se limiter ;icor~sidkrcrles courbes
trans1ati.e~de X , de la forme X - x, avec x E X(c). Ccs courbes forrnent
une farnille algEbrique de courbes trackcs sur S. Cornrne S n'est pas une
variktk abklienne, seuleme~itdeux cas sont possibles (cf. 6.1.1):
 M. RAYNAUD

Bibliographie Euler and the Jacobians

R. Uerger, R. Kiehl, E. Kunz, J.J. Naztold. Differentialrechnung in of Elliptic C-urves
der analytischen Geometric. Springer Lecture Notes in Math. 38,
1969.
F. Bogomolov. Sur I'algkbricitk des reprksentations 1-adiques. C. Andrk Weil
R. Acad. Sc. t. 290, p. 701-704, 1980.
N. Rourbaki. Variktds diffkrentielles e t analytiques; fascicule de
rksultats. IIermanm, 1967. To Igor Rostislavovich Shafarevich
M. Demazure. Lectures on pdivisible groups. Springer Lecture
Notes in Math. 302, 1972. Euler, most universal among mnthematicians of his time and perhaps
J. Dieudonnk et A. Grothendieck. Elkments de Gkomktrie Algkbri- of all times, may be regarded a s the grandfather of modern algebraic
que IT; Etude cohomologique ties faisceaux cohkrents. Pub. IIiES geometry, since he created the theory of elliptic integrals, whence proceeded
No 11, 1961. I
Jacobi's theory of elliptic furictions which in turn gave birth to Riernann's
A. Grothendieck. Les schkmas de Hilbert. Skm. Bourbaki No 221, theory of algebraic functions arid abelian integrals. As there is a direct
1960/61. I line of succession from him to Chebyshev, he can also be counted as the
l3.Mazur et W. Messing. Universal extensions and one dimensional grandfather of It~lssianmathematics, though none of his irnrnediatc pupils
crystalline cohomology. Springer Lecture Notes in Math. 370, in Russia attained to any distinction in mathematics. Of his devotion to
1974. the welfare of science in Russia there can be no doubt, nor of the depth of
M. Raynaud. Geomdtrie analytique rigide. Table rondc d'analyse '
his attachment to that country :md to its language, of which lie acquired a
non archi~nkdienne.Bull. Soc. ~ a t hFrance
. Mkm. No 39-40, 1974. good command during the early ycars of his stay in Saint Petersburg. Thus
M. Raynaud. Courbes sur une variktk abklienne e t points de tor- it seems not inappropriate that he should appear iri a volume dedicated to
sion. A paraitre dans Invent. Math. Igor Shafarevich.
A t the same t,init, since the present contribution is meant to be no more
than a token of affectionate regard and admiration for a colleague whom
Reccived June 8, 1082 I feel privileged to count as a friertd, it may not be amiss for me to inject
Equipe associke au CNRS no 653 into it a personal note. Writing iri 1947 a book on abelian varieties, I had
been vexed a t my inability to show that the jacobian of a curve defined
Professor Michel Raynaud over some field iti itself' defined over the same field. In my attempts to fill
D6partrnent de Mathhmatiques up this gap I discovered that n paper by fIermitr ([2]) contained most of
UniversitC de Paris-Sud l11e formulas r l r d c d for trcaling one fairly typical special case, lliat of a
Bat. 425, Centre d'Orsay curve givrn by 211 equation u 2 = ll'(z),where F is of degree 4 (cf. [3]).
91405 Orsay, Prance IIerrnite had obt:tined his formula in t h e course of an investigation of
the invarianls and covariants of binary biquadralic forms. Let

L '~llusionsabout it, however, he had none. "Un pays oil, quand on parle, on est pendun
is how he once called if, (speaking Lo t l ~ cQueen Dowager of Prussia after his arrival in
Berlin in 1741), acrording to an anrrtlote preserved by CondorceL; cf. L. d u Pasquicr,
Leonard I h k r et de8 a m q l'aris, 1927, pp. 40-41.
f
f
 EULER AND THE JACOUIANS OF I~:I,LIPTICCURVES 355

Notations bcing as above, let k be the ground field; for us it may be any
ficld of characteristic other than 2 or 3; for Euler it was mostly the field of
be such a form; its invariants, in the sense of clavsical invariant theory, are real numbers, occasionally the field of cornplex numbers, and just possibly
given by (though Euler never says so) the field of rational numbers if he had the
possible application to diophantine equations at the back of his mind. P u t
K = k(P), where P is a point of I'. For M = (z, u) on C, p u t z = p(M) as
before, and u = $(M); also p u t y = p ( P - M), v = $ ( P - M). For a given
P u t F ( z ) = F(x, 1) and call C the curve of genus 1 defined by u 2 = F(x). P these arc rational functions on C, the former being defined over k and
One can show t h a t the jacobia~lJ ( C ) of C is the curve r given by t h e latter over K . As x, y are functions of ordrr 2 on C , there is between
them a relation @(x, y) = 0, where @ is a polynomial of degree 2 in each
one of the variables x, y, with coefficients in K. Changing M into P - M
exchanges (z, u) with (y, v); therefore @ is symmetric in z and y. Changing
and t h a t there is a rational mapping of C into r, given by M,Y into - M , -P does not change x and y; therefore changing P into
-1' can a t most multiply with a scalar; by choosing the sralar factor in
@ properly, we may thus assurrle that all its coclficients are polynomials in
r = v(P).
Clearly the two roots of the equation @(x,Y) = 0 are
where g(X,Y), h ( X ,Y ) are the two covariants of t h e form F ( X , Y), ay given y = p ( P - M) = p ( M - P), yt = p ( - P - M ) = P(P + M).
by the classical theory. To this I added in 1954 (loc. cit.) that r and
C may be regarded as the two components of a cornrnutativc algebraic As we have u" F(x), v2 = F(y), and as the mappings M --r M f P
group G(C), with r = J ( C ) as a subgroup of index 2, arid with the leiwe w = ir~variant,whilc the mappings M -+ fY - M change it into
point a t infinity on r as the neutral element. The group-law bcing written -w, one may thus, with Eulrr, interpret the above results by saying that
additively, it induces the usual group-law oq the "Weicrstrass cubic" r, and @(z,y) = 0 is an integral of e i t h r one of the diITercntia1 equations
also mappings ( M , P ) -t M + P of C X f into C and ( M , N ) -+ M + N
of C X C into r. The mapping (4) is then none else than M -+ 2 M .
One defines a function p on G(C) by putting p ( M ) = x for M = (x, u)
on C, and p ( P ) = 6 for P = ( & q ) on f ; we have then - M = (x,-u), or, more accuratcly perhaps, of the equation
(el
-P = -T/), and irl particular p ( - M ) = cp(M), p ( - P ) -- p ( P ) .
I'ut w = $; this is the diUcrentia1 of t h e first kind on C . For any
given point P on f , the mapping M -+ M + P is an autornorpliisrn of
C which leaves w invariant; conversely all such a~~tornorphisms are of that
form. Similarly, tlic autornorphisms of C which change w into -w are the
' it is the "gcnrral integral1', sinrr it contains the arbitrary constant
mappings M -+ P - M. Thus, in an obvious scnsc, the points F of r t = g ( P ) . This is how 1Suler viewed the matter, from the time ( a t the end
may be said to parametrize these two farnilies of autornorphisms of C. As I of 1751) when I~agnano'sProduzioni Matpmatiche drew his attention upon
found in 1954, explicit formulas for all these mappings can easily be derived t h r equation
from IIerlnitels results in [2].
Hati I but known it, I could have found all those forrnulas in a paper of
1765 by Euler ([l(b)]), where I was surprised t o discovcr them not long ago.
This is to be explained here. and others of the mrnc kind.
356 A. WEIL EULER AND T H E JACOBIANS OF ELLIPTIC CURVES 357

Having in fact discovered that the differential equation (5) has the general Thus, for @(x,y) = 0, we have
integral
+ +
x2 y2 c2x2y2= c2 + 2 x y 4 ' ,

where c is an arbitrary constant ([l(a)],p. 63; cf. his letter to Goldbach of

30 May 1752, where this is introduced by the words "Neulich bin ich auf with G = Q2 - P R, p = &I, a = f l ([l(b)], p. 321). At the same time,
curieuse Integrationen verfallen"), and having derived from this the addi- 1 we have
tion and multiplicatiorl formulas for the "lemniscatic" integral $ A, ldcp -. 1dcp
"
,/I-z4 -- -Q +Y =0 5% - Q(x) + R(x)Y = P&G~
2 ax
,
Euler proceeded to extend this to the most general elliptic case by seeking
the general integral of the differential equation
and therefore, differentiatir~gthe equation @(x,y) = 0,

in the form of what he called a 'Lcanonicalequation" @(x,y) = 0, where

is of degree 2 in x and sym~netricin x and y ([](a)], pp. 71-72). In effect, as This shows t h a t = 0 will be an integral of (G) if and only if C di1Ters
we have seen, this amounts to the determination of all the automorphisms from F only by a scalar factor, i.e., if one can write G = pF, where p is
of the curve u2 = F ( x ) which eitlicr leave invariant the diflerential $ or such a factor. This gives
- a set of five equations in n and the coefficient,s
change its sign. Having thus formulated the problem in [l(a)], he solved a,P, etc. of,@, beginning with ,kJ2 - a 7 = Ap ([I@)], p. 323; cf. [l(a)],
it cornplctcly, after a lapse of twelve years, with results which will now be p. 72).
describctd ([l(b)], pp. 321-326, 331, 341-344). Here appears Euler's exhaordinary virtuosity as an algebraist; perhaps
In Euler's notation, put he had acquired some of that skill by handling diophantine equations. I nced
not go here into the details of his dazzling manipulation of the equations
of that problem; suffice it to state the results ([l(b)],p. 326). Put

Write @ with indeterminate coefficients as follows: 0 = $[(b -7)2 - as']

(I am writing 8 where Euler writes M; otherwise all notations arc Euler's).

8 Then, up to an arbitrary scalar factor, thc general solutiori of the systcm
J of equations in question is giver1 by
Diffcrcntiilting the "canonical equation" @(x,y) = 0, one must find a
relation equivalent to (6). Now @(x,y) can be written as a =4(AO - P ) , P = 2140 - C) + 4AD, 7 = 4AE - (0 - c ) ~ ,
S =d2 - C2 + 4(AE + BD),c = 2D(0 C ) + 4BE, s = 4(EO - D ~ ) ,
-

p =40(8 - c )+~160(BD - AE) + 16(/1Il2 - BCD + B ~ E )

where P, Q, R are the polynomials ([l(b)], p. 326). In order to make the transition between Euler's notations
and IIermite's, put
358 A. WEIL EULER AND TIIE JACOBIANS OF ELLIPTIC CURVES 359

also, put O = 4([ + c). Then the last equation in (8) gives the point P = oo (the point a t infinity on r)and therefore the value oo
for 0, while the other sign will then give P' = 2A and thus a finite value
p = 64(4t3 - it - j), for 8. Therefore P, P' correspond respectively to the values of O given by

where, as above, i, j are the invariants of the form F(X, Y).

P u t again

['
For P = ([,q), P' = (6') ql), we have then [ = $ - f , = $ - $. As
to 7, q', they can now be obtained by using either one of the forrnulzm (9)
In order for (7) to determine a transformation (x, u) -+ (y, v) which changes as it applies to the present case; q will be obtained by substituting a, m , b,
% into ,, we must take a = -p. Replacing then q by pq, we can rewrite
dy
for x, u, y in the first one of these formulas, using in the coefficients of Q
(7) as follows: and R the value of O found above; for q' one has first to substitute 6' for O
in those coclficients, then a, -m, b for x, u, y. This completes the solution
found iri 1765 by Euler for a problem first formulated in 1954.

This is therefore the general expression for the auto~norphisrnsof C which

leave w invariant; as expected, they are parametrized by the points References
P = ( & q ) ofr.
r
Not only does Eulcr's paper thus contain the equation for and explicit [I] L. Eu!er: (a) De integratione aequationis differentialis =
+
formulas for the mappings (P,M ) -+ M P, but he also shows in effect 4-
that these mappings make up a simply trausitivc group on C , and he gives (E 251/1753), Opera Ornnia, Seriea I", vol. 20, pp. 58-79;
JG'a
at the same time for~nulasfor thc mapping (M, N ) --+ N - M of C X C into (b) Evolutio generalior formularum comparationi curvarum inser-
r. In ordcr to explain this, let us make explicit the occurrence of O in the vientium (E 311711765), ibid., pp. 318-356.
coeficients of @ by writing @(x, y, 0 ) instead of @(x,Y);this is a polynomial [2] C11. Hermite, Sur la the'orie des fonctions homog6nes ci deux in-
of degree 2 in each one of the three variables x, y, 0. Let A = (a,m), de'termine'es, Premier Me'moire, J. Reine Angew. Math. 52 (1856),
B = (b, n) be two points of C. If P =.B - A, i.e., if (9) maps A onto pp. ILL7 =CEuvres, t. I, pp. 350-371.
B, one must have @(a,b, 0) = 0. Solving this for O, Euler obtains, after a [3] A . Weil, Remarques sur un mimoire d'Hermite, Arch. Math. 5
calculation (not quite an easy one) which he omits: (1954), pp. 197-202 = CEuvres (Coll. Papers), vol. 2, pp. 111--116.

6 = ------ [H(.,
(b - u)'
b) *d m ]
April 23, 1982
where H is the polynomial
Professor Andrk Weil
lf (a, b) = A + R(a + b) + i C ( a 2 + b2) + Dab(a + b) + 13a2b2 School of Mathematics
Institute for Advanced Study
Princeton, New Jersey 08540
([l(b)],p. 343). Clearly these values of O must correspond to the points
P = l3 - A, P I = r l + R on r.
In order to detcrnrine the signs belonging to
tlresc two points, one need only observe that one sign must give for Ll -- A