Professional Documents
Culture Documents
J.S. Milne
October 23, 2004
+
Abstract
This is an introduction to the theory of Shimura varieties, or, in other words, to the
arithmetic theory of automorphic functions and holomorphic automorphic forms.
Contents
Notations and conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Prerequisites . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
Acknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1 Hermitian symmetric domains 8
Brief review of real manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Smooth manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
Brief review of hermitian forms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Complex manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
Hermitian symmetric spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
Curvature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
The three types of hermitian symmetric spaces . . . . . . . . . . . . . . . . . . . . . . . . . 12
Example: Bounded symmetric domains. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
Automorphisms of a hermitian symmetric domain . . . . . . . . . . . . . . . . . . . . . . . . . . 14
The homomorphism u
p
: U
1
Hol(D) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
Cartan involutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
Representations of U
1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
Classication of hermitian symmetric domains in terms of real groups . . . . . . . . . . . . . . . 19
Classication of hermitian symmetric domains in terms of dynkin diagrams . . . . . . . . . . . . 21
2 Hodge structures and their classifying spaces 23
Reductive groups and tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
Flag varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
The projective space P(V ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Grassmann varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Flag varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
Hodge structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
Denition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
The hodge ltration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
+
Copyright c _ 2004 J.S. Milne
1
2 CONTENTS
Hodge structures as representations of S . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
The Weil operator . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Hodge structures of weight 0. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Tensor products of hodge structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Morphisms of hodge structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Hodge tensors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Polarizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
Variations of hodge structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3 Locally symmetric varieties 32
Quotients of hermitian symmetric domains by discrete groups . . . . . . . . . . . . . . . . . . . . 32
Subgroups of nite covolume . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Arithmetic subgroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Brief review of algebraic varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
Algebraic varieties versus complex manifolds . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
The functor from nonsingular algebraic varieties to complex manifolds . . . . . . . . . . . . 36
Necessary conditions for a complex manifold to be algebraic . . . . . . . . . . . . . . . . . 37
Projective manifolds and varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
The theorem of Baily and Borel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
The theorem of Borel . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39
Finiteness of the group of automorphisms of D(I) . . . . . . . . . . . . . . . . . . . . . . . . . 40
4 Connected Shimura varieties 42
Congruence subgroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
Connected Shimura data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
Denition of a connected Shimura variety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
The strong approximation theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
An ad` elic description of D(I) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
Alternative denition of connected Shimura data . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5 Shimura varieties 51
Notations for reductive groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
The real points of algebraic groups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Shimura data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
Shimura varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
Morphisms of Shimura varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
The structure of a Shimura variety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
Zero-dimensional Shimura varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Additional axioms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
Arithmetic subgroups of tori . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
Passage to the limit. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
6 The Siegel modular variety 65
Dictionary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
Symplectic spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
The Shimura datum attached to a symplectic space . . . . . . . . . . . . . . . . . . . . . . . . . 66
The Siegel modular variety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68
Complex abelian varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
A modular description of the points of the Siegel variety . . . . . . . . . . . . . . . . . . . . . . 72
7 Shimura varieties of hodge type 73
CONTENTS 3
8 PEL Shimura varieties 76
Algebras with involution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
Symplectic modules and the associated algebraic groups . . . . . . . . . . . . . . . . . . . . . . 77
Algebras with positive involution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
PEL data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
PEL Shimura varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
PEL modular varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
9 General Shimura varieties 84
Abelian motives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
Shimura varieties of abelian type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85
Classication of Shimura varieties of abelian type . . . . . . . . . . . . . . . . . . . . . . . . . . 86
Shimura varieties not of abelian type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Example: simple Shimura varieties of type A
1
. . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
10 Complex multiplication: the Shimura-Taniyama formula 88
Where we are headed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Review of abelian varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
CM elds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Abelian varieties of CM-type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90
Abelian varieties over a nite eld . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
The Shimura-Taniyama formula. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
The O
E
-structure of the tangent space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
Sketch of the proof the Shimura-Taniyama formula . . . . . . . . . . . . . . . . . . . . . . 96
11 Complex multiplication: the main theorem 97
Review of class eld theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Convention for the (Artin) reciprocity map . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
The reex eld and norm of a CM-type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Statement of the main theorem of complex multiplication . . . . . . . . . . . . . . . . . . . . . . 99
12 Denition of canonical models 101
Models of varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
The reex eld . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Special points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
The homomorphism r
x
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
Denition of a canonical model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 104
Examples: Shimura varieties dened by tori . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Tori . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
CM-tori . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
13 Uniqueness of canonical models 107
Extension of the base eld . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
Uniqueness of canonical models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
The galois action on the connected components . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
14 Existence of canonical models 110
Descent of the base eld . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
The regularity condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
The continuity condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
A sufcient condition for descent . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111
Review of local systems and families of abelian varieties . . . . . . . . . . . . . . . . . . . . . . 111
The Siegel modular variety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
The reex eld . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
The special points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 112
4 CONTENTS
A criterion to be canonical . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
Outline of the proof of the existence of a canonical model . . . . . . . . . . . . . . . . . . . 114
Condition (a) of (14.6). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Condition (b) of (14.6). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
Condition (c) of (14.6) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Simple PEL Shimura varieties of type A or C . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Shimura varieties of hodge type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
Shimura varieties of abelian type . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
General Shimura varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
Final remark: rigidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117
15 Abelian varieties over nite elds 118
Semisimple categories . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
Division algebras; the Brauer group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
Abelian varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Abelian varieties over F
q
, q = p
n
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
Abelian varieties over F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121
Tori and their representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
Afne extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
The afne extension P . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
The local form P
l
of P . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
The
I
-space attached to a fake motive . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126
The isocrystal of a fake motive . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
Abelian varieties of CM-type and fake abelian varieties . . . . . . . . . . . . . . . . . . . . 127
16 The good reduction of Shimura varieties 129
The points of the Shimura variety with coordinates in the algebraic closure of the rational numbers 129
The points of the Shimura variety with coordinates in the reex eld . . . . . . . . . . . . . . . . 129
Hyperspecial subgroups . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
The good reduction of Shimura varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
Denition of the Langlands-Rapoport set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
Denition of the set S() . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Denition of the group I(). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Denition of X
p
(). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
Denition of X
p
(). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
Denition of the Frobenius element . . . . . . . . . . . . . . . . . . . . . . . . . 133
The set S(). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
The admissibility condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
The condition at o. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
The condition at ,= p. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
The condition at p. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
The global condition. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
The Langlands-Rapoport set . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 134
The conjecture of Langlands and Rapoport . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
17 A formula for the number of points 136
Triples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
The triple attached to an admissible pair (, c) . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
The formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138
18 Endnotes 139
Proof of Theorem 5.4 (footnote 39) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 139
Proof of the claim in 5.23 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 140
References 142
CONTENTS 5
Index of denitions 147
6 CONTENTS
Introduction
The arithmetic properties of elliptic modular functions and forms were extensively studied
in the 1800s, culminating in the beautiful Kronecker Jugendtraum. Hilbert emphasized the
importance of extending this theory to functions of several variables in the twelfth of his
famous problems at the International Congress in 1900. The rst tentative steps in this di-
rection were taken by Hilbert himself and his students Blumenthal and Hecke in their study
of what are now called Hilbert (or Hilbert-Blumenthal) modular varieties. As the theory
of complex functions of several variables matured, other quotients of bounded symmet-
ric domains by arithmetic groups were studied (Siegel, Braun, and others). However, the
modern theory of Shimura varieties
1
only really began with the development of the theory
of abelian varieties with complex multiplication by Shimura, Taniyama, and Weil in the
mid-1950s, and with the subsequent proof by Shimura of the existence of canonical mod-
els for certain families of Shimura varieties. In two fundamental articles, Deligne recast
the theory in the language of abstract reductive groups and extended Shimuras results on
canonical models. Langlands made Shimura varieties a central part of his program, both as
a source of representations of galois groups and as tests for the conjecture that all motivic
L-functions are automorphic. These notes are an introduction to the theory of Shimura
varieties from the point of view of Deligne and Langlands. Because of their brevity, many
proofs have been omitted or only sketched.
Notations and conventions
Unless indicated otherwise, vector spaces are assumed to be nite dimensional and free
Z-modules are assumed to be of nite rank. The linear dual Hom(V, k) of a vector space
(or module) V is denoted V
= SO
n
, (GL
n
)
= GL
n
, and GL
n
(1)
consists of the n n
matrices with det > 0. For an algebraic group G over , G()
= G() G(1)
.
Following Bourbaki, I require compact topological spaces to be separated.
Semisimple and reductive groups, whether algebraic or Lie, are required to be con-
nected. A simple algebraic or Lie group is a semisimple group with no connected proper
normal subgroups other than 1 (some authors say almost-simple). For a torus T , X
+
(T)
denotes the character group of T. The inner automorphism dened by an element g is de-
noted ad(g). The derived group of a reductive group G is denoted G
der
(it is a semisimple
group). For more notations concerning reductive groups, see p51. For a nite extension of
elds L F of characteristic zero, the torus over F obtained by restriction of scalars from
G
m
over L is denoted
2
(G
m
)
L{F
.
1
The term Shimura variety was introduced by Langlands (1976, 1977), although earlier Shimura
curve had been used for the varieties of dimension one (Ihara 1968).
2
Thus, (G
m
)
L{F
has character group X
+
((G
m
)
L{F
) = Z
Hom(L,F
al
)
(free Z-module on Hom(L, F
al
) with
CONTENTS 7
Throughout, I use the notations standard in algebraic geometry, which sometimes con-
ict with those used in other areas. For example, if G and G
t
are algebraic groups over a
eld k, then a homomorphism G G
t
means a homomorphism dened over k; if K is
a eld containing k, then G
K
is the algebraic group over K obtained by extension of the
base eld and G(K) is the group of points of G with coordinates in K. If o: k K is a
homomorphism of elds and V is an algebraic variety (or other algebro-geometric object)
over k, then oV has its only possible meaning: apply o to the coefcients of the equations
dening V .
Let Aand B be sets and let be an equivalence relation on A. If there exists a canonical
surjection A B whose bres are the equivalence classes, then I say that B classies the
elements of A modulo or that it classies the -classes of elements of A.
A functor F: A B is fully faithful if the maps Hom
A
(a, a
t
) Hom
B
(Fa, Fa
t
) are
bijective. The essential image of such a functor is the full subcategory of B whose objects
are isomorphic to an object of the form Fa. Thus, a fully faithful functor F: A B is an
equivalence if and only if its essential image is B (Mac Lane 1998, p93).
References
In addition to those listed at the end, I refer to the following of my course notes (available
at www.jmilne.org/math/).
AG: Algebraic Geometry, v4.0, October 30, 2003.
ANT: Algebraic Number Theory, v2.1, August 31, 1998.
CFT: Class Field Theory, v3.1, May 6, 1997.
FT: Fields and galois Theory, v3.0, August 31, 2003.
MF: Modular Functions and Modular Forms, v1.1, May 22, 1997.
Prerequisites
Beyond the mathematics that students usually acquire by the end of their rst year of grad-
uate work (a little complex analysis, topology, algebra, differential geometry,...), I assume
familiarity with some algebraic number theory, algebraic geometry, algebraic groups, and
elliptic modular curves.
Acknowledgements
I thank the Clay Mathematical Institute and the organizers for giving me the opportunity
to lecture on Shimura varieties, the Fields Institute for providing a excellent setting for the
Summer School, and the audience for its lively participation. Also, I thank Lizhen Ji and
Gopal Prasad for their help with references, and F. H ormann and others for alerting me to
errors in earlier versions.
the natural action of Gal(F
al
,F)), and its points in an F-algebra R are (G
m
)
L{F
(R) = (L
F
R)
.
8 1 HERMITIAN SYMMETRIC DOMAINS
1 Hermitian symmetric domains
In this section, I describe the complex manifolds that play the role in higher dimensions of
the complex upper half plane, or, equivalently, the open unit disk:
{z c C [ `(z) > 0} = H
1
z
z-i
zi
>
-
i
z1
z-1
z
D
1
= {z c C [ [z[ - 1}.
This is a large topic, and I can do little more than list the denitions and results that we
shall need.
Brief review of real manifolds
A manifold M of dimension n is a separated topological space that is locally isomorphic to an open
subset of 1
n
and admits a countable basis of open subsets. A homeomorphism from an open subset
of M onto an open subset of 1
n
is called a chart of M.
Smooth manifolds
I use smooth to mean C
o
. A smooth manifold is a manifold M endowed with a smooth structure,
i.e., a sheaf O
M
of 1-valued functions such that (M, O
M
) is locally isomorphic to 1
n
endowed
with its sheaf of smooth functions. For an open U M, the c O
M
(U) are called the smooth
functions on U. A smooth structure on a manifold M can be dened by a family u
: U
1
n
of
charts such that M =
_
U
u
-1
: u
(U
) u
(U
)
are smooth for all , . A continuous map : M N of smooth manifolds is smooth if it is a map
of ringed spaces, i.e., smooth on an open V N implies smooth on
-1
(V ).
Let (M, O
M
) be a smooth manifold, and let O
M,p
be the ring of germs of smooth functions
at p. The tangent space T
p
M to M at p is the 1-vector space of 1-derivations X
p
: O
M,p
1.
If x
1
, . . . , x
n
are local coordinates at p, then
J
Jx
1
, . . . ,
J
Jx
n
is a basis for T
p
M and dx
1
, . . . , dx
n
is
the dual basis.
Let U be an open subset of a smooth manifold M. A smooth vector eld X on U is a family
of tangent vectors X
p
c T
p
(M) indexed by p c U, such that, for any smooth function on an
open subset of U, p . X
p
is smooth. A smooth r-tensor eld on U is a family t = (t
p
)
p=M
of multilinear mappings t
p
: T
p
M T
p
M 1 (r copies of T
p
M) such that, for any smooth
vector elds X
1
, . . . , X
r
on an open subset of U, p . t
p
(X
1
, . . . , X
r
) is a smooth function. A
smooth (r, s)-tensor eld is a family t
p
: (T
p
M)
r
(T
p
M)
s
1 satisfying a similar condition.
Note that to give a smooth (1, 1)-eld amounts to giving a family of endomorphisms t
p
: T
p
M
T
p
M with the property that p .t
p
(X
p
) is a smooth vector eld for any smooth vector eld X.
Ariemannian manifold is a smooth manifold endowed with a riemannian metric, i.e., a smooth
2-tensor eld g such that, for all p c M, g
p
is symmetric and positive denite. In terms of local
coordinates
3
x
1
, . . . , x
n
at p,
g
p
=
g
i,j
(p)dx
i
dx
j
, i.e., g
p
_
J
Jx
i
,
J
Jx
j
_
= g
ij
(p).
3
In this situation, we usually write dx
i
dx
j
for dx
i
dx
j
see Lee 1997, p24 for an explanation of this.
Brief review of hermitian forms 9
A morphism of riemannian manifolds is called an isometry.
A real Lie group
4
G is a smooth manifold endowed with a group structure dened by smooth
maps g
1
, g
2
.g
1
g
2
, g .g
-1
.
Brief review of hermitian forms
To give a complex vector space amounts to giving a real vector space V together with an endo-
morphism J: V V such that J
2
= 1. A hermitian form on (V, J) is an 1-bilinear mapping
( [ ) : V V C such that (Ju[:) = i(u[:) and (:[u) = (u[:). When we write
5
(u[:) = (u, :) i[(u, :), (u, :), [(u, :) c 1, (1)
then and [ are 1-bilinear, and
is symmetric (Ju, J:) = (u, :), (2)
[ is alternating [(Ju, J:) = [(u, :), (3)
[(u, :) = (u, J:), (u, :) = [(u, J:). (4)
As (u[u) = (u, u), ( [ ) is positive denite if and only if is positive denite. Conversely, if
satises (2) (resp. [ satises (3)), then the formulas (4) and (1) dene a hermitian form:
(u[:) = (u, :) i(u, J:) (resp. (u[:) = [(u, J:) i[(u, :)) (5)
Complex manifolds
AC-valued function on an open subset U of C
n
is analytic if it admits a power series expan-
sion in a neighbourhod of each point of U. A complex manifold is a manifold M endowed
with a complex structure, i.e., a sheaf O
M
of C-valued functions such that (M, O
M
) is
locally isomorphic to C
n
with its sheaf of analytic functions. A complex structure on a
manifold M can be dened by a family u
: U
C
n
of charts such that M =
_
U
and
the maps u
u
-1
are analytic for all , . Such a family also makes M into a smooth
manifold denoted M
o
. A continuous map : M N of complex manifolds is analytic if
it is a map of ringed spaces. A riemann surface is a one-dimensional complex manifold.
A tangent vector at a point p of a complex manifold is a C-derivation O
M,p
C. The
tangent spaces T
p
M (M as a complex manifold) and T
p
M
o
(M as a smooth manifold)
can be identied. Explicitly, complex local coordinates z
1
, . . . , z
n
at a point p of M dene
real local coordinates x
1
, . . . , x
n
, y
1
, . . . , y
n
with z
r
= x
r
iy
r
. The real and complex
tangent spaces have bases
J
Jx
1
, . . . ,
J
Jx
n
,
J
Jy
1
, . . . ,
J
Jy
n
and
J
Jz
1
, . . . ,
J
Jz
n
respectively. Under
the natural identication of the two spaces,
J
Jz
r
=
1
2
_
J
Jx
r
i
J
Jy
r
_
.
4
According to a theorem of Lie, this is equivalent to the usual denition in which smooth is replaced by
real-analytic.
5
For example, let V = C, so (z[z
t
) = azz
t
for some a > 0. Then
(x iy)(x
t
iy
t
) = a(xx
t
yy
t
) ia(xy
t
yx
t
),
and so
= a(xx
t
yy
t
), [ = a(xy
t
yx
t
).
10 1 HERMITIAN SYMMETRIC DOMAINS
A C-valued function on an open subset U of C
n
is holomorphic if it is holomorphic
(i.e., differentiable) separately in each variable. As in the one-variable case, is holomor-
phic if and only if it is analytic (Hartogs theorem, Taylor 2002, 2.2.3), and so we can use
the terms interchangeably.
Recall that a C-valued function on U C is holomorphic if and only if it is smooth
(as a function of two real variables) and satises the Cauchy-Riemann condition. This
condition has a geometric interpretation: it requires that d
p
: T
p
U T
((p)
C be C-linear
for all p c U. It follows that a smooth C-valued function on U C
n
is holomorphic if
and only if the maps d
p
: T
p
U T
((p)
C are C-linear for all p c U.
An almost-complex structure on a smooth manifold M is a smooth tensor eld (J
p
)
p=M
,
J
p
: T
p
M T
p
M, such that J
2
p
= 1 for all p, i.e., it is a smoothly varying family of
complex structures on the tangent spaces. A complex structure on a smooth manifold en-
dows it with an almost-complex structure. In terms of complex local coordinates z
1
, . . . , z
n
in a neighbourhood of a point p on a complex manifold and the corresponding real local
coordinates x
1
, . . . , y
n
, J
p
acts by
d
dx
r
.
d
dy
r
,
d
dy
r
.
d
dx
r
. (6)
It follows from the last paragraph that the functor from complex manifolds to almost-
complex manifolds is fully faithful: a smooth map : M N of complex manifolds is
holomorphic (analytic) if the maps d
p
: T
p
M T
(p)
N are C-linear for all p c M. Not
every almost-complex structure on a smooth manifold arises from a complex structure
those that do are said to be integrable. An almost-complex structure J on a smooth man-
ifold is integrable if M can be covered by charts on which J takes the form (6) (because
this condition forces the transition maps to be holomorphic).
6
A hermitian metric on a complex (or almost-complex) manifold M is a riemannian
metric g such that
g(JX, J Y ) = g(X, Y ) for all vector elds X, Y . (7)
According to (5), for each p c M, g
p
is the real part of a unique hermitian form h
p
on
T
p
M, which explains the name. A hermitian manifold (M, g) is a complex manifold with
a hermitian metric, or, in other words, it is a riemannian manifold with a complex structure
such that J acts by isometries.
Hermitian symmetric spaces
A manifold (riemannian, hermitian, . . . ) is said to be homogeneous if its automorphism
group acts transitively. It is symmetric if, in addition, at some point p there is an involution
s
p
(the symmetry at p) having p as an isolated xed point. This means that s
p
is an auto-
morphism such that s
2
p
= 1 and that p is the only xed point of s
p
in some neighbourhood
of p.
6
See Wolf 1984, 8.7.2.
Hermitian symmetric spaces 11
For a riemannian manifold (M, g), the automorphism group is the group Is(M, g) of
isometries. A connected symmetric riemannian manifold is called a symmetric space.
7
For example, 1
n
with the standard metric g
p
=
dx
i
dx
i
is a symmetric space the
translations are isometries, and x .x is a symmetry at 0.
For a hermitian manifold (M, g), the automorphism group is the group Is(M, g) of
holomorphic isometries:
Is(M, g) = Is(M
o
, g) Hol(M) (8)
(intersection inside Aut(M
o
); Hol(M) is the group of automorphisms of M as a complex
manifold). A connected symmetric hermitian manifold is called a hermitian symmetric
space.
8
EXAMPLE 1.1. (a) The complex upper half plane H
1
becomes a hermitian symmetric space
when endowed with the metric
dxdy
y
2
. The action
_
a b
c d
_
z =
az b
cz d
,
_
a b
c d
_
c SL
2
(1), z c H
1
,
identies SL
2
(1),{I} with the group of holomorphic automorphisms of H
1
. For any
x iy c H
1
, x iy =
_
_
y x{
_
y
0 1{
_
y
_
i, and so H
1
is homogeneous. The isomorphism
z . 1,z is a symmetry at i c H
1
, and the riemannian metric
dxdy
y
2
is invariant under the
action of SL
2
(1) and has the hermitian property (7).
(b) The projective line P
1
(C) (= riemann sphere) becomes a hermitian symmetric space
when endowed with the restriction (to the sphere) of the standard metric on 1
3
. The group
of rotations is transitive, and reection along a geodesic (great circle) through a point is a
symmetry. Both of these transformations leave the metric invariant.
(c) Any quotient C, of C by a discrete additive subgroup becomes a hermitian
symmetric space when endowed with the standard metric. The group of translations is
transitive, and z .z is a symmetry at 0.
Curvature.
Recall that, for a plane curve, the curvature at a point p is 1,r where r is the radius of the
circle that best approximates the curve at p. For a surface in 3-space, the principal curva-
tures at a point p are the maximum and minimum of the signed curvatures of the curves
obtained by cutting the surface with planes through a normal at p (the sign is positive
9
or
7
Let (M, g) be a connected riemannian manifold. For each p c M, there is a diffeomorphism on a
neighbourhood of p (the geodesic symmetry at p) that sends sends ;(t ) to ;(t ) for each geodesic ; with
;(0) = p. Geometrically, it is reection along geodesics through p. When the geodesic symmetry at p is an
isometry, M is said to be locally symmetric at p. A symmetry s
p
at p coincides with the geodesic symmetry
at p (see 1.11 below), and conversely (M, g) is (globally) symmetric if, for every p c M, the geodesic
symmetry at p extends to a symmetry s
p
at p.
8
Some authors say globally symmetric riemannian space for symmetric space and globally symmet-
ric hermitian space for hermitian symmetric space.
9
According to my dictionary, positive can mean greater than zero or not negative. I use it only in
the rst sense.
12 1 HERMITIAN SYMMETRIC DOMAINS
negative according as the curve bends towards the normal or away). Although the principal
curvatures depend on the embedding of the surface into 1
3
, their product, the sectional
curvature at p, does not (Gausss Theorema Egregium) and so it is well-dened for any
two-dimensional riemannian manifold. More generally, for a point p on any riemannian
manifold M, one can dene the sectional curvature K(p, E) of the submanifold cut out
by the geodesics tangent to a two-dimensional subspace E of T
p
M. Intuitively, positive
curvature means that the geodesics through a point converge, and negative curvature means
that they diverge. The geodesics in the upper half plane are the half-lines and semicir-
cles orthogonal to the real axis. Clearly, they diverge in fact, this is Poincar es famous
model of noneuclidean geometry in which there are innitely many lines through a point
parallel to any xed line not containing it. More prosaically, one can compute that the
sectional curvature is 1. The Gauss curvature of P
1
(C) is obviously positive, and that of
C, is zero.
The three types of hermitian symmetric spaces
The group of isometries of a symmetric space (M, g) has a natural structure of a Lie
group
10
(Helgason 1978, IV 3.2). For a hermitian symmetric space (M, g), the group
Is(M, g) of holomorphic isometries is closed in the group of isometries of (M
o
, g) and
so is also a Lie group.
There are three families of hermitian symmetric spaces (ibid, VIII; Wolf 1984, 8.7):
Name example simply connected? curvature Is(M, g)
noncompact type H
1
yes negative adjoint, noncompact
compact type P
1
(C) yes positive adjoint, compact
euclidean C, not necessarily zero
A Lie group is adjoint if it is semisimple with trivial centre.
Every hermitian symmetric space, when viewed as hermitian manifold, decomposes
into a product M
0
M
-
M
with M
0
euclidean, M
-
of noncompact type, and M
is a product
of irreducible hermitian symmetric spaces, each of which has a simple isometry group.
We shall be especially interested in the hermitian symmetric spaces of noncompact type
they are called hermitian symmetric domains.
EXAMPLE 1.2 (SIEGEL UPPER HALF SPACE). The Siegel upper half space H
g
of degree
g consists of the symmetric complex g g matrices with positive denite imaginary part,
i.e.,
H
g
= {Z = X iY c M
g
(C) [ X = X
t
, Y > 0}.
Note that the map Z = (z
ij
) . (z
ij
)
j_i
identies H
g
with an open subset of C
g(g1){2
.
The symplectic group Sp
2g
(1) is the group xing the alternating form
g
i=1
x
i
y
-i
10
This was proved by E. Cartan, and extended to all riemannian manifolds by Myers and Steenrod.
Example: Bounded symmetric domains. 13
g
i=1
x
-i
y
i
:
Sp
2g
(1) =
__
A B
C D
_
A
t
C = C
t
A A
t
D C
t
B = I
g
D
t
AB
t
C = I
g
B
t
D = D
t
B
_
.
The group Sp
2g
(1) acts transitively on H
g
by
_
A B
C D
_
Z = (AZ B)(CZ D)
-1
.
The matrix
_
0 -I
g
I
g
0
_
acts as an involution on H
g
, and has iI
g
as its only xed point. Thus,
H
g
is homogeneous and symmetric as a complex manifold, and we shall see in (1.4) below
that H
g
is in fact a hermitian symmetric domain.
Example: Bounded symmetric domains.
A domain D in C
n
is a nonempty open connected subset. It is symmetric if the group
Hol(D) of holomorphic automorphisms of D (as a complex manifold) acts transitively and
for some point there exists a holomorphic symmetry. For example, H
1
is a symmetric
domain and D
1
is a bounded symmetric domain.
THEOREM 1.3. Every bounded domain has a canonical hermitian metric (called
11
the
Bergman(n) metric). Moreover, this metric has negative curvature.
PROOF (SKETCH): Initially, let D be any domain in C
n
. The holomorphic square-integrable
functions : D C form a Hilbert space H(D) with inner product ( [g) =
_
D
gd:.
There is a unique
12
(Bergman kernel) function K: D D C such that
(a) the function z .K(z, () lies in H(D) for each (,
(b) K(z, () = K((, z), and
11
After Stefan Bergmann. When he moved to the United States in 1939, he dropped the second n from his
name.
12
When one ignores convergence questions, the proof is easy. Let k be a second function satisfying the
three conditions. Then
k(z, () =
_
K(z, t )k(t, ()d:(t )
=
_
k((, t )K(z, t )d:(t )
= K(z, (),
which proves the uniqueness. Let
K(z, () =
m
e
m
(z) e
m
(().
Then clearly K(z, () = K((, z), and
=
m
( [e
m
)e
m
=
_
K(, () (()d:(()
(actual equality, not almost-everywhere equality, because the functions are holomorphic).
14 1 HERMITIAN SYMMETRIC DOMAINS
(c) (z) =
_
K(z, () (()d:(() for all c H(D).
For example, for any complete orthonormal set (e
m
)
m=N
in H(D), K(z, () =
m
e
m
(z)
e
m
(() is such a function. If D is bounded, then all polynomial functions on D are square-
integrable, and so certainly K(z, z) > 0 for all z. Moreover, log(K(z, z)) is smooth and
the equations
h =
h
ij
dz
i
dz
j
, h
ij
(z) =
d
2
dz
i
dz
j
log K(z, z),
dene a hermitian metric on D, which can be shown to have negative curvature (Helgason
1978, VIII 3.3, 7.1; Krantz 1982, 1.4).
The Bergman metric, being truly canonical, is invariant under the action Hol(D). Hence,
a bounded symmetric domain becomes a hermitian symmetric domain for the Bergman
metric. Conversely, it is known that every hermitian symmetric domain can be embedded
into some C
n
as a bounded symmetric domain. Therefore, a hermitian symmetric domain
D has a unique hermitian metric that maps to the Bergman metric under every isomorphism
of D with a bounded symmetric domain. On each irreducible factor, it is a multiple of the
original metric.
EXAMPLE 1.4. Let D
g
be the set of symmetric complex matrices such that I
g
Z
t
Z is pos-
itive denite. Note that (z
ij
) . (z
ij
)
j_i
identies D
g
as a bounded domain in C
g(g1){2
.
The map Z . (Z iI
g
)(Z iI
g
)
-1
is an isomorphism of H
g
onto D
g
. Therefore, D
g
is symmetric and H
g
has an invariant hermitian metric: they are both hermitian symmetric
domains.
Automorphisms of a hermitian symmetric domain
LEMMA 1.5. Let (M, g) be a symmetric space, and let p c M. Then the subgroup K
p
of
Is(M, g)
,K
p
M
is an isomorphism of smooth manifolds. In particular, Is(M, g)
acts transitively on M.
PROOF. For any riemannian manifold (M, g), the compact-open topology makes Is(M, g)
into a locally compact group for which the stabilizer K
t
p
of a point p is compact (Helgason
1978, IV 2.5). The Lie group structure on Is(M, g) noted above is the unique such structure
compatible with the compact-open topology (ibid. II 2.6). An elementary argument (e.g.,
MF 1.2) now shows that Is(M, g),K
t
p
M is a homeomorphism, and it follows that
the map a . ap: Is(M, g) M is open. Write Is(M, g) as a nite disjoint union
Is(M, g) =
i
Is(M, g)
a
i
of cosets of Is(M, g)
a
i
p and Is(M, g)
a
j
p are either disjoint or equal, but, as M is connected, they
must all be equal, which shows that Is(M, g)
,K
p
M is a homeomorphism, and it follows that it is a diffeomorphism (Helgason 1978, II
4.3a).
The homomorphism u
p
: U
1
Hol(D) 15
PROPOSITION 1.6. Let (M, g) be a hermitian symmetric domain. The inclusions
Is(M
o
, g) Is(M, g) Hol(M)
give equalities:
Is(M
o
, g)
= Is(M, g)
= Hol(M)
.
Therefore, Hol(M)
,K
p
= M
o
.
PROOF. The rst equality is proved in Helgason 1978, VIII 4.3, and the second can be
proved similarly. The rest of the statement follows from (1.5).
Let H be a connected real Lie group. There need not be an algebraic group G over 1
such that
13
G(1)
).
There is a unique connected algebraic subgroup G of GL(h) such that
G(1)
= Hol(M)
(inside GL(h)).
For such a G,
G(1)
as the
stabilizer of H
1
.
The homomorphism u
p
: U
1
Hol(D)
Let U
1
= {z c C [ [z[ = 1} (the circle group).
THEOREM 1.9. Let D be a hermitian symmetric domain. For each p c D, there exists
a unique homomorphism u
p
: U
1
Hol(D) such that u
p
(z) xes p and acts on T
p
D as
multiplication by z.
13
For example, the (topological) fundamental group of SL
2
(1) is Z, and so SL
2
(1) has many proper
covering groups (even of nite degree). None of them is algebraic.
16 1 HERMITIAN SYMMETRIC DOMAINS
EXAMPLE 1.10. Let p = i c H
1
, and let h: C
SL
2
(1) be the homomorphismz = a
ib .
_
a b
-b a
_
. Then h(z) acts on the tangent space T
i
H
1
as multiplication by z,z, because
d
dz
_
azb
-bza
_
[
i
=
a
2
b
2
(a-bi)
2
. For z c U
1
, choose a square root
_
z c U
1
, and set u(z) = h(
_
z)
mod I. Then u(z) is independent of the choice of
_
z because h(1) = I. Therefore,
u is a well-dened homomorphism U
1
PSL
2
(1) such that u(z) acts on the tangent
space T
i
H
1
as multiplication by z.
Because of the importance of the theorem, I sketch a proof.
PROPOSITION 1.11. Let (M, g) be symmetric space. The symmetry s
p
at p acts as 1 on
T
p
M, and, for any geodesic ; with ;(0) = p, s
p
(;(t )) = ;(t ). Moreover, (M, g) is
(geodesically) complete.
PROOF. Because s
2
p
= 1, (ds
p
)
2
= 1, and so ds
p
acts semisimply on T
p
M with eigenval-
ues 1. Recall that for any tangent vector X at p, there is a unique geodesic ;: I M
with ;(0) = p, ;(0) = X. If (ds
p
)(X) = X, then s
p
; is a geodesic sharing these prop-
erties, and so p is not an isolated xed point of s
p
. This proves that only 1 occurs as an
eigenvalue. If (ds
p
)(X) = X, then s
p
; and t .;(t ) are geodesics through p with
velocity X, and so are equal. For the nal statement, see Boothby 1975, VII 8.4.
By a canonical tensor on a symmetric space (M, g), I mean any tensor canonically
derived from g, and hence xed by any isometry of (M, g).
PROPOSITION 1.12. On a symmetric space (M, g) every canonical r-tensor with r odd is
zero. In particular, parallel translation of two-dimensional subspaces does not change the
sectional curvature.
PROOF. Let t be a canonical r-tensor. Then
t
p
= t
p
(ds
p
)
r
1.11
= (1)
r
t
p
,
and so t = 0 if r is odd. For the second statement, let V be the riemannian connection, and
let R be the corresponding curvature tensor (Boothby 1975, VII 3.2, 4.4). Then VR is an
odd tensor, and so is zero. This implies that parallel translation of 2-dimensional subspaces
does not change the sectional curvature.
PROPOSITION 1.13. Let (M, g) and (M
t
, g
t
) be riemannian manifolds in which paral-
lel translation of 2-dimensional subspaces does not change the sectional curvature. Let
a: T
p
M T
p
0 M
t
be a linear isometry such that K(p, E) = K(p
t
, aE) for every 2-
dimensional subspace E T
p
M. Then exp
p
(X) .exp
p
0 (aX) is an isometry of a neigh-
bourhood of p onto a neighbourhood of p
t
.
PROOF. This follows from comparing the expansions of the riemann metrics in terms of
normal geodesic coordinates. See Wolf 1984, 2.3.7.
PROPOSITION 1.14. If in (1.13) M and M
t
are complete, connected, and simply con-
nected, then there is a unique isometry : M M
t
such that (p) = p
t
and d
p
= a.
PROOF. See Wolf 1984, 2.3.12.
Cartan involutions 17
I now complete the sketch of the proof of Theorem 1.9. Each z with [z[ = 1 denes an
automorphism of (T
p
D, g
p
), and one checks that it preserves sectional curvatures. Accord-
ing to (1.11, 1.12, 1.14), there exists a unique isometry u
p
(z): D D such that du
p
(z)
p
is multiplication by z. It is holomorphic because it is C-linear on the tangent spaces. The
isometry u
p
(z) u
p
(z
t
) xes p and acts as multiplication by zz
t
on T
p
D, and so equals
u
p
(zz
t
).
Cartan involutions
Let G be a connected algebraic group over 1, and let g . g denote complex conjugation
on G(C). An involution 0 of G (as an algebraic group over 1) is said to be Cartan if the
group
G
(0)
(1)
df
= {g c G(C) [ g = 0(g)} (9)
is compact.
EXAMPLE 1.15. Let G = SL
2
, and let 0 = ad
_
0 1
-1 0
_
. For
_
a b
c d
_
c SL
2
(C), we have
0
__
a b
c d
__
=
_
0 1
-1 0
_
_
a b
c d
_
_
0 1
-1 0
_
-1
=
_
d -c
-b a
_
.
Thus,
SL
(0)
2
(1) =
_
_
a b
c d
_
c SL
2
(C) [ d = a, c = b
_
=
_
a b
-b a
_
c GL
2
(C) [ [a[
2
[b[
2
= 1
_
= SU
2
,
which is compact, being a closed bounded set in C
2
. Thus 0 is a Cartan involution for SL
2
.
THEOREM 1.16. There exists a Cartan involution if and only if G is reductive, in which
case any two are conjugate by an element of G(1).
PROOF. See Satake 1980, I 4.3.
EXAMPLE 1.17. Let G be a connected algebraic group over 1.
(a) The identity map on G is a Cartan involution if and only if G(1) is compact.
(b) Let G = GL(V ) with V a real vector space. The choice of a basis for V determines
a transpose operator M . M
t
, and M . (M
t
)
-1
is obviously a Cartan involution. The
theorem says that all Cartan involutions of G arise in this way.
(c) Let G GL(V ) be a faithful representation of G. Then G is reductive if and
only if G is stable under g . g
t
for a suitable choice of a basis for V , in which case the
restriction of g . (g
t
)
-1
to G is a Cartan involution; all Cartan involutions of G arise in
this way from the choice of a basis for V (Satake 1980, I 4.4).
(d) Let 0 be an involution of G. There is a unique real form G
(0)
of G
C
such that
complex conjugation on G
(0)
(C) is g .0(g). Then, G
(0)
(1) satises (9), and we see that
the Cartan involutions of G correspond to the compact forms of G
C
.
18 1 HERMITIAN SYMMETRIC DOMAINS
PROPOSITION 1.18. Let G be a connected algebraic group over 1. If G(1) is compact,
then every nite-dimensional real representation of G GL(V ) carries a G-invariant
positive denite symmetric bilinear form; conversely, if one faithful nite-dimensional real
representation of G carries such a form, then G(1) is compact.
PROOF. Let ,: G GL(V ) be a real representation of G. If G(1) is compact, then its
image H in GL(V ) is compact. Let dh be the Haar measure on H, and choose a positive
denite symmetric bilinear form ( [ ) on V . Then the form
(u[:)
t
=
_
H
(hu[h:)dh
is G-invariant, and it is still symmetric, positive denite, and bilinear. For the converse,
choose an orthonormal basis for the form. Then G(1) becomes identied with a closed set
of real matrices A such that A
t
A = I, which is bounded.
REMARK 1.19. The proposition can be restated for complex representations: if G(1) is
compact then every nite-dimensional complex representation of G carries a G-invariant
positive denite Hermitian form; conversely, if some faithful nite-dimensional complex
representation of G carries a G-invariant positive denite Hermitian form, then G is com-
pact. (In this case, G(1) is a subgroup of a unitary group instead of an orthogonal group.
For a sesquilinear form to be G-invariant means that (gu, g:) = (u, :), g c G(C),
u, : c V .)
Let G be a real algebraic group, and let C be an element of G(1) whose square is
central (so that adC is an involution). A C-polarization on a real representation V of G is
a G-invariant bilinear form such that the form
C
,
(u, :) .(u, C:),
is symmetric and positive denite.
PROPOSITION 1.20. If adC is a Cartan involution of G, then every nite-dimensional real
representation of G carries a C-polarization; conversely, if one faithful nite-dimensional
real representation of G carries a C-polarization, then adC is a Cartan involution.
PROOF. An 1-bilinear form on a real vector space V denes a sesquilinear form
t
on
V (C),
t
: V (C) V (C) C,
t
(u, :) =
C
(u, :).
Moreover,
t
is hermitian (and positive denite) if and only if is symmetric (and positive
denite).
Let ,: G GL(V ) be a real representation of G. For any G-invariant bilinear form
on V ,
C
is G(C)-invariant, and so
t
(gu, g:) =
t
(u, :), all g c G(C), u, : c V (C). (10)
On replacing : with C: in this equality, we nd that
t
(gu, C(C
-1
gC):) =
t
(u, C:), all g c G(C), u, : c V (C), (11)
Representations of U
1
19
which says that
t
C
is invariant under G
(adC)
.
If , is faithful and is a C-polarization, then
t
C
is a positive denite hermitian form,
and so G
(adC)
(1) is compact (1.19): adC is a Cartan involution.
Conversely, if G
(adC)
(1) is compact, then every real representation G GL(V ) car-
ries a G
(adC)
(1)-invariant positive denite symmetric bilinear form (1.18). Similar cal-
culations to the above show that
C
1 is a C-polarization on V .
Representations of U
1
Let T be a torus over a eld k, and let K be a galois extension of k splitting T. To give
a representation , of T on a k-vector space V amounts
14
to giving an X
+
(T)-grading
V (K) =
y=X
(T)
V
y
on V (K) =
df
K
k
V with the property that
o(V
y
) = V
oy
, all o c Gal(K,k), y c X
+
(T).
Here V
y
is the subspace of K
k
V on which T acts through y:
,(t ): = y(t ) :, for : c V
y
, t c T(K).
If V
y
,= 0, we say that y occurs in V .
When we regard U
1
as a real algebraic torus, its characters are z . z
n
, n c Z. Thus,
X
+
(U
1
)
n_0
V (C)
n
=
n>0
V (C)
n
(12)
is an isomorphism. From this discussion, we see that every real representation of U
1
is a
direct sum of representations of the following types:
(a) V = 1 with U
1
acting trivially (so V (C) = V
0
);
(b) V = 1
2
with z = xiy c U
1
(1) acting as
_
x y
-y x
_
n
, n > 0 (so V (C) = V
n
V
-n
).
Classication of hermitian symmetric domains in terms of real groups
The representations of U
1
have the same description whether we regard it as a Lie group or
an algebraic group, and so every homomorphism U
1
GL(V ) of Lie groups is algebraic.
It follows that the homomorphismu
p
: U
1
Hol(D)
= G(1)
= Hol(D)
,K
p
= D where K
p
is the group xing p (see 1.6). For z c U
1
,
u
p
(z) acts on the 1-vector space
Lie(G), Lie(K
p
)
= T
p
D
as multiplication by z, and it acts on Lie(K
p
) trivially. From this, (a) follows.
The symmetry s
p
at p and u
p
(1) both x p and act as 1 on T
p
D (see 1.11); they are
therefore equal (1.14). It is known that the symmetry at a point of a symmetric space gives
a Cartan involution of G if and only if the space has negative curvature (see Helgason 1978,
V2; the real formof G dened by ads
p
is that attached to the compact dual of the symmetric
space). Thus (b) holds.
Finally, if the projection of u(1) into a simple factor of G were trivial, then that factor
would be compact (by (b); see 1.17a), and D would have an irreducible factor of compact
type.
For the converse, let D be the set of G(1)
,K
u
_
u
endows D with the structure of smooth (even real-analytic) manifold. For this structure,
the tangent space to D at u,
T
u
D = Lie(G), Lie(K
u
),
which, because of (a), can be identied with the subspace of Lie(G)
C
on which u(z) acts as
z (see (12)). This endows T
u
D with a C-vector space structure for which u(z), z c U
1
, acts
as multiplication by z. Because D is homogeneous, this gives it the structure of an almost-
complex manifold, which can be shown to integrable (Wolf 1984, 8.7.9). The action of
K
u
on D denes an action of it on T
u
D. Because K
u
is compact, there is a K
u
-invariant
positive denite form on T
u
D (see 1.18), and because J = u(i) c K
u
, any such form will
have the hermitian property (7). Choose one, and use the homogeneity of D to move it
to each tangent space. This will make D into a hermitian symmetric space, which will be
a hermitian symmetric domain because each simple factor of its automorphism group is a
noncompact semisimple group (because of (b,c)).
COROLLARY 1.22. There is a natural one-to-one correspondence between isomorphism
classes of pointed hermitian symmetric domains and pairs (G, u) consisting of a real ad-
joint Lie group and a nontrivial homomorphism u: U
1
G(1) satisfying (a), (b), (c).
EXAMPLE 1.23. Let u: U
1
PSL
2
(1) be as in (1.10). Then u(1) =
_
0 1
-1 0
_
and we saw
in 1.15 that adu(1) is a Cartan involution of SL
2
, hence also of PSL
2
.
Classication of hermitian symmetric domains in terms of dynkin diagrams 21
Classication of hermitian symmetric domains in terms of dynkin dia-
grams
Let G be a simple adjoint group over 1, and let u be a homomorphism U
1
G satisfying
(a) and (b) of Theorem 1.21. By base extension, we get an adjoint group G
C
, which is
simple because it is an inner form of its compact form,
15
and a cocharacter j = u
C
of G
C
satisfying the following condition:
(*) in the action of G
m
on Lie(G
C
) dened by ad j, only the characters
z, 1, z
-1
occur.
PROPOSITION 1.24. The map (G, u) . (G
C
, u
C
) denes a bijection between the sets of
isomorphism classes of pairs consisting of
(a) a simple adjoint group over 1 and a conjugacy class of u: U
1
H satisfying
(1.21a,b), and
(b) a simple adjoint group over C and a conjugacy class of cocharacters satisfying (*).
PROOF. Let (G, j) be as in (b), and let g . g denote complex conjugation on G(C)
relative to the unique compact real form of G (cf. 1.16). There is a real form H of G such
that complex conjugation on H(C) = G(C) is g .j(1) g j(1)
-1
, and u =
df
j[U
1
takes values in H(1). The pair (H, u) is as in (a), and the map (G, j) (H, u) is inverse
to (H, u) .(H
C
, u
C
) on isomorphism classes.
Let G be a simple algebraic group C. Choose a maximal torus T in G and a base (
i
)
i=I
for the roots of G relative to T . Recall, that the nodes of the dynkin diagram of (G, T)
are indexed by I. Recall also (Bourbaki 1981, VI 1.8) that there is a unique (highest) root
=
n
i
i
such that, for any other root
m
i
i
, n
i
_ m
i
all i. An
i
(or the associated
node) is said to be special if n
i
= 1.
Let M be a conjugacy class of nontrivial cocharacters of G satisfying (*). Because all
maximal tori of G are conjugate, M has a representative in X
+
(T) X
+
(G), and because
the Weyl group acts simply transitively on the Weyl chambers (Humphreys 1972, 10.3)
there is a unique representative j for M such that (
i
, j) _ 0 for all i c I. The condition
(*) is that
16
(, j) c {1, 0, 1} for all roots . Since j is nontrivial, not all the values
(, j) can be zero, and so this condition implies that (
i
, j) = 1 for exactly one i c I,
which must in fact be special (otherwise ( , j) > 1). Thus, the M satisfying (*) are in
one-to-one correspondence with the special nodes of the dynkin diagram. In conclusion:
THEOREM 1.25. The isomorphism classes of irreducible hermitian symmetric domains are
classied by the special nodes on connected dynkin diagrams.
The special nodes can be read off from the list of dynkin diagrams in, for example,
Helgason 1978, p477. In the following table, we list the number of special nodes for each
type:
Type A
n
B
n
C
n
D
n
E
6
E
7
E
8
F
4
G
2
n 1 1 3 2 1 0 0 0
15
If G
C
is not simple, say, G
C
= G
1
G
2
, then G = Res
C{R
(G
1
) and any inner form of G is also the
restriction of scalars of a C-group; but such a group can not be compact (look at a subtorus).
16
The j with this property are sometimes said to be minuscule (cf. Bourbaki 1981, pp226227).
22 1 HERMITIAN SYMMETRIC DOMAINS
In particular, there are no irreducible hermitian symmetric domains of type E
8
, F
4
, or G
2
and, up to isomorphism, there are exactly 2 of type E
6
and 1 of type E
7
. It should be noted
that not every simple real algebraic group arises as the automorphism group of a hermitian
symmetric domain. For example, PGL
n
arises in this way only for n = 2.
NOTES. For introductions to smooth manifolds and riemannian manifolds, see Boothby
1975 and Lee 1997. The ultimate source for hermitian symmetric domains is Helgason
1978, but Wolf 1984 is also very useful, and Borel 1998 gives a succinct treatment close
to that of the pioneers. The present account has been inuenced by Deligne 1973a and
Deligne 1979.
23
2 Hodge structures and their classifying spaces
We describe various objects and their parameter spaces. Our goal is a description of hermi-
tian symmetric domains as the parameter spaces for certain special hodge structures.
Reductive groups and tensors
Let G be a reductive group over a eld k of characteristic zero, and let ,: G GL(V ) be
a representation of G. The contragredient or dual ,
dened by
(,
, : c V.
A representation is said to be self-dual if it is isomorphic to its contragredient.
An r-tensor of V is a multilinear map
t : V V k (r-copies of V ).
For an r-tensor t , the condition
t (g:
1
, . . . , g:
r
) = (:
1
, . . . , :
r
), all :
i
c V,
on g denes a closed subgroup of GL(V )
t
of GL(V ). For example, if t is a nondegenerate
symmetric bilinear form V V k, then GL(V )
t
is the orthogonal group. For a set T
of tensors of V ,
_
t=T
GL(V )
t
is called the subgroup of GL(V ) xing the t c T .
PROPOSITION 2.1. For any faithful self-dual representation G GL(V ) of G, there exists
a nite set T of tensors of V such that G is the subgroup of GL(V ) xing the t c T .
PROOF. In Deligne 1982, 3.1, it is shown there exists a possibly innite set T with this
property, but, because G is noetherian as a topological space (i.e., it has the descending
chain condition on closed subsets), a nite subset will sufce.
PROPOSITION 2.2. Let G be the subgroup of GL(V ) xing the tensors t c T. Then
Lie(G) =
g c End(V )
j
t (:
1
, . . . , g:
j
, . . . , :
r
) = 0, all t c T , :
i
c V
_
.
PROOF. The Lie algebra of an algebraic group G can be dened to be the kernel of
G(k[c]) G(k). Here k[c] is the k-algebra with c
2
= 0. Thus Lie(G) consists of
the endomorphisms 1 gc of V (k[c]) such that
t ((1 gc):
1
, (1 gc):
2
, . . .) = t (:
1
, :
2
, . . .), all t c T , :
i
c V.
On expanding this and cancelling, we obtain the assertion.
Flag varieties
Fix a vector space V of dimension n over a eld k.
24 2 HODGE STRUCTURES AND THEIR CLASSIFYING SPACES
The projective space P(V )
The set P(V ) of one-dimensional subspaces L of V has a natural structure of an algebraic
variety: the choice of a basis for V determines a bijection P(V ) P
n-1
, and the structure
of an algebraic variety inherited by P(V ) from the bijection is independent of the choice of
the basis.
Grassmann varieties
Let G
d
(V ) be the set of d-dimensional subspaces of V , some 0 - d - n. Fix a basis
for V . The choice of a basis for W then determines a d n matrix A(W) whose rows
are the coordinates of the basis elements. Changing the basis for W multiplies A(W) on
the left by an invertible d d matrix. Thus, the family of minors of degree d of A(W)
is well-determined up to multiplication by a nonzero constant, and so determines a point
P(W) in P
(
n
d
)-1
. The map W . P(W): G
d
(V ) P
(
n
d
)-1
identies G
d
(V ) with a
closed subvariety of P
(
n
d
)-1
(AG, 5.38). A coordinate-free description of this map is given
by
W .
_
d
W: G
d
(V ) P(
_
d
V ). (13)
Let S be a subspace of V of complementary dimension nd, and let G
d
(V )
S
be the set
of W c G
d
(V ) such that W S = {0}. Fix a W
0
c G
d
(V )
S
, so that V = W
0
S. For any
W c G
d
(V )
S
, the projection W W
0
given by this decomposition is an isomorphism,
and so W is the graph of a homomorphism W
0
S:
n .s == (n, s) c W.
Conversely, the graph of any homomorphism W
0
S lies in G
d
(V )
S
. Thus,
G
d
(V )
S
= Hom(W
0
, S). (14)
When we regard G
d
(V )
S
as an open subvariety of G
d
(V ), this isomorphism identies it
with the afne space A(Hom(W
0
, S)) dened by the vector space Hom(W
0
, S). Thus,
G
d
(V ) is smooth, and the tangent space to G
d
(V ) at W
0
,
T
W
0
(G
d
(V ))
= Hom(W
0
, S)
= Hom(W
0
, V,W
0
). (15)
Flag varieties
The above discussion extends easily to chains of subspaces. Let d = (d
1
, . . . , d
r
) be a
sequence of integers with n > d
1
> > d
r
> 0, and let G
d
(V ) be the set of ags
F : V V
1
V
r
0 (16)
with V
i
a subspace of V of dimension d
i
. The map
G
d
(V )
F|-(V
i
)
i
G
d
i
(V )
i
P(
_
d
i
V )
Hodge structures 25
realizes G
d
(V ) as a closed subset of
i
G
d
i
(V ) (Humphreys 1978, 1.8), and so it is a
projective variety. The tangent space to G
d
(V ) at the ag F consists of the families of
homomorphisms
i
: V
i
V,V
i
, 1 _ i _ r, (17)
satisfying the compatibility condition
i
[V
i1
i1
mod V
i1
.
ASIDE 2.3. A basis e
1
, . . . , e
n
for V is adapted to the ag F if it contains a basis e
1
, . . . , e
j
i
for each V
i
. Clearly, every ag admits such a basis, and the basis then determines the ag.
Because GL(V ) acts transitively on the set of bases for V , it acts transitively on G
d
(V ).
For a ag F, the subgroup P(F) stabilizing F is an algebraic subgroup of GL(V ), and the
map
g .gF
0
: GL(V ),P(F
0
) G
d
(V )
is an isomorphism of algebraic varieties. Because G
d
(V ) is projective, this shows that
P(F
0
) is a parabolic subgroup of GL(V ).
Hodge structures
Denition
For a real vector space V , complex conjugation on V (C) =
df
C
R
V is dened by
z : = z :.
An 1-basis e
1
, . . . , e
m
for V is also a C-basis for V (C) and
a
i
e
i
=
a
i
e
i
.
A hodge decomposition of a real vector space V is a decomposition
V (C) =
p,q=ZZ
V
p,q
such that V
q,p
is the complex conjugate of V
p,q
. A hodge structure is a real vector space
together with a hodge decomposition. The set of pairs (p, q) for which V
p,q
,= 0 is
called the type of the hodge structure. For each n,
pq=n
V
p,q
is stable under complex
conjugation, and so is dened over 1, i.e., there is a subspace V
n
of V such that V
n
(C) =
pq=n
V
p,q
(see AG 14.5). Then V =
n
V
n
is called the weight decomposition of V .
If V = V
n
, then V is said to have weight n.
An integral (resp. rational) hodge structure is a free Z-module of nite rank V (resp.
-vector space) together with a hodge decomposition of V (1) such that the weight decom-
position is dened over .
EXAMPLE 2.4. Let J be a complex structure on a real vector space V , and dene V
-1,0
and
V
0,-1
to be the i and i eigenspaces of J acting on V (C). Then V (C) = V
-1,0
V
0,-1
is a hodge structure of type (1, 0), (0, 1), and every real hodge structure of this type
arises from a (unique) complex structure. Thus, to give a rational hodge structure of type
(1, 0), (0, 1) amounts to giving a -vector space V and a complex structure on V (1),
and to give an integral hodge structure of type (1, 0), (0, 1) amounts to giving a C-vector
space V and a lattice V (i.e., a Z-submodule generated by an 1-basis for V ).
26 2 HODGE STRUCTURES AND THEIR CLASSIFYING SPACES
EXAMPLE 2.5. Let X be a nonsingular projective algebraic variety over C. Then H =
H
n
(X, ) has a Hodge structure of weight n for which H
p,q
H
n
(X, C) is canonically
isomorphic to H
q
(X, D
p
) (Voisin 2002, 6.1.3).
EXAMPLE 2.6. Let (m) be the hodge structure of weight 2m on the vector space .
Thus, ((m))(C) = (m)
-m,-m
. Dene Z(m) and 1(m) similarly.
17
The hodge ltration
The hodge ltration associated with a hodge structure of weight n is
F
: F
p
F
p1
, F
p
=
r_p
V
r,s
V (C).
Note that for p q = n,
F
q
=
s_q
V
s,r
=
s_q
V
r,s
=
r_p
V
r,s
and so
V
p,q
= F
p
F
q
. (18)
EXAMPLE 2.7. For a hodge structure of type (1, 0), (0, 1), the hodge ltration is
(F
-1
F
0
F
2
) = (V (C) V
0,-1
0).
The obvious 1-linear isomorphism V V (C),F
0
denes the complex structure on V
noted in (2.4).
Hodge structures as representations of S
Let S be C
.
The characters of S
C
are the homomorphisms (z
1
, z
2
) . z
p
1
z
q
2
, (r, s) c Z Z. Thus,
X
+
(S) = Z Z with complex conjugation acting as (p, q) . (q, p), and to give a
representation of S on a real vector space V amounts to giving a Z Z-grading of V (C)
such that V
p,q
= V
q,p
for all p, q (see p19). Thus, to give a representation of S on a
real vector space V is the same as to give a hodge structure on V . Following Deligne
1979, 1.1.1.1, we normalize the relation as follows: the homomorphism h: S GL(V )
corresponds to the hodge structure on V such that
h
C
(z
1
, z
2
): = z
-p
1
z
-q
2
: for : c V
p,q
. (19)
17
It would be a little more canonical to take the underlying vector space of (m) to be (2i)
m
because
this makes certain relations invariant under a change of the choice of i =
_
1 in C.
18
This is the transpose of the matrix of a ib acting on C relative to the basis 1, i, but it gives the correct
action on the tangent space, namely, if [z[ = 1, then h(z) acts as z
2
(see 1.10).
Hodge structures 27
In other words,
h(z): = z
-p
z
-q
: for : c V
p,q
. (20)
Note the minus signs! The associated weight decomposition has
V
n
= {: c V [ n
h
(r): = r
n
}, n
h
= h n. (21)
Let j
h
be the cocharacter of GL(V ) dened by
j
h
(z) = h
C
(z, 1). (22)
Then the elements of F
p
h
V are sums of : c V (C) satisfying j
h
(z) : = z
-r
: for some
r _ p.
To give a hodge structure on a -vector space V amounts to giving a homomorphism
h: S GL(V (1)) such that n
h
is dened over .
EXAMPLE 2.8. By denition, a complex structure on a real vector space is a homomor-
phism h: C End
R
(V ) of 1-algebras. Then h[C
: C
rr
0
=p,ss
0
=q
V
r,s
V
r
0
,s
0
.
In terms of representations of S,
(V, h
V
) (W, h
W
) = (V W, h
V
h
W
).
Morphisms of hodge structures
A morphism of Hodge structures is a linear map V W sending V
p,q
into W
p,q
for all
p, q. In other words, it is a morphism (V, h
V
) (W, h
W
) of representations of S.
Hodge tensors
Let R = Z, , or 1, and let (V, h) be an R-hodge structure of weight n. A multilinear
form t : V
r
R is a hodge tensor if the map
V V V R(nr,2)
it denes is a morphism of hodge structures. In other words, t is a hodge tensor if
t (h(z):
1
, h(z):
2
, . . .) = (zz)
-nr{2
t
R
(:
1
, :
2
, . . .), all z c C, :
i
c V (1),
or if
p
i
,=
q
i
=t
C
(:
p
1
,q
1
1
, :
p
2
,q
2
2
, . . .) = 0, :
p
i
,q
i
i
c V
p
i
,q
i
. (24)
Note that, for a hodge tensor t ,
t (C:
1
, C:
2
, . . .) = t (:
1
, :
2
, . . .).
EXAMPLE 2.11. Let (V, h) be a hodge structure of type (1, 0), (0, 1). A bilinear form
t : V V 1 is a hodge tensor if and only if t (Ju, J:) = t (u, :) for all u, : c V .
Polarizations
Let (V, h) be a hodge structure of weight n. A polarization of (V, h) is a hodge tensor
[: V V 1 such that [
C
(u, :) =
df
[(u, C:) is symmetric and positive denite. Then
[ is symmetric or alternating according as n is even or odd, because
[(:, u) = [(C:, Cu) = [
C
(C:, u) = [
C
(u, C:) = [(u, C
2
:) = (1)
n
[(u, :).
More generally, let (V, h) be an R-hodge structure of weight n where R is Z or . A
polarization of (V, h) is a bilinear form [: V V R such that [
R
is a polarization of
(V (1), h).
Variations of hodge structures 29
EXAMPLE 2.12. Let (V, h) be an R-hodge structure of type (1, 0), (0, 1) with R =
Z, , or 1, and let J = h(i). A polarization of (V, h) is an alternating bilinear form
[: V V R such that, for u, : c V (1),
[
R
(Ju, J:) = [(u, :), and
[
R
(u, Ju) > 0 if u ,= 0.
(These conditions imply that [
R
(u, J:) is symmetric.)
EXAMPLE 2.13. Let X be a nonsingular projective variety over C. The choice of an embed-
ding X P
N
determines a polarization on the primitive part of H
n
(X, ) (Voisin 2002,
6.3.2).
Variations of hodge structures
Fix a real vector space V , and let S be a connected complex manifold. Suppose that,
for each s c S, we have a hodge structure h
s
on V of weight n (independent of s). Let
V
p,q
s
= V
p,q
h
s
and F
p
s
= F
p
s
V = F
p
h
s
V .
The family of hodge structures (h
s
)
s=S
on V is said to be continuous if, for xed p and
q, the subspace V
p,q
s
varies continuously with s. This means that the dimension d(p, q) of
V
p,q
s
is constant and the map
s .V
p,q
s
: S G
d(p,q)
(V )
is continuous.
A continuous family of hodge structures (V
p,q
s
)
s
is said to be holomorphic if the hodge
ltration F
s
varies holomorphically with s. This means that the map ,
s .F
s
: S G
d
(V )
is holomorphic. Here d = (. . . , d(p), . . .) where d(p) = dimF
p
s
V =
r_p
d(r, q). Then
the differential of at s is a C-linear map
d
s
: T
s
S T
F
s
(G
d
(V ))
(17)
p
Hom(F
p
s
, V,F
p
s
).
If the image of d
s
is contained in
p
Hom(F
p
s
, F
p-1
s
,F
p
s
),
for all s, then the holomorphic family is called a variation of hodge structures on S.
Now let T be a family of tensors on V including a nondegenerate bilinear form t
0
, and
let d: Z Z N be a function such that
d(p, q) = 0 for almost all p, q;
d(q, p) = d(p, q);
d(p, q) = 0 unless p q = n.
Dene S(d, T) to be the set of all hodge structures h on V such that
30 2 HODGE STRUCTURES AND THEIR CLASSIFYING SPACES
dimV
p,q
h
= d(p, q) for all p, q;
each t c T is a hodge tensor for h;
t
0
is a polarization for h.
Then S(d, T) acquires a topology as a subspace of
d(p,q),=0
G
d(p,q)
(V ).
THEOREM 2.14. Let S
for a suitable V , d,
and T.
PROOF (SKETCH). (a) Let S
= S(d, T)
s
: S
G
d
(V ) is injective. Let G be
the smallest algebraic subgroup of GL(V ) such that
h(S) G, all h c S
(25)
(take G to be the intersection of the algebraic subgroups of GL(V ) with this property),
and let h
o
c S
, gh
o
g
-1
c S
is surjective:
S
= G(1)
h
o
.
The subgroup K
o
of G(1)
xing h
o
is closed, and so G(1)
,K
o
is a smooth (in fact, real
analytic) manifold. Therefore, S
= (G(1)
,K
o
) h
o
= G(1)
,K
o
.
Let g = Lie(G). From S
h
o
G
Ad
g End(V ), we obtain Hodge structures on g and
End(V ). Clearly, g
00
= Lie(K
o
) and so T
h
o
S
= g,g
00
. In the diagram,
T
h
o
S
= g,g
00
-
End(V ), End(V )
00
g(C),F
0
(23)
?
~
=
-
End(V (C)),F
0
(23)
?
~
=
=T
h
o
G
d
(V ).
(26)
the map from top-left to bottom-right is (d)
h
o
, which therefore maps T
h
o
S
onto a com-
plex subspace of T
h
o
G
d
(V ). Since this is true for all h
o
c S
G
d
(V ) in the proof is an embedding of smooth manifolds
(injective smooth map that is injective on tangent spaces and maps S
homeomorphically
onto its image). Therefore, if a smooth map T G
d
(V ) factors into
T
S
G
d
(V ),
then will be smooth. Moreover, if the map T G
d
(V ) is dened by a holomorphic
family of hodge structures on T, and it factors through S
i
t (:
1
, . . . , g:
i
, . . . , :
r
) = 0.
Then L has a hodge structure of weight 0. We say that (V, (V
p,q
)
p,q
, T) is special if L is of type
(1, 1), (0, 0), (1, 1). The family S(d, T)
containing (V, (V
p,q
)
p,q
, T) is a variation of hodge structures
if and only if (H, T) is special.
32 3 LOCALLY SYMMETRIC VARIETIES
3 Locally symmetric varieties
In this section, we study quotients of hermitian symmetric domains by certain discrete
groups.
Quotients of hermitian symmetric domains by discrete groups
PROPOSITION 3.1. Let D be a hermitian symmetric domain, and let I be a discrete sub-
group of Hol(D)
,K
p
D is a homeomorphism,
and so (MF, 2.5):
(a) for any p c D, {g c I [ gp = p} is nite;
(b) for any p c D, there exists a neighbourhood U of p such that, for g c I, gU is
disjoint from U unless gp = p;
(c) for any points p, q c D not in the same I-orbit, there exist neighbourhoods U of p
and V of q such that gU V = 0 for all g c I.
Assume I is torsion free. Then the group in (a) is trivial, and so I acts freely on D.
Endow ID with the quotient topology. If U and V are as in (c) , then U and V
are disjoint neighbourhoods of p and q, and so ID is separated. Let q c ID, and
let p c
-1
(q). If U is as in (b), then the restriction of to U is a homeomorphism
U U, and it follows that ID a manifold.
Dene a C-valued function on an open subset U of ID to be holomorphic if
is holomorphic on
-1
U. The holomorphic functions form a sheaf on ID for which is
a local isomorphism of ringed spaces. Therefore, the sheaf denes a complex structure on
ID for which is a local isomorphism of complex manifolds.
Finally, let : ID M be a map such that is holomorphic, and let be a
holomorphic function on an open subset U of M. Then is holomorphic because
is holomorphic, and so is holomorphic.
When I is torsion free, we often write D(I) for ID regarded as a complex manifold.
In this case, D is the universal covering space of D(I) and I is the group of covering
transformations; moreover, for any point p of D, the map
g .[image under of any path from p to gp]: I
1
(D(I), p)
is an isomorphism (Hatcher 2002, 1.40).
Subgroups of nite covolume
We shall only be interested in quotients of D by big discrete subgroups I of Aut(D)
.
This condition is conveniently expressed by saying that ID has nite volume. By deni-
Arithmetic subgroups 33
tion, D has a riemannian metric g and hence a volume element D: in local coordinates
D =
_
det(g
ij
(x))dx
1
. . . . . dx
n
.
Since g is invariant under I, so also is D, and so it passes to the quotient ID. The
condition is that
_
I\D
D - o.
For example, let D = H
1
and let I = PSL
2
(Z). Then
F = {z c H
1
[ [z[ > 1,
1
2
- +z -
1
2
}
is a fundamental domain for I and
_
I\D
D =
__
F
dxdy
y
2
_
_
o
_
3{2
_
1{2
-1{2
dxdy
y
2
=
_
o
_
3{2
dy
y
2
- o.
On the other hand, the quotient of H
1
by the group of translations z . z n, n c Z, has
innite volume, as does the quotient of H
1
by the trivial group.
A real Lie group G has a left invariant volume element, which is unique up to a positive
constant (cf. Boothby 1975, VI 3.5). A discrete subgroup I of G is said to have nite
covolume if IG has nite volume. For a torsion free discrete subgroup I of Hol(D)
,
an application of Fubinis theorem shows that IHol(D)
H and I
0
G() be as in the denition of arithmetic sub-
group. Because Ker() is compact, is proper (Bourbaki 1989, I 10.3) and, in particular,
closed. Because I
0
is discrete in G(1), there exists an open U G(1)
whose intersec-
tion with I
0
is exactly the kernel of I
0
G(1)
I. Now (G(1)
U) is closed in
H, and its complement intersects I in {1
I
}. Therefore, I is discrete in H. It has nite
covolume because I
0
G(1)
acting freely on
H
1
. Since there are continuous families of such riemann surfaces, this shows that there are
uncountably many discrete cocompact subgroups in PGL
2
(1)
(therefore also in SL
2
(1)),
but there only countably many arithmetic subgroups.
The following (Fields medal) theorem of Margulis shows that SL
2
is exceptional in
this regard: let I be a discrete subgroup of nite covolume in a noncompact simple real
Lie group H; then I is arithmetic unless H is isogenous to SO(1, n) or SU(1, n) (see
Witte 2001, 6.21 for a discussion of the theorem). Note that, because SL
2
(1) is isogenous
to SO(1, 2), the theorem doesnt apply to it.
Brief review of algebraic varieties
Let k be a eld. An afne k-algebra is a nitely generated k-algebra A such that A
k
k
al
is
reduced (i.e., has no nilpotents). Such an algebra is itself reduced, and when k is perfect every
reduced nitely generated k-algebra is afne.
Let A be an afne k-algebra. Dene specm(A) to be the set of maximal ideals in A endowed
with the topology having as basis D( ), D( ) = {m [ , c m}, c A. There is a unique sheaf
of k-algebras O on specm(A) such that O(D( )) = A
(
for all . Here A
(
is the algebra obtained
from A by inverting . Any ringed space isomorphic to a ringed space of the form
Specm(A) = (specm(A), O)
is called an afne variety over k. The stalk at m is the local ring A
m
, and so Specm(A) is a locally
ringed space.
This all becomes much more familiar when k is algebraically closed. When we write A =
k[X
1
, . . . , X
n
],a, the space specm(A) becomes identied with the zero set of a in k
n
endowed with
the zariski topology, and O becomes identied with the sheaf of k-valued functions on specm(A)
locally dened by polynomials.
A topological space V with a sheaf of k-algebras O is a prevariety over k if there exists a nite
covering (U
i
) of V by open subsets such that (U
i
, O[U
i
) is an afne variety over k for all i. A
morphism of prevarieties over k is simply a morphism of ringed spaces of k-algebras. A prevariety
V over k is separated if, for all pairs of morphisms of k-prevarieties , : Z V , the subset of Z
on which and agree is closed. A variety over k is a separated prevariety over k.
36 3 LOCALLY SYMMETRIC VARIETIES
Alternatively, the varieties over k are precisely the ringed spaces obtained from geometrically-
reduced separated schemes of nite type over k by deleting the nonclosed points.
A morphism of algebraic varieties is also called a regular map, and the elements of O(U) are
called the regular functions on U.
For the variety approach to algebraic geometry, see AG, and for the scheme approach, see
Hartshorne 1977.
Algebraic varieties versus complex manifolds
The functor from nonsingular algebraic varieties to complex manifolds
For a nonsingular variety V over C, V (C) has a natural structure as a complex manifold.
More precisely:
PROPOSITION 3.8. There is a unique functor (V, O
V
) . (V
an
, O
V
an ) from nonsingular
varieties over C to complex manifolds with the following properties:
(a) as sets, V = V
an
, every zariski-open subset is open for the complex topology, and
every regular function is holomorphic;
25
(b) if V = A
n
, then V
an
= C
n
with its natural structure as a complex manifold;
(c) if : V W is etale, then
an
: V
an
W
an
is a local isomorphism.
PROOF. A regular map : V W is etale if the map d
p
: T
p
V T
p
W is an isomor-
phism for all p c V . Note that conditions (a,b,c) determine the complex-manifold structure
on any open subvariety of A
n
and also on any variety V that admits an etale map to an open
subvariety of A
n
. Since every nonsingular variety admits a zariski-open covering by such V
(AG, 4.31), this shows that there exists at most one functor satisfying (a,b,c), and suggests
how to dene it.
Obviously, a regular map : V W is determined by
an
: V
an
W
an
, but not every
holomorphic map V
an
W
an
is regular. For example, z . e
z
: C C is not regular.
Moreover, a complex manifold need not arise from a nonsingular algebraic variety, and
two nonsingular varieties V and W can be isomorphic as complex manifolds without being
isomorphic as algebraic varieties (Shafarevich 1994, VIII 3.2). In other words, the functor
V .V
an
is faithful, but it is neither full nor essentially surjective on objects.
REMARK 3.9. The functor V .V
an
can be extended to all algebraic varieties once one has
the notion of a complex manifold with singularities. This is called a complex space. For
holomorphic functions
1
, . . . ,
r
on a connected open subset U of C
n
, let V (
1
, . . . ,
r
)
denote the set of common zeros of the
i
in U; one endows V (
1
, . . . ,
r
) with a natural
structure of ringed space, and then denes a complex space to be a ringed space (S, O
S
)
that is locally isomorphic to one of this form (Shafarevich 1994, VIII 1.5).
25
These conditions require that the identity map V V be a map of ringed spaces (V
an
, O
V
an )
(V, O
V
). This map is universal.
Algebraic varieties versus complex manifolds 37
Necessary conditions for a complex manifold to be algebraic
3.10. Here are two necessary conditions for a complex manifold M to arise from an alge-
braic variety.
(a) It must be possible to embed M as an open submanifold of a compact complex
manfold M
+
in such a way that the boundary M
+
M is a nite union of manifolds
of dimension dimM 1.
(b) If M is compact, then the eld of meromorphic functions on M must have transcen-
dence degree dimM over C.
The necessity of (a) follows from Hironakas theorem on the resolution of singularities,
which shows that every nonsingular variety V can be embedded as an open subvariety of a
complete nonsingular variety V
+
in such a way that the boundary V
+
V is a divisor with
normal crossings (see p40), and the necessity of (b) follows from the fact that, when V is
complete and nonsingular, the eld of meromorphic functions on V
an
coincides with the
eld of rational functions on V (Shafarevich 1994, VIII 3.1).
Here is one positive result: the functor
{projective nonsingular curves over C} {compact riemann surfaces}
is an equivalence of categories (see MF, pp88-91, for a discussion of this theorem). Since
the proper zariski-closed subsets of algebraic curves are the nite subsets, we see that for
riemann surfaces the condition (3.10a) is also sufcient: a riemann surface M is algebraic
if and only if it is possible to embed M in a compact riemann surface M
+
in such a way that
the boundary M
+
M is nite. The maximum modulus principle (Cartan 1963, VI 4.4)
shows that a holomorphic function on a connected compact riemann surface is constant.
Therefore, if a connected riemann surface M is algebraic, then every bounded holomorphic
function on M is constant. We conclude that H
1
does not arise from an algebraic curve,
because the function z .
z-i
zi
is bounded, holomorphic, and nonconstant.
For any lattice in C, the Weierstrass , function and its derivative embed C, into
P
2
(C) (as an elliptic curve). However, for a lattice in C
2
, the eld of meromorphic
functions on C
2
, will usually have transcendence degree - 2, and so C
2
, is not an
algebraic variety.
26
For quotients of C
g
by a lattice , condition (3.10b) is sufcient for
algebraicity (Mumford 1970, p35).
Projective manifolds and varieties
A complex manifold (resp. algebraic variety) is projective if it is isomorphic to a closed
submanifold (resp. closed subvariety) of a projective space. The rst truly satisfying theo-
rem in the subject is the following:
THEOREM 3.11 (CHOW 1949). Every projective complex manifold has a unique structure
of a nonsingular projective algebraic variety, and every holomorphic map of projective
complex manifolds is regular for these structures. (Moreover, a similar statement holds for
complex spaces.)
26
A complex torus C
g
, is algebraic if and only if it admits a riemann form (see 6.7 below). When is
the lattice in C
2
generated by (1, 0), (i, 0), (0, 1), (, ) with nonreal, C
2
,does not admit a riemann form
(Shafarevich 1994, VIII 1.4).
38 3 LOCALLY SYMMETRIC VARIETIES
PROOF. See Shafarevich 1994, VIII 3.1 (for the manifold case).
In other words, the functor V . V
an
is an equivalence from the category of (non-
singular) projective algebraic varieties to the category of projective complex (manifolds)
spaces.
The theorem of Baily and Borel
THEOREM 3.12 (BAILY AND BOREL 1966). Let D(I) = ID be the quotient of a
hermitian symmetric domain by a torsion free arithmetic subgroup I of Hol(D)
. Then
D(I) has a canonical realization as a zariski-open subset of a projective algebraic variety
D(I)
+
. In particular, it has a canonical structure as an algebraic variety.
Recall the proof for D = H
1
. Set H
+
1
= H
1
' P
1
() (rational points on the real axis
plus the point io). Then I acts on H
+
1
, and the quotient IH
+
1
is a compact riemann
surface. One can then show that the modular forms of a sufciently high weight embed
IH
+
1
as a closed submanifold of a projective space. Thus IH
+
1
is algebraic, and as
IH
1
omits only nitely many points of IH
+
1
, it is automatically a zariski-open subset
of IH
+
1
. The proof in the general case is similar, but is much more difcult. Briey,
D(I)
+
= ID
+
where D
+
is the union of D with certain rational boundary components
endowed with the Satake topology; again, the automorphic forms of a sufciently high
weight map ID
+
isomorphically onto a closed subvariety of a projective space, and ID
is a zariski-open subvariety of ID
+
.
For the Siegel upper half space H
g
, the compactication H
+
g
was introduced by Sa-
take (1956) in order to give a geometric foundation to certain results of Siegel (1939), for
example, that the space of holomorphic modular forms on H
g
of a xed weight is nite
dimensional, and that the meromorphic functions on H
g
obtained as the quotient of two
modular forms of the same weight form an algebraic function eld of transcendence de-
gree g(g 1),2 = dimH
g
over C.
That the quotient IH
+
g
of H
+
g
by an arithmetic group I has a projective embedding by
modular forms, and hence is a projective variety, was proved in Baily 1958, Cartan 1958,
and Satake and Cartan 1958.
The construction of H
+
g
depends on the existence of fundamental domains for the arith-
metic group I acting on H
g
. Weil (1958) used reduction theory to construct fundamental
sets (a notion weaker than fundamental domain) for the domains associated with certain
classical groups (groups of automorphisms of semsimple -algebras with, or without, in-
volution), and Satake (1960) applied this to construct compactications of these domains.
Borel and Harish-Chandra developed a reduction theory for general semisimple groups
(Borel and Harish-Chandra 1962; Borel 1962), which then enabled Baily and Borel (1966)
to obtain the above theorem in complete generality.
The only source for the proof is the original paper, although some simplications to the
proof are known.
27
27
For a discussion of later work, see Casselman 1997.
The theorem of Borel 39
REMARK 3.13. (a) The variety D(I)
+
is usually very singular. The boundary D(I)
+
D(I) has codimension _ 2, provided PGL
2
is not a quotient of the -group G giving rise
to I.
(b) The variety D(I)
+
= Proj(
n_0
A
n
) where A
n
is the vector space of automorphic
forms for the n
th
power of the canonical automorphy factor (Baily and Borel 1966, 10.11).
It follows that, if PGL
2
is not a quotient of G, then D(I)
+
= Proj(
n_0
H
0
(D(I), o
n
))
where o is the sheaf of algebraic differentials of maximum degree on D(I). Without
the condition on G, there is a similar description of D(I)
+
in terms of differentials with
logarithmic poles (Brylinski 1983, 4.1.4; Mumford 1977).
(b) When D(I) is compact, Theorem 3.12 follows from the Kodaira embedding the-
orem (Wells 1980, VI 4.1, 1.5). Nadel and Tsuji (1988, 3.1) extended this to those D(I)
having boundary of dimension 0, and Mok and Zhong (1989) give an alternative prove of
Theorem 3.12, but without the information on the boundary given by the original proof.
An algebraic variety D(I) arising as in the theorem is called a locally symmetric va-
riety (or an arithmetic locally symmetric variety, or an arithmetic variety, but not yet a
Shimura variety).
The theorem of Borel
THEOREM 3.14 (BOREL 1972). Let D(I) and D(I)
+
be as in (3.12) in particular, I is
torsion free and arithmetic. Let V be a nonsingular quasi-projective variety over C. Then
every holomorphic map : V
an
D(I)
an
is regular.
The key step in Borels proof is the following result:
LEMMA 3.15. Let D
1
be the punctured disk {z [ 0 - [z[ - 1}. Then every holomorphic
map
28
D
r
1
D
s
1
D(I) extends to a holomorphic map D
rs
1
D(I)
+
(of complex
spaces).
The original result of this kind is the big Picard theorem, which, interestingly, was rst
proved using elliptic modular functions. Recall that the theorem says that if a function
has an essential singularity at a point p c C, then on any open disk containing p,
takes every complex value except possibly one. Therefore, if a holomorphic function
on D
1
omits two values in C, then it has at worst a pole at 0, and so extends to a
holomorphic function D
1
P
1
(C). This can be restated as follows: every holomorphic
function from D
1
to P
1
(C) {3 points} extends to a holomorphic function from D
1
to the
natural compactication P
1
(C) of P
1
(C){3 points}. Over the decades, there were various
improvements made to this theorem. For example, Kwack (1969) replaced P
1
(C) {3
points} with a more general class of spaces. Borel (1972) veried that Kwacks theorem
applies to D(I) D(I)
+
, and extended the result to maps from a product D
r
1
D
s
1
.
Using the lemma, we can prove the theorem. According Hironakas (Fields medal)
theorem on the resolution of singularities (Hironaka 1964; see also Bravo et al. 2002), we
can realize V as an open subvariety of a projective nonsingular variety V
+
in such a way
28
Recall that D
1
is the open unit disk. The product D
r
1
D
s
1
is obtained from D
rs
1
by removing the rst
r coordinate hyperplanes.
40 3 LOCALLY SYMMETRIC VARIETIES
that V
+
V is a divisor with normal crossings. This means that, locally for the complex
topology, the inclusion V V
+
is of the formD
r
1
D
s
1
D
rs
1
. Therefore, the lemma
shows that : V
an
D(I)
an
extends to a holomorphic map V
+an
D(I)
+
, which is
regular by Chows theorem (3.11).
COROLLARY 3.16. The structure of an algebraic variety on D(I) is unique.
PROOF. Let D(I) denote ID with the canonical algebraic structure provided by The-
orem 3.12, and suppose ID = V
an
for a second variety V . Then the identity map
: V
an
D(I) is a regular bijective map of nonsingular varieties, and is therefore an
isomorphism (cf. AG 3.19).
The proof of the theorem shows that the compactication D(I) D(I)
+
has the
following property: for any compactication D(I) D(I)
|
with D(I)
|
D(I) a
divisor with normal crossings, there is a unique regular map D(I)
|
D(I)
+
making
D(I)
|
D(I)
@
@
@R
D(I)
+
?
commute. For this reason, D(I) D(I)
+
is often called the minimal compactication.
Other names: standard, Satake-Baily-Borel, Baily-Borel.
ASIDE 3.17. (a) Theorem 3.14 also holds for singular V in fact, it sufces to show that
becomes regular when restricted to an open dense set of V , which we may take to be the
complement of the singular locus.
(b) Theorem 3.14 denitely fails without the condition that I be torsion free. For
example, it is false for IH
1
= A
1
consider z .e
z
: C C.
Finiteness of the group of automorphisms of D(I)
DEFINITION 3.18. A semisimple group G over is said to be of compact type if G(1) is
compact, and it is of noncompact type if it does not contain a nonzero normal subgroup of
compact type.
A semisimple group over is an almost direct product of its minimal connected normal
subgroups, and it will be of noncompact type if and only if none of these subgroups is of
compact type. In particular, a simply connected or adjoint group is of noncompact type if
and only if it has no simple factor of compact type.
We shall need one last result about arithmetic subgroups.
THEOREM 3.19 (BOREL DENSITY THEOREM). Let G be a semisimple group over of
noncompact type. Then every arithmetic subgroup I of G() is zariski-dense in G.
Finiteness of the group of automorphisms of D(I) 41
PROOF. Borel 1969, 15.12, or Platonov and Rapinchuk 1994, Theorem 4.10, p205.
COROLLARY 3.20. For G as in (3.19), the centralizer of I in G(1) is Z(1), where Z is
the centre of G (as an algebraic group over ).
PROOF. The theorem implies that the centralizer of I in G(C) is Z(C), and Z(1) =
Z(C) G(1).
THEOREM 3.21. Let D(I) be the quotient of a hermitian symmetric domain D by a torsion
free arithmetic group I. Then D(I) has only nitely many automorphisms.
PROOF. As I is a torsion free, D is the universal covering space of ID and I is the
group of covering transformations (see p32). An automorphism : ID ID lifts to
an automorphism : D D. For any ; c I, ;
-1
is a covering transformation, and so
lies in I. Conversely, an automorphism of D normalizing I denes an automorphism of
ID. Thus,
Aut(ID) = N, I, N = normalizer of I in Aut(D).
The corollary implies that the map ad: N Aut(I) is injective, and so N is countable.
Because I is closed in Aut(D), so also is N. Write N as a countable union of its nite
subsets. According to the Baire category theorem (MF 1.3) one of the nite sets must have
an interior point, and this implies that N is discrete. Because I Aut(D) has nite volume
(3.3a), this implies that I has nite index in N.
Alternatively, there is a geometric proof, at least when I is neat. According to Mumford
1977, Proposition 4.2, D(I) is then an algebraic variety of logarithmic general type, which
implies that its automorphism group is nite (Iitaka 1982, 11.12).
ASIDE 3.22. In most of this section we have considered only quotients ID with I torsion
free. In particular, we disallowed I(1)H
1
. Typically, if I has torsion, then ID will be
singular and some of the above statements will fail for ID.
NOTES. Borel 1969, Raghunathan 1972, and (eventually) Witte 2001 contain good ex-
positions on discrete subgroups of Lie groups. There is a large literature on the various
compactications of locally symmetric varieties. For overviews, see Satake 2001 and
Goresky 2003, and for a detailed description of the construction of toroidal compactica-
tions, which, in contrast to the Baily-Borel compactication, may be smooth and projective,
see Ash et al. 1975.
42 4 CONNECTED SHIMURA VARIETIES
4 Connected Shimura varieties
Congruence subgroups
Let G be a reductive algebraic group over . Choose an embedding G GL
n
, and dene
I(N) = G() {g c GL
n
(Z) [ g I
n
mod N}.
For example, if G = SL
2
, then
I(N) =
_
a b
c d
_
c SL
2
(Z) [ ad bc = 1, a, d 1, b, c 0 mod N
_
.
A congruence subgroup of G() is any subgroup containing some I(N) as a subgroup
of nite index. Although I(N) depends on the choice the embedding, this denition does
not (see 4.1 below).
With this terminology, a subgroup of G() is arithmetic if it is commensurable with
I(1). The classical congruence subgroup problem for G asks whether every arithmetic
subgroup of G() is congruence, i.e., contains some I(N). For split simply connected
groups other than SL
2
, the answer is yes (Matsumoto 1969), but
29
SL
2
and all nonsimply
connected groups have many noncongruence arithmetic subgroups (for a discussion of the
problem, see Platonov and Rapinchuk 1994, section 9.5). In contrast to arithmetic sub-
groups, the image of a congruence subgroup under an isogeny of algebraic groups need not
be a congruence subgroup.
30
The ring of nite ad` eles is the restricted topological product
A
(
=
(
I
: Z
I
)
29
That SL
2
(Z) has noncongruence arithmetic subgroups was rst noted in Klein 1880. For a proof that
SL
2
(Z) has innitely many subgroups of nite index that are not congruence subgroups see Sury 2003, 3-4.1.
The proof proceeds by showing that the groups occurring as quotients of SL
2
(Z) by congruence subgroups
(especially by principal congruence subgroups) are of a rather special type, and then exploits the known
structure of SL
2
(Z) as an abstract group to construct many nite quotients not of his type.
30
Let G be a semisimple group over . The arithmetic and congruence subgroups of G() dene topolo-
gies on it, and we denote the corresponding completions by
G and G. Because every arithmetic group is
congruence, the identity map on G() gives a surjective homomorphism
G G, whose kernel C(G) is
called the congruence kernel. This kernel is trivial if and only if all arithmetic subgroups are congruence.
The modern congruence subgroup problem is to compute C(G). For example, the group C(SL
2
) is innite.
Now let G be simply connected, and let G
t
= G,N where N is a nontrivial subgroup of Z(G). Consider
the diagram:
1 C(G)
G G 1
_
1 C(G
t
)
G
t
G
t
1.
It follows from the strong approximation theorem (4.16) that G = G(A
(
), and it follows from (3.2) that the
kernel of is N(), which is nite. On the other hand, the kernel of is N(A
(
), which is innite. Because
Ker() ,= N(), : G() G
t
() doesnt map congruence subgroups to congruence subgroups, and
because C(G
t
) contains a subgroup isomorphic to N(A
(
),N(), G
t
() contains a noncongruence arithmetic
subgroup. See Serre 1967 for more details.
Congruence subgroups 43
where runs over the nite primes of (that is, we omit the factor 1). Thus, A
(
is the
subring of
I
consisting of the (a
I
) such that a
I
c Z
I
for almost all , and it is endowed
with the topology for which
Z
I
is open and has the product topology.
Let V = SpecmA be an afne variety over . The set of points of V with coordinates
in a -algebra R is
V (R) = Hom
Q
(A, R).
When we write
A = [X
1
, . . . , X
m
],a = [x
1
, . . . , x
m
],
the map P .(P (x
1
) , . . . , P(x
m
)) identies V (R) with
{(a
1
, . . . , a
m
) c R
m
[ (a
1
, . . . , a
m
) = 0, \ c a}.
Let Z[x
1
, . . . , x
m
] be the Z-subalgebra of A generated by the x
i
, and let
V (Z
I
) = Hom
Z
(Z[x
1
, . . . , x
m
], Z
I
) = V (
I
) Z
m
I
(inside
m
I
).
This set depends on the choice of the generators x
i
for A, but if A = [y
1
, . . . , y
n
], then
the y
i
s can be expressed as polynomials in the x
i
with coefcients in , and vice versa.
For some d c Z, the coefcients of these polynomials lie in Z[
1
d
], and so
Z[
1
d
][x
1
, . . . , x
m
] = Z[
1
d
][y
1
, . . . , y
n
] (inside A).
It follows that for [ d, the y
i
s give the same set V (Z
I
) as the x
i
s. Therefore,
V (A
(
) =
(V (
I
): V (Z
I
))
is independent of the choice of generators for
31
A.
For an algebraic group G over , we dene
G(A
(
) =
(G(
I
): G(Z
I
))
similarly. For example,
G
m
(A
(
) =
I
: Z
I
) = A
(
.
PROPOSITION 4.1. For any compact open subgroup K of G(A
(
), K G() is a congru-
ence subgroup of G(), and every congruence subgroup arises in this way.
32
PROOF. Fix an embedding G GL
n
. From this we get a surjection [GL
n
] [G] (of
-algebras of regular functions), i.e., a surjection
[X
11
, . . . , X
nn
, T],(det(X
ij
)T 1) [G],
31
In a more geometric language, let : V A
m
Q
be a closed immersion. The Zariski closure V
of V in
A
m
Z
is a model of V at over Spec Z. A different closed immersion gives a different at model V
, but for
some d, the isomorphism (V
)
Q
= V
= (V
)
Q
on generic bres extends to an isomorphism V
over
Spec Z[
1
d
]. For the primes not dividing d, the subgroups V
(Z
I
) and V
(Z
I
) of V (
I
) will coincide.
32
To dene a basic compact open subgroup K of G(A
(
), one has to impose a congruence condition at
each of a nite set of primes. Then I = G() K is obtained from G(Z) by imposing the same congruence
conditions. One can think of I as being the congruence subgroup dened by the congruence condition K.
44 4 CONNECTED SHIMURA VARIETIES
and hence [G] = [x
11
, . . . , x
nn
, t ]. For this presentation of [G],
G(Z
I
) = G(
I
) GL
n
(Z
I
) (inside GL
n
(
I
)).
For an integer N > 0, let
K(N) =
I
K
I
, where K
I
=
_
G(Z
I
) if [ N
{g c G(Z
I
) [ g I
n
mod
r
`
} if r
I
= ord
I
(N).
Then K(N) is a compact open subgroup of G(A
(
), and
K(N) G() = I(N).
It follows that the compact open subgroups of G(A
(
) containing K(N) intersect G()
exactly in the congruence subgroups of G() containing I(N). Since every compact open
subgroup of G(A
(
) contains K(N) for some N, this completes the proof.
REMARK 4.2. There is a topology on G() for which the congruence subgroups form a
fundamental system of neighbourhoods. The proposition shows that this topology coin-
cides with that dened by the diagonal embedding G() G(A
(
).
EXERCISE 4.3. Show that the image in PGL
2
() of a congruence subgroup in SL
2
()
need not be congruence.
Connected Shimura data
DEFINITION 4.4. A connected Shimura datum is a pair (G, D) consisting of a semisimple
algebraic group G over and a G
ad
(1)
Hol(D)
=
Hol(D
t
i
)
and D
t
i
is in natural one-to-one correspondence with the set D
i
of H
i
(1)
-
conjugates of u
i
(see 1.21). The product D
t
of the D
t
i
is a hermitian symmetric domain
on which G(1)
Hol(D)
with compact
kernel. Moreover, there is a natural identication of D
t
=
D
t
i
with D =
D
i
.
Conversely, let (G, D, G(1)
Hol(D)
on D denes an isomorphism H
nc
(1)
=
Hol(D)
, and {u
p
[ p c D} is an H
nc
(1)
(1, u
p
): U
1
H
c
(1) H
nc
(1) [ p c D
_
,
is a G
ad
(1)
.
PROOF. The argument in the proof of (1.5) shows that X is a disjoint union of orbits
G
ad
(1)
= Hol(D),
and G(1)
. As H
c
(1) is connected (Borel 1991,
p277), the last part of the proposition follows from (1.7).
46 4 CONNECTED SHIMURA VARIETIES
Denition of a connected Shimura variety
Let (G, D) be a connected Shimura datum, and regard D as a hermitian symmetric domain
with G(1)
Aut(D)
in Aut(D)
is a congruence subgroup.
REMARK 4.11. An element g of G
ad
()
G
ad
()
are equivalent:
(a)
-1
(I) is a congruence subgroup of G()
;
(b)
-1
(I) contains a congruence subgroup of G()
;
(c) I contains the image of a congruence subgroup of G()
.
Therefore, the varieties ID with I a congruence subgroup of G()
(G, D).
PROOF. (a) ==(b). Obvious.
(b) ==(c). Let I
t
be a congruence subgroup of G()
contained in
-1
(I). Then
I (
-1
(I)) (I
t
).
(c) == (a). Let I
t
be a congruence subgroup of G()
such that I (I
t
), and
consider
-1
(I)
-1
(I
t
) I
t
.
Because (I
t
) is arithmetic (3.2), it is of nite index in I, and it follows that
-1
(I
t
) is
of nite index in
-1
(I). Because Z() I
t
-1
(I
t
) and Z() is nite (Z is the
centre of G), I
t
is of nite index in
-1
(I
t
). Therefore, I
t
is of nite index in
-1
(I),
which proves that
-1
(I) is congruence.
The strong approximation theorem 47
REMARK 4.13. The homomorphism : G()
G
ad
()
(G, D).
EXAMPLE 4.14. (a) G = SL
2
, D = H
1
. Then Sh
:F-R
B
F,
1
and each B
F,
1 is isomorphic either to the usual quaternions H or to M
2
(1). Let G be
the semisimple algebraic group over such that
G() = Ker(Nm: B
).
Then
G(1) - H
1
H
1
SL
2
(1) SL
2
(1) (27)
where H
1
= Ker(Nm: H
is not dense in
33
A
(
.
(b) PGL
2
: the determinant denes surjections
PGL
2
()
,
2
PGL
2
(A
(
) A
(
,A
2
(
and
,
2
is not dense in A
(
,A
2
(
.
(c) G of compact type: because G(Z) is discrete in G(1) (see 3.3), it is nite, and so it
is not dense in G(
_
ID
[x]|-[x,1]
G()D G(A
(
),K.
As K is open, G(A
(
),K is discrete, and so the upper map is a homeomorphism of D onto
its image, which is open. It follows easily that the lower map is a homeomorphism.
33
Let (a
I
)
I
c
Z
I
A
(
and let S be a nite set. If
that is
close to a
I
for c S and an -adic unit for , c S. But such an a is an -adic unit for all , and so equals 1.
This yields a contradiction.
An ad` elic description of D(I) 49
What happens when we pass to the inverse limit over I? The obvious map
D lim
ID,
is injective because each I acts freely on D and
_
I = {1}. Is the map surjective? The
example
Z lim
Z,mZ =
Z
is not encouraging it suggests that lim
K
G()D G(A
(
),K = G()D G(A
(
) (29)
(ad` elic topology on G(A
(
)).
Before proving this, we need a lemma.
LEMMA 4.20. Let G be a topological group acting continuously on a topological space
X, and let (G
i
)
i=I
be a directed family of subgroups of G. The canonical map X,
_
G
i
lim
X,G
i
is injective if the G
i
are compact, and it is surjective if in addition the orbits of
the G
i
in X are separated.
PROOF. We shall use that a directed intersection of nonempty compact sets is nonempty,
which has the consequence that a directed inverse limit of nonempty compact sets is nonempty.
Assume that each G
i
is compact, and let x, x
t
c X. For each i, let
G
i
(x, x
t
) = {g c G
i
[ xg = x
t
}.
If x and x
t
have the same image in lim
X,G
i
, then the G
i
(x, x
t
) are all nonempty. Since
each is compact, their intersection is nonempty. For any g in the intersection, xg = x
t
,
which shows that x and x
t
have the same image in X,
_
G
i
.
Now assume that each orbit is separated and hence compact. For any (x
i
G
i
)
i=I
c
lim
X,G
i
, lim
x
i
G
i
is nonempty. If x c lim
x
i
G
i
, then x
_
G
i
maps to (x
i
G
i
)
i=I
.
PROOF OF 4.19. Let (x, a) c D G(A
(
), and let K be a compact open subgroup of
G(A
(
). In order to be able to apply the lemma, we have to show that the image of the orbit
(x, a)K in G()DG(A
(
) is separated for K sufciently small. Let I = G()aKa
-1
we may assume that I is torsion free (3.5). There exists an open neighbourhood V of x
such that gV V = 0 for all g c I {1} (see the proof of 3.1). For any (x, b) c (x, a)K,
g(V aK) (V bK) = 0 for all
34
g c G() {1}, and so the images of V Ka and
V Kb in G()D G(A
(
) separate (x, a) and (x, b).
34
Let g c G(), and suppose that g(V aK) (V bK) ,= 0. Then
gaK = bK = aK
and so g c G() aKa
-1
= I. As gV V ,= 0, this implies that g = 1.
50 4 CONNECTED SHIMURA VARIETIES
ASIDE 4.21. (a) Why replace the single coset space on the left of (28) with the more
complicated double coset space on the right? One reason is that it makes transparent that
(in this case) there is an action of G(A
(
) on the inverse system (ID)
I
, and hence, for
example, on
lim
H
i
(ID, ).
Another reason will be seen presently we use double cosets to dene Shimura varieties.
Double coset spaces are pervasive in work on the Langlands program.
(b) The inverse limit of the D(I) exists as a scheme it is even locally noetherian and
regular (cf. 5.30 below).
Alternative denition of connected Shimura data
Recall that S is the real torus such that S(1) = C
r|-r
1
C
z|-z{z
U
1
0
arises from an exact sequence of tori
0 G
m
u
S U
1
0.
Let H be a semisimple real algebraic group with trivial centre. A homomorphism u: U
1
H denes a homomorphism h: S H by the rule h(z) = u(z,z), and U
1
will act on
Lie(H)
C
through the characters z, 1, z
-1
if and only if S acts on Lie(H)
C
through the
characters z,z, 1, z,z. Conversely, let h be a homomorphism S H for which S acts
on Lie(H)
C
through the characters z,z, 1, z,z. Then n(G
m
) acts trivially on Lie(H)
C
,
which implies that h is trivial on n(G
m
) because the adjoint representation H Lie(H)
is faithful. Thus, h arises from a u.
Now let G be a semisimple algebraic group over . From the above remark, we see that
to give a G
ad
(1)
-conjugacy class X
of homomorphisms h: S G
ad
R
satisfying the following conditions:
SV1: for h c X
) consisting of a semisim-
ple algebraic group over and a G
ad
(1)
= G()G(1)
. For example, GL
2
()
V
k,o
k
al
where o runs through the k
0
-embeddings of k into k
al
.
37
There is a natural action of G
ad
on G for which ad(g) (g c G(k)) acts as x . gxg
-1
. This explains
why we denote this last map by ad(g). The adjoint representation Ad: G Lie(G) denes an isomorphism
of G,Z onto Ad(G), which explains why we denote G,Z by G
ad
and call it the adjoint group of G. Finally,
G itself is called an adjoint group if G = G
ad
.
52 5 SHIMURA VARIETIES
Here T (a torus) is the largest commutative quotient of G, and Z
t
=
df
Z G
der
(a nite
algebraic group) is the centre of G
der
.
The real points of algebraic groups
PROPOSITION 5.1. For a surjective homomorphism : G H of algebraic groups over
1, G(1)
H(1)
is surjective.
PROOF. The map (1): G(1)
H(1)
.
Note that G(1) H(1) need not be surjective. For example, G
m
x|-x
n
G
m
is
surjective as a map of algebraic groups, but the image of G
m
(1)
n
G
m
(1) is G
m
(1)
or
G
m
(1) according as n is even or odd. Also SL
2
PGL
2
is surjective, but the image of
SL
2
(1) PGL
2
(1) is PGL
2
(1)
.
For a simply connected algebraic group G, G(C) is simply connected as a topological
space, but G(1) need not be. For example, SL
2
(1) is not simply connected.
THEOREM 5.2 (CARTAN 1927). For a simply connected group G over 1, G(1) is con-
nected.
PROOF. See Platonov and Rapinchuk 1994, Theorem 7.6, p407.
COROLLARY 5.3. For a reductive group G over 1, G(1) has only nitely many connected
components (for the real topology).
38
PROOF. Because of (5.1), an exact sequence of real algebraic groups
1 N G
t
G 1 (30)
with N Z(G
t
) gives rise to an exact sequence
0
(G
t
(1))
0
(G(1)) H
1
(1, N).
Let
G be the universal covering group of G
der
. As G is an almost direct product of Z =
Z(G) and G
der
, there is an exact sequence (30) with G
t
= Z
G and N nite. Now
0
(
G(1)) = 0 because
G is simply connected,
0
(Z(1)) is nite because Z
)
d
has 2
d
connected components, and each of PGL
2
(1)
and GL
2
(1) has 2 connected components.
38
This also follows from the theorem of Whitney 1957: for an algebraic variety V over 1, V (1) has only
nitely many connected components (for the real topology) see Platonov and Rapinchuk 1994, Theorem
3.6, p119.
Shimura data 53
THEOREM 5.4 (REAL APPROXIMATION). For any connected algebraic group G over ,
G() is dense in G(1).
PROOF. See Platonov and Rapinchuk 1994, Theorem 7.7, p415.
39
Shimura data
DEFINITION 5.5. AShimura datumis a pair (G, X) consisting of a reductive group G over
and a G(1)-conjugacy class X of homomorphisms h: S G
R
satisfying the conditions
SV1, SV2, and SV3 (see p50).
Note that, in contrast to a connected Shimura datum, G is reductive (not semisimple),
the homomorphisms h have target G
R
(not G
ad
R
), and X is the full G(1)-conjugacy class
(not a connected component).
EXAMPLE 5.6. Let G = GL
2
(over ) and let X be the set of GL
2
(1)-conjugates of the
homomorphism h
o
: S GL
2R
, h
o
(a ib) =
_
a b
-b a
_
. Then (G, X) is a Shimura datum.
Note that there is a natural bijection X C 1, namely, h
o
. i and gh
o
g
-1
. gi.
More intrinsically, h - z if and only if h(C
) is the stabilizer of z in GL
2
(1) and h(z)
acts on the tangent space at z as multiplication by z,z (rather than z,z).
PROPOSITION 5.7. Let G be a reductive group over 1. For a homomorphism h: S
G, let h be the composite of h with G G
ad
. Let X be a G(1)-conjugacy class of
homomorphisms S G, and let X be the G
ad
(1)-conjugacy class of homomorphisms
S G
ad
containing the h for h c X.
(a) The map h . h: X X is injective and its image is a union of connected compo-
nents of X.
(b) Let X
in G(1) is G(1)
(i.e., gX
= X
== g c G(1)
).
PROOF. (a) A homomorphism h: S G is determined by its projections to T and G
ad
,
because any other homomorphism with the same projections will be of the form he for
some regular map e: S Z
t
and e is trivial because S is connected and Z
t
is nite. The
elements of X all have the same projection to T , because T is commutative, which proves
that h . h: X X is injective. For the second part of the statement, use that G
ad
(1)
G
ad
(1)
is
surjective.
(b) The rst assertion is obvious. In (a) we showed that
0
(X)
0
(X). The stabilizer
in G
ad
(1) of [X
] is G
ad
(1)
in G(1).
COROLLARY 5.8. Let (G, X) be a Shimura datum, and let X
be a connected component
of X regarded as a G(1)
G
ad
(1),G
ad
(1)
are injective, and the second can be identied with the rst once an
h c X has been chosen. In general, the maps will not be surjective unless H
1
(1, Z) = 0.
Shimura varieties
Let (G, X) be a Shimura datum.
LEMMA 5.11. For any connected component X
G(A
(
) G()X G(A
(
)
is a bijection.
PROOF. Because G() is dense in G(1) (see 5.4) and G(1) acts transitively on X, every
x c X is of the form qx
c X
G(A
(
). If [x, a] = [x
t
, a
t
] in G()X
G(A
(
), then
x
t
= qx, a
t
= qa, some q c G().
Because x and x
t
are both in X
, q stabilizes X
X G(A
(
).
Shimura varieties 55
LEMMA 5.12. For any open subgroup K of G(A
(
), the set G()
G(A
(
),K is nite.
PROOF. Since G()
G() G
ad
(1)
G
ad
(1) is injective and the second group is
nite (5.3), it sufces to showthat G()G(A
(
),K is nite. Later (Theorem5.17) we shall
show that this follows from the strong approximation theorem if G
der
is simply connected,
and the general case is not much more difcult.
For K a compact open subgroup K of G(A
(
), consider the double coset space
Sh
K
(G, X) = G()X G(A
(
),K
in which G() acts on X and G(A
(
) on the left, and K acts on G(A
(
) on the right:
q(x, a)k = (qx, qak), q c G(), x c X, a c G(A
(
), k c K.
LEMMA 5.13. Let C be a set of representatives for the double coset space G()
G(A
(
),K,
and let X
g=C
I
g
X
where I
g
is the subgroup gKg
-1
G()
of G()
G()
G(A
(
),K
is injective,
40
and that G()
G(A
(
),K is the disjoint union of the images of these
maps.
41
Thus, the rst statement follows from (5.11). The second statement can be proved
in the same way as the similar statement in (4.18).
Because I
g
is a congruence subgroup of G(), its image in G
ad
() is arithmetic (3.2),
and so (by denition) its image in Aut(X
will also
be neat and hence torsion free. Then I
g
X
, g c C, k c K. Now [x, a] = [q
-1
x, g],
which lies in the image of I
g
X
. Suppose [x, g] = [x
t
, g
t
], g
t
, g c C. Then x
t
= qx and g
t
= qgk for
some q c G()
>0
.
THEOREM 5.17. Assume G
der
is simply connected. For K sufciently small, the natural
map
G()X G(A
(
),K T()
|
T(A
(
),v(K)
42
Or dened by or associated with, but not associated to.
43
The Shimura varieties with simply connected derived group are the most important if one knows
everything about them, then one knows everything about all Shimura varieties (because the remainder are
quotients of them). However, there are naturally occurring Shimura varieties for which G
der
is not simply
connected, and so we should not ignore them.
The structure of a Shimura variety 57
denes an isomorphism
0
(Sh
K
(G, X))
= T()
|
T(A
(
),v(K).
Moreover, T()
|
T(A
(
),v(K) is nite, and the connected component over [1] is canon-
ically isomorphic IX
) T()
|
. The natural map in the
theorem is
G()X G(A
(
),K
5.11
= G()
G(A
(
),K
[x,g]|-[(g)]
T()
|
T(A
(
),v(K).
The theorem gives a diagram
G()X G(A
(
),K
IX
,
T()
|
T(A
(
),v(K)
?
[1]
?
in which T()
|
T(A
(
),v(K) is nite and discrete, the left hand map is continuous and
onto with connected bres, and IX
= G
der
(1) Z(1).
PROOF. Because G
der
is simply connected, G
der
(1) is connected (5.2) and so G
der
(1)
G(1)
. Hence G(1)
G
der
(1) Z(1). For the converse, we use the exact commutative
diagram:
1 Z
t
(1)
z|-(z
1
,z)
Z(1) G
der
(1)
(z,g)|-zg
G(1) H
1
(1, Z
t
)
_
_
_
_
(z,g)|-g
_
_
_
_
1 Z
t
(1) G
der
(1) G
ad
(1) H
1
(1, Z
t
).
As G
der
G
ad
is surjective, so also is G
der
(1) G
ad
(1)
== g .0 in H
1
(1, Z
t
)
== g lifts to Z(1) G
der
(1)
== g c Z(1) G
der
(1)
LEMMA 5.19. Let H be a simply connected semisimple algebraic group H over .
(a) For every nite prime, the group H
1
(
I
, H) = 0.
(b) The map H
1
(, H)
l_o
H
1
(
l
, H) is injective (Hasse principle).
PROOF. (a) See Platonov and Rapinchuk 1994, Theorem 6.4, p284.
(b) See ibid., Theorem 6.6, p286.
58 5 SHIMURA VARIETIES
Both statements fail for groups that are not simply connected.
LEMMA 5.20. Assume G
der
is simply connected, and let t c T(). Then t c T()
|
if and
only if t lifts to an element of G()
.
PROOF. Lemma 5.19 implies that the vertical arrow at right in the following diagram is
injective:
44
1 G
der
() G()
T() H
1
(, G
der
)
_
injective
1 G
der
(1) G(1)
T(1) H
1
(1, G
der
)
Let t c T()
|
. By denition, the image t
R
of t in T(1) lifts to an element z c Z(1)
G(1). From the diagram, we see that this implies that t maps to the trivial element in
H
1
(, G
der
) and so it lifts to an element g c G(). Now g
R
z
-1
.t
R
t
-1
R
= 1 in T(1),
and so g
R
c G
der
(1) z G
der
(1) Z(1) G(1)
. Therefore, g c G()
.
Let t be an element of T() lifting to an element a of G()
. According to 5.18,
a
R
= gz for some g c G
der
(1) and z c Z(1). Now a
R
and z map to the same element in
T(1), namely, to t
R
, and so t c T()
|
The lemma allows us to write
T()
|
T(A
(
),v(K) = v(G()
)T(A
(
),v(K).
We now study the bre over [1] of the map
G()
G(A
(
),K
[x,g]|-[(g)]
v(G()
)T(A
(
),v(K).
Let g c G(A
(
). If [v(g)] = [1]
K
, then v(g) = v(q)v(k) some q c G()
and k c K. It
follows that v(q
-1
gk
-1
) = 1, that q
-1
gk
-1
c G
der
(A
(
), and that g c G()
G
der
(A
(
)
K. Hence every element of the bre over [1] is represented by an element (x, a) with
a c G
der
(A
(
). But, according to the strong approximation theorem (4.16), G
der
(A
(
) =
G
der
() (K G
der
(A
(
)), and so the bre over [1] is a quotient of X
; in particular,
it is connected. More precisely, it equals IX
in G
ad
()
(G
der
, X
) = (IX
).
The study of the bre over [t ] will be similar once we show that there exists an a c
G(A
(
) mapping to t (so that the bre is nonempty). This follows from the next lemma.
44
The group H
1
(, G
der
) is dened to be the set of continuous crossed homomorphisms Gal(
al
,)
G
der
(
al
) modulo the relation which identies two crossed homomorphisms that differ by a principal crossed
homomorphism. It is a set with a distinguished element e, represented by any principal crossed homomor-
phism. An element of T() lifts to an element of G() if and only if it maps to the distinguished class in
H
1
(, G
der
).
The structure of a Shimura variety 59
LEMMA 5.21. Assume G
der
is simply connected. Then the map v: G(A
(
) T(A
(
) is
surjective and sends compact open subgroups to compact open subgroups.
PROOF. We have to show:
(a) the homomorphism v: G(
I
) T(
I
) is surjective for all nite ;
(b) the homomorphism v: G(Z
I
) T(Z
I
) is surjective for almost all .
(a) For each prime , there is an exact sequence
1 G
der
(
I
) G(
I
)
T(
I
) H
1
(
I
, G
der
)
and so (5.19a) shows that v: G(
I
) T(
I
) is surjective.
(b) Extend the homomorphism G T to a homomorphism of group schemes G T
over Z[
1
N
] for some integer N. After N has been enlarged, this map will be a smooth
morphism of group schemes and its kernel G
t
will have nonsingular connected bres. On
extending the base ring to Z
I
, [ N, we obtain an exact sequence
0 G
t
I
G
I
T
I
0
of group schemes over Z
I
such that v is smooth and (G
t
I
)
F
`
is nonsingular and connected.
Let P c T
I
(Z
I
), and let Y = v
-1
(P) G
I
. We have to show that Y(Z
I
) is nonempty. By
Langs lemma (Springer 1998, 4.4.17), H
1
(F
I
, (G
t
I
)
F
`
) = 0, and so
v: G
I
(F
I
) T
I
(F
I
)
is surjective. Therefore Y(F
I
) is nonempty. Because Y is smooth over Z
I
, an argument as
in the proof of Newtons lemma (e.g., ANT 7.22) now shows that a point Q
0
c Y(F
I
) lifts
to a point Q c Y(Z
I
).
It remains to show that T()
|
T(A
(
),v(K) is nite. Because T()
|
has nite index
in T(), it sufces to prove that T()T(A
(
),v(K) is nite. But v(K) is open, and so
this follows from the next lemma.
LEMMA 5.22. For any torus T over , T()T(A
(
) is compact.
PROOF. Consider rst the case T = G
m
. Then
T(A
(
),T(
Z) = A
(
,
I nite
I
,Z
I
ord
`
~
=
I nite
Z,
which is the group of fractional ideals of Z. Therefore,
(
,
(
=
. Hence
(
is a quotient of
Z
, which is compact.
For a number eld F, the same argument using the niteness of the class number of F
shows that F
F,(
is compact. Here A
F,(
=
nite
(F
: O
).
An arbitrary torus T over will split over some number eld, say, T
F
- G
dim(T)
m
. Then
T(F)T(A
F,(
) - (F
F,(
)
dim(T)
, which is compact, and T()T(A
(
) is a closed
subset of it.
60 5 SHIMURA VARIETIES
REMARK 5.23. One may ask whether the bre over [1] equals
IX
= G
der
()X
G
der
(A
(
),K G
der
(A
(
), I = K G
der
(),
rather than quotient of X
(G
der
, X
), we include quotients IX
in which
I is an arithmetic subgroup of G
ad
()
T(1) denes a Shimura variety Sh(T, {h}) in this case the conditions SV1,2,3 are vac-
uous. For any compact open K T(A
(
),
Sh
K
(T, {h}) = T(){h} T(A
(
),K
= T()T(A
(
),K
(nite discrete set). We should extend this denition a little. Let Y be a nite set on which
T(1),T(1)
0
(Sh
K
(G, X))
= Sh
(K)
(T, Y ).
In other words, the set of connected components of the Shimura variety is a zero-dimensional
Shimura variety (as promised).
Additional axioms
The weight homomorphismn
X
is a homomorphismG
m
G
R
over 1 of algebraic groups
that are dened over . It is therefore dened over
al
. Some simplications to the theory
occur when some of the following conditions hold:
SV4 The weight homomorphism n
X
: G
m
G
R
is dened over (we then say that the
weight is rational).
45
See the endnote.
Arithmetic subgroups of tori 61
SV5 The group Z() is discrete in Z(A
(
).
SV6 The torus Z
. Then, B
Q
F =
B
F,
1 where : runs
over the embeddings of F into 1. Thus,
B
Q
1 - H H M
2
(1) M
2
(1)
G(1) - H
GL
2
(1) GL
2
(1)
h(a ib) = 1 1
_
a b
-b a
_
_
a b
-b a
_
n(r) = 1 1 r
-1
I
2
r
-1
I
2
Let X be the G(1)-conjugacy class of h. Then (G, X) satises SV1 and SV2, and so it
is a Shimura datum if B splits
47
at at least one real prime of F. Let I = Hom(F,
al
) =
Hom(F, 1), and let I
nc
be the set of : such that B
F,
1 is split. Then n is dened over
the subeld of
al
xed by the automorphisms of
al
stabilizing I
nc
. This eld is always
totally real, and it equals if and only if I = I
nc
.
Arithmetic subgroups of tori
Let T be a torus over , and let T(Z) be an arithmetic subgroup of T(), for example,
T(Z) = Hom(X
+
(T), O
L
)
Gal(L{Q)
,
where L is some galois splitting eld of T. The congruence subgroup problem is known
to have a positive answer for tori (Serre 1964, 3.5), i.e., every subgroup of T(Z) of nite
index contains a congruence subgroup. Thus the topology induced on T() by that on
T(A
(
) has the following description: T(Z) is open, and the induced topology on T(Z) is
the pronite topology. In particular,
T() is discrete == T(Z) is discrete == T(Z) is nite.
EXAMPLE 5.25. (a) Let T = G
m
. Then T(Z) = {1}, and so T() is discrete in T(A
(
).
This, of course, can be proved directly.
48
46
In my view, the extra generality obtained by omitting it is spurious, but Deligne disagrees with me.
47
That is, becomes isomorphic to M
2
(1).
48
It is easy to write down an open subgroup of A
(
whose intersection with
is {1}.
62 5 SHIMURA VARIETIES
(b) Let T() = {a c [
_
1]
T such
that T
L
=
_
ty=-y
Ker(y: T
L
G
m
). Then T() is discrete in T(A
(
) if and only if T
is split, i.e., if and only if the largest subtorus of T that splits over 1 is already split over
.
Passage to the limit.
Let K be a compact open subgroup of G(A
(
), and let Z()
-
be the closure of Z() in
Z(A
(
). Then Z() K = Z()
-
K (in G(A
(
)) and
Sh
K
(G, X) =
df
G()X (G(A
(
),K)
=
G()
Z()
_
X (G(A
(
),Z() K)
=
G()
Z()
_
X (G(A
(
),Z()
-
K).
THEOREM 5.28. For any Shimura datum (G, X),
lim
K
Sh
K
(G, X) =
G()
Z()
_
X (G(A
(
),Z()
-
).
When SV5 holds,
lim
K
Sh
K
(G, X) = G() X G(A
(
).
Passage to the limit. 63
PROOF. The rst equality can be proved by the same argument as (4.19), and the second
follows from the rst (cf. Deligne 1979, 2.1.10, 2.1.11).
49
REMARK 5.29. Put S
K
= Sh
K
(G, X). For varying K, the S
K
form a variety (scheme)
with a right action of G(A
(
) in the sense of Deligne 1979, 2.7.1. This means the following:
(a) the S
K
form an inverse system of algebraic varieties indexed by the compact open
subgroups K of G(A
(
) (if K K
t
, there is an obvious quotient map S
K
0 S
K
);
(b) there is an action , of G(A
(
) on the system (S
K
)
K
dened by isomorphisms (of
algebraic varieties) ,
K
(a): S
K
S
g
1
Kg
(on points, ,
K
(a) is [x, a
t
] .[x, a
t
a]);
(c) for k c K, ,
K
(k) is the identity map; therefore, for K
t
normal in K, there is an
action of the nite group K,K
t
on S
K
0 ; the variety S
K
is the quotient of S
K
0 by the
action of K,K
t
.
REMARK 5.30. When we regard the Sh
K
(G, X) as schemes, the inverse limit of the system
Sh
K
(G, X) exists
50
:
S = lim
Sh
K
(G, X).
This is a scheme over C, not(!) of nite type, but it is locally noetherian and regular (cf.
Milne 1992, 2.4). There is a right action of G(A
(
) on S, and, for K a compact open
subgroup of G(A
(
),
Sh
K
(G, X) = S,K
(Deligne 1979, 2.7.1). Thus, the system (Sh
K
(G, X))
K
together with its right action of
G(A
(
) can be recovered from S with its right action of G(A
(
). Moreover,
S(C)
= lim
Sh
K
(G, X)(C) = lim
G()X G(A
(
),K.
49
The proof of Theorem 5.28 in Deligne 1979 reads (in its entirety): Laction de G(),Z() sur X
_
G(A
(
),Z()
-
_
est propre. Ceci permet le passage ` a la limite sur K.
Properness implies that the quotient of X
_
G(A
(
),Z()
-
_
by G(),Z() is separated (Bourbaki 1989,
III 4.2), and hence Lemma 4.20 applies. Presumably, the action is proper, but I dont know a proof that the
quotient is separated even in the easier case (4.19). Heres how the obvious argument goes.
We want to check that G()D G(A
(
) is separated. Choose a compact open subgroup U of G(A
(
)
such that the (congruence) subgroup I = U G(A
(
) is torsion free (3.6). We have to prove that distinct
points [x, a] and [y, b] of G()D G(A
(
) are separated by open neighbourhoods. Because G() is
dense in G(A
(
), we may suppose a, b c U. If x and y are not in the same I-orbit, then there exist open
neighbourhoods V
x
of x and V
y
of y such that gV
x
V
y
= 0 for all g c I (see the proof of 3.1). Then
g(V
x
Ua)(V
y
Ub) = 0 for all g c G(), and so the images of V
x
Ua and V
y
Ub in G()DG(A
(
)
separate [x, a] and [y, b]. When x and y lie in the same I-orbit, we may suppose y = x. There exists an
open neighbourhood V of x such that gV V = 0 for all g c I {1}, and so g(V Ua) (V Ub) = 0
for all g c G() {1} provided ba
-1
c U (but what if it isnt).
50
Let (A
i
)
i=I
be a direct system of commutative rings indexed by a directed set I, and let A = lim
A
i
.
Then, for any scheme X,
Hom(X, Spec A)
= Hom(A, I(X, O
X
))
= lim
Hom(A
i
, I(X, O
X
))
= lim
Hom(X, Spec A
i
).
(For the rst and third isomorphisms, see Hartshorne 1977, II, Exercise 2.4; the middle isomorphism is the
denition of direct limit). This shows that Spec A is the inverse limit of the inverse system (Spec A
i
)
i=I
in
the category of schemes. More generally, inverse limits of schemes in which the transition morphisms are
afne exist, and can be constructed in the obvious way.
64 5 SHIMURA VARIETIES
NOTES. Axioms SV1, SV2, SV3, and SV4 are respectively the conditions (2.1.1.1), (2.1.1.2),
(2.1.1.3), and (2.1.1.4) of Deligne 1979. Axiom SV5 is weaker than the condition (2.1.1.5)
ibid., which requires that adh(i) be a Cartan involution on (G,n(G
m
))
R
, i.e., that (Z
,n(G
m
))
R
be anisotropic.
65
6 The Siegel modular variety
In this section, we study the most important Shimura variety, namely, the Siegel modular
variety.
Dictionary
Let V be an 1-vector space. Recall (2.4) that to give a C-structure J on V is the same as
to give a hodge structure h
J
on V of type (1, 0), (0, 1). Here h
J
is the restriction to C
of the homomorphism
a bi .a bJ: C End
R
(V ).
For the hodge decompostion V (C) = V
-1,0
V
-1,0
,
V
-1,0
V
0,-1
J acts as i i
h
J
(z) acts as z z
Let [ be a nondegenerate 1-bilinear alternating form on V . A direct calculation
51
shows that
[(Ju, J:) = [(u, :) == [(zu, z:) = [z[
2
[(u, :) for all z c C.
Let [
J
(u, :) = [(u, J:). Then
[(Ju, J:) = [(u, :) == [
J
is symmetric
and
[(Ju, J:) = [(u, :) and
[
J
is positive denite
(2.12)
==
[ is a polarization of the
hodge structure (V, h
J
).
Symplectic spaces
Let k be a eld of characteristic ,= 2, and let (V, [) be a symplectic space of dimension
2n over k, i.e., V is a k-vector space of dimension 2n and [ is a nondegenerate alternating
form [. A subspace W of V is totally isotropic if [(W, W) = 0. A symplectic basis of
51
If [(Ju, J:) = [(u, :), then
[((a bi)u, (a bi):) = [(au, a:) [(au, bi:) [(biu, a:) [(biu, bi:) (biadditivity)
= [(au, a:) [(aiu, b:) [(biu, a:) [(bu, b:) (assumption)
= a
2
[(u, :) b
2
[(u, :) (bilinearity).
66 6 THE SIEGEL MODULAR VARIETY
V is a basis (e
i
)
1_i_n
such that
52
[(e
i
, e
-i
) = 1 for 1 _ i _ n,
[(e
i
, e
j
) = 0 for j ,= i.
LEMMA 6.1. Let W be a totally isotropic subspace of V . Then any basis of W can be
extended to a symplectic basis for V . In particular, V has symplectic bases (and two
symplectic spaces of the same dimension are isomorphic).
PROOF. Standard.
53
Thus, a maximal totally isotropic subspace of V will have dimension n. Such subspaces
are called lagrangians.
Let GSp([) be the group of symplectic similitudes of (V, [), i.e., the group of auto-
morphisms of V preserving [ up to a scalar. Thus
GSp([)(k) = {g c GL(V ) [ [(gu, g:) = v(g) [(u, :) some v(g) c k
}.
Dene Sp([) by the exact sequence
1 Sp([) GSp([)
G
m
1.
Then GSp([) has derived group Sp([), centre G
m
, and adjoint group GSp([),G
m
=
Sp([), I.
For example, when V has dimension 2, there is only one nondegenerate alternating form
on V up to scalars, which must therefore be preserved up to scalars by any automorphism,
and so GSp([) = GL
2
and Sp([) = SL
2
.
The group Sp([) acts simply transitively on the set of symplectic bases: if (e
i
) and
(
i
) are bases of V , then there is a unique g c GL
2n
(k) such that ge
i
=
i
, and if
(e
i
) and (
i
) are both symplectic, then g c Sp([).
The Shimura datum attached to a symplectic space
Fix a symplectic space (V, [) over , and let G = GSp([) and S = Sp([) = G
der
.
Let J be a complex structure on V (1) such that [(Ju, J:) = [(u, :). Then J c
S(1), and h
J
(z) lies in G(1) (and in S(1) if [z[ = 1) see the dictionary. We say that
52
Equivalently, such that the matrix of [ with respect to (e
i
) has 1 down the second diagonal, and zeros
elsewhere:
_
[(e
i
, e
j
)
_
1_i,j_g
=
_
0 I
t
g
I
t
g
0
_
, I
t
g
=
_
_
1
1
_
_
.
53
Certainly, the second statement is true when n = 1. We assume it inductively for spaces of dimension
_ 2n 2. Let W be totally isotropic, and let W
t
be a subspace of V such that V = W
J
W
t
. Then
W
= V,W
J
= W
t
identies W
t
with the dual of W. Let e
1
, . . . , e
m
be a basis for W, and let e
-1
, . . . , e
-m
be the dual basis in W
t
. Then (e
i
)
1_i_m
is a symplectic basis for W W
t
. By induction (W W
t
)
J
has
a symplectic basis (e
i
)
m1_i_n
, and then (e
i
)
1_i_n
is a symplectic basis for V .
The Shimura datum attached to a symplectic space 67
J is positive (resp. negative) if [
J
(u, :) =
df
[(u, J:) is positive denite (resp. negative
denite).
Let X
(resp. X
-
) denote the set of positive (resp. negative) complex structures on
V (1), and let X = X
. X
-
. Then G(1) acts on X according to the rule
(g, J) .gJg
-1
,
and the stabilizer in G(1) of X
is
G(1)
in fact, (e
i
)
i
is an orthonormal basis for [
J
. Conversely,
if J c X
, then J has this description relative to any orthonormal basis for the positive
denite form [
J
. The map from symplectic bases to X
G(1). Then h
gJg
1(z) =
gh
J
(z)g
-1
. Thus J .h
J
identies X with a G(1)-conjugacy class of homomorphisms
h: C
= V
-1,0
and V
-
= V
0,-1
, so that V (C) = V
V
-
with
h(z) acting on V
and V
-
as multiplication by z and z respectively. Then
55
Hom(V (C), V (C)) = Hom(V
, V
) Hom(V
, V
-
) Hom(V
-
, V
) Hom(V
-
, V
-
)
h(z) acts as 1 z,z z,z 1
The Lie algebra of G is the subspace
Lie(G) = { c Hom(V, V ) [ [( (u), :) [(u, (:)) = 0},
of End(V ), and so SV1 holds.
(SV2). We have to show that adJ is a Cartan involution on G
ad
. But, J
2
= 1 lies in
the centre of S(1) and [ is a J-polarization for S
R
in the sense of (1.20), which shows
that adJ is a Cartan involution for S.
(SV3). In fact, G
ad
is -simple, and G
ad
(1) is not compact.
(SV4). For r c 1
, n
h
(r) acts on both V
-1,0
and V
0,-1
as : . r:. Therefore, n
X
is the homomorphism G
mR
GL(V (1)) sending r c 1
to multplication by r. This is
dened over .
(SV5). The centre of G is G
m
, and
is discrete in A
(
(see 5.25).
(SV6). The centre of G is split already over .
We often write (G([), X([)) for the Shimura datumdened by a symplectic space (V, [),
and (S([), X([)
and X
-
.
55
Recall that the group GL(V ) acts on Hom(V, V ) according to the rule
(:) = ( (
-1
:)), c GL(V ), c Hom(V, V ), : c V.
68 6 THE SIEGEL MODULAR VARIETY
EXERCISE 6.2. (a) Show that for any h c X([), v(h(z)) = zz. [Hint: for nonzero
:
c V
and :
-
c V
-
, compute [
C
(h(z):
, h(z):
-
) in two different ways.]
(b) Show that the choice of a symplectic basis for V identies X
with H
g
as an
Sp([)-set (see 1.2).
The Siegel modular variety
Let (G, X) = (G([), X([)) be the Shimura datum dened by a symplectic space (V, [)
over . The Siegel modular variety attached to (V, [) is the Shimura variety Sh(G, X).
Let V (A
(
) = A
(
Q
V . Then G(A
(
) is the group of A
(
-linear automorphisms of
V (A
(
) preserving [ up to multiplication by an element of A
(
.
Let K be a compact open subgroup of G(A
(
), and let H
K
be the set of triples ((W, h), s, jK)
where
(W, h) is a rational hodge structure of type (1, 0), (0, 1);
s is a polarization for (W, h);
jK is a K-orbit of A
(
-linear isomorphisms V (A
(
) W(A
(
) sending [ to an
A
(
-multiple of s.
An isomorphism
((W, h), s, jK) ((W
t
, h
t
), s
t
, j
t
K)
of triples is an isomorphism b: (W, h) (W
t
, h
t
) of rational hodge structures such that
b(s) = cs
t
some c c
and b j = j
t
mod K.
Note that to give an element of H
K
amounts to giving a symplectic space (W, s) over
, a complex structure on W that is positive or negative for s, and jK. The existence
of j implies that dimW = dimV , and so (W, s) and (V, [) are isomorphic. Choose an
isomorphism a: W V sending [ to a
-multiple of s. Then
ah =
df
(z .a h(z) a
-1
)
lies in X, and
V (A
(
)
)
W(A
(
)
a
V (A
(
)
lies in G(A
(
). Any other isomorphism a
t
: W V sending [ to a multiple of s differs
from a by an element of G(), say, a
t
= q a with q c G(). Replacing a with a
t
only
replaces (ah, a j) with (qah, qa j). Similarly, replacing j with jk replaces (ah, a j)
with (ah, a jk). Therefore, the map
(W . . .) .[ah, a j]
K
: H
K
G()X G(A
(
),K
is well-dened.
PROPOSITION 6.3. The set Sh
K
(G, X) classies the triples in H
K
modulo isomorphism.
More precisely, the map (W, . . .) .[ah, a j]
K
denes a bijection
H
K
,-G()X G(A
(
),K.
Complex abelian varieties 69
PROOF. It is straightforward to check that the map sends isomorphic triples to the same
class,
56
and that two triples are isomorphic if they map to the same class.
57
The map is onto
because [h, g] is the image of ((V, h), [, gK).
Complex abelian varieties
An abelian variety A over a eld k is a connected projective algebraic variety over k
together with a group structure given by regular maps. A one-dimensional abelian variety
is an elliptic curve. Happily, a theorem, whose origins go back to Riemann, reduces the
study of abelian varieties over C to multilinear algebra.
Recall that a lattice in a real or complex vector space V is the Z-module generated by an
1-basis for V . For a lattice in C
n
, make C
n
, into a complex manifold by endowing it
with the quotient structure.
58
A complex torus is a complex manifold isomorphic to C
n
,
for some lattice in C
n
.
Note that C
n
is the universal covering space of M = C
n
, with as its group of
covering transformations, and
1
(M, 0) = (Hatcher 2002, 1.40). Therefore, (ib. 2A.1)
H
1
(M, Z)
= (31)
and (Greenberg 1967, 23.14)
H
1
(M, Z)
= Hom(
_
n
, Z),
i.e., H
n
(M, Z) is canonically isomorphic to the set of n-alternating forms Z.
PROOF. From (32), we see that
_
n
H
1
(M, Z)
=
_
n
Hom(, Z).
Since
59
_
n
Hom(, Z)
= Hom(
_
n
, Z),
56
Suppose b: (W, h)
~
(W
t
, h
t
) sends s to a
-multiple of t
t
and is such that b j = j
t
k for some
k c K. Choose an isomorphism a
t
: W
t
V sending s
t
to a
_
n
Z
determined by
(
1
. .
n
, :
1
:
n
) = det(
i
(:
j
))
is nondegenerate (since it is modulo p for every p see Bourbaki 1958, ~8).
70 6 THE SIEGEL MODULAR VARIETY
we see that it sufces to show that cup-product denes an isomorphism
_
n
H
1
(M, Z) H
n
(M, Z). (33)
Let T be the class of topological manifolds M whose cohomology groups are free Z-
modules of nite rank and for which the maps (33) are isomorphisms for all n. Certainly,
the circle S
1
is in T (its cohomology groups are Z, Z, 0, . . .), and the K unneth formula
(Hatcher 2002, 3.16 et seq.) shows that if M
1
and M
2
are in T , then so also is M
1
M
2
.
As a topological manifold, C
n
, - (S
1
)
2n
, and so M is in T .
PROPOSITION 6.5. A linear map : C
n
C
n
0
such that ()
t
denes a holomorphic
map C
n
, C
n
0
,
t
sending 0 to 0, and every holomorphic map C
n
, C
n
0
,
t
sending 0 to 0 is of this form (for a unique ).
PROOF. The map C
n
C
n
0
C
n
0
,
t
is holomorphic, and it factors through C
n
,. Be-
cause C, has the quotient structure, the resulting map C
n
, C
n
0
,
t
is holomorphic.
Conversely, let : C, C,
t
be a holomorphic map such that (0) = 0. Then C
n
and
C
n
0
are universal covering spaces of C
n
, and C
n
0
,
t
, and a standard result in topology
(Hatcher 2002, 1.33, 1.34) shows that lifts uniquely to a continuous map : C
n
C
n
0
such that (0) = 0:
C
n
C
n
0
_
C
n
,
C
n
0
,
t
.
Because the vertical arrows are local isomorphisms, is automatically holomorphic. For
any o c , the map z . (z o) (z) is continuous and takes values in
t
C.
Because C
n
is connected and
t
is discrete, it must be constant. Therefore, for each j,
J
Jz
j
is a doubly periodic function, and so denes a holomorphic function C
n
, C
n
0
,
which must be constant (because C
n
, is compact). Write as an n
t
-tuple (
1
, . . . ,
n
0 )
of holomorphic functions
i
in n variables. Because
i
(0) = 0 and
J
i
Jz
j
is constant for each
j, the power series expansion of
i
at 0 is of the form
a
ij
z
j
. Now
i
and
a
ij
z
j
are
holomorphic functions on C
n
that coincide on a neighbourhood of 0, and so are equal on
the whole of C
n
. We have shown that
(z
1
, . . . , z
n
) = (
a
1j
z
j
, . . . ,
a
n
0
j
z
j
).
ASIDE 6.6. The proposition shows that every holomorphic map : C
n
, C
n
0
,
t
such
that (0) = 0 is a homomorphism. A similar statement is true for abelian varieties over
any eld k: a regular map : A B of abelian varieties such that (0) = 0 is a ho-
momorphism (AG, 5.36). For example, the map sending an element to its inverse is a
homomorphism, which implies that the group law on A is commutative. Also, the group
law on an abelian variety is uniquely determined by the zero element.
Let M = C
n
, be a complex torus. The isomorphism 1
= C
n
denes a complex
structure J on 1. A riemann form for M is an alternating form [: Z such
that [
R
(Ju, J:) = [
R
(u, :) and [
R
(u, Ju) > 0 for u ,= 0. A complex torus C
n
, is said
to be polarizable if there exists a riemann form.
Complex abelian varieties 71
THEOREM 6.7. The complex torus C
n
, is projective if and only if it is polarizable.
PROOF. See Mumford 1970, Chapter I, (or Murty 1993, 4.1, for the if part). Alterna-
tively, one can apply the Kodaira embedding theorem (Voisin 2002, Th. 7.11, 7.2.2).
Thus, by Chows theorem (3.11), a polarizable complex torus is a projective algebraic
variety, and holomorphic maps of polarizable complex tori are regular. Conversely, it is
easy to see that the complex manifold associated with an abelian variety is a complex
torus: let Tgt
0
A be the tangent space to A at 0; then the exponential map Tgt
0
A A(C)
is a surjective homomorphism of Lie groups with kernel a lattice , which induces an
isomorphism (Tgt
0
A),
= C
n
endows
Z
1 with
a complex structure, and hence endows
= H
1
(M, Z) with an integral hodge structure
of weight 1. Note that a riemann form for M is nothing but a polarization of the integral
hodge structure .
THEOREM 6.8 (RIEMANNS THEOREM).
60
The functor A .H
1
(A, Z) is an equivalence
from the category AV of abelian varieties over C to the category of polarizable integral
hodge structures of type (1, 0), (0, 1).
PROOF. We have functors
AV
A|-A
an
{category of polarizable complex tori}
M|-H
1
(M,Z)
{category of polarizable integral hodge structures of type (1, 0), (0, 1)}.
The rst is fully faithful by Chows theorem (3.11), and it is essentially surjective by The-
orem 6.7; the second is fully faithful by Proposition 6.5, and it is obviously essentially
surjective.
Let AV
0
be the category whose objects are abelian varieties over C and whose mor-
phisms are
Hom
AV
0(A, B) = Hom
AV
(A, B) .
COROLLARY 6.9. The functor A .H
1
(A, ) is an equivalence from the category AV
0
to
the category of polarizable rational hodge structures of type (1, 0), (0, 1).
PROOF. Immediate consequence of the theorem.
REMARK 6.10. Recall that in the dictionary between complex structures J on a real vector
space V and hodge structures of type (1, 0), (0, 1),
(V, J)
= V (C),V
-1,0
= V (C),F
0
.
Since the hodge structure on H
1
(A, 1) is dened by the isomorphismTgt
0
(A)
= H
1
(A, 1),
we see that
Tgt
0
(A)
= H
1
(A, C),F
0
(34)
(isomorphism of complex vector spaces).
60
In fact, it should be called the theorem of Riemann, Frobenius, Weierstrass, Poincar e, Lefschetz, et al.
(see Shafarevich 1994, Historical Sketch, 5), but Riemanns theorem is shorter.
72 6 THE SIEGEL MODULAR VARIETY
A modular description of the points of the Siegel variety
Let M
K
be the set of triples (A, s, jK) in which A is an abelian variety over C, s is an
alternating form on H
1
(A, ) such that s or s is a polarization on H
1
(A, ), and j is
an isomorphism V (A
(
) V
(
(A
(
) sending [ to a multiple of s by an element of A
(
.
An isomorphism from one triple (A, s, jK) to a second (A
t
, s
t
, j
t
K) is an isomorphism
A A
t
(as objects in AV
0
) sending s to a multiple of s
t
by an element of
and jK to
j
t
K.
THEOREM 6.11. The set Sh
K
(G, X) classies the triples (A, s, jK) in M
K
modulo iso-
morphism, i.e., there is a canonical bijection M
K
,-G()X G(A
(
),K.
PROOF. Combine (6.9) with (6.3).
73
7 Shimura varieties of hodge type
In this section, we examine one important generalization of Siegel modular varieties.
DEFINITION 7.1. A Shimura datum (G, X) is of hodge type if there exists a symplectic
space (V, [) over and an injective homomorphism ,: G G([) carrying X into
X([). The Shimura variety Sh(G, X) is then said to be of hodge type. Here (G([), X([))
denotes the Shimura datum dened by (V, [).
The composite of , with the character v of G([) is a character of G, which we again
denote by v. Let (r) denote the vector space with G acting by rv, i.e., g : = v(g)
r
:.
For each h c X, ((r), h v) is a rational hodge structure of type (r, r) (apply 6.2a),
and so this notation is consistent with that in (2.6).
LEMMA 7.2. There exist multilinear maps t
i
: V V (r
i
), 1 _ i _ n, such that
G is the subgroup of G([) xing the t
i
.
PROOF. According to Deligne 1982, 3.1, there exist tensors t
i
in V
r
i
V
s
i
such that
this is true. But [ denes an isomorphism V
= V
(1)), and so
V
r
i
V
s
i
= V
(r
i
s
i
)
(r
i
)
= Hom(V
(r
i
s
i
)
, (r
i
)).
Let (G, X) be of hodge type. Choose an embedding of (G, X) into (G([), X([)) for
some symplectic space (V, [) and multilinear maps t
1
, . . . , t
n
as in the lemma. Let H
K
be
the set of triples ((W, h), (s
i
)
0_i_n
, jK) in which
(W, h) is a rational hodge structure of type (1, 0), (0, 1),
s
0
is a polarization for (W, h),
s
1
, . . . , s
n
are multilinear maps s
i
: W W (r
i
), and
jK is a K-orbit of isomorphisms V (A
(
) W(A
(
) sending [ onto an A
(
-multiple
of s
0
and each t
i
to s
i
,
satisfying the following condition:
(*) there exists an isomorphism a: W V sending s
0
to a
-multiple of [,
s
i
to t
i
each i _ 1, and h onto an element of X.
An isomorphismfromone triple (W, . . .) to a second (W
t
, . . .) is an isomorphism(W, h)
(W
t
, h
t
) of rational hodge structures sending s
0
to a
-multiple of s
t
0
, s
i
to s
t
i
for i > 0,
and jK to j
t
K.
PROPOSITION 7.3. The set Sh
K
(G, X)(C) classies the triples in H
K
modulo isomor-
phism.
PROOF. Choose an isomorphism a: W V as in (*), and consider the pair (ah, a j).
By assumption ah c X and a j is a symplectic similitude of (V (A
(
), [) xing the t
i
,
and so (ah, a j) c X G(A
(
). The isomorphism a is determined up to composition
with an element of G() and j is determined up to composition with an element of K. It
follows that the class of (ah, aj) in G()X G(A
(
),K is well-dened. The proof that
(W, . . .) . [ah, a j]
K
gives a bijection from the set of isomorphism classes of triples in
H
K
onto Sh
K
(G, X)(C) is now routine (cf. the proof of 6.3).
74 7 SHIMURA VARIETIES OF HODGE TYPE
Let t : V V (r) (m-copies of V ) be a multilinear form xed by G, i.e.,
such that
t (g:
1
, . . . , g:
m
) = v(g)
r
t (:
1
, . . . , :
m
), for all :
1
, . . . , :
m
c V , g c G().
For h c X, this equation shows that t denes a morphism of hodge structures (V, h)
m
= Hom(
_
m
V, ).
We say that t c H
2r
(A, ) is a hodge tensor for A if the corresponding map
V
2r
_
2r
V (r)
is a morphism of Hodge structures.
Let (G, X) (G([), X([)) and t
1
, . . . , t
n
be as above. Let M
K
be the set of triples
(A, (s
i
)
0_i_n
, jK) in which
A is a complex abelian variety,
s
0
is a polarization for the rational hodge structure H
1
(A, ),
s
1
, . . . , s
n
are hodge tensors for A or its powers, and
jK is a K-orbit of A
(
-linear isomorphisms V (A
(
) V
(
(A) sending [ onto an
A
(
-multiple of s
0
and each t
i
to s
i
,
satisfying the following condition:
(**) there exists an isomorphism a: H
1
(A, ) V sending s
0
to a
-
multiple of [, s
i
to t
i
each i _ 1, and h to an element of X.
An isomorphism from one triple (A, (s
i
)
i
, jK) to a second (A
t
, (s
t
i
), j
t
K) is an isomor-
phism A A
t
(as objects of AV
0
) sending s
0
to a multiple of s
t
0
by an element of
,
each s
i
to s
t
i
, and j to j
t
modulo K.
THEOREM 7.4. The set Sh
K
(G, X)(C) classies the triples in M
K
modulo isomorphism.
PROOF. Combine Propositions 7.3 and 6.9.
The problem with Theorem 7.4 is that it is difcult to check whether a triple satises
the condition (**). In the next section, we show that when the hodge tensors are endomor-
phisms of the abelian variety, then it is sometimes possible to replace (**) by a simpler
trace condition.
REMARK 7.5. When we write A(C) = C
g
,, then (see 6.4),
H
m
(A, )
= Hom(
_
m
, )
Now C
= T T where T = Tgt
0
(A). Therefore,
H
m
(A, C)
= Hom(
_
m
(C), C)
= Hom(
pq=m
_
p
T
_
q
T, C)
=
pq=m
H
p,q
75
where
H
p,q
= Hom(
_
p
T
_
q
T, C).
This rather ad hoc construction of the Hodge structure on H
m
does agree with the usual
construction (2.5) see Mumford 1970, Chapter I. A hodge tensor on A is an element of
H
2r
(A, ) H
r,r
(intersection inside H
2r
(A, C)).
The Hodge conjecture predicts that all hodge tensors are the cohomology classes of alge-
braic cycles with -coefcients. For r = 1, this is known even over Z. The exponential
sequence
0 Z O
A
z|-exp(2iz)
O
A
0
gives a cohomology sequence
H
1
(A, O
A
) H
2
(A, Z) H
2
(A, O
A
).
The cohomology group H
1
(A, O
A
) classies the divisors on A modulo linear equivalence,
i.e., Pic(A)
= H
1
(A, O
A
), and the rst arrow maps a divisor to its cohomology class. A
class in H
2
(A, Z) maps to zero in H
2
(A, O
A
) = H
0,2
if and only if it maps to zero in its
complex conjugate H
2,0
. Therefore, we see that
Im(Pic(A)) = H
2
(A, Z) H
1,1
.
76 8 PEL SHIMURA VARIETIES
8 PEL Shimura varieties
Throughout this section, k is a eld of characteristic zero. Bilinear forms are always non-
degenerate.
Algebras with involution
By a k-algebra I mean a ring B containing k in its centre and nite dimensional over k. A
k-algebra A is simple if it contains no two-sided ideals except 0 and A. For example, every
matrix algebra M
n
(D) over a division algebra D is simple, and conversely, Wedderburns
theorem says that every simple algebra is of this form (CFT, IV 1.9). Up to isomorphism, a
simple k-algebra has only one simple module (ibid, IV 1.15). For example, up to isomor-
phism, D
n
is the only simple M
n
(D)-module.
Let B = B
1
B
n
be a product of simple k-algebras (a semisimple k-algebra). A
simple B
i
-module M
i
becomes a simple B-module when we let B act through the quotient
map B B
i
. These are the only simple B-modules, and every B-module is a direct sum
of simple modules. A B-module M denes a k-linear map
b .Tr
k
(b[M): B k
which we call the trace map of M.
PROPOSITION 8.1. Let B be a semisimple k-algebra. Two B-modules are isomorphic if
and only if they have the same trace map.
PROOF. Let B
1
, . . . , B
n
be the simple factors of B, and let M
i
be a simple B
i
-module.
Then every B-module is isomorphic to a direct sum
j
r
j
M
j
with r
j
M
j
the direct sum of
r
j
copies of M
i
. We have to show that the trace map determines the multiplicities r
j
. But
for e
i
= (0, . . . , 0, 1
i
, 0, . . .),
Tr
k
(e
i
[
r
j
M
j
) = r
i
dim
k
M
i
.
REMARK 8.2. The lemma fails when k has characteristic p, because the trace map is
identically zero on pM.
An involution of a k-algebra B is a k-linear map b . b
+
: B B such that (ab)
+
=
b
+
a
+
and b
++
= b. Note that then 1
+
= 1 and so c
+
= c for c c k.
PROPOSITION 8.3. Let k be an algebraically closed eld, and let (B, +) be a semisimple
k-algebra with involution. Then (B, +) is isomorphic to a product of pairs of the following
types:
(A) M
n
(k) M
n
(k), (a, b)
+
= (b
t
, a
t
);
(C) M
n
(k), b
+
= b
t
;
(BD) M
n
(k), b
+
= J b
t
J
-1
, J =
_
0 -I
I 0
_
.
Symplectic modules and the associated algebraic groups 77
PROOF. The decomposition B = B
1
B
r
of B into a product of simple algebras
B
i
is unique up to the ordering of the factors (Farb and Dennis 1993, 1.13). Therefore, +
permutes the set of B
i
, and B is a product of semisimple algebras with involution each of
which is either (i) simple or (ii) the product of two simple algebras interchanged by +.
Let (B, +) be as in (i). Then B is isomorphic to M
n
(k) for some n, and the Noether-
Skolem theorem (CFT, 2.10) shows that b
+
= u b
t
u
-1
for some u c M
n
(k). Then
b = b
++
= (u
t
u
-1
)
-1
b(u
t
u
-1
) for all b c B, and so u
t
u
-1
lies in the centre k of M
n
(k).
Denote it by c, so that u
t
= cu. Then u = u
t t
= c
2
u, and so c
2
= 1. Therefore, u
t
= u,
and u is either symmetric or skew-symmetric. Relative to a suitable basis, u is I or J, and
so (B, +) is of type (C) or (BD).
Let (B, +) be as in (ii). Then + is an isomorphism of the opposite of the rst factor
onto the second. The Noether-Skolem theorem then shows that (B, +) is isomorphic to
M
n
(k) M
n
(k)
opp
with the involution (a, b) .(b, a). Now use that a -a
t
: M
n
(k)
opp
=
M
n
(k) to see that (B, +) is of type (A).
The following is a restatement of the proposition.
PROPOSITION 8.4. Let (B, +) and k be as in (8.3). If the only elements of the centre of B
invariant under + are those in k, then (B, +) is isomorphic to one of the following:
(A) End
k
(W) End
k
(W
), (a, b)
+
= (b
t
, a
t
);
(C) End
k
(W), b
+
the transpose of b with respect to a symmetric bilinear form on W;
(BD) End
k
(W), b
+
the transpose of b with respect to an alternating bilinear form on W.
Symplectic modules and the associated algebraic groups
Let (B, +) be a semisimple k-algebra with involution +, and let (V, [) be a symplectic
(B, +)-module, i.e., a B-module V endowed with an alternating k-bilinear form [: V
V k such that
[(bu, :) = [(u, b
+
:) for all b c B, u, : c V. (35)
Let F be the centre of B, and let F
0
be the subalgebra of invariants of + in F. Assume
that B and V are free over F and that for all k-homomorphisms ,: F
0
k
al
, (B
F
0
,p
k
al
, +) is of the same type (A), (C), or (BD). This will be the case, for example, if F is a
eld. Let G be the subgroup of GL(V ) such that
G() = {g c Aut
B
(V ) [ [(gu, g:) = j(g)[(u, :) some j(g) c k
},
and let
G
t
= Ker(j) Ker(det).
EXAMPLE 8.5. (Type A.) Let F be k k or a eld of degree 2 over k, and let B =
End
F
(W) equipped with the involution + dened by a hermitian form
61
: W W F.
61
There is a unique involution of F xing k, which we again denote +. To say that is hermitian means
that it is F-linear in one variable and satises (n, :) = (:, n)
+
.
78 8 PEL SHIMURA VARIETIES
Then (B, +) is of type A. Let V
0
be an F-vector space, and let [
0
be a skew-hermitian form
V
0
V
0
F. The bilinear form [ on V = W
F
V
0
dened by
[(n :, n
t
:
t
) = Tr
F{k
((n, n
t
)[
0
(:, :
t
)) (36)
is alternating and satises (35): (V, [) is a symplectic (B, +)-module. Let C = End
B
(V )
(the centralizer of B in End
F
(V )). Then C is stable under the involution + dened by [,
and
G(k) = {c c C
[ cc
+
c k
} (37)
G
t
(k) = {c c C
[ cc
+
= 1, det(c) = 1}. (38)
In fact, C
= End
F
(V
0
) and + is transposition with respect to [
0
. Therefore, G is the group
of symplectic similitudes of [
0
whose multiplier lies in k, and G
t
is the special unitary
group of [
0
.
Conversely, let (B, +) be of type A, and assume
(a) the centre F of B is of degree 2 over k (so F is a eld or k k);
(b) B is isomorphic to a matrix algebra over F (when F is a eld, this just means that
B is simple and split over F).
Then I claim that (B, +, V, [) arises as in the last paragraph. To see this, let W be a
simple B-module condition (b) implies that B
= End
F
(W) and that + is dened by a
hermitian form: W W F. As a B-module, V is a direct sum of copies of W, and so
V = W
F
V
0
for some F-vector space V
0
. Choose an element of F k whose square
is in k. Then
+
= , and
[(:, :
t
) = Tr
F{k
_
(:, :
t
)
_
for a unique hermitian form : V V F (Deligne 1982, 4.6), which has the property
that (b:, :
t
) = (:, b
+
:
t
). The form (:, :
t
) . (:, :
t
) is skew-hermitian, and can
be
62
written = [
0
with [
0
skew-hermitian on V
0
. Now [, , [
0
are related by
(36).
EXAMPLE 8.6. (Type C.) Let B = End
k
(W) equipped with the involution + dened by a
symmetric bilinear form : W W k. Let V
0
be a k-vector space, and let [
0
be an
alternating form V
0
V
0
k. The bilinear form [ on V = W V
0
dened by
[(n :, n
t
:
t
) = (n, n
t
)[
0
(:, :
t
)
is alternating and satises (35). Let C = End
B
(V ). Then C is stable under the involution
+ dened by [, and G(k) and G
t
(k) are described by the equations (37) and (38). In fact,
C
= End
k
(V
0
) and + is transposition with respect to [
0
. Therefore G = GSp(V
0
, [
0
) and
G
t
= Sp(V
0
, [
0
). Every system (B, +, V, [) with B simple and split over k arises in this
way (cf. 8.5).
62
Probably the easiest way to prove things like this is use the correspondence between involutions on
algebras and (skew-)hermitian forms (up to scalars) see Knus et al. 1998, I 4.2. The involution on End
F
(V )
dened by [ stabilizes C and corresponds to a skew-hermitian form on V
0
.
Algebras with positive involution 79
PROPOSITION 8.7. For (B, +) of type A or C, the group G is reductive (in particular,
connected), and G
t
is semisimple and simply connected.
PROOF. It sufces to prove this after extending the scalars to the algebraic closure of k.
Then (B, +, V, [) decomposes into quadruples of the types considered in Examples 8.5
and 8.6, and so the proposition follows from the calculations made there.
REMARK 8.8. Assume B is simple, and let m be the reduced dimension of V ,
m =
dim
F
(V )
[B: F]
1
2
.
In case (A), G
t
Q
al
- (SL
m
)
[F
0
:Q]
and in case (C), G
t
Q
al
- (Sp
m
)
[F
0
:Q]
.
REMARK 8.9. In case (BD), the group G is not connected (G
t
is a special orthogonal
group) although its identity component is reductive.
Algebras with positive involution
Let C be a semisimple 1-algebra with an involution +, and let V be a C-module. In the next
proposition, by a hermitian form on V we mean a symmetric bilinear form[: V V 1
satisfying (35)
63
. Such a form is said to be positive denite if [(:, :) > 0 for all nonzero
: c V .
PROPOSITION 8.10. Let C be a semisimple algebra over 1. The following conditions on
an involution + of C are equivalent:
(a) some faithful C-module admits a positive denite hermitian form;
(b) every C-module admits a positive denite hermitian form;
(c) Tr
C{R
(c
+
c) > 0 for all nonzero c c C.
PROOF. (a) == (b). Let V be a faithful C-module. Then every C-module is a direct
summand of a direct sum of copies of V (see p76). Hence, if V carries a positive denite
hermitian form, then so does every C-module.
(b) == (c). Let V be a C-module with a positive denite hermitian form ( [ ), and
choose an orthonormal basis e
1
, . . . , e
n
for V . Then
Tr
R
(c
+
c[V ) =
i
(e
i
[c
+
ce
i
) =
i
(ce
i
[ce
i
),
which is > 0 unless c acts as the zero map on V . On applying this remark with V = C,
we obtain (c).
(c) == (a). The condition (c) is that the hermitian form (c, c
t
) . Tr
C{R
(c
+
c
t
) on C is
positive denite.
DEFINITION 8.11. An involution satisfying the equivalent conditions of (8.10) is said to be
positive.
63
Strictly, a hermitian form should be a form : V V C that is C-linear in one of the variables and
satises (y, x) = (x, y)
+
. But then [ = Tr
C{R
is a symmetric bilinear form satisfying (35), and
(V, ) .(V, [) is an equivalence of categories.
80 8 PEL SHIMURA VARIETIES
PROPOSITION 8.12. Let B be a semisimple 1-algebra with a positive involution + of type
A or C. Let (V, [) be a symplectic (B, +)-module, and let C be the centralizer of B
in End
R
(V ). Then there exists a homomorphism of 1-algebras h: C C, unique up to
conjugation by an element c of C
with cc
+
= 1, such that
h(z) = h(z)
+
and
u, : .[(u, h(i):) is positive denite and symmetric.
PROOF. To give an h satisfying the conditions amounts to giving an element J (= h(i)) of
C such that
J
2
= 1, [(Ju, J:) = [(u, :), [(:, J:) > 0 if : ,= 0. (39)
Suppose rst that (B, +) is of type A. Then (B, +, V, [) decomposes into systems arising
as in (8.5). Thus, we may suppose B = End
F
(W), V = W V
0
, etc., as in (8.5). We
then have to classify the J c C
= End
C
(V
0
) satisfying (39) with [ replaced by [
0
. There
exists a basis (e
j
) for V
0
such that
([
0
(e
j
, e
k
))
j,k
= diag(i, . . . , i
r
, i, . . . , i), i =
_
1.
Dene J by J(e
j
) = [
0
(e
j
, e
j
)e
j
. Then J satises the required conditions, and it is
uniquely determined up to conjugation by an element of the unitary group of [
0
. This
proves the result for type A, and type C is similar. (For more details, see Zink 1983,
3.1).
REMARK 8.13. Let (B, +) and (V, [) be as in the proposition. For an h satisfying the
conditions of the proposition, dene
t (b) = Tr
C
(b[V,F
0
h
V ), b c B.
Then, t is independent of the choice of h, and in fact depends only on the isomorphism
class of (V, [) as a B-module. Conversely, (V, [) is determined up to B-isomorphism by
its dimension and t . For example, if V = W
C
V
0
, , [
0
, etc. are as in the above proof,
then
Tr
k
(b[V ) = r Tr
k
(b[W),
and r and dimV
0
determine (V
0
, [
0
) up to isomorphism. Since W and are determined
(up to isomorphism) by the requirement that W be a simple B-module and be a hermitian
form giving + on B, this proves the claim for type A.
PEL data
Let B be a simple -algebra with a positive involution + (meaning that it becomes positive
on B
Q
1), and let (V, [) be a symplectic (B, +)-module. Throughout this subsection,
we assume that (B, +) is of type A or C.
PROPOSITION 8.14. There is a unique G(1)-conjugacy class X of homomorphisms h: S
G
R
such that each h c X denes a complex structure on V (1) that is positive or negative
for [. The pair (G, X) satises the conditions SV14.
PEL Shimura varieties 81
PROOF. The rst statement is an immediate consequence of (8.12). The composite of h
with G G([) lies in X([), and therefore satises SV1, SV2, SV4. As h is nontrivial,
SV3 follows from the fact that G
ad
is simple.
DEFINITION 8.15. The Shimura data arising in this way are called simple PEL data of
type A or C.
The simple refers to the fact that (for simplicity), we required B to be simple (which
implies that G
ad
is simple).
REMARK 8.16. Let b c B, and let t
b
be the tensor (x, y) .[(x, by) of V . An element
g of G([) xes t
b
if and only if it commutes with b. Let b
1
, . . . , b
s
be a set of generators
for B as a -algebra. Then (G, X) is the Shimura datum of hodge type associated with the
system (V, {[, t
b
1
, . . . , t
b
s
}).
PEL Shimura varieties
THEOREM 8.17. Let (G, X) be a simple PEL datumof type A or Cassociated with (B, +, V, [)
as in the last subsection, and let K be a compact open subgroup of G(A
(
). Then Sh
K
(G, X)(C)
classies the isomorphism classes of quadruples (A, s, i, jK) in which
A is a complex abelian variety,
s is a polarization of the hodge structure H
1
(A, ),
i is a homomorphism B End(A) , and
jK is a K-orbit of B A
(
-linear isomorphisms j: V (A
(
) H
1
(A, ) A
(
sending [ to an A
(
-multiple of s,
satisfying the following condition:
(**) there exists a B-linear isomorphism a: H
1
(A, ) V sending s to a
-multiple of [.
PROOF. In view of the dictionary b - t
b
between endomorphisms and tensors (8.16),
Theorem 7.4 shows that Sh
K
(G, X)(C) classies the quadruples (A, i, t, jK) with the ad-
ditional condition that ah c X, but ah denes a complex structure on V (1) that is positive
or negative for [, and so (8.14) shows that ah automatically lies in X.
Let (G, X) be the Shimura datum arising from (B, +) and (V, [). For h c X, we have
a trace map
b .Tr(b[V (C),F
0
h
): B C.
Since this map is independent of the choice of h in X, we denote it by Tr
X
.
REMARK 8.18. Consider a triple (A, s, i, jK) as in the theorem. The existence of the
isomorphism a in (**) implies that
(a) s(bu, :) = s(u, b
+
:), and
(b) Tr(i(b)[Tgt
0
A) = Tr
X
(b) for all b c B C.
The rst is obvious, because [ has this property, and the second follows from the B-
isomorphisms
Tgt
0
(A)
(34)
= H
1
(A, C),F
0
a
V (C),F
0
h
.
82 8 PEL SHIMURA VARIETIES
We now divide the type A in two, depending on whether the reduced dimension of V is
even or odd.
PROPOSITION 8.19. For types Aeven and C, the condition (**) of Theorem 8.17 is implied
by conditions (a) and (b) of (8.18).
PROOF. Let W = H
1
(A, ). We have to show that there exists a B-linear isomorphism
: W V sending s to a
al
-isomorphism: V (
al
) W(
al
) sending
t to a
al
-multiple of [. For o c Gal(
al
,) write o = a
o
with a
o
c G(
al
). Then
o . a
o
is a one-cocycle. If its class in H
1
(, G) is trivial, say, a
o
= a
-1
oa, then
a
-1
is xed by all o c Gal(
al
,), and is therefore dened over .
Thus, it remains to show that the class of (a
o
) in H
1
(, G) is trivial. The existence of
j shows that the image of the class in H
1
(
I
, G) is trivial for all nite primes , and (8.13)
shows that its image in H
1
(1, G) is trivial, and so the statement follows from the next two
lemmas.
LEMMA 8.20. Let G be a reductive group with simply connected derived group, and let
T = G,G
der
. If H
1
(, T)
l_o
H
1
(
l
, T) is injective, then an element of H
1
(, G)
that becomes trivial in H
1
(
l
, G) for all l is itself trivial.
PROOF. Because G
der
is simply connected, H
1
(
l
, G
der
) = 0 for l ,= oand H
1
(, G
der
)
H
1
(1, G
der
) is injective (5.19). Using this, we obtain a commutative diagram with exact
rows
T() H
1
(, G
der
) H
1
(, G) H
1
(, T)
_
injective
_
injective
_
G(1) T(1) H
1
(1, G
der
)
l
H
1
(
l
, G)
l
H
1
(
l
, T).
If an element c of H
1
(, G) becomes trivial in all H
1
(
l
, G), then a diagram chase shows
that it arises from an element c
t
of H
1
(, G
der
) whose image c
t
o
in H
1
(1, G
der
) maps to
the trivial element in H
1
(1, G). The image of G(1) in T(1) contains T(1)
(see 5.1),
and the real approximation theorem (5.4) shows that T() T(1)
= T(1). Therefore,
there exists a t c T() whose image in H
1
(1, G
der
) is c
t
o
. Then t . c
t
in H
1
(, G
der
),
which shows that c is trivial.
LEMMA 8.21. Let (G, X) be a simple PEL Shimura datum of type Aeven or C, and let
T = G,G
der
. Then H
1
(, T)
l_o
H
1
(
l
, T) is injective.
PROOF. For G of type Aeven, T = Ker((G
m
)
F
Nm
F=k
(G
m
)
F
0
) G
m
. The group
H
1
(, G
m
) = 0, and the map on H
1
s of the rst factor is
F
0
, NmF
0
, NmF
.
This is injective (CFT, VIII 1.4).
For G of type C, T = G
m
, and so H
1
(, T) = 0.
PEL modular varieties 83
PEL modular varieties
Let B be a semisimple algebra over with a positive involution +, and let (V, [) be a sym-
plectic (B, +)-module. Let K be a compact open subgroup of G(A
(
). There exists an al-
gebraic variety M
K
over C classifying the isomorphism classes of quadruples (A, s, i, jK)
satisfying (a) and (b) of (8.18) (but not necessarily condition (**)), which is called the PEL
modular variety attached to (B, +, V, [). In the simple cases (Aeven) and (C), Proposition
8.17 shows that M
K
coincides with Sh
K
(G, X), but in general it is a nite disjoint union
of Shimura varieties.
NOTES. The theory of Shimura varieties of PEL-type is worked out in detail in several
papers of Shimura, for example, Shimura 1963, but in a language somewhat different from
ours. The above account follows Deligne 1971c, ~~5,6. See also Zink 1983 and Kottwitz
1992, ~~14.
84 9 GENERAL SHIMURA VARIETIES
9 General Shimura varieties
Abelian motives
Let Hod() be the category of polarizable rational hodge structures. It is an abelian subcat-
egory of the category of all rational hodge structures closed under the formation of tensor
products and duals.
Let V be a variety over C whose connected components are abelian varieties, say V =
V
i
with V
i
an abelian variety. Recall
64
that for manifolds M
1
and M
2
,
H
r
(M
1
. M
2
, )
= H
r
(M
1
, ) H
r
(M
2
, ).
For each connected component V
of V ,
H
+
(V
, )
=
_
H
1
(V
, )
= Hom
Q
(
_
H
1
(V
, ), )
(see 6.4). Therefore, H
+
(V, ) acquires a polarizable hodge structure fromthat on H
1
(V, ).
We write H
+
(V, )(m) for the hodge structure H
+
(V, ) (m) (see 2.6).
Let (W, h) be a rational Hodge structure. An endomorphism e of (W, h) is an idempo-
tent if e
2
= e. Then
(W, h) = Im(e) Im(1 e)
(direct sum of rational hodge structures).
An abelian motive over C is a triple (V, e, m) in which V is a variety over C whose
connected components are abelian varieties, e is an idempotent in End(H
+
(V, )), and
m c Z. For example, let A be an abelian variety; then the projection
H
+
(A, ) H
i
(A, ) H
+
(A, )
is an idempotent e
i
, and we denote (A, e
i
, 0) by h
i
(A).
Dene Hom((V, e, m), (V
t
, e
t
, m
t
)) to be the set of maps H
+
(V, ) H
+
(V
t
, ) of
the forme
t
e with a homomorphismH
+
(V, ) H
+
(V
t
, ) of degree d = m
t
m.
Moreover, dene
(V, e, m) (V
t
, e
t
, m) = (V . V
t
, e e
t
, m)
(V, e, m) (V
t
, e
t
, m) = (V V
t
, e e
t
, mm
t
)
(V, e, m)
= (V, e
t
, d m) if V is purely d-dimensional.
For an abelian motive (V, e, m) over C, let H(V, e, m) = eH
+
(V, )(m). Then (V, e, m) .
H(V, e, m) is a functor from the category of abelian motives AM to Hod() commuting
with , , and
. We say that a rational hodge structure is abelian if it is in the essen-
tial image of this functor, i.e., if it is isomorphic to H(V, e, m) for some abelian motive
(V, e, m). Every abelian hodge structure is polarizable.
64
The set of singular simplexes in M is the disjoint union of the similar sets for M
1
and M
2
. Therefore,
the complex of singular cochains for M is the direct sum of the similar complexes for M
1
and M
2
.
Shimura varieties of abelian type 85
PROPOSITION 9.1. Let Hod
ab
() be the full subcategory of Hod() of abelian hodge
structures. Then Hod
ab
() is the smallest strictly full subcategory of Hod() contain-
ing H
1
(A, ) for each abelian variety A and closed under the formation of direct sums,
subquotients, duals, and tensor products; moreover, H: AM Hod
ab
() is an equivalence
of categories.
PROOF. Straightforward from the denitions.
For a description of the essential image of H, see Milne 1994, 1.27.
Shimura varieties of abelian type
Recall (~6) that a symplectic space (V, [) over denes a connected Shimura datum
(S([), X([)
into X([)
.
(b) A connected Shimura datum(H, X
i
)
of primitive abelian type and an isogeny
i
H
i
H carrying
i
X
i
into X.
(b) A Shimura datum (G, X) is of abelian type if (G
der
, X
) is of abelian type.
(c) The (connected) Shimura variety attached to a (connected) Shimura datumof abelian
type is said to be of abelian type.
PROPOSITION 9.3. Let (G, X) be a Shimura datum, and assume
(a) the weight n
X
is rational SV4 and Z(G)
(
-
multiple of s
0
and each t
i
to s
i
,
86 9 GENERAL SHIMURA VARIETIES
satisfying the following condition:
(**) there exists an isomorphism a: H(A) V sending s
0
to a
-multiple
of [, each s
i
to t
i
, and h onto an element of X.
PROOF. With A replaced by a hodge structure, this can be proved by an elementary argu-
ment (cf. 6.3, 7.3), but (9.3) shows that the hodge structures arising are abelian, and so can
be replaced by abelian motives (9.1). For more details, see Milne 1994, Theorem 3.31.
Classication of Shimura varieties of abelian type
Deligne (1979) classies the connected Shimura data of abelian type. Let (G, X
) be a
connected Shimura datum with G simple. If G
ad
is of type A, B, or C, then (G, X
) is
of abelian type. If G
ad
is of type E
6
or E
7
, then (G, X
) may or may not be of abelian type. There are two problems that may
arise.
(a) Let G be the universal covering group of G
ad
. There may exist homomorphisms
(G, X
) (S([), X([)
into X([)
. Then (G,N
t
, X
n-1
D
n
(1): (n _ 4)
n-2
H
H
H
H
H
-
n-1
D
n
(n): - (n _ 4)
n-2
H
H
H
H
H
n
D
n
(n1): Same as D
n
(n) by with
n-1
and
n
interchanged (rotation about the horizontal
axis).
Nodes marked by squares are special (p21), and nodes marked by stars correspond to
symplectic representations. The number in parenthesis indicates the position of the special
node. As is explained in ~1, the projection of X
=I
c
H
=I
nc
GL
2
(1).
(a) If B = M
2
(F), then (G, X) is of PEL-type, and Sh
K
(G, X) classies isomorphism
classes of quadruples (A, i, t, jK) in which Ais an abelian variety of dimension d = [F: ]
and i is a homomorphism homomorphism i: F End(A) . These Shimura varieties
are called Hilbert (or Hilbert-Blumenthal) varieties, and whole books have been written
about them.
(b) If B is a division algebra, but I
c
= 0, then (G, X) is again of PEL-type, and
Sh
K
(G, X) classies isomorphism classes of quadruples (A, i, t, jK) in which A is an
abelian variety of dimension 2[F: ] and i is a homomorphism i: B End(A) . In
this case, the varieties are projective. These varieties have also been extensively studied.
(c) If B is a division algebra and I
c
,= 0, then (G, X) is of abelian type, but the weight is
not dened over . Over 1, the weight map n
X
sends a c 1to the element of (F 1)
:F-R
1 with component 1 for : c I
c
and component a for : c I
nc
. Let T be the torus
over with T() = F
. Then n
X
: G
m
T
R
is dened over the subeld L of
whose xed group is the subgroup of Gal(,) stabilizing I
c
I
c
. I
nc
. On choosing a
rational representation of G, we nd that Sh
K
(G, X) classies certain isomorphism classes
of hodge structures with additional structure, but the hodge structures are not motivic
they do not arise in the cohomology of algebraic varieties (they are not rational hodge
structures).
65
65
Summary: MV=modular variety; SV=shimura variety; =rational weight.
{SV of abelian type} {SV}
' '
{PEL SV} {SV of hodge type} {SV of abelian type, } {SV, }
'
{PEL MV simple type A,C} {PEL MV}
88 10 COMPLEX MULTIPLICATION: THE SHIMURA-TANIYAMA FORMULA
10 Complex multiplication: the Shimura-Taniyama for-
mula
Where we are headed
Let V be a variety over . For any o c Gal(
al
,) and P c V (
al
), the point oP c
V (
al
). For example, if V is the subvariety of A
n
dened by equations
(X
1
, . . . , X
n
) = 0, c [X
1
, . . . , X
n
],
then
(a
1
, . . . , a
n
) = 0 == (oa
1
, . . . , oa
n
) = 0
(apply o to the rst equality). Therefore, if we have a variety V over
al
that we suspect is
actually dened over , then we should be able to describe an action of Gal(
al
,) on its
points V (
al
).
Let E be a number eld contained in C, and let Aut(C,E) denote the group of automor-
phisms of C (as an abstract eld) xing the elements of E. Then a similar remark applies:
if a variety V over C is dened by equations with coefcients in E, then Aut(C,E) will
act on V (C). Now, I claim that all Shimura varieties are dened (in a natural way) over
specic number elds, and so I should be able to describe an action of a big subgroup of
Aut(C,) on their points. If, for example, the Shimura variety is of hodge type, then there
is a set M
K
whose elements are abelian varieties plus additional data and a map
(A, . . .) .P(A, . . .): M
K
Sh
K
(G, X)(C)
whose bres are the isomorphism classes in M
K
. On applying o c Aut(C,) to the
coefcients of the polynomials dening A, . . ., we get a new triple (oA, . . .) which may
or may not lie in M
K
. When it does we dene
o
P(A, . . .) to be P(oA, . . .). Our task
will be to show that, for some specic eld E, this does give an action of Aut(C,E) on
Sh
K
(G, X) and that the action does arise from a model of Sh
K
(G, X) over E.
For example, for P c I(1)H
1
,
o
P is the point such that j(
o
P) = o (j(P)). If j were
a polynomial with coefcients in Z (rather than a power series with coefcients in Z), we
would have j(oP) = oj(P) with the obvious meaning of oP, but this is denitely false (if
o is not complex conjugation, then it is not continuous, nor even measurable).
You may complain that our description of the action of Aut(C,E) on Sh(G, X)(C)
is not explicit, but I contend that there can not exist a completely explicit description of
the action. What are the elements of Aut(C,E)? To construct them, we can choose a
transcendence basis B for C over E, choose a permutation of the elements of B, and
extend the resulting automorphism of (B) to its algebraic closure C. But proving the
existence of transcendence bases requires the axiom of choice (e.g., FT, 8.13), and so we
can have no explicit description of, or way of naming, the elements of Aut(C,E), and
hence no completely explicit description of the action is possible.
However, all is not lost. Abelian class eld theory names the elements of Gal(E
ab
,E),
where E
ab
is a maximal abelian extension of E. Thus, if we suspect that a point P has
coordinates in E
ab
, the action of Aut(C,E) on it will factor through Gal(E
ab
,E), and we
may hope to be able to describe the action of Aut(C,E) explicitly. This the theory of
complex multiplication allows us to do for certain special points P.
Review of abelian varieties 89
Review of abelian varieties
The theory of abelian varieties is very similar to that of elliptic curves just replace E
with A, 1 with g (the dimension of A), and, whenever E occurs twice, replace one copy
with the dual A
of A.
Thus, for any m not divisible by the characteristic of the ground eld k,
A(k
al
)
m
- (Z,mZ)
2g
. (40)
Here A(k
al
)
m
consists of the elements of A(k
al
) killed by m. Hence, for ,= char(k),
T
I
A
df
= lim
A(k
al
)
I
n
is a free Z
I
-module of rank 2g, and
V
I
(A)
df
= T
I
A
Z
`
I
is a
I
-vector space of dimension 2g. In characteristic zero, we set
T
(
A =
T
I
A = lim
m
A(k
al
)
m
,
V
(
A = T
(
Z
=
(V
I
A: T
I
A) (restricted topological product).
They are, respectively, a free
Z-module of rank 2g and a free A
(
-module of rank 2g. The
galois group Gal(k
al
,k) acts continuously on these modules.
For an endomorphism a of an abelian variety A, there is a unique monic polynomial
P
a
(T) with integer coefcients (the characteristic polynomial of a) such that [P
a
(n)[ =
deg(a n) for all n c Z. Moreover, P
a
is the characteristic polynomial of a acting on V
I
A
( ,= char(k)).
For an abelian variety A over a eld k, the tangent space Tgt
0
(A) to A at 0 is a vector
space over k of dimension g. As we noted in ~6, when k = C, the exponential map denes
a surjective homomorphism Tgt
0
(A) A(C) whose kernel is a lattice in Tgt
0
(A). Thus
A(C)
m
=
1
m
,
= ,m, and
T
I
A
=
Z
Z
I
, V
I
A
=
Z
I
, T
(
A =
Z
Z, V
(
A =
Z
A
(
. (41)
An endomorphism a of A denes a C-linear endomorphism (da)
0
= of Tgt
0
(A) such
that () (see 6.5), and P
a
(T) is the characteristic polynomial of on .
For abelian varieties A, B, Hom(A, B) is a torsion free Z-module of nite rank. We let
AV(k) denote the category of abelian varieties and homomorphisms over k and AV
0
(k) the
category with the same objects but with
Hom
AV
0
(k)
(A, B) = Hom
0
(A, B) = Hom
AV(k)
(A, B) .
An isogeny of abelian varieties is a surjective homomorphism with nite kernel. A homo-
morphism of abelian varieties is an isogeny if and only if it becomes an isomorphism in the
category AV
0
. Two abelian varieties are said to be isogenous if there is an isogeny from
one to the other this is an equivalence relation.
90 10 COMPLEX MULTIPLICATION: THE SHIMURA-TANIYAMA FORMULA
An abelian variety A over a eld k is simple if it contains no nonzero proper abelian
subvariety. Every abelian variety is isogenous to a product of simple abelian varieties. If A
and B are simple, then every nonzero homomorphism from Ato B is an isogeny. It follows
that End
0
(A) is a division algebra when A is simple and a semisimple algebra in general.
NOTES. For a detailed account of abelian varieties over algebraically closed elds, see
Mumford 1970, and for a summary over arbitrary elds, see Milne 1986.
CM elds
A number eld E is a CM (or complex multiplication) eld if it is a quadratic totally imag-
inary extension of a totally real eld F. Let a . a
+
denote the nontrivial automorphism
of E xing F. Then ,(a
+
) = ,(a) for every ,: E C. We have the following picture:
E
Q
1 - C C
[ [
F
Q
1 - 1 1
(42)
The involution + is positive (in the sense of 8.11), because we can compute Tr
E
Q
R{F
Q
R
(b
+
b)
on each factor on the right, where it becomes Tr
C{R
(zz) = 2[z[
2
> 0. Thus, we are in the
PEL situation considered in ~8.
Let E be a CM-eld with largest real subeld F. Each embedding of F into 1 will
extend to two conjugate embeddings of E into C. A CM-type for E is a choice of one
element from each conjugate pair {, }. In other words, it is a subset Hom(E, C)
such that
Hom(E, C) = . (disjoint union, = { [ c }).
Because E is quadratic over F, E = F[] with a root of a polynomial X
2
aX b.
On completing the square, we obtain an such that
2
c F
. Then
+
= . Such an
element of E is said to be totally imaginary (its image in C under every embedding is
purely imaginary).
Abelian varieties of CM-type
Let E be a CM-eld of degree 2g over . Let A be an abelian variety of dimension g over
C, and let i be a homomorphism E End
0
(A). If
Tr(i(a) [ Tgt
0
(A)) =
=
(a), all a c E, (43)
for some CM-type of E, then (A, i) is said to be of CM-type (E, ).
REMARK 10.1. (a) In fact, (A, i) will always be of CM-type for some . Recall (p71) that
A(C)
= Tgt
0
(A), with a lattice in Tgt
0
(A) (so 1
= Tgt
0
(A)). Moreover,
= H
1
(A, )
1
= H
1
(A, 1),
= Tgt
0
(A)
C = H
1
(A, C)
= H
-1,0
H
0,-1
= Tgt
0
(A) Tgt
0
(A).
Abelian varieties of CM-type 91
NowH
1
(A, ) is a one-dimensional vector space over E, and so H
1
(A, C)
:E-C
C
where C
=
C
, a .(a)
=
.
PROPOSITION 10.2. The image (O
E
) of O
E
in C
,(O
E
)
is an abelian variety A
: E
End
0
(A
, i
).
PROOF. We have
O
E
Z
1
= O
E
Z
Q
1
= E
Q
1
er|-(...,re,...)
~
=
C
,
and so (O
E
) is a lattice in C
.
To show that the quotient is an abelian variety, we have to exhibit a riemann form (6.7).
Let be a totally imaginary element of E. The weak approximation theorem allows us to
choose so that `() > 0 for c , and we can multiply it by an integer (in N) to make
it an algebraic integer. Dene
[(u, :) = Tr
E{Q
(u:
+
), u, : c O
E
.
Then [(u, :) c Z. The remaining properties can be checked on the right of (42). Here [
takes the form [ =
=
[
, where
[
(u, :) = Tr
C{R
(
u :),
= (), u, : c C.
Because is totally imaginary,
[
(u, :) =
(u: u:) c 1,
from which it follows that [
(u, u) = 0, [
(iu, i:) = [
is an abelian variety.
An element c O
E
acts on C
. The map C
,(O
E
) denes an isomorphism C
=
Tgt
0
(C
) Tgt
0
(A
, i
) is of CM-
type (E, ).
Finally, let (A, i) be of CM-type (E, ). The condition (43) means that Tgt
0
(A) is
isomorphic to C
as an E
Q
C-module. Therefore, A(C) is a quotient of C
by a lattice
such that is stable under the action of E on C
,
N
C
,N C
,(O
E
).
92 10 COMPLEX MULTIPLICATION: THE SHIMURA-TANIYAMA FORMULA
PROPOSITION 10.3. Let (A, i) be an abelian variety of CM-type (E, ) over C. Then
(A, i) has a model over
al
, uniquely determined up to isomorphism.
PROOF. Let k D be algebraically closed elds of characteristic zero. For an abelian
variety A over k, the torsion points in A(k) are zariski dense, and the map on torsion
points A(k)
tors
A(D)
tors
is bijective (see (40)), and so every regular map A
D
W
D
(W a variety over k) is xed by the automorphisms of D,k and is therefore dened over k
(AG 14.7; see also 13.1 below). It follows that A .A
D
: AV(k) AV(D) is fully faithful.
It remains to show that every abelian variety (A, i) of CM-type over Carises from a pair
over
al
. The polynomials dening Aand i have coefcients in some subring Rof Cthat is
nitely generated over
al
. According to the Hilbert Nullstellensatz, a maximal ideal m of
Rwill have residue eld
al
, and the reduction of (A, i) mod m is called a specialization of
(A, i). Any specialization (A
t
, i
t
) of (A, i) to a pair over
al
with A
t
nonsingular will still
be of CM-type (E, ), and therefore (see 10.2) there exists an isogeny (A
t
, i
t
)
C
(A, i).
The kernel H of this isogeny is a subgroup of A
t
(C)
tors
= A
t
(
al
)
tors
, and (A
t
,H, i) will
be a model of (A, i) over
al
.
REMARK 10.4. The proposition implies that, in order for an elliptic curve A over C to be
of CM-type, its j-invariant must be algebraic.
66
Let A be an abelian variety of dimension g over a subeld k of C, and let i: E
End
0
(A) be a homomorphism with E a CM-eld of degree 2g. Then Tgt
0
(A) is a k-
vector space of dimension g on which E acts k-linearly, and, provided k is large enough to
contain all conjugates of E, it will decompose into one-dimensional k-subspaces indexed
by a subset of Hom(E, k). When we identify with a subset of Hom(E, C), it becomes
a CM-type, and we again say (A, i) is of CM-type (E, ).
Let A be an abelian variety over a number eld K. We say that A has good reduction
at P if it extends to an abelian scheme over O
K,P
, i.e., a smooth proper scheme over O
K,P
with a group structure. In down-to-earth terms this means the following: embed A as a
closed subvariety of some projective space P
n
K
; for each polynomial P(X
0
, . . . , X
n
) in the
homogeneous ideal a dening A P
n
K
, multiply P by an element of K so that it (just)
lies in O
K,P
[X
0
, . . . , X
n
] and let P denote the reduction of P modulo P; the Ps obtained
in this fashion generate a homogeneous a ideal in k[X
0
, . . . , X
n
] where k = O
K
,P; the
abelian variety A has good reduction at P if it is possible to choose the projective embed-
ding of A so that the zero set of a is an abelian variety A over k. Then A is called the
reduction of A at P. It can be shown that, up to a canonical isomorphism, A is indepen-
dent of all choices. For ,= char(k), V
I
(A)
= V
I
(A). There is an injective homorphism
End(A) End(A) compatible with V
I
(A)
= V
I
(A) (both are reduction maps).
66
Consider the curve
E : Y
2
(j 1728)XY = X
3
36(j 1728)
2
X (j 1728)
3
where j c C is transcendental. Specializing E to
al
amounts to replacing j with an algebraic number, say,
j
t
, in the equation. Since E has j-invariant j, and the specialized curve E
t
has j-invariant j
t
, we see that
E
t
C
is not isomorphic to E.
Abelian varieties over a nite eld 93
PROPOSITION 10.5. Let (A, i) be an abelian variety of CM-type (E, ) over a number
eld K C, and let P be a prime ideal in O
K
. Then, after possibly replacing K by a
nite extension, A will have good reduction at P.
PROOF. We use the N eron (alias, Ogg-Shafarevich) criterion (Serre and Tate 1968, Theo-
rem 1):
an abelian variety over a number eld K has good reduction at P if for some
prime ,= char(O
K
,P), the inertia group I at P acts trivially on T
I
A.
In our case, V
I
A is a free E
Q
I
-module of rank 1 because H
1
(A
C
, ) is a one-
dimensional vector space over E and V
I
A
= H
1
(A
C
, )
I
(see (41)). Therefore,
E
Q
I
is its own centralizer in End
Q
`
(V
I
A) and the representation of Gal(
al
,) on
V
I
A has image in (E
I
)
. But such
a subgroup will have a pro- subgroup of nite index. Since I has a pro-p subgroup of
nite index (ANT, 7.5), this shows that image of I is nite. After K has been replaced by
a nite extension, the image of I will be trivial, and N erons criterion applies.
Abelian varieties over a nite eld
Let F be an algebraic closure of the eld F
p
of p-elements, and let F
q
be the subeld of F
with q = p
m
elements. An element a of F lies in F
q
if and only if a
q
= a. Recall that, in
characteristic p, (X Y )
p
= X
p
Y
p
. Therefore, if (X
1
, . . . , X
n
) has coefcients in
F
q
, then
(X
1
, . . . , X
n
)
q
= (X
q
1
, . . . , X
q
n
), (a
1
, . . . , a
n
)
q
= (a
q
1
, . . . , a
q
n
), a
i
c F.
In particular,
(a
1
, . . . , a
n
) = 0 == (a
q
1
, . . . , a
q
n
) = 0, a
i
c F.
PROPOSITION 10.6. There is a unique way to attach to every variety V over F
q
a regular
map
V
: V V such that
(a) for any regular map : V W,
V
=
W
;
(b) _
A
n is the map (a
1
, . . . , a
n
) .(a
q
1
, . . . , a
q
n
).
PROOF. For an afne variety V = SpecmA, dene
V
be the map corresponding to the
F
q
-homomorphism x .x
q
: A A. The rest of the proof is straightforward.
The map
V
is called the Frobenius map of V .
THEOREM 10.7 (WEIL 1948). For an abelian variety A over F
q
, End
0
(A) is a nite-
dimensional semisimple -algebra with
A
in its centre. For every embedding ,: [
A
]
C, [,(
A
)[ = q
1
2
.
PROOF. See, for example, Milne 1986, 19.1.
94 10 COMPLEX MULTIPLICATION: THE SHIMURA-TANIYAMA FORMULA
If A is simple, [
A
] is a eld (p89), and
A
is an algebraic integer in it (p89). An
algebraic integer such that [,()[ = q
1
2
for all embeddings ,: [] C is called a Weil
q-integer (formerly, Weil q-number).
For a Weil q-integer ,
,() ,() = q = ,() ,(q,), all ,: [] C,
and so ,(q,) = ,(). It follows that the eld ,([_]) is stable under complex con-
jugation and that the automorphism of [] induced by complex conjugation sends to
q, and is independent of ,. This implies that [] is a CM-eld (the typical case), , or
[
_
p].
LEMMA 10.8. Let and
t
be Weil q-integers lying in the same eld E. If ord
() =
ord
(
t
) for all :[p, then
t
= ( for some root of 1 in E.
PROOF. As noted above, there is an automorphism of [] sending to q,. Therefore
q, is also an algebraic integer, and so ord
= [
t
[
= 1 for all :[o. But in the course of proving the unit theorem, one shows that
such a unit has to be root of 1 (ANT, 5.6).
The Shimura-Taniyama formula.
LEMMA 10.9. Let (A, i) be an abelian variety of CM-type (E, ) over a number eld
k C having good reduction at P O
k
to (A, {) over O
k
,P = F
q
. Then the Frobenius
map
A
of A lies in {(E).
PROOF. Let =
A
. It sufces to check that lies in {(E) after tensoring
67
with
I
. As
we saw in the proof of (10.5), V
I
A is a free E
Q
I
-module of rank 1. It follows that V
I
A
is also a free E
Q
I
-module of rank 1 (via {). Therefore, any endomorphism of V
I
A
commuting with the action of E
I
will lie in E
I
.
Thus, from (A, i) and a prime P of k at which A has good reduction, we get a Weil
q-integer c E.
THEOREM 10.10 (SHIMURA-TANIYAMA).
68
In the situation of the lemma, assume that k
is galois over and contains all conjugates of E. Then for all primes : of E dividing p,
ord
()
ord
(q)
=
[ H
[
[H
[
(44)
where H
= {,: E k [ ,
-1
(P) = p
() ord
(
+
) = ord
(q). (45)
Moreover,
ord
(
+
) = ord
()
and
H
= H
.
Therefore, (44) is consistent with (45):
ord
()
ord
(q)
ord
(
+
)
ord
(q)
(44)
=
[ H
[ [ H
[
[H
[
=
[(' ) H
[
[H
[
= 1.
In fact, (44) is the only obvious formula for ord
i
R,p
r
i
i
with each p
i
an ideal in R, and the set of pairs (p
i
, r
i
) is uniquely
determined by M. Dene
70
[M[
R
=
p
r
i
i
. For example, for R = Z, M is a nite abelian
group and [M[
Z
is the ideal in Z generated by the order of M.
For Dedekind domains R S with S nite over R, there is a norm homomorphism
sending fractional ideals of S to fractional ideals of R (ANT, p58). It is compatible with
norms of elements, and
Nm(P) = p
((P{p)
, P prime, p = P R
Clearly,
[S,A[
R
= Nm(A) (46)
since this is true for prime ideals, and both sides are multiplicative.
PROPOSITION 10.12. Let A be an abelian variety of dimension g over F
q
, and let i be a
homomorphism from the ring of integers O
E
of a eld E of degree 2g over into End(A).
Then
[ Tgt
0
A[
O
E
= (
A
).
PROOF. Omitted (for a scheme-theoretic proof, see Giraud 1968, Th eor` eme 1).
69
Let
t
arise from second model (A
t
, i
t
). Then (A
t
, i
t
) will become E-isogenous to (A, i) over a nite
extension k
t
of k (see 10.2), from which it follows that
(
=
t(
for the degree of the residue eld
extension.
70
Better, the rst statement shows that the K-group of the category of nitely generated torsion R-modules
is canonically isomorphic to the group of fractional ideals of R, and so [M[
R
denotes the class of M in the
K-group.
96 10 COMPLEX MULTIPLICATION: THE SHIMURA-TANIYAMA FORMULA
Sketch of the proof the Shimura-Taniyama formula
We return to the situation of the Theorem 10.10. After replacing A with an isogenous
variety, we may assume i(O
E
) End(A). By assumption, there exists an abelian scheme
A over O
k,P
with generic bre A and special bre an abelian variety A. Because A is
smooth over O
k,P
, the relative tangent space of A,O
k,P
is a free O
k,P
-module T of rank
g endowed with an action of O
E
such that
T
O
k,P
k = Tgt
0
(A), T
O
k,P
O
k,P
,P = Tgt
0
(A).
Therefore,
()
10.12
=
Tgt
0
A
O
E
=
T
O
k,P
(O
k,P
,P)
O
E
. (47)
For simplicity, assume
71
that (p) =
df
PZ is unramied in E. Then the isomorphism
of E-modules
T
O
k,P
k - k
induces an isomorphism of O
E
-modules
T - O
k,P
(48)
In other words, T is a direct sum of copies of O
k,P
indexed by the elements of , and O
E
acts on the
th
copy through the map
O
E
O
k
O
k,P
.
As O
k
,P
= O
k,P
,P (ANT, 3.11), the contribution of the
th
copy to () in (47) is
[O
k
,P[
O
E
(46)
=
-1
(Nm
k{E
P).
Thus,
() =
-1
(Nm
k{E
P). (49)
It is only an exercise to derive (44) from (49).
NOTES. The original formulation of the Shimura-Taniyama theorem is in fact (49). It
is proved in Shimura and Taniyama 1961, III.13, in the unramied case using spaces of
differentials rather than tangent spaces. The proof sketched above is given in detail in
Giraud 1968, and there is a proof using p-divisible groups in Tate 1969, ~5. See also
Serre 1968, pII-28.
71
This, in fact, is the only case we need, because it sufces for the proof of the main theorem in ~10, which
in turn implies the Shimura-Taniyama formula.
97
11 Complex multiplication: the main theorem
Review of class eld theory
Classical class eld theory classies the abelian extensions of a number eld E, i.e., the
galois extensions L,E such Gal(L,E) is commutative. Let E
ab
be the composite of all
the nite abelian extensions of E inside some xed algebraic closure E
al
of E. Then E
ab
is an innite galois extension of E.
According to class eld theory, there exists a continuous surjective homomorphism (the
reciprocity or Artin map)
rec
E
: A
E
Gal(E
ab
,E)
such that, for every nite extension L of E contained in E
ab
, rec
E
gives rise to a commu-
tative diagram
E
E
rec
E
onto
Gal(E
ab
,E)
_
o|-o[L
E
E
, Nm
L{E
(A
L
)
rec
L=E
~
=
Gal(L,E).
It is determined by the following two properties:
(a) rec
L{E
(u) = 1 for every u = (u
) c A
E
such that
i) if : is unramied in L, then u
is a unit,
ii) if : is ramied in L, then u
> 0.
(b) For every prime : of E unramied in L, the id` ele
= (1, . . . , 1,
E
, and use
the weak approximation theorem to choose an a c E
that is close to
E
is contained in the kernel of rec
E
. In particular, the identity component of
[o
E
is con-
tained in the kernel, and so, when E is totally imaginary, rec
E
factors through E
E,(
.
For E = , the reciprocity map factors through
{} A
(
, and every element in
this quotient is uniquely represented by an element of
Z
(
. In this case, we get the
98 11 COMPLEX MULTIPLICATION: THE MAIN THEOREM
diagram
rec
Q
~
=
Gal(
ab
,)
_
[(
N
]
_restrict
(Z,NZ)
[a]|-(
N
|-
a
N
)
~
=
Gal([(
N
],)
(50)
which commutes with an inverse. This can be checked by writing an id` ele in the form
au as above, but it is more instructive to look at an example. Let p be a prime not dividing
N, and let
= p (1, . . . , 1, p
-1
p
, 1, . . .) c Z A
(
= A
(
.
Then c
Z
E
, write
art
E
() = rec
E
()
-1
. (51)
Now, the diagram (50) commutes. In other words,
art
Q
(y(o)) = o, for o c Gal(
ab
,),
where y is the cyclotomic character Gal(
ab
,)
Z
, which is characterized by
o( = (
y(o)
, ( a root of 1 in C
.
The reex eld and norm of a CM-type
Let (E, ) be a CM-type.
DEFINITION 11.1. The reex eld E
+
of (E, ) is the subeld of
al
characterized by any
one of the following equivalent
72
conditions:
(a) o c Gal(
al
,) xes E
+
if and only if o = ; here o = {o [ c };
(b) E
+
is the eld generated over by the elements
=
(a), a c E;
72
If o c Gal(
al
,) permutes the s, then clearly it xes all elements of the form
=
(a). Con-
versely, if
=
(a) =
=
(o)(a) for all a c E
=
(a).
Statement of the main theorem of complex multiplication 99
(c) E
+
is the smallest subeld k of
al
such that there exists a k-vector space V with an
action of E for which
Tr
k
(a[V ) =
=
(a), all a c E.
Let V be an E
+
-vector space with an action of E such that Tr
E
(a[V ) =
=
(a)
for all a c E. We can regard V as an E
+
Q
E-space, or as an E-vector space with a E-
linear action of E
+
. The reex norm is the homomorphism
73
N
: (G
m
)
E
{Q
(G
m
)
E{Q
such that
N
(a) = det
E
(a[V ), all a c E
+
.
This denition is independent of the choice of V because V is unique up to an isomorphism
respecting the actions of E and E
+
.
Let (A, i) be an abelian variety of CM-type (E, ) dened over C. According to
(11.1c) applied to Tgt
0
(A), any eld of denition of (A, i) contains E
+
.
Statement of the main theorem of complex multiplication
A homomorphism o: k D of elds denes a functor V .oV , .o, extension of
the base eld from varieties over k to varieties over D. In particular, an abelian variety A
over k equipped with a homomorphism i: E End
0
(A) denes a similar pair o(A, i) =
(oA,
o
i) over D. Here
o
i: E End(oA) is dened by
o
i(a) = o(i(a)).
Apoint P c A(k) gives a point oP c A(D), and so o denes a homomorphismo: V
(
(A)
V
(
(oA) provided that k and D are algebraically closed (otherwise one would have to
choose an extension of k to a homomorphism k
al
D
al
).
THEOREM 11.2. Let (A, i) be an abelian variety of CM-type (E, ) over C, and let o c
Aut(C,E
+
). For any s c A
,(
with art
E
(s) = o[E
+ab
, there is a unique E-linear
isogeny : A oA such that (N
E,(
. When the choice of is changed, then a is
changed only by an element of E
: Gal(
al
,E
+
) A
E,(
,E
,
73
One can show that E
+
is again a CM-eld, and that an embedding of E into
al
denes a CM-type on
E
+
. The reex norm is usually dened in terms of
+
but we will not need it.
100 11 COMPLEX MULTIPLICATION: THE MAIN THEOREM
which one checks to be a homomorphism. The map therefore factors through Gal(E
+ab
,E
+
),
and so, when composed with the reciprocity map art
E
, it gives a homomorphism
j: A
,(
,E
+
A
E,(
,E
.
We have to check that j is the homomorphism dened by N
(a)
-1
.
(b) The theorem is a statement about the E-isogeny class of (A, i) if : (A, i)
(B, j) is an E-linear isogeny, and satises the conditions of the theorem for (A, i), then
(o)
-1
satises the conditions for (B, j).
ASIDE 11.4. What happens in (11.2) when o is not assumed to x E
+
? This also is known,
thanks to Deligne and Langlands. For a discussion of this, and much else concerning
complex multiplication, see my notes Milne 1979.
101
12 Denition of canonical models
We attach to each Shimura datum (G, X) an algebraic number eld E(G, X), and we
dene the canonical model of Sh(G, X) to be an inverse system of varieties over E(G, X)
that is characterized by reciprocity laws at certain special points.
Models of varieties
Let k be a subeld of a eld D, and let V be a variety over D. A model of V over k (or a
k-structure on V ) is a variety V
0
over k together with an isomorphism : V
0D
V . We
often omit the map and regard a model as a variety V
0
over k such that V
0D
= V .
Consider an afne variety V over C and a subeld k of C. An embedding V A
n
C
denes a model of V over k if the ideal I(V ) of polynomials zero on V is generated
by polynomials in k[X
1
, . . . , X
n
], because then I
0
=
df
I(V ) k[X
1
, . . . , X
n
] is a radical
ideal, k[X
1
, . . . , X
n
],I
0
is an afne k-algebra, and V (I
0
) A
n
k
is a model of V . Moreover,
every model (V
0
, ) arises in this way because every model of an afne variety is afne.
However, different embeddings in afne space will usually give rise to different models.
For example, the embeddings
A
2
C
(x,y)
(x,y)
V (X
2
Y
2
1)
(x,y)
-
(x,y{
_
2)
A
2
C
dene the -structures
X
2
Y
2
= 1, X
2
2Y
2
= 1
on the curve X
2
Y
2
= 1. These are not isomorphic.
Similar remarks apply to projective varieties.
In general, a variety over C will not have a model over a number eld, and when it
does, it will have many. For example, an elliptic curve E over C has a model over a
number eld if and only if its j-invariant j(E) is an algebraic number, and if Y
2
Z =
X
3
aXZ
2
bZ
3
is one model of E over a number eld k (meaning, a, b c k), then
Y
2
Z = X
3
ac
2
XZ
2
bc
3
Z
3
is a second, which is isomorphic to the rst only if c is a
square in k.
The reex eld
For a reductive group G over and a subeld k of C, we write C(k) for the set of G(k)-
conjugacy classes of cocharacters of G
k
dened over k:
C(k) = G(k) Hom(G
m
, G
k
).
A homomorphism k k
t
induces a map C(k) C(k
t
); in particular, Aut(k
t
,k) acts on
C(k
t
).
LEMMA 12.1. Assume G splits over k, so that it contains a split maximal torus T , and let
W be the Weyl group N
G(k)
(T),C
G(k)
(T) of T. Then the map
W Hom(G
m
, T
k
) G(k) Hom(G
m
, G
k
)
is bijective.
102 12 DEFINITION OF CANONICAL MODELS
PROOF. As any two maximal split tori are conjugate (Springer 1998, 15.2.6), the map is
surjective. Let j and j
t
be cocharacters of T that are conjugate by an element of G(k),
say, j = ad(g) j
t
with g c G(k). Then ad(g)(T) and T are both maximal split
tori in the centralizer
74
C of j(G
m
), which is a connected reductive group (ibid., 15.3.2).
Therefore, there exists a c c C(k) such that ad(cg)(T) = T. Now cg normalizes T and
ad(cg) j
t
= j, which proves that j and j
t
are in the same W-orbit.
Let (G, X) be a Shimura datum. For each x c X, we have a cocharacter j
x
of G
C
:
j
x
(z) = h
xC
(z, 1).
A different x c X will give a conjugate j
x
, and so X denes an element c(X) of C(C).
Neither Hom(G
m
, T
Q
al ) nor W changes when we replace C with the algebraic closure
al
of in C, and so the lemma shows that c(X) contains a j dened over
al
and that the
G(
al
)-conjugacy class of j is independent of the choice of j. This allows us to regard
c(X) as an element of C(
al
).
DEFINITION 12.2. The reex (or dual) eld E(G, X) is the eld of denition of c(X),
i.e., it is the xed eld of the subgroup of Gal(
al
,) xing c(X) as an element of C(
al
)
(or stabilizing c(X) as a subset of Hom(G
m
, G
Q
al )).
Note that the reex eld a subeld of C.
REMARK 12.3. (a) Any subeld k of
al
splitting G contains E(G, X). This follows from
the lemma, because W Hom(G
m
, T) does not change when we pass from k to
al
. If
follows that E(G, X) has nite degree over .
(b) If c(X) contains a j dened over k, then k E(G, X). Conversely, if G is quasi-
split over k and k E(G, X), then c(X) contains a j dened over k (Kottwitz 1984,
1.1.3).
(c) Let (G, X)
i
(G
t
, X
t
) be an inclusion of Shimura data. Suppose o xes c(X), and
let j c c(X). Then oj = g j g
-1
for some g c G(
al
), and so, for any g
t
c G
t
(
al
),
o(g
t
(i j) g
t-1
) = (og
t
)(i(g)) i j (i(g))
-1
(og
t
)
-1
c c(X
t
).
Hence o xes c(X
t
), and we have shown that
E(G, X) E(G
t
, X
t
).
EXAMPLE 12.4. (a) Let T be a torus over , and let h be a homomorphismS T
R
. Then
E(T, h) is the eld of denition of j
h
, i.e., the smallest subeld of C over which j
h
is
dened.
(b) Let (E, ) be a CM-type, and let T be the torus (G
m
)
E{Q
, so that T() = E
and
T(1) = (E
Q
1)
= (C
, (e r) .((e) r)
=
.
74
Certainly T C. Let t c T(k
al
) and a c G
m
(k
al
). Then
gtg
-1
j(a) = gt j
t
(a) g
-1
= g j
t
(a) tg
-1
= j(a) gtg
-1
,
and so gTg
-1
C.
Special points 103
Dene h
: C
is
C
T(C)
= (C
(C
z . (z, . . . , z, 1, . . . , 1)
Therefore, oj
= j
splits over a CM-eld (i.e., SV4 and SV6 hold), the special
points and special pairs are called CM points and CM pairs.
76
REMARK 12.6. Let T be a maximal torus of G such that T(1) xes x, i.e., such that
ad(t ) h
x
= h
x
for all t c T(1). Because T
R
is its own centralizer in G
R
, this implies that
h
x
(C
1
= C 1. Then G(1) acts on H
1
by
_
a b
c d
_
z =
az b
cz d
.
Suppose that z c C 1 generates a quadratic imaginary extension E of . Using the -
basis {1, z} for E, we obtain an embedding E M
2
(), and hence a maximal subtorus
(G
m
)
E{Q
G. As (G
m
)
E{Q
(1) xes z, this shows that z is special. Conversely, if z c H
1
is special, then [z] is a eld of degree 2 over .
The homomorphism r
x
Let T be a torus over and let j be a cocharacter of T dened over a nite extension E
of . For Q c T(E), the element
p:E-Q
al
,(Q) of T(
al
) is stable under Gal(
al
,)
and hence lies in T(). Let r(T, j) be the homomorphism (G
m
)
E{Q
T such that
r(T, j)(P) =
p:E-Q
al
,(j(P)), all P c E
. (52)
Let (T, x) (G, X) be a special pair, and let E(x) be the eld of denition of j
x
. We
dene r
x
to be the homomorphism
A
E(x)
r(T,)
T(A
Q
)
project
T(A
Q,(
). (53)
Let a c A
E(x)
, and write a = (a
o
, a
(
) c (E(x)
Q
1)
E(x),(
; then
r
x
(a) =
p:E-Q
al
,(j
x
(a
(
)).
Denition of a canonical model
For a special pair (T, x) (G, X), we have homomorphisms ((51),(53)),
art
E(x)
: A
E(x)
Gal(E(x)
ab
,E(x))
r
x
: A
E(x)
T(A
(
).
DEFINITION 12.8. Let (G, X) be a Shimura datum, and let K be a compact open subgroup
of G(A
(
). A model M
K
(G, X) of Sh
K
(G, X) over E(G, X) is canonical if, for every
special pair (T, x) (G, X) and a c G(A
(
), [x, a]
K
has coordinates in E(x)
ab
and
o[x, a]
K
= [x, r
x
(s)a]
K
, (54)
for all
77
o c Gal(E(x)
ab
,E(x))
s c A
E(x)
_
with art
E(x)
(s) = o.
In other words, M
K
(G, X) is canonical if every automorphism o of C xing E(x) acts on
[x, a]
K
according to the rule (54) where s is any id` ele such that art
E(x)
(s) = o[E(x)
ab
.
77
If q c G() and qx = x, then [x, qa]
K
= [x, a]
K
, and so, according to (54), we should have
[x, r
x
(s)qa]
K
= [x, r
x
(s)a]
K
. Following Deligne 1979, 2.2.4, I leave it to the reader to check this.
Examples: Shimura varieties dened by tori 105
REMARK 12.9. Let (T
1
, x) and (T
2
, x) be special pairs in (G, X) (with the same x). Then
(T
1
T
2
, x) is also a special pair, and if the condition in (54) holds for one of (T
1
T
2
, x),
(T
1
, x), or (T
2
, x), then it holds for all three. Therefore, in stating the denition, we could
have considered only special pairs (T, x) with, for example, T minimal among the tori
such that T
R
contains h
x
(S).
DEFINITION 12.10. Let (G, X) be a Shimura datum.
(a) A model of Sh(G, X) over a subeld k of C is an inverse system M(G, X) =
(M
K
(G, X))
K
of varieties over k endowed with a right action of G(A
(
) such that M(G, X)
C
=
Sh(G, X) (with its G(A
(
) action).
(b) A model M(G, X) of Sh(G, X) over E(G, X) is canonical if each M
K
(G, X) is
canonical.
Examples: Shimura varieties dened by tori
For a eld k of characteristic zero, the functor V . V (k
al
) is an equivalence from the
category of zero-dimensional varieties over k to the category of nite sets endowed with
a continuous action of Gal(k
al
,k). Continuous here just means that the action factors
through Gal(L,k) for some nite galois extension L of k contained in k
al
. In particular,
to give a zero-dimensional variety over an algebraically closed eld of characteristic zero
is just to give a nite set. Thus, a zero-dimensional variety over C can be regarded as a
zero-dimensional variety over
al
, and to give a model of V over a number eld E amounts
to giving a continuous action of Gal(
al
,) on V (C).
Tori
Let T be a torus over , and let h be a homomorphism S T
R
. Then (T, h) is a Shimura
datum, and E =
df
E(T, h) is the eld of denition of j
h
. In this case
Sh
K
(T, h) = T(){h} T(A
(
),K
is a nite set (see 5.22), and (54) denes a continuous action of Gal(E
ab
,E) on Sh
K
(T, h).
This action denes a model of Sh
K
(T, h) over E, which, by denition, is canonical.
CM-tori
Let (E, ) be a CM-type, and let (T, h
) = E
+
, and r(T, j
): (G
m
)
E
{Q
(G
m
)
E{Q
is the reex norm N
.
Let K be a compact open subgroup of T(A
(
). The Shimura variety Sh
K
(T, h
) clas-
sies isomorphism classes of triples (A, i, jK) in which (A, i) is an abelian variety over
C of CM-type (E, ) and j is an E A
(
-linear isomorphism V (A
(
) V
(
(A). An
isomorphism (A, i, jK) (A
t
, i
t
, j
t
K) is an E-linear isomorphism A A
t
in AV
0
(C)
sending jK to j
t
K. To see this, let V be a one-dimensional E-vector space. The action
of E on V realizes T as a subtorus of GL(V ). If (A, i) is of CM-type (E, ), then there
106 12 DEFINITION OF CANONICAL MODELS
exists an E-homomorphism a: H
1
(A, ) V carrying h
A
to h
= T
E
(A
(
).
The map (A, i, j) .[g] gives the bijection.
In (10.3) and its proof, we showed that the functor (A, i) .(A
C
, i
C
) denes an equiv-
alence from the category of abelian varieties over
al
of CM-type (E, ) to the similar
category over C (the abelian varieties are to be regarded as objects of AV
0
). Therefore,
Sh
K
(T
E
, h
(s)g]
K
, art
E
(s) = o[E
+
.
PROPOSITION 12.11. The map (A, i, j) .[aj]
K
: M
K
Sh
K
(T
E
, h
) commutes with
the actions of Gal(
al
,E
+
).
PROOF. Let (A, i, j) c M
K
map to [aj]
K
for an appropriate isomorphisma: H
1
(A, )
V , and let o c Gal(
al
,E
+
). According to the main theorem of complex multiplica-
tion (11.2), there exists an isomorphism : A oA such that (N
(s) x) = ox
for x c V
(
(A), where s c A
E
is such that art
E
(s) = o[E
+
. Then o(A, i, j) .
[a H
1
()
=1
o j]
K
. But
V
(
()
=1
o = N
(s) = r
h
(s),
and so
[a H
1
()
-1
o j]
K
= [r
h
(s) (a j)]
K
as required.
NOTES. Our denitions coincide with those of Deligne 1979, except that we have corrected
a sign error there (it is necessary to delete inverse in ibid. 2.2.3, p269, line 10, and in
2.6.3, p284, line 21).
78
We are using that V
(
(A)
= H
1
(A, )
Q
A
(
see (41).
107
13 Uniqueness of canonical models
In this section, I sketch a proof that a Shimura variety has at most one canonical model (up
to a unique isomorphism).
Extension of the base eld
PROPOSITION 13.1. Let k be a subeld of an algebraically closed eld D of characteristic
zero. If V and W are varieties over k, then a regular map V
D
W
D
commuting with the
actions of Aut(D,k) on V (D) and W(D) arises from a unique regular map V W. In
other words, the functor
V .V
D
action of Aut(D,k) on V (D)
is fully faithful.
PROOF. See AG 14.7. [The rst step is to show that the D
Aut(D{k)
= k, which requires
Zorns lemma in general.]
COROLLARY 13.2. A variety V over k is uniquely determined (up to a unique isomor-
phism) by V
D
and the action of Aut(D,k) on V (D).
Uniqueness of canonical models
Let (G, X) be a Shimura datum.
LEMMA 13.3. There exists a special point in X.
PROOF (SKETCH). Let x c X, and let T be a maximal torus in G
R
containing h
x
(C).
Then T is the centralizer of any regular element z of Lie(T). If z
0
c Lie(G) is chosen
sufciently close to z, then the centralizer T
0
of z
0
in G will be a maximal torus in G
(Borel 1991, 18.1, 18.2), and T
0
will become conjugate
79
to T over 1:
T
0R
= gTg
-1
, some g c G(1).
Now h
gx
(S) =
df
ghg
-1
(S) T
0R
, and so gx is special.
LEMMA 13.4 (KEY LEMMA). For any nite extension L of E(G, X) in C, there exists a
special point x
0
such that E(x
0
) is linearly disjoint from L.
PROOF. See Deligne 1971b, 5.1. [The basic idea is the same as that of the proof of 13.6
above, but requires the Hilbert irreducibility theorem.]
If G = GL
2
, the lemma just says that, for any nite extension L of in C, there exists
a quadratic imaginary extension E over linearly disjoint from L. This is obvious for
example, take E = [
_
p] for any prime p unramied in L.
79
Any element sufciently close to a regular element will also be regular, which implies that T
0
is a maxi-
mal torus. Not all maximal tori in G
{R
are conjugate rather, they fall into several connected components,
from which the second statement can be deduced.
108 13 UNIQUENESS OF CANONICAL MODELS
LEMMA 13.5. For any x c X, {[x, a]
K
[ a c G(A
(
)} is dense in Sh
K
(G, X) (in the zariski
topology).
PROOF. Write
Sh
K
(G, X)(C) = G()X (G(A
(
),K)
and note that the real approximation theorem (5.4) implies that G()x is dense in X for
the complex topology, and, a fortiori, the zariski topology.
Let g c G(A
(
), and let K and K
t
be compact open subgroups such that K
t
g
-1
Kg.
Then the map T (g)
[x, a]
K
.[x, ag]
K
0 : Sh
K
(C) Sh
K
0 (C)
is well-dened.
THEOREM 13.6. If Sh
K
(G, X) and Sh
K
0 (G, X) have canonical models over E(G, X),
then T (g) is dened over E(G, X).
PROOF. After (13.1), it sufces to show that o(T (g)) = T (g) for all automorphisms o of
C xing E(G, X). Let x
0
c X be special. Then E(x
0
) E(G, X) (see 12.3b), and we
rst show that o(T (g)) = T (g) for those os xing E(x
0
). Choose an s c A
E
0
such that
art(s) = o[E(x
0
)
ab
. For a c G(A
(
),
[x
0
, a]
K
T (g)
-
[x
0
, ag]
K
0
[x
0
, r
x
0
(s)a]
K
?
o
T (g)
-
[x
0
, r
x
0
(s)ag]
K
0
?
o
commutes. Thus, T (g) and o(T (g)) agree on {[x
0
, a] [ a c G(A
(
)}, and hence on all of
Sh
K
by Lemma 13.5. We have shown that o(T (g)) = T (g) for all o xing the reex eld
of any special point, but Lemma 13.4 shows that these os generate Aut(C,E(G, X)).
THEOREM 13.7. (a) A canonical model of Sh
K
(G, X) (if it exists) is unique up to a unique
isomorphism.
(b) If, for all compact open subgroups K of G(A
(
), Sh
K
(G, X) has a canonical model,
then so also does Sh(G, X), and it is unique up to a unique isomorphism.
PROOF. (a) Take K = K
t
and g = 1 in (13.6).
(b) Obvious from (13.6).
In more detail, let (M
K
(G, X), ) and (M
t
K
(G, X),
t
) be canonical models of Sh
K
(G, X)
over E(G, X). Then the composite
M
K
(G, X)
C
Sh
K
(G, X)
01
M
t
K
(G, X)
C
is xed by all automorphisms of C xing E(G, X), and is therefore dened over E(G, X).
REMARK 13.8. In fact, one can prove more. Let a: (G, X) (G
t
, X
t
) be a morphism of
Shimura data, and suppose Sh(G, X) and Sh(G
t
, X
t
) have canonical models M(G, X) and
M(G
t
, X
t
). Then the morphism Sh(a): Sh(G, X) Sh(G
t
, X
t
) is dened over E(G, X)
E(G
t
, X
t
).
The galois action on the connected components 109
The galois action on the connected components
A canonical model for Sh
K
(G, X) will dene an action of Aut(C,E(G, X)) on the set
0
(Sh
K
(G, X)). In the case that G
der
is simply connected, we saw in ~5 that
0
(Sh
K
(G, X))
= T()Y T(A
(
),v(K)
where v: G T is the quotient of G by G
der
and Y is the quotient of T(1) by the image
T(1)
|
of Z(1) in T(1). Let h = v h
x
for any x c X. Then j
h
is certainly dened over
E(G, X). Therefore, it denes a homomorphism
r = r(T, j
h
): A
E(G,X)
T(A
Q
).
The action of o c Aut(C,E(G, X)) on
0
(Sh
K
(G, X)) can be described as follows: let
s c A
E(G,X)
be such that art
E(G,X)
(s) = o[E(G, X)
ab
, and let r(s) = (r(s)
o
, r(s)
(
) c
T(1) T(A
(
); then
o[y, a]
K
= [r(s)
o
y, r(s)
(
a]
K
, for all y c Y, a c T(A
(
). (56)
When we use (56) to dene the notion a canonical model of a zero-dimensional Shimura
variety, we can say that
0
of the canonical model of Sh
K
(G, X) is the canonical model of
Sh(T, Y ).
If o xes a special x
0
mapping to y, then (56) follows from (54), and a slight improve-
ment of (13.4) shows that such os generate Aut(C,E(G, X)).
NOTES. The proof of uniqueness follows Deligne 1971b, ~3, except that I am more un-
scrupulous in my use of the Zorns lemma.
110 14 EXISTENCE OF CANONICAL MODELS
14 Existence of canonical models
Canonical models are known to exist for all Shimura varieties. In this section, I explain
some of the ideas that go into the proof.
Descent of the base eld
Let k be a subeld of an algebraically closed eld D of characteristic zero, and let A =
Aut(D,k). In (13.1) we observed that the functor
{varieties over k} {varieties V over Daction of A on V (D)},
is fully faithful. In this subsection, we nd conditions on a pair (V, ) that ensure that it is
in the essential image of the functor, i.e., that it arises from a variety over k. We begin by
listing two necessary conditions.
The regularity condition
Obviously, the action should recognize that V (D) is not just a set, but rather the set of
points of an algebraic variety. Recall that, for o c A, oV is obtained from V by applying
o to the coefcients of the polynomials dening V , and oP c (oV )(D) is obtained from
P c V (D) by applying o to the coordinates of P.
DEFINITION 14.1. An action of A on V (D) is regular if the map
oP .o P: (oV )(D) V (D)
is a regular isomorphism for all o.
A priori, this is only a map of sets. The condition requires that it be induced by a regular
map
o
: oV V . If (V, ) arises from a variety over k, then oV = V and oP = o P,
and so the condition is clearly necessary.
REMARK 14.2. (a) When regular, the maps
o
are automatically isomorphisms provided
V is nonsingular.
(b) The maps
o
satisfy the cocycle condition
o
o
r
=
or
. Conversely, every family
(
o
)
o=A
of regular isomorphisms satisfying the cocycle condition arises from an action of
A satisfying the regularity condition. Such families (
o
)
o=A
are called descent data, and
normally one expresses descent theory in terms of them rather than actions of A.
The continuity condition
DEFINITION 14.3. An action of A on V (D) is continuous if there exists a subeld L of
D nitely generated over k and a model V
0
of V over L such that the action of Aut(D,L)
on V (D) dened by V
0
is .
More precisely, the condition requires that there exist a model (V
0
, ) of V over L such
that (oP) = o (P) for all P c V
0
(D) and o c Aut(C,L). Clearly this condition is
necessary.
Review of local systems and families of abelian varieties 111
PROPOSITION 14.4. A regular action of A on V (D) is continuous if there exist points
P
1
, . . . , P
n
c V (D) such that
(a) the only automorphism of V xing every P
i
is the identity map;
(b) there exists a subeld L of D nitely generated over k such that o P
i
= P
i
for all
o xing L.
PROOF. Let (V
0
, ) be a model of V over a subeld L of D nitely generated over k.
After possibly enlarging L, we may assume that
-1
(P
i
) c V
0
(L) and that o P
i
= P
i
for
all o xing L (because of (b)). For such a o,
o
and (o)
-1
are regular maps oV V
sending oP
i
to P
i
for each i, and so they are equal (because of (a)). Hence
(oP) =
o
((o)(oP)) =
o
(o((P))) = o (P)
for all P c V
0
(D), and so the action of Aut(C,L) on V (D) dened by (V
0
, ) is .
A sufcient condition for descent
THEOREM 14.5. If V is quasiprojective and is regular and continuous, then (V, ) arises
from a variety over k.
PROOF. This is a restatement of the results of Weil 1956a (see Milne 1999, 1.1).
COROLLARY 14.6. The pair (V, ) arises from a variety over k if
(a) V is quasiprojective,
(b) is regular, and
(c) there exists points P
1
, . . . , P
n
in V (D) satisfying the conditions (a) and (b) of (14.4).
PROOF. Immediate from (14.5) and (14.6).
For an elementary proof of the corollary, not using the results of Weil 1956a, see AG
14.27.
Review of local systems and families of abelian varieties
Let S be a topological manifold. A local system of Z-modules on S is a sheaf F on S that
is locally isomorphic to the constant sheaf Z
n
(n c N).
Let F be a local system of Z-modules on S, and let o c S. There is an action of
1
(S, o) on F
o
that can be described as follows: let ;: [0, 1] S be a loop at o; because
[0, 1] is simply connected, there is an isomorphism from ;
+
F to the constant sheaf de-
ned by a group M say; when we choose such an isomorphism, we obtain isomorphisms
(;
+
F)
i
M for all i c [0, 1]; now (;
+
F)
i
= F
;(i)
and ;(0) = o = ;(1), and so we get
two isomorphisms F
o
M; these isomorphisms differ by an automorphism of F
o
, which
depends only the homotopy class of ;.
PROPOSITION 14.7. If S is connected, then F .(F
o
, ,
o
) denes an equivalence from the
category of local systems of Z-modules on S to the category of nitely generated Z-modules
endowed with an action of
1
(S, o).
112 14 EXISTENCE OF CANONICAL MODELS
PROOF. This is well known; cf. Deligne 1970, I 1.
Let F be a local system of Z-modules on S. Let :
S S be the universal covering
space of S, and choose a point o c
S. We can identiy
+
F with the constant sheaf dened
by F
(o)
. Suppose that we have a hodge structure h
s
on F
s
1 for every s c S. We say
that F, together with the hodge structures, is a variation of integral hodge structures on
S if s . h
(s)
(hodge structure on F
(o)
1) is a variation of hodge structures on
S. A
polarization of a variation of hodge structures (F, (h
s
)) is a pairing [: F F Z such
that [
s
is a polarization of (F
s
, h
s
) for every s.
Let V be a nonsingular algebraic variety over C. A family of abelian varieties over V is
a regular map : A V of nonsingular varieties plus a regular multiplication A
V
A
Aover V such that the bres of are abelian varieties of constant dimension (in a different
language, A is an abelian scheme over V ).
THEOREM 14.8. Let V be a nonsingular variety over C. There is an equivalence (A, ) .
(R
1
+
Z)
GL(H
1
(A, 1)) describing the natural complex structure on H
1
(A, 1) (see ~6).
80
PROPOSITION 14.10. An abelian variety A over C is CM if and only if there exists a torus
T GL(H
1
(A, )) such that h
A
(C
) T(1).
PROOF. See Mumford 1969, ~2, or Deligne 1982, ~3.
COROLLARY 14.11. If (A, s, jK) . [x, a]
K
under M
K
Sh
K
(G, X), then A is of
CM-type if and only if x is special.
PROOF. Recall that if (A, s, jK) .[x, a]
K
, then there exists an isomorphismH
1
(A, )
V sending h
A
to h
x
. Thus, the statement follows from the proposition.
A criterion to be canonical
We now dene an action of Aut(C) on M
K
. Let (A, s, jK) c M
K
. Then s c H
2
(A, )
is a hodge tensor, and therefore equals r[D] for some r c
o
s is a polarization for H
1
(A, ). We dene o(A, s, jK) to be (oA,
o
s,
o
jK) with
o
j
as in (55).
PROPOSITION 14.12. Suppose that Sh
K
has a model M
K
over for which the map
M
K
M
K
(C)
commutes with the actions of Aut(C). Then M
K
is canonical.
PROOF. For a special point [x, a]
K
corresponding to an abelian variety A with complex
multiplication by a eld E, the condition (54) is an immediate consequence of the main
theorem of complex multiplication (cf. 12.11). For more general special points, it also
follows from the main theorem of complex multiplication, but not quite so immediately.
80
If A(C) = C
g
,, then
H
1
(A, Z) = , H
1
(A, ) = , H
1
(A, 1) = 1
= C
g
.
114 14 EXISTENCE OF CANONICAL MODELS
Outline of the proof of the existence of a canonical model
Since the action of Aut(C) on M
K
preserves the isomorphism classes, from the map
M
K
Sh
K
(C), we get an action of Aut(C) on Sh
K
(C). If this action satises the
conditions of hypotheses of Corollary 14.6, then Sh
K
(G, X) has a model over , which
Proposition 14.12 will show to be canonical.
Condition (a) of (14.6). We know that Sh
K
(G, X) is quasi-projective from (3.12).
Condition (b) of (14.6). We have to show that the map
oP .o P: o Sh
K
(C)
(
Sh
K
(C)
is regular. It sufces to do this for K small, because if K
t
K, then Sh
K
0 (G, X) is a
quotient of Sh
K
(G, X).
Recall (5.17) that
0
(Sh
K
)
=
>0
A
(
,v(K). Let c c
>0
A
(
,v(K), and let Sh
t
K
be the corresponding connected component of Sh
K
. Then Sh
t
K
= I
t
X
where I
t
=
G() K
t
for some conjugate K
t
of K (see 5.17, 5.23)
Let (A, s, jK) c M
K
and choose an isomorphism a: H
1
(A, ) V sending s to a
multiple of [. Then the image of (A, s, jK) in
>0
A
(
,v(K) is represented by v(a j)
where a j: V (A
(
) V (A
(
) is to be regarded as an element of G(A
(
). Write M
t
K
for
the set of triples with v(a j) c c. Dene H
t
K
similarly.
The map M
K
>0
A
(
,v(K) is equivariant for the action of Aut(C) when we let
Aut(C) act on
>0
A
(
,v(K) through the cyclotomic character, i.e.,
o[] = [y(o)] where y(o) c
Z
, (
y(o)
= o(, ( a root of 1.
Write X
(I
t
) for I
t
X
(I
t
)) be the
algebraic variety obtained from X
(I
t
) by change of base eld o: C C. Consider the
diagram:
X
_
X
(I
ot
)
(
o(X
(I
t
))
M
ot
K
o
M
t
K
The map o sends (A, . . .) to o(A, . . .), and the map
o
is the map of sets oP . o P.
The two maps are compatible. The map U o(X
(I
t
)) is the universal covering space
of the complex manifold (o(X
(I
t
)))
an
.
Fix a lattice in V that is stable under the action of I
t
. From the action of I
t
on ,
we get a local system of Z-modules M on X
(I
t
) (see 14.7), which, in fact, is a polarized
integral variation of hodge structures F. According to Theorem 14.8, this variation of
hodge structures arises from a polarized family of abelian varieties : A X
(I
t
). As
is a regular map of algebraic varieties, we can apply o to it, and obtain a polarized family
of abelian varieties o : oA o(X
(I
t
)). Then (R
1
(o )
+
Z)
is a polarized integral
Simple PEL Shimura varieties of type A or C 115
hodge structure on o(X
(I
t
)). On pulling this back to U and tensoring with , we obtain
a variation of polarized rational hodge structures over the space U, whose underlying local
system can identied with the constant sheaf dened by V . When this identication is done
correctly, each u c U denes a complex structure on V that is positive for [, i.e., a point
x of X
, and the map u .x makes the diagram commute. Now (2.15) shows that u .x
is holomorphic. It follows that
o
is holomorphic, and Borels theorem (3.14) shows that it
is regular.
Condition (c) of (14.6) For any x c X, the set {[x, a]
K
[ a c G(A
(
)} has the property
that only the identity automorphism of Sh
K
(G, X) xes its elements (see 13.5). But, there
are only nitely many automorphisms of Sh
K
(G, X) (see 3.21), and so a nite sequence
of points [x, a
1
], . . . , [x, a
n
] will have this property. When we choose x to be special, the
main theorem of complex multiplication (11.2) tells us that o [x, a
i
] = [x, a
i
] for all o
xing some xed nite extension of E(x), and so condition (c) holds for these points.
Simple PEL Shimura varieties of type A or C
The proof is similar to the Siegel case. Here Sh
K
(G, X) classies quadruples (A, i, s, jK)
satisfying certain conditions. One checks that if o xes the reex eld E(G, X), then
o(A, i, s, jK) lies in the family again (see 12.7). Again the special points correspond to CM
abelian varieties, and the Shimura-Taniyama theoremshows that, if Sh
K
(G, X) has a model
M
K
over E(G, X) for which the action of Aut(C,E(G, X)) on M
K
(C) = Sh
K
(G, X)(C)
agrees with its action on the quadruples, then it is canonical.
Shimura varieties of hodge type
In this case, Sh
K
(G, X) classies isomorphism classes of triples (A, (s
i
)
0_i_n
, jK) where
the s
i
are hodge tensors. A proof similar to that in the Siegel case will apply once we have
dened
o
s for s a hodge tensor on an abelian variety.
If the Hodge conjecture is true, then s is the cohomology class of some algebraic cycle
Z on A (i.e., formal -linear combination of integral subvarieties of A). Then we could
dene
o
s to be the cohomology class of oZ on oA. Unfortunately, a proof of the Hodge
conjecture seems remote, even for abelian varieties. Deligne succeeded in dening
o
s
without the Hodge conjecture. It is important to note that there is no natural map between
H
n
(A, ) and H
n
(oA, ) (unless o is continuous, and hence is the identity or complex
conjugation). However, there is a natural isomorphism o: H
n
(A, A
(
) H
n
(oA, A
(
)
coming from the identication
H
n
(A, A
(
)
= Hom(
n
_
, A
(
)
= Hom(
n
_
(A
(
), A
(
)
= Hom(
n
_
V
(
A, A
(
)
(or, equivalently, from identifying H
n
(A, A
(
) with etale cohomology).
THEOREM 14.13. Let s be a hodge tensor on an abelian variety A over C, and let s
A
f
be
the image of s the A
(
-cohomology. For any automorphism o of C, there exists a hodge
tensor
o
s on oA (necessarily unique) such that (
o
s)
A
f
= o(s
A
f
).
116 14 EXISTENCE OF CANONICAL MODELS
PROOF. This is the main theorem of Deligne 1982. [Interestingly, the theory of locally
symmetric varieties is used in the proof.]
As an alternative to using Delignes theorem, one can apply the following result (note,
however, that the above approach has the advantage of giving a description of the points of
the canonical model with coordinates in any eld containing the reex eld).
PROPOSITION 14.14. Let (G, X) (G
t
, X
t
) be an inclusion of Shimura data; if Sh(G
t
, X
t
)
has canonical model, so also does Sh(G, X).
PROOF. This follows easily from 5.16.
Shimura varieties of abelian type
Deligne (1979, 2.7.10) denes the notion of a canonical model of a connected Shimura
variety Sh
be a connected component
of X. Then Sh(G, X) has a canonical model if and only if Sh
(G
der
, X
) has a canonical
model.
PROOF. See Deligne 1979, 2.7.13.
Thus, for example, if (G
1
, X
1
) and (G
2
, X
2
) are Shimura data such that (G
der
1
, X
1
) -
(G
der
2
, X
2
), and one of Sh (G
1
, X
1
) or Sh(G
2
, X
2
) has a canonical model, then they both
do.
The next result is more obvious (ibid. 2.7.11).
PROPOSITION 14.16. (a) Let (G
i
, X
i
) (1 _ i _ m) be connected Shimura data. If
each connected Shimura variety Sh
(G
i
, X
i
) has a canonical model M
(G
i
, X
i
), then
i
M
(G
i
, X
i
) is a canonical model for Sh
i
G
i
,
i
X
i
).
(b) Let (G
1
, X
1
) (G
2
, X
2
) be an isogeny of connected Shimura data. If Sh
(G
1
, X
1
)
has a canonical model, then so also does Sh
(G
2
, X
2
).
More precisely, in case (b) of the theorem, let G
ad
()
1
and G
ad
()
2
be the completions
of G
ad
()
and G
2
()
(G
2
, X
2
) is the quotient of the
canonical model for Sh
(G
2
, X
2
) by the kernel of G
ad
()
1
G
ad
()
2
.
We can now prove the existence of canonical models for all Shimura varieties of abelian
type. For a connected Shimura variety of primitive type, the existence follows from (14.15)
and the existence of canonical models for Shimura varieties of hodge type (see above).
Now (14.16) proves the existence for all connected Shimura varieties of abelian type, and
(14.16) proves the existence for all Shimura varieties of abelian type.
REMARK 14.17. The above proof is only an existence proof: it gives little information
about the canonical model. For the Shimura varieties it treats, Theorem 9.4 can be used to
construct canonical models and give a description of the points of the canonical model in
any eld containing the reex eld.
General Shimura varieties 117
General Shimura varieties
There is an approach that proves the existence of canonical models for all Shimura varieties,
and is largely independent of that discussed above except that it assumes the existence
81
of
canonical models for Shimura varieties of type A
1
(and it uses (14.15) and (14.16)).
The essential idea is the following. Let (G, X) be a connected Shimura datum with G
the group over obtained from a simple group H over a totally real eld F by restriction
of scalars.
Assume rst that H splits over a CM-eld of degree 2 over F. Then there exist many
homomorphisms H
i
H from groups of type A
1
into H. From this, we get many
inclusions
Sh
(G
i
, X
i
) Sh
(G, X)
where G
i
is the restriction of scalars of H
i
. From this, and the existence of canonical
models for the Sh
(G
i
, X
i
), it is possible to prove the existence of the canonical model for
Sh
(G, X).
In the general case, there will be a totally real eld F
t
containing F and such that H
F
0
splits over a CM-eld of degree 2 over F. Let G
+
be the restriction of scalars of H
F
0 . Then
there is an inclusion (G, X) (G
+
, X
+
) of connected Shimura data, and the existence of a
canonical model for Sh
(G
+
, X
+
) implies the existence of a canonical model for Sh
(G, X)
(cf. 14.14).
For the details, see Borovoi 1984, 1987 and Milne 1983.
Final remark: rigidity
One might expect that if one modied the condition (54), for example, by replacing r
x
(s)
with r
x
(s)
-1
, then one would arrive at a modied notion of canonical model, and the same
theorems would hold. This is not true: the condition (54) is the only one for which canon-
ical models can exist. In fact, if G is adjoint, then the Shimura variety Sh(G, X) has no
automorphisms commuting with the action of G(A
(
) (Milne 1983, 2.7), from which it fol-
lows that the canonical model is the only model of Sh(G, X) over E(G, X), and we know
that for the canonical model the reciprocity law at the special points is given by (54).
NOTES. The concept of a canonical model characterized by reciprocity laws at special
points is due to Shimura, and the existence of such models was proved for major families
by Shimura, Miyake, and Shih. Shimura recognized that to have a canonical model it is
necessary to have a reductive group, but for him the semisimple group was paramount:
in our language, given a connected Shimura datum (H, Y ), he asked for Shimura datum
(G, X) such that (G
der
, X
M
s
i
,r
i
(End(e
i
)).
Thus, the category M is described up to equivalence by:
(a) the set (M) of isomorphism classes of simple objects in M;
(b) for each o c , the isomorphism class [D
o
] of the endomorphism algebra D
o
of a
representative of o.
We call ((M), ([D
o
])
o=(M)
) the numerical invariants of M.
Division algebras; the Brauer group
We shall need to understand what the set of isomorphism classes of division algebras over
a eld F look like.
Recall the denitions: by an F-algebra, we mean a ring A containing F in its centre
and nite dimensional as F-vector space; if F equals the centre of A, then A is called a
central F-algebra; a division algebra is an algebra in which every nonzero element has an
inverse; an F-algebra A is simple if it contains no two-sided ideals other than 0 and A.
By a theorem of Wedderburn, the simple F-algebras are the matrix algebras over division
F-algebras.
Division algebras; the Brauer group 119
EXAMPLE 15.1. (a) If F is algebraically closed or nite, then
82
every central division
algebra is isomorphic to F.
(b) Every central division algebra over 1 is isomorphic either to 1 or to the (usual)
quaternion algebra:
H = C Cj, j
2
= 1, jzj
-1
= z (z c C).
(c) Let F be a p-adic eld (nite extension of
p
), and let be a prime element of
O
F
. Let L be an unramied extension eld of F of degree n, and let o denote the
Frobenius generator of Gal(L,F) o acts as x . x
p
on the residue eld. For
each i, 1 _ i _ n, dene
D
i,n
= LLa La
n-1
, a
n
=
i
, aza
-1
= o(z) (z c L).
Then D
i,n
is a central simple algebra over F, which is a division algebra if and only
if gcd(i, n) = 1. Every central division algebra over F is isomorphic to D
i,n
for
exactly one relatively prime pair (i, n) (CFT, IV 4.2).
If B and B
t
are central simple F-algebras, then so also is B
F
B
t
(CFT, 2.8). If D
and D
t
are central division algebras, then Wedderburns theorem shows that D
F
D
t
-
M
r
(D
tt
) for some r and some central division algebra D
tt
well-dened up to isomorphism,
and so we can set
[D][D
t
] = [D
tt
].
This law of composition is obviously, and [F] is an identity element. Let D
opp
denote the
opposite algebra to D (the same algebra but with the multiplication reversed: a
opp
b
opp
=
(ba)
opp
). Then (CFT, IV 2.9)
D
F
D
opp
= End
F-linear
(D) - M
r
(F),
and so [D][D
opp
] = [F]. Therefore, the isomorphism classes of central division algebras
over F (equivalently, the isomorphism classes of central simple algebras over F) form a
group, called the Brauer group of F.
EXAMPLE 15.2. (a) The Brauer group of an algebraically closed eld or a nite eld is
zero.
(b) The Brauer group 1 has order two: Br(1)
=
1
2
Z,Z.
(c) For a p-adic eld F, the map [D
n,i
] .
i
n
mod Z is an isomorphism Br(F)
= ,Z.
(d) For a number eld F and a prime :, write inv
inv
v
,Z 0.
is exact.
82
If F is algebraically closed, then each element of a central division algebra over F generates a eld of
nite degree over F, and so lies in F. For the proof in the nite case, see CFT, IV 4.1.
120 15 ABELIAN VARIETIES OVER FINITE FIELDS
Statement (d) is shown in the course of proving the main theorem of class eld theory
by the cohomological approach (CFT, VIII 2.2). It says that to give a division algebra over
F (up to isomorphism) is the same as to give a family (i
) c
nite
,Z
real
1
2
Z,Z
such that
= 0.
The key tool in computing Brauer groups is an isomorphism
Br(F)
= H
2
(F, G
m
)
df
= H
2
(Gal(F
al
,F), F
al
)
df
= lim
H
2
(Gal(L,F), L
).
The last limit is over the elds L F
al
of nite degree and galois over . This isomor-
phism can be most elegantly dened as follows. Let D be a central simple division of
degree n
2
over F, and assume
83
that D contains a subeld L of degree n over F and galois
over F. Then each c D normalizing L denes an element x . x
-1
of Gal(L,F),
and the Noether-Skolem theorem (CFT, IV 2.10) shows that every element of Gal(L,F)
arises in this way. Because L is its own centralizer (ibid., 3.4), the sequence
1 L
N(L) Gal(L,F) 1
is exact. For each o c Gal(L,F), choose an s
o
c N(L) mapping to o, and let
s
o
s
r
= d
o,r
s
or
, d
o,r
c L
.
Then (d
o,r
) is a 2-cocycle whose cohomology class is independent of the choice of the
family (s
o
). Its class in H
2
(Gal(L,F), L
) H
2
(F, G
m
) is the cohomology class of
[D].
EXAMPLE 15.3. Let Lbe the completion of
un
p
(equal to the eld of fractions of the ring of
Witt vectors with coefcients in F), and let o be the automorphism of L inducing x .x
p
on its residue eld. An isocrystal is a nite dimensional L-vector space V equipped with
a o-linear isomorphism F: V V . The category Isoc of isocrystals is a semisimple
p
-
linear category with (Isoc) = , and the endomorphism algebra of a representative of
the isomorphism class z is a division algebra over
p
with invariant z. If z _ 0, z = r,s,
gcd(r, s) = 1, s > 0, then E
z
can be taken to be (
p
,(T
r
p
s
))
Q
p
L, and if z - 0, E
z
can be taken to be the dual of E
-z
. See Demazure 1972, Chap. IV.
Abelian varieties
Recall (p89) that AV
0
(k) is the category whose objects are the abelian varieties over k,
but whose homs are Hom
0
(A, B) = Hom(A, B) . It follows from results of Weil that
AV
0
(k) is a semisimple -category with the simple abelian varieties (see p89) as its simple
objects. Amazingly, when k is nite, we know its numerical invariants.
Abelian varieties over F
q
, q = p
n
Recall that a Weil q-integer is an algebraic integer such that, for every embedding ,: []
C, [,[ = q
1
2
. Two Weil q-integers and
t
are conjugate if there exists an isomorphism
[] [
t
] sending to
t
.
83
This will always be true when F is a p-adic or number eld, but is not true (or, at least, no known to be
true) for other elds. In the general case, it becomes true after D has been replaced by M
r
(D) for some r.
Abelian varieties 121
THEOREM 15.4 (HONDA-TATE). The map A .
A
denes a bijection from (AV(F
q
))
to the set of conjugacy classes of Weil q-integers. For any simple A, the centre of D =
df
End
0
(A) is F = [
A
], and for a prime : of F,
inv
(D) =
_
_
_
1
2
if : is real
ord
v
(
A
)
ord
v
(q)
[F
:
p
] if :[p
0 otherwise.
Moreover, 2 dimA = [D: F]
1
2
[F: ].
In fact, [] can only have a real prime if =
_
p
n
. Let W
1
(q) be the set of Weil
q-integers in
al
C. Then the theorem gives a bijection
(AV
0
(F
q
)) IW
1
(q), I = Gal(
al
,).
NOTES. Except for the statement that every
A
arises from an A, the theorem is due to
Tate. That every Weil q-integer arises from an abelian variety was proved (using 10.10) by
Honda. See Tate 1969 for a discussion of the theorem.
Abelian varieties over F
We shall need a similar result for an algebraic closure F of F
p
.
If is a Weil p
n
-integer, then
m
is a Weil p
mn
-integer, and so we have a homomor-
phism .
m
: W
1
(p
n
) W
1
(p
nm
). Dene
W
1
= lim
W
1
(p
n
).
If c W
1
is represented by
n
c W
1
(p
n
), then
m
n
c W
1
(p
nm
) also represents , and
[
n
] [
m
n
]. Dene {} to be the eld of smallest degree over generated by a
representative of .
Every abelian variety over F has a model dened over a nite eld, and if two abelian
varieties over a nite eld become isomorphic over F, then they are isomorphic already over
a nite eld. Let A be an abelian variety over F
q
. When we regard A as an abelian variety
over F
q
m, then the Frobenius map is raised to the m
th
-power (obviously):
A
F
q
m
=
m
A
.
Let A be an abelian variety dened over F, and let A
0
be a model of A over F
q
. The
above remarks show that s
A
(:) =
df
ord
v
(
A
0
)
ord
v
(q)
is independent of the choice of A
0
. More-
over, for any ,: [
A
0
]
al
, the I-orbit of the element
A
of W
1
represented by ,
A
0
depends only on A.
THEOREM 15.5. The map A . I
A
denes a bijection (AV
0
(F)) IW
1
. For any
simple A, the centre of D =
df
End
0
(A) is isomorphic to F = {
A
}, and for any prime :
of F,
inv
(D) =
_
_
_
1
2
if : is real
s
A
(:) [F
:
p
] if :[p
0 otherwise.
PROOF. This follows from the Honda-Tate theorem and the above discussion.
122 15 ABELIAN VARIETIES OVER FINITE FIELDS
Our goal in the remainder of this section is to give an elementary construction of a
semisimple -category that contains, in a natural way, a category of fake abelian varieties
over F with the same numerical invariants as AV
0
(F).
For the remainder of this section F is a eld of characteristic zero.
Tori and their representations
Let T be a torus over F split by a galois extension L,F with galois group I. As we
noted on p19, to give a representation , of T on an F-vector space V amounts to giving
an X
+
(T)-grading V (L) =
y=X
(T)
V
y
of V (L) with the property that oV
y
= V
oy
for
all o c I and y c X
+
(T). In this, L,F can be an innite galois extension.
PROPOSITION 15.6. Let I = Gal(F
al
,F). The category of representations Rep(T) of T
on F-vector spaces is semisimple. The set of isomorphism classes of simple objects is in
natural one-to-one correspondence with the orbits of I acting on X
+
(T), i.e., (Rep(T)) =
IX
+
(T). If V
Iy
is a simple object corresponding to Iy, then dim(V
Iy
) is the order of
Iy, and
End(V
y
) - F(y)
where F(y) is the xed eld of the subgroup I(y) of I xing y.
PROOF. Follows easily from the preceding discussion.
REMARK 15.7. Let y c X
+
(T), and let I(y) and F(y) be as in the proposition. Then
Hom(F(y), F
al
)
= I, I(y), and so X
+
((G
m
)
F(y){F
) = Z
I{ I(y)
. The map
n
o
o .
n
o
oy: Z
I{ I(y)
X
+
(T)
denes a homorphism
T (G
m
)
F(y){F
. (58)
From this, we get a homomorphism of cohomology groups
H
2
(F, T) H
2
(F, (G
m
)
F(y){F
).
But Shapiros lemma (CFT, II 1.11) shows that H
2
(F, (G
m
)
F(y){F
)
= H
2
(F(y), G
m
),
which is the Brauer group of F(y). On composing these maps, we get a homomorphism
H
2
(F, T) Br(F(y)). (59)
The proposition gives a natural construction of a semisimple category M with (M) =
IN, where N is any nitely generated Z-module equipped with a continuous action of
I. However, the simple objects have commutative endomorphism algebras. To go further,
we need to look at new type of structure.
Afne extensions 123
Afne extensions
Let L,F be a Galois extension of elds with Galois group I, and let G be an algebraic
group over F. In the following, we consider only extensions
1 G(L) E I 1
in which the action of I on G(L) dened by the extension is the natural action, i.e.,
if e
o
.o, then e
o
ge
-1
o
= og (e
o
c E, o c I, g c T(F
al
)).
For example, there is always the split extension E
G
=
df
G(L) I.
An extension E is afne if its pull-back to some open subgroup of I is split. Equiva-
lently, if for the o in some open subgroup of I, there exist e
o
.o such that e
or
= e
o
e
r
.
We sometimes call such an E an L,F-afne extension with kernel G.
Consider an extension
1 T E I 1
with T commutative. If E is afne, then it is possible to choose the e
o
s so that the 2-
cocycle d: I I T(L) dened by
e
o
e
r
= d
o,r
e
o
e
r
, d
o,r
c T(F
al
).
is continuous. Thus, in this case E denes a class cl (E) c H
2
(F, T).
A homomorphism of afne extensions is a commutative diagram
1 G
1
(L) E
1
I 1
_
_
_
1 G
2
(L) E
2
I 1
such that the restriction of the homomorphism to G
1
(L) is dened by a homomorphism
of algebraic groups (over L). A morphism
t
of homomorphisms E
1
E
2
is an
element of g of G
2
(L) such that ad(g) =
t
, i.e., such that
g (e) g
-1
=
t
(e), all e c E
1
.
For a vector space V over F, let E
V
be the split afne extension dened by the algebraic
group GL(V ). A representation of an afne extension E is a homomorphism E E
V
.
REMARK 15.8. To give a representation of E
G
on E
V
is the same as to give a represen-
tation of G on V . More precisely, the functor Rep(G) Rep(E
G
) is an equivalence of
categories. The proof of this uses that H
1
(F, GL(V )) = 1.
PROPOSITION 15.9. Let E be an L,F-afne extension whose kernel is a torus T split
by L. The category Rep(E) is a semisimple F-category with (Rep(E)) = IX
+
(T).
Let V
Iy
be a simple representation of E corresponding to Iy c IX
+
(T). Then, D =
End(V
I
) has centre F(y), and its class in Br(F(y)) is the image of cl (E) under the
homomorphism (59).
124 15 ABELIAN VARIETIES OVER FINITE FIELDS
PROOF. Omitted (but it is not difcult).
We shall also need to consider afne extensions in which the kernel is allowed to be a
protorus, i.e., the limit of an inverse system of tori. For T = lim
T
i
, X
+
(T) = lim
X
+
(T
i
),
and T . X
+
(T) denes an equivalence from the category of protori to the category of
free Z-modules with a continuous action of I. Here continuous means that every element
of the module is xed by an open subgroup of I. Let L = F
al
. By an afne extension
with kernel T , we mean an exact sequence
1 T(F
al
) E I 1
whose push-out
1 T
i
(F
al
) E
i
I 1
by T(F
al
) T
i
(F
al
) is an afne extension in the previous sense. A representation of such
an extension is dened exactly as before.
REMARK 15.10. Let
L
L
t
F
I
F
t
I
0
be a diagram of elds in which L
t
,F
t
is Galois with group I
t
. From an L,F-afne exten-
sion
1 G(L) E I 1
with kernel G we obtain an L
t
,F
t
-afne extension
1 G(L
t
) E
t
I
t
1
with kernel G
F
0 by pulling back by o . o[L: I
t
I and pushing out by G(L)
G(L
t
)).
EXAMPLE 15.11. Let
un
p
be a maximal unramied extension of
p
, and let L
n
be the
subeld of
un
p
of degree n over
p
. Let I
n
= Gal(L
n
,
p
), let D
1,n
be the division
algebra in (15.1c), and let
1 L
n
N(L
n
) I
n
1
be the corresponding extension. Here N(L
n
) is the normalizer of L
n
in D
1,n
:
N(L
n
) =
0_i_n-1
L
n
a
i
.
This is an L
n
,
p
-afne extension with kernel G
m
. On pulling back by I I
n
and
pushing out by L
n
un
p
, we obtain a
un
p
,
p
-afne extension D
n
with kernel G
m
.
From a representation of D
n
we obtain a vector space V over
un
p
equipped with a o-linear
map F (the image of (1, a) is (F, o)). On tensoring this with the completion L of
un
p
, we
obtain an isocrystal that can be expressed as a sum of E
z
s with z c
1
n
Z.
The afne extension P 125
Note that there is a canonical section to N(L
n
) I
n
, namely, o
i
.a
i
, which denes
a canonical section to D
n
I.
There is a homomorphism D
nm
D
n
whose restriction to the kernel is multiplication
by m. The inverse limit of this system is a
un
p
,
p
-afne extension D with kernel G =
df
lim
G
m
. Note that X
+
(G) = lim
1
n
Z,Z = . There is a natural functor from Rep(D) to
the category of isocrystals, which is faithful and essentially surjective on objects but not
full. We call D the Dieudonn e afne extension.
The afne extension P
Let W(p
n
) be the subgroup of
al
generated by W
1
(p
n
), and let W = lim
W(p
n
). Then
W is a free Z-module of innite rank with a continuous action of I = Gal(
al
,). For
c W, we dene {} to be the smallest eld generated by a representative of . If is
represented by
n
c W(p
n
) and [,(
n
)[ = (p
n
)
m{2
, we say that has weight m and we
write
s
(:) =
ord
(
n
)
ord
(q)
.
THEOREM 15.12. Let P be the protorus over with X
+
(P) = W. Then there exists an
afne extension
1 P(
al
) P I 1
such that
(a) (Rep(P)) = IW;
(b) for c W, let D() = End(V
I
) where V
I
is a representation corresponding to
I; then D() is isomorphic to the division algebra D with centre {} and the
invariants
inv
(D) =
_
_
_
(
1
2
)
ut()
if : is real
s
(:) [{}
:
p
] if :[p
0 otherwise.
Moreover, P is unique up to isomorphism.
PROOF. Let c() denote the class in Br({}) of the division algebra D in (b). To prove
the result, we have to show that there exists a unique class in H
2
(, P) mapping to c()
in Br({}) for all :
c .(c()): H
2
(, P)
(59)
I=I\W
Br({}).
This is an exercise in galois cohomology, which, happily, is easier than it looks.
We call a representation of P a fake motive over F, and a fake abelian variety if its
simple summands correspond to c IW
1
. Note that the category of fake abelian varieties
is a semisimple -category with the same numerical invariants as AV
0
(F).
126 15 ABELIAN VARIETIES OVER FINITE FIELDS
The local form P
l
of P
Let l be a prime of , and choose a prime n
l
of
al
dividing l . Let
al
l
be the algebraic
closure of
l
in the completion of
al
at n
l
. Then I
l
=
df
Gal(
al
l
,
l
) is a closed subgroup
of I =
df
Gal(
al
,), and we have a diagram
al
al
l
l
.
I
l
(60)
From P we obtain a
al
l
,
l
-afne extension P(l ) by pulling back by I
l
I and pushing
out by P(
al
) P(
al
l
) (cf. 15.10).
The
I
-space attached to a fake motive
Let ,= p, obe a prime of .
PROPOSITION 15.13. There exists a continuous homomorphism (
I
making
I
I
`
[[
1
-
P(
al
I
)
-
P()
-
I
I
-
1
commute.
PROOF. To prove this, we have to show that the cohomology class of P in H
2
(, P) maps
to zero in H
2
(
I
, P), but this is not difcult.
Fix a homomorphism (
I
: I
I
P() as in the diagram. Let ,: P E
V
be a fake
motive. From ,, we get a homomorphism
,(): P() GL(V (
al
I
)) I
I
.
For o c I
I
, let (,() (
I
)(o) = (e
o
, o). Because (
I
is a homomorphism, the automor-
phisms e
o
of V (
al
I
) satisfy
e
o
oe
r
= e
or
, o, t c I
I
,
and so
o : = e
o
(o:)
is an action of I
I
on V (
al
I
), which one can check to be continuous. Therefore (AG, 14.13),
V
I
(,) =
df
V (
al
I
)
I
`
is a
I
-structure on V (
al
I
). In this way, we get a functor , . V
I
(,)
from the category of fake motives over F to vector spaces over
I
.
The (
I
can be chosen in such a way that the spaces V
I
(,) contain lattices
I
(,) that are
well-dened for almost all ,= p, which makes it possible to dene
V
p
(
(,) =
I,=p,o
(V
I
(,):
I
(,)).
It is a free module over A
p
(
=
df
I,=p,o
(
I
: Z
I
).
The afne extension P 127
The isocrystal of a fake motive
Choose a prime n
p
of
al
dividing p, and let
un
p
and
al
p
denote the subelds of the com-
pletion of
al
at n
p
. Then I
p
=
df
Gal(
al
p
,
p
) is a closed subgroup of I =
df
Gal(
al
,)
and I
un
p
=
df
Gal(
un
p
,
p
) is a quotient of I
p
.
PROPOSITION 15.14. (a) The afne extension P(p) arises by pull-back and push-out from
a
un
p
,
p
-afne extension P(p)
un
.
(b) There is a homomorphism of
un
p
,
p
-extensions D P(p)
un
whose restriction to
the kernels, G P
Q
p
, corresponds to the map on characters .s
(n
p
): W .
PROOF. (a) This follows from the fact that the image of the cohomology class of P in
H
2
(I
p
, P(
al
p
)) arises from a cohomology class in H
2
(I
un
p
, P(
un
p
)).
(b) This follows from the fact that the homomorphism H
2
(
p
, G) H
2
(
p
, P
Q
p
)
sends the cohomology class of D to that of P(p)
un
.
In summary:
1 G
m
(
un
p
) D I
un
p
1
_
_
_
_
1 P(
un
p
) P(p)
un
I
un
p
1
_
.
.
.
_
1 P(
al
p
) P(p) I
p
1
A fake motive ,: P E
V
gives rise to a representation of P(p), which arises from a
representation of P(p)
un
(cf. the argument in the preceding subsubsection). On composing
this with the homomorphism D P(p)
un
, we obtain a representation of D, which gives
rise to an isocrystal D(,) as in (15.11).
Abelian varieties of CM-type and fake abelian varieties
We saw in (10.5) that an abelian variety of CM-type over
al
denes an abelian variety
over F. Does it also dene a fake abelian variety? The answer is yes.
PROPOSITION 15.15. Let T be a torus over split by a CM-eld, and let j be a cochar-
acter of T such that j |j is dened over (here | is complex conjugation). Then there
is a homomorphism, well dened up to isomorphism,
: P E
T
.
PROOF. Omitted.
Let A be an abelian variety of CM-type (E, ) over
al
, and let T = (G
m
)
E{Q
. Then
denes a cocharacter j
-multiple of an ample
divisor, and j is a K-orbit of isomorphisms V (A
(
) V
(
(A) sending [ to an A
(
-multiple
of the pairing dened by s. Here NS(A) is the N` eron-Severi group of A (divisor classes
modulo algebraic equivalence).
On the other hand, I do not know a description of Sh
K
(
al
) when, for example, G
ad
has
factors of type E
6
or E
7
or mixed type D. In these cases, the proof of the existence of a
canonical model is quite indirect.
The points of the Shimura variety with coordinates in the reex eld
Over E = E(G, X) the following additional problem arises. Let A be an abelian variety
over
al
. Suppose we know that oA is isomorphic to A for all o c Gal(
al
,E). Does this
imply that A is dened over E? Choose an isomorphism
o
: oA A for each o xing E.
A necessary condition that the
o
arise from a model over E is that they satisfy the cocycle
condition:
o
o
r
=
or
. Of course, if the cocycle condition fails for one choice of the
o
s, we can try another, but there is an obstruction to obtaining a cocycle which lies in the
cohomology set H
2
(Gal(
al
,E), Aut(A)).
Certainly, this obstruction would vanish if Aut(A) were trivial. One may hope that the
automorphism group of an abelian variety (or motive) plus data in the family classied by
Sh
K
(G, X) is trivial, at least when K is small. This is so when condition SV5 holds, but
not otherwise.
In the Siegel case, the centre of G is G
m
and so SV5 holds. Therefore, provided K is
sufciently small, for any eld Lcontaining E(G, X), Sh
K
(L) classies triples (A, s, jK)
satisfying the same conditions as when L =
al
. Here A an abelian variety over L,
s c NS(A) , and j is an isomorphism V (A
(
) V
(
(A) such that jK is stable under
the action of Gal(L
al
,L).
In the Hilbert case (4.14), the centre of G is (G
m
)
F{Q
for F a totally real eld and
SV5 fails: F
is not discrete in A
F,(
because every nonempty open subgroup of A
F,(
will contain innitely many units. In this case, one has a description of Sh
K
(L) when L is
algebraically closed, but otherwise all one can say is that Sh
K
(L) = Sh
K
(L
al
)
Gal(L
al
{L)
.
130 16 THE GOOD REDUCTION OF SHIMURA VARIETIES
Hyperspecial subgroups
The modular curve I
0
(N)H
1
is dened over , and it has good reduction at the primes
not dividing the level N and bad reduction at the others. Before explaining what is known
in general, we need to introduce the notion of a hyperspecial subgroup.
DEFINITION 16.1. Let G be a reductive group over (over
p
will do). A subgroup
K G(
p
) is hyperspecial if there exists a at group scheme G over Z
p
such that
G
Q
p
= G (i.e., G extends G to Z
p
);
G
F
p
is a connected reductive group (necessarily of the same dimension as G because
of atness);
G(Z
p
) = K.
EXAMPLE 16.2. Let G = GSp(V, [). Let be a lattice in V (
p
), and let K
p
be the
stabilizer of . Then K
p
is hyperspecial if the restriction of [ to takes values in
Z
p
and is perfect (i.e., induces an isomorphism
p
must be a product of matrix algebras over
unramied extensions of
p
. When this condition holds, the description of the hyperspecial
groups is similar to that in the Siegel case.
There exists a hyperspecial subgroup in G(
p
) if and only if G is unramied over
p
,
i.e., quasisplit over
p
and split over an unramied extension.
For the remainder of this section we x a hyperspecial subgroup K
p
G(
p
), and
we write Sh
p
(G, X) for the family of varieties Sh
K
p
K
p
(G, X) with K
p
running over the
compact open subgroups of G(A
p
(
). The group G(A
p
(
) acts on the family Sh
p
(G, X).
The good reduction of Shimura varieties
Roughly speaking, there are two reasons a Shimura variety may have bad reduction at a
prime dividing p: the reductive group itself may be ramied at p or p may divide the level.
For example, the Shimura curve dened by a quaternion algebra B over will have bad
reduction at a prime p dividing the discriminant of B, and (as we noted above) I
0
(N)H
1
has bad reduction at a prime dividing N. The existence of a hyperspecial subgroup K
p
forces G to be unramied at p, and by considering only the varieties Sh
K
p
K
p
(G, X) we
avoid the second problem.
THEOREM 16.4. Let Sh
p
(G, X) be the inverse system of varieties over E(G, X) dened
by a Shimura datum (G, X) of abelian type and a hyperspecial subgroup K
p
G(
p
).
Then, except possibly for some small set of primes p depending only on (G, X), Sh
p
(G, X)
has canonical good reduction at every prime p of E(G, X) dividing p, .
REMARK 16.5. Let E
p
be the completion of E at p, let
O
p
be the ring of integers in E
p
,
and let k(p) be the residue eld
O
p
,p.
Denition of the Langlands-Rapoport set 131
(a) By Sh
p
(G, X) having good reduction p, we mean that the inverse system
(Sh
K
p
K
p
(G, X))
K
p, K
p
G(A
p
(
) compact open, K
p
xed,
extends to an inverse system of at schemes S
p
= (S
K
p) over
O
p
whose reduction modulo
p is an inverse system of varieties (Sh
K
p
K
p
(G, X))
K
p over k(p) such that, for K
p
K
tp
sufciently small,
Sh
K
p
K
p
Sh
K
0p
K
p
is an etale map of smooth varieties. We require also that the action of G(A
p
(
) on Sh
p
extends to an action on S
p
.
(b) A variety over E
p
may not have good reduction to a smooth variety over k(p)
this can already be seen for elliptic curves and, when it does it will generally have good
reduction to many different smooth varieties, none of which is obviously the best. For
example, given one good reduction, one can obtain another by blowing up a point in its
closed bre. By Sh
p
(G, X) having canonical good reduction at p, I mean that, for any
formally smooth scheme T over
O
p
,
Hom
O
p
(T, lim
K
p
S
K
p)
= Hom
E
p
(T
E
p
, lim
K
p
Sh
K
p
K
p
). (61)
A smooth scheme is formally smooth, and an inverse limit of schemes etale over a smooth
scheme is formally smooth. As lim
S
K
p is formally smooth over
O
p
, (61) characterizes it
uniquely up to a unique isomorphism (by the Yoneda lemma).
(c) In the Siegel case, Theorem 16.4 was proved by Mumford (his Fields medal theo-
rem; Mumford 1965). In this case, the S
K
p and Sh
K
p
K
p
are moduli schemes. The PEL-case
can be considered folklore in that several authors have deduced it from the Siegel case and
published sketches of proof, the most convincing of which is in Kottwitz 1992. In this
case, S
p
(G, X) is the zariski closure of Sh
p
(G, X) in S
p
(G([), X([)) (cf. 5.16), and it is
a moduli scheme. The hodge case
85
was proved by Vasiu (1999) except for a small set of
primes. In this case, S
p
(G, X) is the normalization of the zariski closure of Sh
p
(G, X) in
S
p
(G([), X([)). The case of abelian type follows easily from the hodge case.
(d) That Sh
p
should have good reduction when K
p
is hyperspecial was conjectured in
Langlands 1976, p411. That there should be a canonical model characterized by a condition
like that in (b) was conjectured in Milne 1992, ~2.
Denition of the Langlands-Rapoport set
Let (G, X) be a Shimura datum for which SV4,5,6 hold, and let
Sh
p
(C) = Sh(C),K
p
= lim
K
p
Sh
K
p
K
p
(G, X)(C).
85
Over the reex eld, Shimura varieties of Hodge type are no more difcult than Shimura varieties of
PEL-type, but when one reduces modulo a prime they become much more difcult for two reasons: general
tensors are more difcult to work with than endomorphisms, and little is known about Hodge tensors in
characteristic p.
132 16 THE GOOD REDUCTION OF SHIMURA VARIETIES
For x c X, let I(x) be the subgroup G() xing x, and let
S(x) = I(x)X
p
(x) X
p
(x), X
p
(x) = G(A
p
(
), X
p
(x) = G(
p
),K
p
.
One sees easily that there is a canonical bijection of sets with G(A
p
(
)-action
S(x) Sh
p
(C)
where the left hand side is the disjoint union over a set of representatives for G()X.
This decomposition has a modular interpretation. For example, in the case of a Shimura
variety of hodge type, the set S(x) classies the family of isomorphism classes of triples
(A, (s
i
), jK) with (A, (s
i
)) isomorphic to a xed pair.
Langlands and Rapoport (1987, 5e) conjecture that Sh
p
(F) has a similar description ex-
cept that now the left hand side runs over a set of isomorphism classes of homomorphisms
: P E
G
. Recall that an isomorphism from one to a second
t
is an element g of
G(
al
) such that
t
(p) = g (p) g
-1
, all p c P.
Such a should be thought of as a pre fake abelian motive with tensors. Specically,
if we x a faithful representation G GL(V ) and tensors t
i
for V such that G is the
subgroup of GL(V ) xing the t
i
, then each gives a representation P GL(V (
al
)) I
(i.e., a fake abelian motive) plus tensors.
Denition of the set S()
We now x a homomorphism : P E
G
and dene a set S() equipped with a right
action of G(A
p
(
) and a commuting Frobenius operator .
Denition of the group I(). The group I() is dened to be the group of automor-
phisms of ,
I() = {g c G(
al
) [ ad(g) = }.
Note that if ,: G GL(V ) is a faithful representation of G, then , is a fake motive
and I() Aut(, ) (here we have abbreviated , 1 to ,).
Denition of X
p
(). Let be a prime ,= p, o. We choose a prime n
I
of
al
dividing ,
and dene
al
I
and I
I
I as on p126.
Regard I
I
as an
al
I
,
I
-afne extension with trivial kernel, and write
I
for the homo-
morphism
o .(1, o): I
I
E
G
(), E
G
() = G(
al
I
) I
I
.
From we get a homomorphism (): P() E
G
(), and, on composing this with
the homomorphism (
I
: I
I
P() provided by (15.13), we get a second homomorphism
I
I
E
G
().
Dene
X
I
() = Isom(
I
, (
I
()).
Denition of the Langlands-Rapoport set 133
Clearly, Aut(
I
) = G(
I
) acts on X
I
() on the right, and I() acts on the left. If X
I
()
is nonempty, then the rst action makes X
I
() into a principal homogeneous space for
G(
I
).
Note that if ,: G GL(V ) is a faithful representation of G, then
X
I
() Isom(V (
I
), V
I
(, )). (62)
By choosing the (
I
judiciously (cf. p126), we obtain compact open subspaces of the
X
I
(), and we can dene X
p
() to be the restricted product of the X
I
(). If nonempty, it
is a principal homogeneous space for G(A
p
(
).
Denition of X
p
(). We choose a prime n
p
of
al
dividing p, and we use the notations
of p127. We let L denote the completion of
un
p
, and we let O
L
denote the ring of integers
in L(it is the ring of Witt vectors with coefcients in F). We let o Frobenius automorphism
of
un
p
or L that acts as x .x
p
on the residue eld.
From and (15.14), we have homomorphisms
D P(p)
un
(p)
un
G(
un
p
) I
un
p
.
For some n, the composite factors through D
n
. There is a canonical element in D
n
mapping
to o, and we let (b, o) denote its image in G(
un
p
) I
un
p
. The image b() of b in G(L)
is well-dened up to o-conjugacy, i.e., if b()
t
also arises in this way, then b()
t
= g
-1
b() og.
Note that if ,: G GL(V ) is a faithful representation of G, then D( ,) is V (L)
with F acting as : .b()o:.
Recall p102 that we have a well-dened G(
al
)-conjugacy class c(X) of cocharacters
of G
Q
al . We can transfer this to conjugacy class of cocharacters of G
Q
al
p
, which contains
an element j dened over
un
p
(see 12.3; G splits over
un
p
because we are assuming it
contains a hyperspecial group). Let
C
p
= G(O
L
) j(p) G(O
L
).
Here we are writing G(O
L
) for G(O
L
) with G as in the denition of hyperspecial.
Dene
X
p
() = {g c G(L),G(O
L
) [ g
-1
b() g c C
p
}.
There is a natural action of I() on this set.
Denition of the Frobenius element . For g c X
p
(), dene
(g) = b() ob() o
m-1
b() o
m
g
where m = [E
:
p
].
The set S(). Let
S() = I()X
p
() X
p
().
Since I() acts on both X
p
() and X
p
(), this makes sense. The group G(A
p
(
) acts on
S() through its action on X
p
() and acts through its action on X
p
().
134 16 THE GOOD REDUCTION OF SHIMURA VARIETIES
The admissibility condition
The homomorphisms : P E
G
contributing to the Langlands-Rapoport set must satisfy
an admissibility condition at each prime plus one global condition.
The condition at o. Let E
o
be the extension
1 C
E
o
I
o
1, I
o
= Gal(C,1) = (|)
associated with the quaternion algebra H, and regard it as an afne extension with kernel
G
m
. Note that E
o
= C
. C
j and jzj
-1
= z.
From the diagram (60) with l = o, we obtain a C,1-afne extension
1 P(C) P(o) I
o
1.
LEMMA 16.6. There is a homomorphism (
o
: E
o
P(o) whose restriction to the ker-
nels, G
m
.P
C
, corresponds to the map on characters .nt ().
PROOF. This follows from the fact that the homomorphism H
2
(I
o
, G
m
) H
2
(I
o
, P
R
)
sends the cohomology class of E
o
to that of P(o).
LEMMA 16.7. For any x c X, the formulas
x
(z) = (n
X
(z), 1),
x
(j) = (j
x
(1)
-1
, |)
dene a homomorphism E
o
P(o). Replacing x with a different point, replaces the
homomorphism with an isomorphic homomorphism.
PROOF. Easy exercise.
Write
X
for the isomorphism class of homomorphisms dened in (16.7). Then the
admissibility condition at ois that (
o
(o) c
X
.
The condition at ,= p. The admissibility condition at ,= p is that the set X
I
() be
nonempty, i.e., that (
I
() be isomorphic to
I
.
The condition at p. The condition at p is that the set X
p
() be nonempty.
The global condition. Let v: G T be the quotient of G by its derived group. From X
we get a conjugacy class of cocharacters of G
C
, and hence a well dened cocharacter j of
T . Under our hypotheses on (G, X), j satises the conditions of (15.15), and so denes a
homomorphism
: P E
T
. The global condition is that v be isomorphic to
.
The Langlands-Rapoport set
The Langlands-Rapoport set
LR(G, X) =
_
S()
where the disjoint union is over a set of representatives for the isomorphism classes of
admissible homomorphism : P E
G
. There are commuting actions of G(A
p
(
) and of
the Frobenius operator on LR(G, X).
The conjecture of Langlands and Rapoport 135
The conjecture of Langlands and Rapoport
CONJECTURE 16.8 (LANGLANDS AND RAPOPORT 1987). Let (G, X) be a Shimura da-
tum satisfying SV4,5,6 and such that G
der
is simply connected, and let K
p
be a hyperspecial
subgroup of G(
p
). Let p be a prime of E(G, X) dividing p, and assume that Sh
p
has
canonical good reduction at p. Then there is a bijection of sets
LR(G, X) Sh
p
(G, X)(F) (63)
compatible with the actions G(A
p
(
) and the Frobenius elements.
REMARK 16.9. (a) The conditions SV5 and SV6 are not in the original conjecture I
included them to simplify the statement of the conjecture.
(b) There is also a conjecture in which one does not require SV4, but this requires that
P be replaced by a more complicated afne extension
86
Q.
(c) The conjecture as originally stated is denitely wrong without the assumption that
G
der
is simply connected. However, when one replaces the admissible homomorphisms
in the statement with another notion, that of special homomorphisms, one obtains a state-
ment that should be true for all Shimura varieties. In fact, it is known that the statement
with G
der
simply connected then implies the general statement (see Milne 1992, ~4, for the
details and a more precise statement).
(d) It is possible to state, and prove, a conjecture similar to (16.8) for zero-dimensional
Shimura varieties. The map (G, X) (T, Y ) (see p60) denes a map of the associated
Langlands-Rapoport sets, and we should add to the conjecture that
LR(G, X) Sh
p
(G, X)(F)
_
LR(T, Y ) Sh
p
(T, Y )(F)
commutes.
86
This is done in the original paper of Langlands and Rapoport, but their denition is of Q is wrong. For a
correct denition, see Pfau 1996.
136 17 A FORMULA FOR THE NUMBER OF POINTS
17 A formula for the number of points
A reader of the last two sections may be sceptical of the value of a description like (63),
even if proved. In this section we briey explain how it leads to a very explicit, and useful,
formula for the number of points on the reduction of a Shimura variety with values in a
nite eld.
Throughout, (G, X) is a Shimura datum satisfying SV4,5,6 and K
p
is a hyperspecial
subgroup of G(
p
). We assume that G
der
simply connected and that Sh
p
(G, X) has canon-
ical good reduction at a prime p[p of the reex eld E = E(G, X). Other notations are as
in the last section; for example, L
n
is the subeld of
un
p
of degree n over
p
and L is the
completion of
un
p
. We x a eld F
q
k(p) F
p
, q = p
n
.
Triples
We consider triples (;
0
; ;, ) where
;
0
is a semisimple element of G() that is contained in an elliptic torus of G
R
(i.e.,
a torus that is anisotropic modulo the centre of G
R
),
; = (;())
I,=p,o
is an element of G(A
p
(
) such that, for all , ;() becomes conjugate
to ;
0
in G(
al
I
),
is an element of G(L
n
) such that
N
df
= o . . . o
n-1
,
becomes conjugate to ;
0
in G(
al
p
).
Two triples (;
0
; ;, ) and (;
t
0
; ;
t
,
t
) are said to be equivalent, (;
0
; ;, ) (;
t
0
; ;
t
,
t
), if
;
0
is conjugate to ;
t
0
in G(), ;() is conjugate to ;
t
() in G(
I
) for each ,= p, o, and
is o-conjugate to
t
in G(L
n
).
Given such a triple (;
0
; ;, ), we set:
I
0
= G
;
0
, the centralizer of ;
0
in G; it is connected and reductive;
I
o
= the inner form of I
0R
such that I
o
,Z(G) is anisotropic;
I
I
= the centralizer of ;() in G
Q
`
;
I
p
= the inner form of G
Q
p
such that I
p
(
p
) = {x c G(L
n
) [ x
-1
ox = }.
We need to assume that the triple satises the following condition:
(*) there exists an inner form I of I
0
such that I
Q
`
is isomorphic to I
I
for
all (including p and o).
Because ;
0
and ;
I
are stably conjugate, there exists an isomorphism a
I
: I
0,Q
al
`
I
I,Q
al
`
,
well-dened up to an inner automorphism of I
0
over
al
I
. Choose a system (I, a, (j
I
))
consisting of a -group I, an inner twisting a: I
0
I (isomorphism over
al
), and iso-
morphisms j
I
: I
Q
`
I
I
over
I
for all , unramied for almost all , such that j
I
a and a
I
differ by an inner automorphism our assumption (*) guarantees the existence of such a
system. Moreover, any other such system is isomorphic to one of the form(I, a, (j
I
adh
I
))
where (h
I
) c I
ad
(A).
The triple attached to an admissible pair (, c) 137
Let dx denote the Haar measure on G(A
p
(
) giving measure 1 to K
p
. Choose a Haar
measure di
p
on I(A
p
(
) that gives rational measure to compact open subgroups of I(A
p
(
),
and use the isomorphisms j
I
to transport it to a measure on G(A
p
(
)
;
(the centralizer of
; in G(A
p
(
)). The resulting measure does not change if (j
I
) is modied by an element
of I
ad
(A). Write dx for the quotient of dx by di
p
. Let be an element of the Hecke
algebra H of locally constant K-bi-invariant -valued functions on G(A
(
), and assume
that =
p
p
where
p
is a function on G(A
p
(
) and
p
is the characteristic function of
K
p
in G(
p
) divided by the measure of K
p
. Dene
O
;
(
p
) =
_
G(A
p
f
)
\G(A
p
f
)
p
(x
-1
;x) dx
Let dy denote the Haar measure on G(L
n
) giving measure 1 to G(O
L
n
). Choose a
Haar measure di
p
on I(
p
) that gives rational measure to the compact open subgroups, and
use j
p
to transport the measure to I
p
(
p
). Again the resulting measure does not change
if j
p
is modied by an element of I
ad
(
p
). Write dy for the quotient of dy by di
p
.
Proceeding as on p133, we choose a cocharacter jin c(X) well-adapted to the hyperspecial
subgroup K
p
and dened over L
n
, and we let be the characteristic function of the coset
G(O
L
n
) j(p) G(O
L
n
). Dene
TO
() =
_
I(Q
p
)\G(L
n
)
(y
-1
o(y))dy
Since I,Z(G) is anisotropic over 1, and since we are assuming SV5, I() is a discrete
subgroup of I(A
p
(
), and we can dene the volume of I()I(A
(
). It is a rational number
because of our assumption on di
p
and di
p
. Finally, dene
I(;
0
; ;, ) = I(;
0
; ;, )(
p
, r) = vol(I()I(A
(
)) O
;
(
p
) TO
(
r
).
The integral I(;
0
; ;, ) is independent of the choices made, and
(;
0
; ;, ) (;
t
0
; ;
t
,
t
) == I(;
0
; ;, ) = I(;
t
0
; ;
t
,
t
).
The triple attached to an admissible pair (, c)
An admissible pair (, ;
0
) is an admissible homomorphism : P E
G
and a ; c I
()
such that ;
0
x =
r
x for some x c X
p
(). Here r = [k(p): F
p
]. An isomorphism
(, ;
0
) (
t
, ;
t
0
) of admissible pairs is an isomorphism
t
sending ; to ;
t
, i.e., it
is a g c G(
al
) such that
ad(g) =
t
, ad(g)(;) = ;
t
.
Let (T, x) (G, X) be a special pair. Because of our assumptions on (G, X), the
cocharacter j
x
of T satises the conditions of (15.15) and so denes a homomorphism
x
: P E
T
. Langlands and Rapoport (1987, 5.23) show that every admissible pair is
isomorphic to a pair (, ;) with =
x
and ; c T(). For such a pair (, ;), b() is
represented by a c G(L
n
) which is well-dened up to conjugacy.
Let ; be the image of ;
0
in G(A
p
(
). Then the triple (;
0
; ;, ) satises the conditions
in the last subsection. A triple arising in this way from an admissible pair will be called
effective.
138 17 A FORMULA FOR THE NUMBER OF POINTS
The formula
For a triple (;
0
. . .), the kernel of
H
1
(, I
0
) H
1
(, G)
l
H
1
(
l
, I
0
)
is nite we denote its order by c(;
0
).
THEOREM 17.1. Let (G, X) be a Shimura datum satisfying the hypotheses of (16.8). If that
conjecture is true, then
# Sh
p
(F
q
) =
(;
0
;;,)
c(;
0
) I(;
0
; ;, ) (64)
where the sum is over a set of representatives for the effective triples.
PROOF. See Milne 1992, 6.13.
NOTES. Early versions of (64) can be found in papers of Langlands, but the rst precise
general statement of such a formula is in Kottwitz 1990. There Kottwitz attaches a coho-
mological invariant (;
0
; ;, ) to a triple (;
0
; ;, ), and conjectures that the formula (64)
holds if the sum is taken over a set of representatives for the triples with = 1 (ibid. ~3).
Milne (1992, 7.9) proves that, among triples contributing to the sum, = 1 if and only if
the triple is effective, and so the conjecture of Langlands and Rapoport implies Kottwitzs
conjecture.
87
On the other hand, Kottwitz (1992) proves his conjecture for Shimura vari-
eties of simple PEL type A or C unconditionally (without however proving the conjecture
of Langlands and Rapoport for these varieties).
87
At least in the case that the weight is rational Kottwitz does not make this assumption.
139
18 Endnotes
Proof of Theorem 5.4 (footnote 39)
First Proof: A torus T over a eld k is said to be quasisplit if it is a product of tori
of the form Res
F{k
G
m
. For such a torus T , Shapiros lemma and Hilberts Theorem 90
imply that H
1
(k, T) = 0, and the weak approximation theorem in algebraic number theory
implies that real approximation holds.
Let T be a torus over , and let F be a nite galois extension of k splitting T with
galois group I. From an x c X
+
(T) xed by ^ I, we get a homomorphismZ[I,^]
X
+
(T), [;] .;x. On applying this observation to enough xs, we get an exact sequence
0 M
2
M
1
X
+
(T) 0 (*)
of I-modules with M
1
a nite direct sum of modules of the form Z[I,^] (varying ^) and
M
Z
1
X
+
(T)
Z
surjective for all subgroups ^ of I. It follows from the cohomology
sequence of (*) that H
1
(^, M
2
) = 0 for all subgroups ^ of I. The sequence (*) is the
cocharacter sequence of an exact sequence of tori
0 T
2
T
1
T 0 (**)
with T
1
quasisplit. The cohomology sequence of (**) is an exact sequence
T
1
(1) T(1) H
1
(C,1, T
2
).
But
H
1
(C,1, T
2
)
= H
-1
Tate
((|), M
2
)
= H
1
((|), M
2
) = 0
where | denotes complex conjugation (the rst isomorphism is Tate-Nakayama, and the
second is the periodicity of the cohomology of cyclic groups). Therefore, T
1
(1) maps onto
T(1), and so the real approximation theorem for T follows from that for T
1
.
Let S be a group of multiplicative type over (i.e., S is commutative and S
is a
torus). Then X
+
(S) is a quotient of a direct sum of modules of the form Z[I,^] (as in the
preceding paragraph), and correspondingly there is an exact sequence
0 S T
1
T
2
0
with T
1
quasisplit. From the diagram
T
1
() T
2
() H
1
(, S) 0
_dense
_
T
1
(1)
onto T
2
(R)
C
T
2
(1) H
1
(1, S) 0
we see that H
1
(, S) H
1
(1, S) is surjective.
Finally, let G be a reductive group with centre Z. Choose a surjection T Z
with
T a quasisplit torus (cf. (**) above). As G = G
der
Z
_dense
_
onto
_
injective
G
t
(1) T(1)
onto G(R)
C
G(1) H
1
(1, S) H
1
(1, G
t
)
we see that the real approximation theorem holds for G (the injectivity of the arrow at right
is the Hasse principle for G
t
, Platonov and Rapinchuk 1994, Theorem 6.6, p286).
Second Proof (from a letter of G. Prasad, Sept 1, 1987). To prove that G() is
dense in G(1), what you need is a result of H. Matsumoto, which is reproved in Borel-
Tits Groupes r eductifs, Publ. Math. IHES no 27, as Th eor` eme 14.4, according to which
given a maximal 1-split torus S of G, S(1) meets every connected component of G(1).
Now we observe that there is a maximal torus T dened over which contains a maximal
1-split torus of G: To prove this, we will make use of the fact that the closure of G()
contains G(1)
l=2,3,...,o
H
1
(
l
, Z
t
) is injective. Then, for any sufciently small compact open sub-
group K of G(A
(
),
G()
K Z() G
der
().
PROOF. (Cf. the proof of 5.20.) Consider the diagram:
1 Z
t
() Z() G
der
() G() H
1
(, Z
t
)
_
_
_
_
_
_
_
1 Z
t
() G
der
() G
ad
() H
1
(, Z
t
).
Let q c G()
I[d
K
I
, then it maps to zero in
l nite
H
1
(
l
, Z
t
). Because of the Hasse principle, this implies that g maps to zero in
H
1
(, Z
t
), and therefore lies in Z() G
der
().
142 REFERENCES
References
Ash, A., Mumford, D., Rapoport, M. and Tai, Y. (1975), Smooth compactication of locally symmetric vari-
eties, Math. Sci. Press, Brookline, Mass.
Baily, Jr., W. L. and Borel, A. (1966), Compactication of arithmetic quotients of bounded symmetric do-
mains, Ann. of Math. (2) 84, 442528.
Baily, W. L. (1958), Satakes compactication of V
n
, Amer. J. Math. 80, 348364.
Boothby, W. M. (1975), An introduction to differentiable manifolds and Riemannian geometry, Academic
Press, New York-London.
Borel, A. (1962), Ensembles fondamentaux pour les groupes arithm etiques, in Colloq. Th eorie des Groupes
Alg ebriques (Bruxelles, 1962), Librairie Universitaire, Louvain, pp. 2340.
Borel, A. (1969), Introduction aux groupes arithm etiques, Publications de lInstitut de Math ematique de
lUniversit e de Strasbourg, XV. Actualit es Scientiques et Industrielles, No. 1341, Hermann, Paris.
Borel, A. (1972), Some metric properties of arithmetic quotients of symmetric spaces and an extension
theorem, J. Differential Geometry 6, 543560.
Borel, A. (1991), Linear algebraic groups, Vol. 126 of Graduate Texts in Mathematics, Springer-Verlag, New
York.
Borel, A. (1998), Semisimple groups and Riemannian symmetric spaces, Vol. 16 of Texts and Readings in
Mathematics, Hindustan Book Agency, New Delhi.
Borel, A. and Harish-Chandra (1962), Arithmetic subgroups of algebraic groups, Ann. of Math. (2) 75, 485
535.
Borovo, M. V. (1984), Langlands conjecture concerning conjugation of connected Shimura varieties, Se-
lecta Math. Soviet. 3(1), 339.
Borovo, M. V. (1987), The group of points of a semisimple group over a totally real-closed eld, in Problems
in group theory and homological algebra (Russian), Matematika, Yaroslav. Gos. Univ., pp. 142149.
Bourbaki, N. (1958),
El ements de math ematique. I: Les structures fondamentales de lanalyse. Fascicule
VII. Livre II: Alg` ebre. Chapitre 3: Alg` ebre multilin eaire, Nouvelle edition. Actualit es Scientiques et
Industrielles, No. 1044, Hermann, Paris.
Bourbaki, N. (1981), Groupes et Alg` ebres de Lie. Chapitres 4,5 et 6, Elements of Mathematics, Masson,
Paris.
Bourbaki, N. (1989), General topology. Chapters 14, Elements of Mathematics, Springer-Verlag, Berlin.
Bravo, A., Encinas, S. and Villamayor, O. (2002), A Simplied Proof of Desingularization and Applica-
tions. arXiv:math.AG/0206244.
Brylinski, J.-L. (1983), 1-motifs et formes automorphes (th eorie arithm etique des domaines de Siegel), in
Conference on automorphic theory (Dijon, 1981), Vol. 15 of Publ. Math. Univ. Paris VII, Univ. Paris
VII, Paris, pp. 43106.
Cartan, H. (1958), Prolongement des espaces analytiques normaux, Math. Ann. 136, 97110.
Cartan, H. (1963), Elementary theory of analytic functions of one or several complex variables,
Editions
Scientiques Hermann, Paris.
Casselman, W. A. (1997), Geometric rationality of Satake compactications, in Algebraic groups and Lie
groups, Vol. 9 of Austral. Math. Soc. Lect. Ser., Cambridge Univ. Press, Cambridge, pp. 81103.
Chow, W.-L. (1949), On compact complex analytic varieties, Amer. J. Math. 71, 893914.
Deligne, P. (1970),
Equations diff erentielles ` a points singuliers r eguliers, Springer-Verlag, Berlin.
REFERENCES 143
Deligne, P. (1971a), Th eorie de Hodge. II, Inst. Hautes
Etudes Sci. Publ. Math. 40, 557.
Deligne, P. (1971b), Travaux de Shimura, in S eminaire Bourbaki, 23` eme ann ee (1970/71), Exp. No. 389,
Springer, Berlin, pp. 123165. Lecture Notes in Math., Vol. 244.
Deligne, P. (1971c), Travaux de Shimura. Preliminary unpublished version of Deligne 1971b.
Deligne, P. (1973a), Espaces hermitiens sym etriques, Spring 1973, Transcription available at
www.jmilne.org/math/.
Deligne, P. (1973b), Les constantes des equations fonctionnelles des fonctions L, in Modular functions of
one variable, II (Proc. Internat. Summer School, Univ. Antwerp, Antwerp, 1972), Springer, Berlin,
pp. 501597. Lecture Notes in Math., Vol. 349.
Deligne, P. (1979), Vari et es de Shimura: interpr etation modulaire, et techniques de construction de mod` eles
canoniques, in Automorphic forms, representations and L-functions (Proc. Sympos. Pure Math., Ore-
gon State Univ., Corvallis, Ore., 1977), Part 2, Proc. Sympos. Pure Math., XXXIII, Amer. Math. Soc.,
Providence, R.I., pp. 247289.
Deligne, P. (1982), Hodge cycles on abelian varieties (notes by J.S. Milne), in Hodge cycles, motives, and
Shimura varieties, Vol. 900 of Lecture Notes in Mathematics, Springer-Verlag, Berlin, pp. 9100.
Available at www.jmilne.org/math/.
Demazure, M. (1972), Lectures on p-divisible groups, Springer-Verlag, Berlin.
Farb, B. and Dennis, R. K. (1993), Noncommutative algebra, Vol. 144 of Graduate Texts in Mathematics,
Springer-Verlag, New York.
Giraud, J. (1968), Remarque sur une formule de Shimura-Taniyama, Invent. Math. 5, 231236.
Goresky, M. (2003), Compactications of modular varieties. Preliminary version of article for this volume.
Greenberg, M. J. (1967), Lectures on algebraic topology, W. A. Benjamin, Inc., New York-Amsterdam.
Hartshorne, R. (1977), Algebraic geometry, Springer-Verlag, New York.
Hatcher, A. (2002), Algebraic topology, Cambridge University Press, Cambridge. Available at
http://www.math.cornell.edu/ hatcher/.
Helgason, S. (1978), Differential geometry, Lie groups, and symmetric spaces, Vol. 80 of Pure and Applied
Mathematics, Academic Press Inc., New York.
Hironaka, H. (1964), Resolution of singularities of an algebraic variety over a eld of characteristic zero. I,
II, Ann. of Math. (2) 79 (1964), 109203; ibid. (2) 79, 205326.
Humphreys, J. E. (1972), Introduction to Lie algebras and representation theory, Springer-Verlag, New York.
Humphreys, J. E. (1978), Introduction to Lie algebras and representation theory, Vol. 9 of Graduate Texts in
Mathematics, Springer-Verlag, New York.
Ihara, Y. (1968), On congruence monodromy problems. Vol. 1, Lecture Notes, No. 1, Department of Mathe-
matics, University of Tokyo, Tokyo.
Iitaka, S. (1982), Algebraic geometry, Vol. 76 of Graduate Texts in Mathematics, Springer-Verlag, New York.
Klein, F. (1880), Zur Theorie der elliptischen Modulfunktionen, Math. Ann. 17, 6270.
Knus, M.-A., Merkurjev, A., Rost, M. and Tignol, J.-P. (1998), The book of involutions, Vol. 44 of American
Mathematical Society Colloquium Publications, American Mathematical Society, Providence, RI.
Kottwitz, R. E. (1984), Shimura varieties and twisted orbital integrals, Math. Ann. 269(3), 287300.
Kottwitz, R. E. (1990), Shimura varieties and z-adic representations, in Automorphic forms, Shimura va-
rieties, and L-functions, Vol. I (Ann Arbor, MI, 1988), Vol. 10 of Perspect. Math., Academic Press,
Boston, MA, pp. 161209.
144 REFERENCES
Kottwitz, R. E. (1992), Points on some Shimura varieties over nite elds, J. Amer. Math. Soc. 5(2), 373
444.
Krantz, S. G. (1982), Function theory of several complex variables, John Wiley & Sons Inc., New York.
Kwack, M. H. (1969), Generalization of the big Picard theorem, Ann. of Math. (2) 90, 922.
Langlands, R. P. (1976), Some contemporary problems with origins in the Jugendtraum, in Mathematical
developments arising from Hilbert problems (Proc. Sympos. Pure Math., Vol. XXVIII, Northern Illinois
Univ., De Kalb, Ill., 1974), Amer. Math. Soc., Providence, R. I., pp. 401418.
Langlands, R. P. (1977), Shimura varieties and the Selberg trace formula, Canad. J. Math. 29(6), 1292
1299.
Langlands, R. P. and Rapoport, M. (1987), Shimuravariet aten und Gerben, J. Reine Angew. Math. 378, 113
220.
Lee, J. M. (1997), Riemannian manifolds, Vol. 176 of Graduate Texts in Mathematics, Springer-Verlag, New
York.
Mac Lane, S. (1998), Categories for the working mathematician, Vol. 5 of Graduate Texts in Mathematics,
second edn, Springer-Verlag, New York.
Matsumoto, H. (1969), Sur les sous-groupes arithm etiques des groupes semi-simples d eploy es, Ann. Sci.