Gerard van der Geer

Ben Moonen

ABELIAN VARIETIES
(PRELIMINARY VERSION OF THE FIRST CHAPTERS)
Contents

Notation and conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

1. Definitions and basic examples . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2. Line bundles and divisors on abelian varieties . . . . . . . . . . . . . . . . . . 17
§ 1. The theorem of the square . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
§ 2. Projectivity of abelian varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
§ 3. Projective embeddings of abelian varieties . . . . . . . . . . . . . . . . . . . . . . 26
3. Basic theory of group schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
§ 1. Definitions and examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
§ 2. Elementary properties of group schemes . . . . . . . . . . . . . . . . . . . . . . . 34
§ 3. Cartier duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
§ 4. The component group of a group scheme . . . . . . . . . . . . . . . . . . . . . . . 43
4. Quotients by group schemes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
§ 1. Categorical quotients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
§ 2. Geometric quotients, and quotients by finite group schemes . . . . . . . . . . . . . 52
§ 3. FPPF quotients . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
§ 4. Finite group schemes over a field . . . . . . . . . . . . . . . . . . . . . . . . . . . 66
5. Isogenies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
§ 1. Definition of an isogeny, and basic properties . . . . . . . . . . . . . . . . . . . . . 72
§ 2. Frobenius and Verschiebung . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
§ 3. Density of torsion points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6. The Picard scheme of an abelian variety . . . . . . . . . . . . . . . . . . . . . 87
§ 1. Relative Picard functors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
§ 2. Digression on graded bialgebras . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
§ 3. The dual of an abelian variety . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
7. Duality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
§ 1. Formation of quotients and the descent of coherent sheaves . . . . . . . . . . . . . 98
§ 2. Two duality theorems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
§ 3. Further properties of Pic0X/k . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
§ 4. Applications to cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
§ 5. The duality between Frobenius and Verschiebung . . . . . . . . . . . . . . . . . . 110
8. The Theta group of a line bundle . . . . . . . . . . . . . . . . . . . . . . . . . 113
§ 1. The theta group G (L) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113
§ 2. Descent of line bundles over homomorphisms . . . . . . . . . . . . . . . . . . . . . 116
§ 3. Theta groups of non-degenerate line bundles . . . . . . . . . . . . . . . . . . . . . 118
§ 4. Representation theory of non-degenerate theta groups . . . . . . . . . . . . . . . . 122
9. The cohomology of line bundles . . . . . . . . . . . . . . . . . . . . . . . . . . 126
10. Tate modules, p-divisible groups, and the fundamental group . . . . . . . . 141
§ 1. Tate-!-modules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

i
§ 2. The p-divisible group . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 145
§ 3. The algebraic fundamental group—generalities . . . . . . . . . . . . . . . . . . . . 149
§ 4. The fundamental group of an abelian variety . . . . . . . . . . . . . . . . . . . . . 154
11. Polarizations and Weil pairings . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
§ 1. Polarizations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159
§ 2. Pairings . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
§ 3. Existence of polarizations, and Zarhin’s trick . . . . . . . . . . . . . . . . . . . . . 169
§ 4. Polarizations associated to line bundles on torsors . . . . . . . . . . . . . . . . . . 174
§ 5. Symmetric line bundles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
12. The endomorphism ring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
§ 1. First basic results about the endomorphism algebra . . . . . . . . . . . . . . . . . 180
§ 2. The characteristic polynomial of an endomorphism . . . . . . . . . . . . . . . . . 184
§ 3. The Rosati involution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
§ 4. The Albert classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
13. The Fourier transform and the Chow ring . . . . . . . . . . . . . . . . . . . . 192
§ 1. The Chow ring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
§ 2. The Hodge bundle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 196
§ 3. The Fourier transform of an abelian variety . . . . . . . . . . . . . . . . . . . . . 199
§ 4. Decomposition of the diagonal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204
§ 5. Motivic decomposition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
14. Jacobian Varieties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
§ 1. The Jacobian variety of a curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
§ 2. Comparison with the g-th symmetric power of C . . . . . . . . . . . . . . . . . . . 221
§ 3. Universal line bundles and the Theta divisor . . . . . . . . . . . . . . . . . . . . . 224
§ 4. Riemann’s Theorem on the Theta Divisor . . . . . . . . . . . . . . . . . . . . . . 231
§ 5. Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233
§ 6. A universal property—the Jacobian as Albanese . . . . . . . . . . . . . . . . . . . 235
§ 7. Any Abelian Variety is a Factor of a Jacobian . . . . . . . . . . . . . . . . . . . . 236
§ 8. The Theorem of Torelli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237
§ 9. The Criterion of Matsusaka-Ran . . . . . . . . . . . . . . . . . . . . . . . . . . . . 239
15. Dieudonné theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 245
§ 1. Dieudonné theory for finite commutative group schemes and for p-divisible groups 245
§ 2. Classification up to isogeny . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 246
§ 3. The Newton polygon of an abelian variety . . . . . . . . . . . . . . . . . . . . . . 262
16. Abelian Varieties over Finite Fields . . . . . . . . . . . . . . . . . . . . . . . . 265
§ 1. The eigenvalues of Frobenius . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
§ 2. The Hasse-Weil-Serre bound for curves . . . . . . . . . . . . . . . . . . . . . . . . 272
§ 3. The theorem of Tate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
§ 4. Corollaries of Tate’s theorem, and the structure of the endomorphism algebra . . . 281
§ 5. Abelian varieties up to isogeny and Weil numbers . . . . . . . . . . . . . . . . . . 289
§ 6. Isomorphism classes contained in an isogeny class . . . . . . . . . . . . . . . . . . 293
§ 7. Elliptic curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 296
§ 8. Newton polygons of abelian varieties over finite fields . . . . . . . . . . . . . . . . 302
§ 9. Ordinary abelian varieties over a finite field . . . . . . . . . . . . . . . . . . . . . 303

ii
Appendix A. Algebra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 308
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321

iii
Notation and conventions.

(0.1) In general, k denotes an arbitrary field, k̄ denotes an algebraic closure of k, and ks a

separable closure.

(0.2) If A is a commutative ring, we sometimes simply write A for Spec(A). Thus, for instance,
by an A-scheme we mean a scheme over Spec(A). If A → B is a homomorphism of rings and X
is an A-scheme then we write XB = X ×A B rather than X ×Spec(A) Spec(B).

(0.3) If X is a scheme then we write |X| for the topological space underlying X and OX for its
structure sheaf. If f : X → Y is a morphism of schemes we write |f |: |X| → |Y | and f ! : OY →
f∗ OX for the corresponding map on underlying spaces, resp. the corresponding homomorphism
of sheaves on Y . If x ∈ |X| we write k(x) for the residue field. If X is an integral scheme we
write k(X) for its field of rational functions.
If S is a scheme and X and T are S-schemes then we write X(T ) for the set of T -valued
points of X, i.e., the set of morphisms of S-schemes T → X. Often we simply write XT for
the base change of X to T , i.e., XT := X ×S T , to be viewed as a T -scheme via the canonical
morphism XT → T .

(0.4) If k is a field then by a variety over k we mean a separated k-scheme of finite type which
is geometrically integral. Recall that a k-scheme is said to be geometrically integral if for some
algebraically closed field K containing k the scheme XK is irreducible and reduced. By EGA
IV, (4.5.1) and (4.6.1), if this holds for some algebraically closed overfield K then XK is integral
for every field K containing k. A variety of dimension 1 (resp. 2, resp. n ! 3) is called a curve
(resp. surface, resp. n-fold).
By a line bundle (resp. a vector bundle of rank d) on a scheme X we mean a locally free
OX -module of rank 1 (resp. of rank d). By a geometric vector bundle of rank d on X we mean a
group scheme π: V → X over X for which there exists a affine open covering X = ∪Uα such that
the restriction of V to each Uα is isomorphic to Gda over Uα . In particular this means that we
∼
have isomorphisms of Uα -schemes ϕα : π −1 (Uα ) −→ Uα × Ad , such that all transition morphisms
ϕβ ◦ ϕ−1
tα,β : Uα,β × Ad −−−−−α−→ Uα,β × Ad

are linear automorphisms of Uα,β × Ad over Uα,β := Uα ∩ Uβ ; this last condition means that
tα,β is given by a O(Uα,β )-linear automorphism of O(Uα,β )[x1 , . . . , xd ]. For d = 1 we obtain the
notion of a geometric line bundle.
If V is a geometric vector bundle of rank d on X then its sheaf of sections is a vector
! bundle∨ of"
rank d. Conversely, if E is a vector bundle of rank d on X then the scheme V := Spec Sym(E )
has a natural structure of a geometric vector bundle of rank d. These two constructions are
quasi-inverse to each other and establish an equivalence between vector bundles and geometric
vector bundles.

(0.5) In our definition of an étale morphism of schemes we follow EGA; this means that we only
require the morphism to be locally of finite type. Note that in some literature étale morphisms
are assumed to be quasi-finite. Thus, for instance, if S is a scheme and I is an index set, the
#
disjoint union i∈I S is étale over S according to our conventions, also if the set I is infinite.

–1–
(0.6) If K is a number field then by a prime of K we mean an equivalence class of valuations
of K. See for instance Neukirch [1], Chap. 3. The finite primes of K are in bijection with
the maximal ideals of the ring of integers OK . An infinite prime corresponds either to a real
embedding K $→ R or to a pair {ι, ῑ} of complex embeddings K $→ C.
If v is a prime of K, we have a corresponding homomorphism ordv : K ∗ → R and a normal-
ized absolute value || ||v . If v is a finite prime then we let ordv be the corresponding valuation,
normalized such that ordv (K ∗ ) = Z, and we define || ||v by
$
(qv )−ordv (x) if x '= 0,
||x||v :=
0 if x = 0,

where qv is the cardinality of the residue field at v. If v is an infinite prime then we let
$
|ι(x)| if v corresponds to a real embedding ι: K → R,
||x||v =
|ι(x)|2 if v corresponds to a pair of complex embeddings {ι, ῑ},
! "
and we define ordv by the rule ordv (x) := − log |ι(x)| . Here | |: C → R!0 is given by |a + bi| =
√
a 2 + b2 .

–2–
 Definition. Let p be a prime number. We say that a scheme X has characteristic p if the
unique morphism X → Spec(Z) factors through Spec(Fp ) $→ Spec(Z). This is equivalent to the
requirement that p · f = 0 for every open U ⊂ X and every f ∈ OX (U ). We say that a scheme X
has characteristic 0 if X → Spec(Z) factors through Spec(Q) $→ Spec(Z). This is equivalent to
the requirement that n ∈ OX (U )∗ for every n ∈ Z \ {0} and every open U ⊂ X.
Note that if X → Y is a morphism of schemes and Y has characteristic p (with p a prime
number or p = 0) then X has characteristic p, too.

The absolute Frobenius. Let p be a prime number. Let Y be a scheme of characteristic p.

Then we have a morphism FrobY : Y → Y , called the absolute Frobenius morphism of Y ; it is
given by
(a) FrobY is the identity on the underlying topological space |Y |;
(b) Frob!Y : OY → OY is given on sections by f *→ f p .
To describe FrobY in another way, consider a covering {Uα } of Y by affine open subsets, say
Uα = Spec(Aα ). The endomorphism of Aα given by f *→ f p defines a morphism Frobα : Uα →
Uα . On the intersections Uα ∩ Uβ the morphisms Frobα and Frobβ agree, and by gluing we
obtain the absolute Frobenius morphism FrobY of Y . Note that Frobα is none other than the
absolute Frobenius morphism of the scheme Uα .
One readily verifies that for any morphism f : X → Y of schemes of characteristic p we have
a commutative diagram Frob
X −−−−X→ X
 

f&
f
& (1)
Frob
Y −−−−Y→ Y .

The relative Frobenius. Let us now consider the relative situation, i.e., we fix a base
scheme S and consider schemes over S. If π: X → S is an S-scheme then in general the absolute
Frobenius morphism FrobX is not a morphism of S-schemes, unless for instance S = Spec(Fp ).
To remedy this we define π (p) : X (p/S) → S to be the pull-back of π: X → S via FrobS : S → S.
Thus, by definition we have X (p/S) = S ×FrobS ,S X and we have a cartesian diagram
h
X (p/S) −−−−→ X
 
 π
π &
(p)
& (2)
Frob
S
S −−−−→ S .

If there is no risk of confusion we often write X (p) for X (p/S) ; note however that in general this
scheme very much depends on the base scheme S over which we are working.
As the diagram (2) is cartesian, the commutative diagram (1), applied with Y = S, gives a
commutative diagram (nog aanpassen)

X
+ FX/S
W
X (p/S) −−→ X (3)
 
 π
π (p) & &
FrobS
S −−−−→ S .

–3–
The morphism of S-schemes FX/S : X → X (p/S) is called the relative Frobenius morphism of X
over S. By its definition, FX/S is a morphism of S-schemes (in other words, π (p) ◦ FX/S = π)
and W ◦ FX/S is the absolute Frobenius of X.
! "
Example. Suppose S = Spec(R) and X = Spec R[t1 , . . . , tm ]/I for some ideal I =
(p)
(f1 , . . . , fn ) ⊂ R[t1 , . . . , tm ]. Let fi ∈ R[t1 , . . . , tm ] be the polynomial obtained from fi by
raising all coefficients (but not the variables!) to the pth power. ! Thus, if, in" multi-index
' (p) ' p α
notation, fi = cα tα then fi = cα t . Then X (p) = Spec R[t1 , . . . , tm ]/I (p) with I (p) =
(p) (p)
(f1 , . . . , fn ), and the relative Frobenius morphism FX/S : X → X (p) is given on rings by the
homomorphism
R[t1 , . . . , tm ]/I (p) −→ R[t1 , . . . , tm ]/I
with r *→ r for all r ∈ R and tj *→ tpj . Note that this is a well-defined homomorphism.

The morphism W : X (p) → X that appears in (3) does not have a standard name in the
literature. As one easily checks (see Exercise ??), FrobX/S ◦ W : X (p) → X (p) equals the absolute
Frobenius morphism of X (p) . Since an absolute Frobenius morphism is the identity on the
∼
underlying topological space, it follows that FX/S : X → X (p) induces a homeomorphism |X| −→
|X (p) |.
Formation of the relative Frobenius morphism is compatible with base change. This state-
ment means the following. Let π: X → S be an S-scheme. Let T → S be another scheme over S,
and consider the morphism πT : XT → T obtained from π by base-change. The first observation
is that (XT )(p/T ) is canonically isomorphic to (X (p/S) )T . Identifying the two schemes, the rela-
tive Frobenius FXT /T of XT over T is equal to the pull-back (FX/S )T of the relative Frobenius
of X over S. Proofs of these assertions are left to the reader.
The absolute and relative Frobenii can be iterated. For the absolute Frobenius this is
immediate: FrobnY : Y → Y is simply the nth iterate of FrobY . The nth iterate of the relative
n
n
Frobenius is a morphism FX/S : X → X (p /S) . Its definition is an easy generalization of the
n n
definition of FX/S . Namely, we define π (p ) : X (p /S) → S as the pull-back of π: X → S via
FrobnS . Then FrobnX factors as
n
FX/S (n)
n
/S) h
X −−−−→ X (p −−−→ X
n
with π (p )◦ n
FX/S = π. Alternatively,
2 ! "(p/S) 3 ! 2 "(p/S)
X (p /S) = X (p/S) , X (p /S)
= X (p /S) , etc.,
and
( FX/S FX (p) /S 2
F (pn−1 ) n
)
n X /S
FX/S = X −−−−→ X (p) −−−−−→ X (p ) −→ · · · −−−−−−−−→ X (p ) .

The geometric Frobenius. Suppose S = Spec(Fq ), with q = pn . If X is an S-scheme

then the nth iterate of the absolute Frobenius morphism FrobnX : X → X is a morphism of
S-schemes. In fact, FrobnX = FX/S
n
. We refer to πX := FrobnX as the geometric Frobenius of X.
More generally, suppose that S is a scheme over Spec(Fq ). If X is an S-scheme then by
an Fq -structure on X we mean a scheme X0 → Spec(Fq ) together with an isomorphism of
S-schemes X0 ⊗Fq S ∼ = X. In practice we usually encounter this notion in the situation that
S = Spec(K), where Fq ⊂ K is a field extension. Given an Fq -structure on X, the geometric
Frobenius morphism πX0 induces, by extension of scalars, a morphism πX : X → X; we again
refer to this morphism as the geometric Frobenius of X (relative to the given Fq -structure).

–4–
Chapter I. Definitions and basic examples.

An abelian variety is a complete algebraic variety whose points form a group, in such a way that
the maps defining the group structure are given by morphisms. It is the analogue in algebraic
geometry of the concept of a compact complex Lie group. To give a more precise definition of
a abelian variety we take a suitable definition of a group and translate it into the language of
complete varieties.

(1.1) Definition. A group consists of a non-empty set G together with maps

m: G × G → G (the group law) and i: G → G (the inverse)

and a distinguished element

e∈G (the identity element)

such that we have the following equalities of maps.

(i) Associativity: m ◦ (m × idG ) = m ◦ (idG × m): G × G × G −→ G.
(ii) Defining property of the identity element:

m ◦ (e × idG ) = j1 : {e} × G −→ G , and

m ◦ (idG × e) = j2 : G × {e} −→ G ,

∼ ∼
where j1 and j2 are the canonical identifications {e} × G −→ G and G × {e} −→ G,
respectively, and where we write e for the inclusion map {e} !→ G.
(iii) Left and right inverse:

e ◦ π = m ◦ (idG × i) ◦ ∆G = m ◦ (i × idG ) ◦ ∆G : G −→ G ,

where π: G → {e} is the constant map and ∆G is the diagonal map.

Written out in diagrams, we require the commutativity of the following diagrams.
(i) Associativity:
idG ×m
G × G × G −−− −−→ G × G
 

m×idG "
m
"
G×G −−−−−→ G .
m

(ii) Identity element:

e×id id ×e
{e} × G −−−−G→ G×G G × {e} −−G
−−→ G × G
j1 % &m and j2 % &m
G G .

DefBasEx, 15 september, 2011 (812)

–5–
(iii) Two-sided inverse:
π π
G −→ {e} G −→ {e}
   
 e  e
(idG ,i)" " and (i,idG )" "
G×G −→ G G×G −→ G .
m m

To simplify notation, one often simply writes the symbol G instead of the quadruple
(G, m, i, e), assuming it is clear what m, i and e are.
Adapting this definition to the category of varieties, we obtain the definition of a group
variety.

(1.2) Definition. A group variety over a field k is a k-variety X together with k-morphisms

m: X × X → X (the group law) and i: X → X (the inverse)

and a k-rational point

e ∈ X(k) (the identity element)
such that we have the following equalities of morphisms:
(i) m ◦ (m × idX ) = m ◦ (idX × m): X × X × X −→ X.
(ii) m ◦ (e × idX ) = j1 : Spec(k) × X −→ X and m ◦ (idX × e) = j2 : X × Spec(k) −→ X, where
∼ ∼
j1 : Spec(k) × X −→ X and j2 : X × Spec(k) −→ X are the canonical isomorphisms.
(iii) e ◦ π = m ◦ (idX × i) ◦ ∆X/k = m ◦ (i × idX ) ◦ ∆X/k : X −→ X, where π: X → Spec(k) is the
structure morphism.
Note that, since we are working with varieties, checking equality of two morphisms (as in
(i), (ii) and (iii)) can be done on k-rational points.
If X is a group variety then the set X(k) of k-rational points naturally inherits the structure
of a group. More generally, if T is any k-scheme then the morphisms m, i and e induce a group
structure on the set X(T ) of T -valued points of X. In this way, the group variety X defines
a contravariant functor from the category of k-schemes to the category of groups. In practice
it is often most natural to use this “functorial” point of view; we shall further discuss this
in Chapter III.
We can now define the main objects of study in this book.

(1.3) Definition. An abelian variety is a group variety which, as a variety, is complete.

As we shall see, the completeness condition is crucial: abelian varieties form a class of group
varieties with very special properties.
A group is a homogeneous space over itself, either via left or via right translations. We
have this concept here too.

(1.4) Definition. Let X be a group variety over a field k, and let x ∈ X(k) be a k-rational
point. We define the right translation tx : X → X and the left translation t#x : X → X to be the
compositions
# idX ×x m $
tx = X ∼ = X ×k Spec(k) −−−−→ X ×k X −→ X ,
and # $
x×idX m
t#x = X ∼
= Spec(k) ×k X −−−−→ X ×k X −→ X .

–6–
On points, these maps are given by tx (y) = m(y, x) and t#x (y) = m(x, y).
More generally, if T is a scheme over Spec(k) and x ∈ X(T ) is a T -valued point of X then
we define the right and left translations tx : XT → XT and t#x : XT → XT (with XT := X ×k T )
to be the compositions

# idXT ×xT m $
tx = XT ∼
= XT ×T T −−−−−−→ XT ×T XT −→ XT ,

and
# xT ×idXT m $
t#x = XT ∼
= T ×T XT −−−−−−→ XT ×T XT −→ XT ,
where we write xT : T → XT for the morphism (x, idT ): T → X ×k T = XT .

Figure 1.

Given a k-scheme T and two points x, y ∈ X(T ), one easily verifies that ty ◦ tx = tm(x,y)
and t#x ◦ t#y = t#m(x,y) . In particular, it follows that ti(x) = t−1 # # −1
x and ti(x) = (tx ) .
Geometrically, the fact that a group variety X is a principal homogenous space over itself
has the consequence that X, as a variety over k, “looks everywhere the same”. As a consequence
we obtain that group varieties are smooth and have a trivial tangent bundle.

(1.5) Proposition. Let X be a group variety over a field k. Then X is smooth over k. If
we write TX,e for the tangent space at the identity element, there is a natural isomorphism
TX/k ∼= TX,e ⊗k OX . This induces natural isomorphisms ΩnX/k ∼
= (∧n TX,e
∨
) ⊗k OX . In particular,
g ∼
if g = dim(X) then ΩX/k = OX .
Proof. Since X is a variety, the smooth locus sm(X/k) ⊂ X is open and dense. It is also stable
under all translations. Since these make X into a homogenous space over itself, it follows that
sm(X/k) = X.
# $
Set S = Spec k[ε]/(ε2 ) . Let XS := X ×k S, where the morphism S → Spec(k) is given
on rings by a +→ a + 0 · ε. We may think of XS as a “thickened” version of X. Tangent vectors
τ ∈ TX,e correspond to S-valued points τ̃ : S → X which reduce to e: Spec(k) → X modulo ε.
(See Exercise 1.3.) A vector field on X is given by an automorphism XS → XS which reduces
to the identity on X. To a tangent vector τ we can thus associate the vector field ξ(τ ) given by

–7–
the right translation tτ̃ . The map TX,e → Γ(X, TX/k ) given by τ +→ ξ(τ ) is k-linear and induces
a homomorphism α: TX,e ⊗k OX → TX/k .
The claim is that α is an isomorphism. As it is a homomorphism between locally free
OX -modules of the same rank, it suffices to show that α is surjective. If x ∈ X is a closed point
then the map

(αx mod mx ): TX,e ⊗k k(x) −→ (TX/k )x ⊗OX,x k(x) = TX,x

is the map TX,e → TX,x induced on tangent spaces by tx , which is an isomorphism. Applying the
Nakayama Lemma, it follows that the map on stalks αx : TX,e ⊗k OX,x → (TX/k )x is surjective.
As this holds for all closed points x, it follows that α is surjective.
The last assertion of the proposition now follows from the identities Ω1X/k = TX/k
∨
and
n n 1
ΩX/k = ∧ ΩX/k . !
With notations as in the proof, note that tτ̃ commutes with all left translations. It follows
that the vector field ξ(τ ) is left invariant, i.e., for every left translation t# we have t#∗ ξ(τ ) = ξ(τ ).
The map τ +→ ξ(τ ) identifies TX,e with the space of left invariant vector fields on X. In case X
is an abelian variety, so that Γ(X, OX ) = k, we can restate (1.5) by saying that all global vector
fields are left invariant.
If X is an abelian variety, these are the only global vector fields on X, since Γ(X, OX ) = k.

(1.6) Corollary. Any morphism from P1 to a group variety is constant.

Proof. Consider a morphism ϕ: P1 → X, with X a group variety. If ϕ is non-constant then
its image C ⊂ X is unirational, hence C is a rational curve. Replacing ϕ by the morphism
C% → X (where C % is the normalization of C), we are reduced to the case that the morphism
ϕ is birational onto its image. Then there exists a point y ∈ P1 such that the map on tan-
gent spaces Ty ϕ: Ty P1 → Tϕ(y) X is non-zero. Since Ω1X/k is free we then can find a global
1-form ω ∈ Γ(X, Ω1X/k ) such that ϕ∗ ω does not vanish at y. Since Γ(P1 , Ω1P1 /k ) = 0 this is a
contradiction. !
Before we give examples, let us introduce some notation. Consider a smooth complete
curve C over a field k. Note that by a curve we mean a variety of dimension 1; in particular,
C is assumed to be geometrically reduced and irreducible. By a (Weil) divisor on C we mean a
finite formal linear combination D = m1 P1 + · · · + mr Pr , where P1 , . . . , Pr are mutually distinct
closed points of C and where m1 , . . . , mr are integers. The degree of such a divisor is defined to
be deg(D) := m1 · [k(P1 ) : k] + · · · + mr · [k(Pr ) : k]. If f ∈ k(C)∗ is a non-zero rational function
on C, we have an associated divisor div(f ) of degree zero; such divisors are called principal.
Two divisors D1 and D2 are said to be linearly equivalent, notation D1 ∼ D2 , if they differ by
a principal divisor. The divisor class group Cl(C) is then defined to be the group of divisors
modulo linear equivalence, with group law induced by addition of divisors. Associating to a
divisor its degree gives a homomorphism deg: Cl(C) → Z. We set Cl0 (C) := Ker(deg), the class
group of degree zero divisors on C (over k).
A divisor D = m1 P1 + · · · + mr Pr is said to be effective,# notation $ D " 0, if all coefficients
mi are in Z!0 . Given a divisor D on C, write L(D) = Γ C, OC (D) for the k-vector # $space of
rational functions f on C such that div(f ) + D " 0. Also we write )(D) = dimk L(D) . Recall
that the theorem of Riemann-Roch says that

)(D) − )(K − D) = deg(D) + 1 − g ,

–8–
where K is the canonical divisor class and g is the genus of C.
With these notations, we turn to elliptic curves, the classical examples of abelian varieties,
and at the origin of the whole theory.

(1.7) Example. We define an elliptic curve to be a complete, non-singular curve of genus 1

over a field k, together with a k-rational point. Let E be such a curve, and let P ∈ E(k)
# be $the
distinguished rational point. The Riemann-Roch theorem tells us that )(nP ) := dimk L(nP ) =
n for n " 1.
We have L(P ) #= k. $Choose a basis 1, x of L(2P ) and extend it to a basis 1, x, y of
L(3P ). Since dimk L(6P ) = 6, the seven elements 1, x, y, x2 , xy, y 2 , x3 ∈ L(6P ) satisfy a
linear relation. Looking at pole orders, we see that the terms y 2 and x3 must both occur with
a non-zero coefficient, and possibly after rescaling x and y by a unit we may assume that there
is a relation of the form

y 2 + a1 xy + a3 y = x3 + a2 x2 + a4 x + a6 with ai ∈ k . (1)

The functions x and y define a rational map

# $
E ##$ P2 bya +→ 1 : x(a) : y(a) for a ,= P .

This rational map extends to an embedding of E into P2 which sends P to (0 : 1 : 0). It realizes
E as the non-singular cubic curve in P2 given by the affine equation (1), called a Weierstrass
equation for E. The non-singularity of this curve can be expressed by saying that a certain
expression in the coefficients ai , called the discriminant of the equation, is invertible. It is easily
seen from (1) that the image of P is a flex point, i.e., a point where the tangent has a threefold
intersection with the curve. (Alternatively, this is obvious from the fact that the embedding
E !→ P2 is given by the linear system |3P |.)
In order to define the structure of an abelian variety on E, let us first show that the map

α: E(k) → Cl0 (E) given by Q +→ [Q − P ]

# # #
$ If α(Q) = α(Q ) while Q ,= Q , then Q and Q are linearly equivalent and
is a bijection.
#
dimk L(Q) " 2, which contradicts # Riemann-Roch.
$ Thus α is injective. Conversely, if A is a
divisor of degree zero then dimk (L(A + P ) = 1, so there exists an effective divisor of degree 1
which is linearly equivalent to A + P . This divisor is necessarily a point, say Q, and α(Q) = [A].
This shows that α is a bijection.
We obtain a group structure on E(k) by transporting the natural group structure on Cl0 (E)
via α. Clearly, if k ⊂ K is a field extension then the group laws obtained on E(k) and EK (K) =
E(K) are compatible, in the sense that the natural inclusion E(k) ⊂ E(K) is a homomorphism.
The point P is the identity element for the group law.
The group law just defined has the following geometric interpretation. Here we shall write
A ⊕ B for the group law and .A for the inverse.

(1.8) Lemma. Let K be a field containing k. Let A, B and C be K-rational points of E. Then
A ⊕ B ⊕ C = P in the group E(K) if and only if A, B and C are the three intersection points
of EK with a line.
Proof. By construction, A ⊕ B ⊕ C = P means that A ⊕ B ⊕ C is linearly equivalent to 3P .
The lemma is therefore a reformulation of the fact that the embedding E !→ P2 is given by the
linear system |3P |. !

–9–
 The addition of K-rational points is now given as follows. To add A and B one takes the line
through A and B (by which we mean the tangent line to E at A if A = B). This line intersects E
in a third point R (possibly equal to A or B). Note that if A and B are K-rational then so is R.
Then one takes the line through R and P , which intersects E in a third point S. This is the
sum of A and B. To see this, note that by the lemma we have the relations: A ⊕ B ⊕ R = P
and R ⊕ P ⊕ S = P . Since P is the identity element we get A ⊕ B = S, as claimed. Similarly,
the inverse of an element A is the third intersection point of E with the line through A and P .

Figure 2.

We claim that the group structure on E(K) comes from the structure of a group variety
on E. In other words: we want to show that there exist morphisms m: E ×E → E and i: E → E
such that the group structure on E(K) is the one induced by m and i. To see this, let k ⊂ K
again be a field extension. If A, B ∈ E(K) then R = .(A ⊕ B) is the third intersection
point of E with the line through A and B. Direct computation shows that if we work on an
affine open subset U ⊂ P2 containing A and B then the projective coordinates of R can be
expressed as polynomials, with coefficients in k, in the coordinates of A and B. This shows that
(A, B) +→ .(A ⊕ B) is given by a morphism ϕ: E × E → E. Taking B = P we find that A +→ .A
is given by a morphism i: E → E, and composing ϕ and i we get the addition morphism m.
Explicit formulas for i and m can be found in Silverman [1], Chapter III, §2.
We conclude that the quadruple (E, m, i, P ) defines an abelian variety of dimension 1
over k. As we have seen, abelian varieties have a trivial tangent bundle. Therefore, if X is
a 1-dimensional abelian variety, it has genus 1: abelian varieties of dimension 1 are elliptic
curves.
To get a feeling for the complexity of elliptic curves we take E to be the elliptic curve over Q
given by the Weierstrass equation y 2 + y = x3 − x, with origin P∞ = (0 : 1 : 0). Let Q be the
rational point (−1, −1). If for n = 1, . . . , 20 we plot the coordinates of n · Q = Q ⊕ · · · ⊕ Q as
rational numbers, or even if we just plot the absolute value of the numerator of the x-coordinate
we find a parabola shape which indicates that the “arithmetic complexity” of the point n · Q

– 10 –
 1
6
20
1357
8385
12551561
1849037896
4881674119706
2786836257692691
79799551268268089761
280251129922563291422645
54202648602164057575419038802
3239336802390544740129153150480400
1425604881483182848970780090473397497201
596929565407758846078157850477988229836340351
1356533706384096591887827693333962338847777347485221
2389750519110914018630990937660635435269956452770356625916
47551938020942325784141569050513811957803129798534598981096547726
43276783438948886312588030404441444313405755534366254416432880924019065
66655479518893093532610447590226207125008330695731551720689810858664307580428417

Figure 3.

grows quadratically in n; see Figure 3.

(1.9) Example. Now we try to generalize the above example, taking a curve of genus 2. So,
let C be a smooth projective curve of genus g = 2 over a field k. Then C is a hyperelliptic
curve and can be described as a double covering π: C → P1k of the projective line. Let i be the
hyperelliptic involution of C. Consider the surface C × C, on which we have an involution ι
given by (a, b) +→ (b, a). The
# quotient
$ S 2 C = (C × C)/ι is a non-singular surface. The image of
the anti-diagonal ∆− = { a, i(a) : a ∈ C} on S 2 C is a curve D which is isomorphic to C/i = P1
and has self-intersection number 21 (∆− )2 = (2 − 2g)/2 = −1, i.e., D is an exceptional curve. By
elementary theory of algebraic surfaces we can blow D down, obtaining a non-singular projective
surface S.
We claim that there is a bijection
∼
α: S(k) −→ Cl0 (C) .

To define α we look at the map S 2 C(k) → Cl0 (C) given by (a, b) +→ [a + b] − [K], where
[K] is the canonical divisor class. Since [a + i(a)] = [K] for every a ∈ C, this map factors
through the contraction of D and we get a map α: S(k) → Cl0 (C). If a + b ∼ c + d with
a ,= c ,= b then )(a + b) " 2, and it follows by Riemann-Roch that [a + b] = [K]. (Indeed,
we have )(a + b) − )(K − a − b) = 1, so )(a + b) " 2 implies that the degree zero divisor
K − a − b is effective, hence equivalent to zero.) This shows that α is injective. It is surjective
by Riemann-Roch.
Transporting the natural group structure on Cl0 (C) via α, we obtain a group structure on
S(k). The formation of this group structure is compatible with field extensions k ⊂ K. The
identity element of S(k) is the point obtained by contracting D.
We claim that the addition and inverse #on S(k) are $ given by morphisms. For the inverse
this is easy: the inverse of (a, b) is given by i(a), i(b) , again using that x + i(x) ∼ K for all
points x. To see that addition is given by a morphism, consider the projection π: C 5 → C 4 onto
the first four factors. This map has four natural sections (p1 , p2 , p3 , p4 ) +→ (p1 , . . . , p4 , pi ), and
this defines a relative effective divisor D of degree 4 on C 5 over C 4 . Let K be a fixed canonical
divisor on the last factor C. By the Riemann-Roch theorem for the curve C over the function
field k(C 4 ) the divisor D − K is linearly equivalent to an effective divisor of degreee 2 on C
over k(C 4 ). It follows that D is linearly equivalent to a divisor of the form E + π ∗ (G), with E

– 11 –
a relative effective divisor of degree 2 and G a divisor on C 4 . The restriction of E to a fibre
{P = (p1 , p2 , p3 , p4 )} × C is a of degree 2, hence determines a point ψ(P ) of S 2 C. This gives a
map ψ: C 4 → S 2 C which is clearly a morphism. If β: S 2 C → S is the blowing-down of ∆− then
the composition C 4 → S 2 C → S factors through β × β. The resulting morphism S × S → S is
precisely the addition on S.

The preceding two examples suggest that, given a smooth projective curve C over a field
k, there should exist an abelian variety whose points parametrize the degree zero divisor classes
on C. If C has a k-rational point then such an abelian variety indeed exists (as we shall see
later), though the construction will not be as explicit and direct as in the above two examples.
The resulting abelian variety is called the jacobian of the curve.

(1.10) Example. In this example we work over the field k = C. Consider a complex vector
space V of finite dimension n. For an additive subgroup L ⊂ V the following conditions are
equivalent: (i) L ⊂ V is discrete and co-compact, i.e., the euclidean topology on V induces
the discrete topology on L and the quotient X := V /L is compact for the quotient topology;
(ii) the natural map L ⊗Z R → V is bijective; (iii) there is an R-basis e1 , . . . , e2n of V such that
L = Ze1 + · · · + Ze2n . A subgroup satisfying these conditions is called a lattice in V .
Given a lattice L ⊂ V , the quotient X naturally inherits the structure of a compact (complex
analytic) Lie group. Lie groups of this form are called complex tori. (This usage of the word
torus is not to be confused with its meaning in the theory of linear algebraic groups.)
Let us first consider the case n = 1. By a well-known theorem of Riemann, every compact
Riemann surface is algebraic. Since X has genus 1, it can be embedded as a non-singular cubic
curve in P2C , see (1.7). If ϕ: X !→ P2C is such an embedding, write E = ϕ(X) and P = ϕ(0).
We see that (E, P ) is an elliptic curve (taking P to be the identity element). The structure of
a group variety on E as defined in (1.7) is the same as the group structure on X, in the sense
∼
that ϕ: X −→ E an is an isomorphism of Lie groups.
For n " 2 it is not true that any n-dimensional complex torus X = V /L is algebraic;
in fact, “most” of them are not. What is true, however, is that every abelian variety over C
can analytically be described as a complex torus. In this way, complex tori provide “explicit”
examples of abelian varieties. We will return to this theme in (?.?).

The group structure of an abelian variety imposes strong conditions on the geometry of the
underlying variety. The following lemma is important in making this explicit.

(1.11) Rigidity Lemma. Let X, Y and Z be algebraic varieties over a field k. Suppose that X
is complete. If f : X × Y → Z is a morphism with the property that, for some y ∈ Y (k), the fibre
X × {y} is mapped to a point z ∈ Z(k) then f factors through the projection prY : X × Y → Y .

Proof. We may assume that k = k. Choose a point x0 ∈ X(k), and define a morphism g: Y → Z
by g(y) = f (x0 , y). Our goal is to show that f = g ◦ prY . As X × Y is reduced it suffices to
prove this on k-rational points.
Let U ⊂ Z be an affine open neighbourhood# of z. Since $ X is complete, the projection
prY : X ×Y → Y is a closed map, so that V := prY f −1 (Z −U ) is closed in Y . By construction,
if P ∈/ V then f (X × {P }) ⊂ U . Since X is complete and U is affine, this is possible only if f
is constant on X × {P }. This shows that f = g ◦ prY on the non-empty open set X × (Y − V ).
Because X × Y is irreducible, it follows that f = g ◦ prY everywhere. !

– 12 –
(1.12) Definition. Let (X, mX , iX , eX ) and (Y, mY , iY , eY ) be group varieties. A morphism
f : X → Y is called a homomorphism if

f ◦ mX = mY ◦ (f × f ) .

If this holds then also f (eX ) = eY and f ◦ iX = iY ◦ f .

The rigidity of abelian varieties is illustrated by the fact that up to a translation every
morphism is a homomorphism:

(1.13) Proposition. Let X and Y be abelian varieties and let f : X → Y be a morphism.

Then f is the composition f = tf (eX ) ◦ h of a homomorphism h: X → Y and a translation tf (eX )
over f (eX ) on Y .
# $
Proof. Set y := iY f (eX ) , and define h := ty ◦ f . By construction we have h(eX ) = eY . Consider
the composite morphism
# $
(h ◦ mX )× iY ◦ mY ◦ (h×h)
Y m
g := (X × X −−−−−−−−−−−−−−−−−−→ Y × Y −−−
→ Y ).

(To understand what this morphism does: if we use the additive notation for the group structures
on X and Y then g is given on points by g(x, x# ) = h(x + x# ) − h(x# ) − h(x).) We have

g({eX } × X) = g(X × {eX }) = {eY } .

By the Rigidity Lemma this implies that g factors both through the first and through the
second projection X × X → X, hence g equals the constant map with value eY . This means
that h ◦ mX = mY ◦ (h × h), i.e., h is a homomorphism. !

(1.14) Corollary. (i) If X is a variety over a field k and e ∈ X(k) then there is at most one
structure of an abelian variety on X for which e is the identity element.
(ii) If (X, m, i, e) is an abelian variety then the group structure on X is commutative, i.e.,
m ◦ s = m: X × X → X, where s: X × X → X × X is the morphism switching the two factors.
In particular, for every k-scheme T the group X(T ) is abelian.
Proof. (i) If (X, m, i, e) and (X, n, j, e) are abelian varieties then m and n are equal when
restricted to X × {e} and {e} × X. Applying (1.11) to m ◦ (m, i ◦ n): X × X → X, which is
constant when restricted to X × {e} and {e} × X, we get m = n. This readily implies that i = j
too.
(ii) By the previous proposition, the map i: X → X is a homomorphism. This implies that
the group structure is abelian. !

(1.15) Remark. It is worthwile to note that in deriving the commutativity of the group the
completeness of the variety is essential. Examples of non-commutative group varieties are linear
algebraic groups (i.e., matrix groups) like GLn for n > 1, the orthogonal groups On for n > 1
and symplectic groups Sp2n .

(1.16) Notation. From now on we shall mostly use the additive notation for abelian varieties,
writing x + y for m(x, y), writing −x for i(x), and 0 for e. Since abelian varieties are abelian
as group varieties, we no longer have to distinguish between left and right translations. Also

– 13 –
we can add homomorphisms: given two homomorphisms of abelian varieties f , g: X → Y , we
define f + g to be the composition

f + g := mY ◦ (f, g): X −→ Y × Y −→ Y ,

and we set −f := f ◦ iX = iY ◦ f . This makes the set Hom (X, Y ) of homomorphisms of X to

Y into an abelian group.
As we have seen, also Hom /k (X, Y ) = Y (X)—the set of X-valued points of Y —has
a natural structure of an abelian group. By Prop. (1.13), Hom (X, Y ) is just the sub-
group of Hom /k (X, Y ) consisting of those morphisms f : X → Y such that f (0X ) = 0Y ,
and Hom /k (X, Y ) = Hom (X, Y ) × Y (k) as groups. We shall adopt the convention that
Hom(X, Y ) stands for Hom (X, Y ). If there is a risk of confusion we shall indicate what we
mean by a subscript “AV” or “Sch/k”.
We close this chapter with another result that can be thought of as a rigidity property of
abelian varieties.

(1.17) Theorem. Let X be an abelian variety over a field k. If V is a smooth k-variety then
any rational map f : V ##$ X extends to a morphism V → X.
Proof. We may assume that k = k, for if a morphism Vk → Xk is defined over k on some dense
open subset of Vk , then it is defined over k. Let U ⊆ V be the maximal open subset on which
f is defined. Our goal is to show that U = V .
If P ∈ |V | is a point of codimension 1 then the local ring OV,P# is a $dvr, because V is
regular. By the valuatative criterion for properness the map f : Spec k(V ) → X extends to
a morphism Spec(OV,P ) → X. Because X is locally of finite type over k, this last morphism
extends to a morphism Y → X for some open Y ⊂ V containing P . (Argue on rings.) Hence
codimX (X \ U ) " 2.
Consider the rational map F : V × V ##$ X given on points by (v, w) +→ f (v) − f (w). Let
W ⊂ V × V be the domain of definition of F . We claim that f is defined at a point v ∈ V (k)
if and only if F is defined at (v, v). In the “only if” direction this# is immediate,
$ as clearly
U × U ⊆ W . For the converse, suppose F is defined at (v, v). Then V × {v} ∩ W is an open
subset#of V ∼= V $×{v} containing v. Hence we can choose a point u ∈ U (k) such that (u, v) ∈ W .
Then {u} × V ∩ W is an open subset of V ∼ = {u} × V containing v, on which f is defined
because we have the relation f (w) = f (u) − F (u, w).
Our job is now to show that F extends over the diagonal ∆ ⊂ V × V . Consider the
homomorphism on function fields F $ : k(X) → k(V × V ). Note that F maps ∆ ∩ W to 0 ∈ X.
It follows that F is regular at a point (v, v) ∈ ∆(k) if and only if F $ maps OX,0 ⊂ k(X) into
OV ×V,(v,v) . Suppose that f is not regular at some point v ∈ V (k), and choose an element
ϕ ∈ OX,0 with F $ (ϕ) ∈/ OV ×V,(v,v) . Let D be the polar divisor of F $ (ϕ), i.e.,
& # $
D= ordP F $ (ϕ) · [P ]
# $
where the sum runs over all codimension 1 points P ∈ |V × V | with ordP F $ (ϕ) < 0. If (w, w)
is a k-valued point in ∆ ∩ |D| then F $ (ϕ) is not in OV ×V,(w,w) , hence F is not regular at (w, w).
But V × V is a regular scheme, so D ⊂ V × V is locally a principal divisor. Then also ∆ ∩ |D|
is locally defined, inside ∆, by a single equation, and it follows that ∆ ∩ |D| has codimension
% 1 in ∆. Hence f is not regular on a subset of V of codimension % 1, contradicting our earlier
conclusion that codimX (X\U ) " 2. !

– 14 –
Exercises.
(1.1) Let X1 and X2 be varieties over a field k.
(i) If X1 and X2 are given the structure of a group variety, show that their product X1 × X2
naturally inherits the structure of a group variety.
(ii) Suppose Y := X1 × X2 carries the structure of an abelian variety. Show that X1 and X2
each have a unique structure of an abelian variety such that Y = X1 × X2 as abelian
varieties.
(1.2) Check that an equivalent definition of an abelian variety is the following. An abelian
variety is a 4-tuple (X, m, i, e) consisting of a complete k-variety X together with morphisms

m: X × X → X , i: X → X and e: Spec(k) → X

such for every k-scheme T the set of T -valued points X(T ) together with the maps m, i and e
forms a group, and such that the map T +→ X(T ) defines a contravariant functor from the
category of k-schemes to the category of groups.
(1.3) Let X be a variety over
# a$ field k. Write k[ε] for the ring of dual numbers over k (i.e.,
ε = 0), and let S := Spec k[ε] . Write Aut(1) (XS /S) for the group of automorphisms of XS
2

over S which reduce to the identity on the special fibre X !→ XS .

(i) Let x be a k-valued point of X (thought of either as a morphism of k-schemes x: Spec(k) →
X or as a point x ∈ |X| with k(x) = k). Show that the tangent space TX,x := (mx /m2x )∗ is
in natural bijection with the space of k[ε]-valued points of X which reduce to x modulo ε.
(Cf. HAG, Chap. II, Exercise 2.8.)
(ii) Suppose X = Spec(A) is affine. It is immediate from the definitions that

H 0 (X, TX/k ) ∼
= Homk (Ω1A/k , A) ∼
= Derk (A, A) .

Use this to show that H 0 (X, TX/k ) is a naturally isomorphic with Aut(1) (XS /S).
(iii) Show, by taking an affine covering and using (ii), that for arbitrary variety X we have a
natural isomorphism
∼
h: H 0 (X, TX/k ) −→ Aut(1) (XS /S) .
(iv) Suppose X is a group variety over k. If x ∈ X(k) and τ : S → X is a tangent vector at x,
check that the associated global vector field ξ := h−1 (tτ ) is right-invariant, meaning that
t∗y ξ = ξ for all y ∈ X.
(1.4) A ring variety over a field k is a commutative group variety (X, +, 0) over k, together
with a ring multiplication morphism X ×k X → X written as (x, y) +→ x · y, and a k-rational
point 1 ∈ X(k), such that the ring multiplication is associative: x · (y · z) = (x · y) · z, distributive:
x · (y + z) = (x · y) + (x · z), and 1 is a 2-sided identity element: 1 · x = x = x · 1. Show that the
only complete ring variety is a point. (In fact, you do not need the identity element for this.)
(1.5) Let E be the elliptic curve over Q given by the Weierstrass equation y 2 + y = x3 − x,
with origin P∞ = (0 : 1 : 0). Let Q be the rational point (−1, −1). Compute m · Q for m = 2, 4
and 10.
(1.6) Let X1 , X2 , Y1 and Y2 be abelian varieties over a field k. Show that

Hom (X1 × X2 , Y1 × Y2 )
∼
= Hom (X1 , Y1 ) × Hom (X1 , Y2 ) × Hom (X2 , Y1 ) × Hom (X2 , Y2 ) .

– 15 –
Does a similar statement hold if we everywhere replace “Hom ” by “Hom ”?

Notes. If one wishes to go back to classical antiquity one may put the origin of the theory of abelian varieties with
Diophantos (± 200 – ± 284) who showed how to construct a third rational solution of certain cubic equations in
two unknowns from two given ones. The roots in a not so distant past may be layed with Giulio Carlo Fagnano
(1682–1766) and others who considered addition laws for elliptic integrals. From this the theory of elliptic
functions was developed. The theory of elliptic functions played a major role in 19th century mathematics. Niels
Henrik Abel (1802–1829), after which our subject is named, had a decisive influence on its development. Other
names that deserve to be mentioned are Adrien-Marie Legendre (1752–1833), Carl-Friedrich Gauss (1777–1855)
and Carl Gustav Jacobi (1804–1851).
Bernhard Riemann (1826–1866) designed a completely new theory of abelian functions in which the algebraic
curve was no longer the central character, but abelian integrals and their periods and the associated complex
torus. The theory of abelian functions was further developed by Leopold Kronecker (1823–1891), Karl Weierstrass
(1815–1897) and Henri Poincaré (1854–1912). After Emile Picard (1856– 1941) abelian functions were viewed as
the meromorphic functions on a complex abelian variety.
It was André Weil (1906–1998) who made the variety the central character of the subject when he developed
a theory of abelian varieties over arbitrary fields; he was motivated by the analogue of Emil Artin (1898–1962)
of the Riemann hypothesis for curves over finite fields and the proof by Helmut Hasse (1898–1979) for genus 1.
See Weil [2]. David Mumford (1937) recasted the theory of Weil in terms of Grothendieck’s theory of schemes.
His book MAV is a classic. We refer to Klein [1] and Dieudonné [2] for more on the history of our subject. The
Rigidity Lemma is due to Mumford.

– 16 –
Chapter II. Line bundles and divisors on abelian varieties.

In this chapter we study divisors on abelian varieties. One of the main goals is to prove that
abelian varieties are projective. The Theorem of the Square (2.9) plays a key role. Since abelian
varieties are nonsingular, a Weil divisor defines a Cartier divisor and a line bundle, and we
∼
have a natural isomorphism Cl(X) −→ Pic(X). We shall mainly work with line bundles, but
sometimes (Weil) divisors are more convenient.
The following abuse of notation will prove handy: if L is a line bundle on a product variety
X × Y then we shall write Lx for the restriction of L to {x} × Y and, similarly, Ly denotes the
restriction L|X×{y} . Here of course x shall always be a point of X and y a point of Y .
In this chapter, varieties shall always be varieties over some ground field k, which in most
cases shall not be mentioned.

§ 1. The theorem of the square.

(2.1) Theorem. Let X and Y be varieties. Suppose X is complete. Let L and M be two line
bundles on X × Y . If for all closed points y ∈ Y we have Ly ∼
= My there exists a line bundle N
∼ ∗
on Y such that L = M ⊗ p N , where p = prY : X × Y → Y is the projection onto Y .
Proof. This is a standard fact of algebraic geometry. A proof using cohomology runs as follows.
Since Ly ⊗My−1 is the trivial bundle and Xy is complete, the space of sections H 0 (Xy , Ly ⊗My−1 )
is isomorphic to k(y), the residue field of y. This implies that p∗ (L ⊗ M −1 ) is locally free of
rank one, hence a line bundle (see MAV, §5 or HAG, Chap. III, § 12). We shall prove that the
natural map
α: p∗ p∗ (L ⊗ M −1 ) → L ⊗ M −1
is an isomorphism. If we restict to a fibre we find the map
OXy ⊗ Γ(Xy , OXy ) → OXy
which is an isomorphism. By Nakayama’s Lemma, this implies that α is surjective and by
comparing ranks we conclude that it is an isomorphism. !
As an easy consequence we find a useful prinicple.

(2.2) See-saw Principle. If, in addition to the assumptions of (2.1), we have Lx = Mx for
some point x ∈ X then L ∼
= M.
Proof. We have L ∼ = M ⊗ pr∗Y N . Over {x} × Y this gives Lx ∼
= Mx ⊗ (pr∗Y N )x . Therefore,
(pr∗Y N )x is trivial, and this implies that N is trivial. !

(2.3) Lemma. Let X and Y be varieties, with X complete. For a line bundle L on X × Y , the
set {y ∈ Y | Ly is trivial} is closed in Y .
Proof. If M is a line bundle on a complete variety then M is trivial if and only if both H 0 (M )
and H 0 (M −1 ) are non-zero. Hence
{y ∈ Y | Ly is trivial} = {y ∈ Y | h0 (Ly ) > 0} ∩ {y ∈ Y | h0 (L−1
y ) > 0} . (1)

LineBund, 15 september, 2011 (812)

– 17 –
But the functions y (→ h0 (Ly ) and y (→ h0 (L−1 ) are upper semi-continuous on Y ; see MAV, § 5
or HAG, Chap. III, Thm. 12.8. So the two sets in the right hand side of (1) are closed in Y . !
Actually, there is a refinement of this which says the following.

(2.4) Lemma. Let X be a complete variety over a field k, let Y be a k-scheme, and let L be
a line bundle on X × Y . Then there exists a closed subscheme Y0 "→ Y which is the maximal
subscheme of Y over which L is trivial; i.e., (i) the restriction of L to X × Y0 is the pull back
(under prY0 ) of a line bundle on Y0 , and (ii) if ϕ: Z → Y is a morphism such that (idX × ϕ)∗ (L)
is the pullback of a line bundle on Z under p∗Z then ϕ factors through Y0 .
For the proof we refer to MAV, §10. In Chapter 6 we shall discuss Picard schemes; once
we know the existence and some properties of PicX/k the assertion of the lemma is a formal
consequence. (See (6.4).)
The following theorem is again a general fact from algebraic geometry and could be accepted
as a black box. As it turns out, it is of crucial importance for the theory of abelian varieties. In
view of its importance we give a proof.

(2.5) Theorem. Let X and Y be complete varieties and let Z be a connected, locally noetherian
scheme. Let x ∈ X(k), y ∈ Y (k), and let z be a point of Z. If L is a line bundle on X × Y × Z
whose restriction to {x} × Y × Z, to X × {y} × Z and to X × Y × {z} is trivial then L is trivial.
Proof. We follow the proof given by Mumford in MAV. We view L as a family of line bundles
on X × Y parametrized by Z. Let Z % be the maximal closed subscheme of Z over which L is
trivial, as discussed above. We have z ∈ Z % . We shall show that Z % = Z by showing that Z % is
an open subscheme and using the connectedness of Z.
Let ζ be a point of Z % . Write m for the maximal ideal of the local ring OZ,ζ and I ⊂ OZ,ζ
for the ideal defining (the germ of) Z % . We have to show that I = (0). Suppose not. By Krull’s
Theorem (here we use that Z is locally noetherian) we have ∩n mn = (0), hence there exists a
positive integer n such that I ⊂ mn , I *⊂ mn+1 . Put a1 = (I, mn+1 ), and choose an ideal a2 with

mn+1 ⊂ a2 ⊂ (I, mn+1 ) = a1 and dimk(ζ) (a1 /a2 ) = 1 .

(Note that such ideals exist.) Let Zi ⊂ Spec(OZ,ζ ) be the closed subscheme defined by the ideal
ai (i = 1, 2). We will show that the restriction of L to X × Y × Z2 is trivial. This implies that
Z2 is contained in Z % , which is a contradiction, since I *⊂ a2 .
Write Li for the restriction of L to X × Y × Zi . By construction, L1 is trivial; choose a
trivializing global section s. The inclusion Z1 "→ Z2 induces a restriction map Γ(L2 ) → Γ(L1 ).
We claim: L2 is trivial if and only if s can be lifted to a global section of L2 . To see this,
suppose first that we have a lift s% . The schemes X × Y × Z1 and X × Y × Z2 have the same
underlying point sets. If s% (P ) = 0 for some point P then also s(P ) = 0, but this contradicts
the assumption that s is a trivialization of L1 . Hence s% is nowhere zero, and since L2 is locally
free of rank 1 this implies that s% trivializes L2 . Conversely, if L2 is trivial then the restriction
map Γ(L2 ) → Γ(L1 ) is just Γ(OZ2 ) → Γ(OZ1 ) and this is surjective.
The obstruction for lifting s to a global section of L2 is an element ξ ∈ H 1 (X × Y, OX×Y ).
We know that the restrictions of L2 to {x} × Y × Z2 and to X × {y} × Z2 are trivial. Writing
i1 = (idX , y): X "→ X × Y and i2 = (x, idY ): Y "→ X × Y , this means that ξ has trivial image
under i∗1 : H 1 (X × Y, OX×Y ) → H 1 (X, OX ) and under i∗2 : H 1 (X × Y, OX×Y ) → H 1 (Y, OY ). But

– 18 –
the map (i∗1 , i∗2 ) gives a (Künneth) isomorphism
∼
H 1 (X × Y, OX×Y ) −→ H 1 (X, OX ) ⊕ H 1 (Y, OY ) ,

hence ξ = 0 and s can be lifted. !

(2.6) Remarks. (i) In the theorem as stated we require x and y to be k-rational points of X,
resp. Y . We use this when we define the maps i1 and i2 . But in fact the theorem holds without
these assumptions. The point is that if k ⊂ K is a field extension then a line bundle M on a
k-variety V is trivial if and only if the line bundle MK on VK is trivial. (See Exercise (2.1).)
Hence we may first pass to a bigger field K over which the points x and y (and even the point
z ∈ Z) become rational.
(ii) The previous theorem gives a strong general result about line bundles on a product of
three complete varieties. Note that the analogous statement for line bundles on a product of two
complete varieties is false in general. More precisely, suppose X and Y are complete k-varieties
and L is a line bundle on X × Y . If there exist points x ∈ X and y ∈ Y such that Lx ∼ = OY
∼ ∼
and Ly = OX then it is not true in general that L = OX×Y . For instance, take X = Y to be an
elliptic curve, and consider the divisor
! " ! "
D = ∆X − {0} × X − X × {0}

where ∆X ⊂ X × X is the diagonal. Note that L = OX×X (D) restricts to the trivial bundle on
{0} × X and on X × {0}. (Use that the divisor 1 · 0 (= 1 · eX ) on X is linearly equivalent to a
divisor whose support does not contain 0.) But L is certainly not the trivial bundle: if it were,
L|{P }×X = OX (P − eX ) ∼
= OX for all points P ∈ X. But then there is a function f on X with
one zero and one pole and X would have to be a rational curve, which we know it is not.
Theorem (2.5), together with the previous remark, is a reflection of the quadratic character
of line bundles which comes out clearer as follows. If f (x) = ax2 + bx + c is a quadratic function
on the real line then

f (x + y + z) − f (x + y) − f (x + z) − f (y + z) + f (x) + f (y) + f (z)

is constant. The analogue of this for line bundles on abelian varieties is the celebrated Theorem
of the Cube. First a notational convention. If X is an abelian variety and I = {i1 , . . . , ir } ⊂
{1, 2, . . . , n} then we write

pI : X n → X , or pi1 ···ir : X n → X ,

for the morphism sending (x1 , x2 , . . . , xn ) to xi1 +· · ·+xir . Thus, for example, pi is the projection
onto the ith factor, p12 = p1 + p2 , etc. With these notations we have the following important
corollary to the theorem.

(2.7) Theorem of the Cube. Let L be a line bundle on X. Then the line bundle
# 1+#I
Θ(L) := p∗I L⊗(−1)
I⊂{1,2,3}

= p∗123 L ⊗ p∗12 L−1 ⊗ p∗13 L−1 ⊗ p∗23 L−1 ⊗ p∗1 L ⊗ p∗2 L ⊗ p∗3 L

– 19 –
on X × X × X is trivial.
Proof. Restriction of Θ(L) to {0} × X × X gives the bundle

m∗ L ⊗ p∗2 L−1 ⊗ p∗3 L−1 ⊗ m∗ L−1 ⊗ OX×X ⊗ p∗2 L ⊗ p∗3 L

which is obviously trivial. Similarly for X × {0} × X and X × X × {0}. By (2.5) the result
follows. !
We could sharpen the corollary by saying that Θ(L) is canonically trivial, see Exercise (2.2).

(2.8) Corollary. Let Y be a scheme and let X be an abelian variety. For every triple f , g, h
of morphisms Y → X and for every line bundle L on X, the bundle

(f + g + h)∗ L ⊗ (f + g)∗ L−1 ⊗ (f + h)∗ L−1 ⊗ (g + h)∗ L−1 ⊗ f ∗ L ⊗ g ∗ L ⊗ h∗ L

on Y is trivial.
Proof. Consider (f, g, h): Y → X × X × X and use (2.7). !
Another important corollary is the following.

(2.9) Theorem of the Square. Let X be an abelian variety and let L be a line bundle on X.
Then for all x, y ∈ X(k),
t∗x+y L ⊗ L ∼
= t∗x L ⊗ t∗y L .
More generally, let T be a k-scheme and write LT for the pull-back of L to XT . Then
! "
t∗x+y LT ⊗ LT ∼
= t∗x LT ⊗ t∗y LT ⊗ pr∗T (x + y)∗ L ⊗ x∗ L−1 ⊗ y ∗ L−1

for all x, y ∈ X(T ).

Proof. In the first formulation, this is immediate from (2.8) by taking for f the identity on X and
for g and h the constant maps with images x and y. For the general form, take f = prX : XT =
X ×k T → X, take g = x ◦ prT and h = y ◦ prT . Then

f + g = prX ◦ tx , f + h = prX ◦ ty , g + h = (x + y) ◦ prT

and
f + g + h = prX ◦ tx+y
Now again apply (2.8). !
The theorem allows the following interpretation. (Compare this with what we have seen in
Examples (1.7) and (1.9).)

(2.10) Corollary. Let L be a line bundle on an abelian variety X. Let Pic(X) be the group
of isomorphism classes of line bundles on X. Then the map ϕL : X(k) → Pic(X) given by
x (→ [t∗x L ⊗ L−1 ] is a homomorphism.
Proof. Immediate from (2.9). !

(2.11) Remark. The homomorphisms ϕL will play a very important role in the theory. In
later chapters (see in particular Chapters 6 and 7) we shall introduce the dual X t of an abelian

– 20 –
variety X, and we shall interprete ϕL as a homomorphism X → X t . The homomorphisms
λ: X → X t that are (geometrically) of the form ϕL for an ample line bundle L are called
polarizations; see Chapter 11.
At this point, let us already caution the reader that there is a sign convention in the theory
that can easily lead to misunderstanding. In the theory of elliptic curves one usually describes
line bundles of degree 0 (which is what the dual elliptic curve is about!) in the form OE (P − O).
More precisely: if E is an elliptic curve with origin O then the map P (→ OE (P − O) gives an
∼
isomorphism E −→ E t = Pic0E/k . This map is not the polarization associated to the ample line
bundle L = OE (O); rather it is! minus that" map. In general, if D is a divisor on an abelian
variety X then t∗x OX (D) is OX (t−x (D) = OX (D − x), not OX (D + x). So if L = OE (O) on
an elliptic curve E, the map ϕL is given on points by P (→ OE (O − P ).
The same remark applies to the theory of Jacobians (see in particular Chapter 14). If C is
a smooth projective curve over a field k, and if P0 ∈ C(k) is a k-rational point then we have a
natural morphism ϕ from C to its Jacobian variety J = Jac(C) := Pic0C/k . In most literature one
considers the map C → J given on points by P (→ OC (P − P0 ). However, we have a canonical
principal polarization on J (see again Chapter 14 for further details), and in connection with
this it is more natural to consider the morphism ϕ: C → J given by P (→ OC (P0 − P ).
Let X be an abelian variety. For every n ∈ Z we have a homomorphism [n] = [n]X : X → X
called “multiplication by n”. For n " 1, it sends x ∈ X(k) to x + · · · + x (n terms); for
n = −m # −1 we have [n]X = iX ◦ [m]X . If there is no risk of confusion, we shall often simply
write n for [n]; in particular this includes the abbreviations 1 for [1] = idX , 0 for [0] (the constant
map with value 0), and −1 for [−1] = −idX . The effect of n on line bundles is described by the
following result.

(2.12) Corollary. For every line bundle L on an abelian variety X we have

n∗ L ∼
= Ln(n+1)/2 ⊗ (−1)∗ Ln(n−1)/2 .

Proof. Set f = n, g = 1, and h = −1. Applying (2.8), one finds that

n∗ L ⊗ (n + 1)∗ L−1 ⊗ (n − 1)∗ L−1 ⊗ n∗ L ⊗ L ⊗ (−1)∗ L

is trivial, i.e.,
n∗ L2 ⊗ (n + 1)∗ L−1 ⊗ (n − 1)∗ L−1 ∼
= (L ⊗ (−1)∗ L)−1 .

The assertion now follows by induction, starting from the cases n = −1, 0, 1. !
In particular, if the line bundle L is symmetric, by which we mean that (−1)∗ L ∼ = L,
then we find that n∗ L ∼
2
= Ln for all n. For instance, if M is an arbitrary line bundle then
L+ := M ⊗ (−1)∗ M is symmetric. Similarly, L− := M ⊗ (−1)∗ M −1 is an example of an anti-
symmetric line bundle, i.e., a line bundle L for which (−1)∗ L ∼
= L−1 ; for such line bundles we
have n∗ L ∼
= Ln for all n. Note the contrast between the quadratic effect of n∗ in the symmetric
case and the linear effect in the anti-symmetric case. Further note that with the notation just
introduced we have M 2 ∼ = L+ ⊗ L− ; so we find that the square of a line bundle can be written
as the product of a symmetric and an anti-symmetric part. This is a theme we shall explore in
much greater detail in later chapters.

– 21 –
§ 2. Projectivity of abelian varieties.

We now turn to the question whether abelian varieties are projective. As it turns out the answer
is “yes”. We give two proofs of this. A fairly short proof is given in (2.26); the Theorem of the
Square plays a key role in this argument. The other proof we give is longer—it takes up most of
this section—but along the way we shall obtain a number of results that are interesting in their
own right. We think that Proposition (2.20) is particularly remarkable.
We shall need a couple of facts about group schemes. Since these form the main objects of
study of the next two chapters, we shall simply use what we need, and refer forward to the next
chapter for a precise explanation. Most of what is needed in this chapter can be summarized as
follows.

(2.13) Fact. Let X be an abelian variety over a field k. Suppose Y "→ X is a closed subgroup
scheme. If Y 0 is the connected component of Y containing the origin then Y 0 is an open and
closed subgroup scheme of Y and Y 0 is geometrically irreducible. If furthermore k is perfect
0
then the reduced underlying scheme Yred "→ X is an abelian subvariety of X.
For the proof of this statement, see Prop. (3.17) and Exercise (3.2).

(2.14) Remark. The fact just stated is weaker than what is actually true. Namely, the con-
0
clusion that Yred "→ X is an abelian subvariety of X holds true without the assumption that
the base field k is perfect. We shall see this in Prop. (5.31), once we have more theory at our
disposal. If we already knew the stronger version of the above fact at this stage, it would sim-
plify some of the arguments that we shall give. For instance, in the rest of this chapter we shall
sometimes work over k and then later draw conclusions that are valid over an arbitrary field.
The reason for this detour is that, at this stage, we can apply (2.13) only over a perfect field.
Suppose X = A × B is an abelian variety which is a product of positive dimensional
abelian varieties A and B, and suppose M is a line bundle on A. If prA : A × B → A is the
projection onto A then the bundle L := pr∗A M is invariant under translation over the points of
{0A } × B ⊂ X. Obviously, L is not ample. This suggests that if L is a line bundle on X which
is invariant under many translations, then L might not be ample.

(2.15) Definition. Let L be a line bundle on an abelian variety X. On X × X we define the

Mumford line bundle Λ(L) by

Λ(L) := m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1 .

As we shall see, Λ(L) is a very useful bundle. The restriction of Λ(L) to a vertical fibre
{x} × X and to a horizontal fibre X × {x} is t∗x L ⊗ L−1 . In particular, Λ(L) is trivial on {0} × X
and on X × {0}.

(2.16) Definition. With the above notations, we define K(L) ⊆ X as the maximal closed
subscheme (in the sense of (2.4)) such that Λ(L)|X×K(L) is trivial over K(L), i.e., such that
Λ(L)|X×K(L) ∼
= pr∗2 M for some line bundle M on K(L).
It follows from the universal property in (2.4) that the formation of K(L) is compatible
with base-change. In particular, if k ⊂ k % is a field extension, writing L% for the pull-back of L
to X ×k k % , we have K(L% ) = K(L) ×k k % .

– 22 –
 Roughly speaking, a point belongs to K(L) if L is invariant under translation by this point.
A more precise statement is given by the following lemma.

(2.17) Lemma. Let T be a k-scheme and x: T → X a T -valued point of X.

(i) The morphism x factors through K(L) if and only if t∗x LT ⊗ L−1
T is the pull-back of a
line bundle on T .
(ii) If t∗x LT ⊗ L−1 ∼ ∗ ∼ ∗
T = prT M then M = x L.
(iii) We have Λ(L)|X×K(L) ∼ = OX×K(L) .
In (iii), note that a priori we only knew that Λ(L)|X×K(L) is the pull-back of a line bundle
on K(L).
Proof. As usual, LT denotes the pull-back of L via the projection prX : XT → X. Since
prX ◦ tx : XT → XT → X is equal to the composition m ◦ (idX ×x): XT = X ×k T → X ×k X → X,
we find
t∗x LT ∼
= (idX × x)∗ m∗ L .
Note that we can write LT as LT = (idX × x)∗ p∗1 L. This gives

t∗x LT ⊗ L−1 ∼ ∗ ∗ ∗
T = (idX × x) Λ(L) ⊗ (idX × x) p2 L
= (idX × x)∗ Λ(L) ⊗ (pr∗T x∗ L) .

Using the defining properties of K(L) as given in Lemma (2.4), the assertion of (i) readily follows
from this formula.
For (ii) note that t∗x LT ⊗ L−1 ∗
T restricts to x L on {0} × T .
For (iii), take T = K(L), and let x: K(L) → X be the inclusion. Then
! ∗ "
t∗x LT ⊗ L−1 ∼ ∗ −1
T = m L ⊗ p1 L |X×K(L)
! "
= Λ(L)|X×K(L) ⊗ p∗2 L |X×K(L)
= p∗2 M ⊗ p∗2 (L|K(L) ) for some line bundle M on K(L),

whereas x∗ L = L|K(L) . Now apply (ii) to find that M = OK(L) . !

(2.18) Proposition. The subscheme K(L) is a subgroup scheme of X.

Proof. Strictly speaking we have not yet defined the notion of a subgroup scheme; see Def-
inition (3.7) below. With that definition the proposition boils down to the statement that
K(L)(T ) ⊂ X(T ) is a subgroup, for any k-scheme T . This follows from (i) of the Lemma
together with the Theorem of the Square. !
The following lemma shows that an ample line bundle is invariant under only finitely many
translations.

(2.19) Lemma. If L is ample then K(L) is a finite group scheme.

Proof. Without loss of generality we may assume that k is algebraically closed. Set Y :=
K(L)0red ⊂ X which, as we noted in (2.13), is an abelian subvariety of X. The restriction L% of
L to Y is again ample. By (iii) of Lemma (2.17) the bundle Λ(L% ) on Y × Y is trivial. Pulling
this bundle back to Y via (1, −1): Y → Y × Y gives that L% ⊗ (−1)∗ L% is trivial on Y . But L%
is ample, hence (−1)∗ L% and L% ⊗ (−1)∗ L% are ample too. It follows that dim(Y ) = 0. Hence
K(L) is finite. !

– 23 –
 We would like to have a converse to this fact. To obtain this we first prove the following
remarkable result.

(2.20) Proposition. Let X be an abelian variety over an algebraically closed field k. Let
f : X → Y be a morphism of k-varieties. For x ∈ X, let Cx denote the connected component of
the fibre over f (x) such that x ∈ Cx , and write Fx for the reduced scheme underlying Cx . Then
F0 is an abelian subvariety of X and Fx = tx (F0 ) = x + F0 for all x ∈ X(k).
Proof. Consider the morphism ϕ: X × Fx → Y obtained by restricting f ◦ m to X × Fx . Clearly
ϕ({0} × Fx ) = {f (x)}. Since Fx is complete and connected, the Rigidity Lemma (1.11) implies
that ϕ maps the fibres {z}×Fx to a point. In particular, we find that f (y −x+Fx ) = f (y) for all
x, y ∈ X(k). Putting y = z, x = 0 gives z + F0 ⊆ Fz ; putting y = 0, x = z gives −z + Fz ⊆ F0 .
This shows that Fz = z + F0 .
To see that F0 is a subgroup scheme of X we take a geometric point a ∈ F0 (k). Then
obviously Fa = F0 so that a + F0 = Fa = F0 . Since F0 is reduced, it follows that F0 is a
subgroup scheme of X. By (2.13) it is then an abelian subvariety. !
To illustrate the proposition, suppose X is a simple abelian variety (over k = k), meaning
that it does not have any non-trivial abelian subvarieties. Then the conclusion is that every
morphism from X to another k-variety is either constant or finite. So the proposition puts
strong restrictions on the geometry of abelian varieties.
We give another interpretation of F0 . For this, let D be an effective divisor on X and
let L = OX (D) be the corresponding line bundle. We claim that linear system |2D| has no
base-points, i.e., the sections of L⊗2 define a morphism of X to projective space. To see this we
have to show that for every geometric point y of X there exists an element E ∈ |2D| that does
not contain y. Now the Theorem of the Square tells us that the divisors of the form

t∗x D + t∗−x D (2)

belong to |2D|. It is easy to see that given y there exists a geometric point x such that y does
not belong to the support of the divisor (2). This means that the map ϕ: X → P(Γ(X, L⊗2 )∗ )
defined by the sections of L⊗2 is a morphism. Note that we also have a morphism

f : X → P = |2D|, x (→ t∗x D + t∗−x D .

The relation between ϕ and f shall be discussed in ??.

We now again assume that k = k. For an effective divisor D on X we define the reduced
closed subscheme H(D) ⊂ X by
$ % &
H(D)(k) = x ∈ X(k) % t∗x D = D .

By t∗x D = D we here mean equality of divisors, not of divisor classes. Clearly H(D) is a subgroup
scheme of X.

(2.21) Lemma. Assume k = k and let L be an effective line bundle on the abelian variety X.
Let f : X → Pn be the morphism defined by the sections of L⊗2 . As in (2.20) let F0 be the
reduced connected fibre of f containing 0. Then H(D)0 = F0 = K(L)0red , where the superscript
“ 0 ” denotes the connected component containing 0.

– 24 –
Proof. Let x ∈ F0 . It follows from (2.20) that f ◦ tx = f . Hence if s ∈ Γ(X, L⊗2 ) then s and t∗x s
have the same zero divisor. We apply this to s = t2 , where t is a section of L with divisor D.
This gives t∗x D = D, i.e., x ∈ H(D). This shows that F0 ⊆ H(D), and since F0 is connected
we find F0 ⊆ H(D)0 . Next, it is obvious that H(D)0 is contained in K(L)0red . To prove that
K(L)0red ⊆ F0 , write L% for the restriction of L to K(L)0red . By (2.13), K(L)0red is an abelian
subvariety of X. Clearly it suffices to show that L% is trivial. Now L% , hence also (−1)∗ L% , has
a non-trivial global section. On the other hand, (−1)∗ L% ∼ = (L% )−1 , as we have seen already in
%
the proof of (2.19). Hence L is trivial. !
As we shall see in the next chapters, there exists a quotient X % := X/F0 which is again an
abelian variety. The Stein factorisation of the morphism f is given by X → → X % → Pn , and L is
%
the pull-back of a bundle on X .

(2.22) Proposition. Let L be a line bundle on an abelian variety X which has a non-zero
global section. If K(L) is a finite group scheme then L is ample.
Proof. We may work over an algebraic closure of k. (Note that if a line bundle L becomes ample
after extension of the ground field then it is already ample.) Let D be the divisor of the given
section. By (2.21) the fibre F0 is reduced to a point and by (2.20) it follows that f is quasi-finite.
Since f is also proper, it is finite. By general theory (see HAG, Chap. III, Exercise 5.7), if the
sections of L⊗2 define a finite morphism X → Pn then L is ample. !

(2.23) Corollary. Let D be an effective divisor on an abelian variety X over an algebraically

closed field. Set L = OX (D). Then the following are equivalent:
(a) H(D) is finite,
(b) K(L) is finite,
(c) L is ample.
For later use we introduce some terminology.

(2.24) Definition. A line bundle L on an abelian variety is said to be non-degenerate if K(L)

is finite.
So, an effective line bundle is non-degenerate if and only if it is ample.

(2.25) Theorem. An abelian variety is a projective variety.

Proof. We first prove this for k = k. Choose a quasi-affine open subset U ⊂ X such that
'
X \ U = ∪i∈I Di for certain prime divisors Di . Set D = i∈I Di . By the preceding results it
suffices to show that H(D) is finite. If x ∈ H(D) then tx transforms U into itself. Assuming—
as we may—that 0 ∈ U , we find that H(D) is contained in U . But H(D) is proper, since
F0 = H(D)0 (as in (2.21)). It follows that H(D) is finite.
If k is arbitrary, we first choose an ample divisor D ⊂ Xk . Then D is defined over a finite
extension k % of k. If k % is Galois over k (which we may assume if k % /k is separable) then
(
σ
D̃ := D
σ∈Gal(k" /k)
m
is an ample divisor on Xk which descends to X. If k % /k is purely inseparable such that αp ∈ k
for all α ∈ k % then pm ·D is an ample divisor which descends to X (clear from working at charts).
Combination of these two cases gives the theorem. !

– 25 –
(2.26) We give another proof of the theorem. Choose a collection of prime divisors D1 , . . . , Dn ,
all containing 0, such that the (scheme-theoretic) intersection ∩ni=1 Di reduces to the single closed
'n
point 0. Set D = i=1 Di . We claim that 3D is a very ample divisor. To prove this we may
pass to an algebraic closure of the ground field, so we will now assume that k = k.
First let us show that the linear system |3D| separates points. Thus, given points P *= Q
of X we want to find a divisor ∆, linearly equivalent to 3D, with P ∈ Supp(∆) but Q ∈ / Supp(∆).
The divisor we take shall be of the form
n
(
∆= t∗ai Di + t∗bi Di + t∗−ai −bi Di (3)
i=1

for certain points ai , bi ∈ X. Note that by the Theorem of the Square, any divisor of this form
is linearly equivalent to 3D. As P *= Q and ∩Di = {0}, one of the Di does not contain P − Q.
Say it is D1 . Take a1 = P , and choose the points b1 , ai and bi (for 2 # i # n) such that Q is
not in the support of
n
(
t∗b1 D1 + t∗−P −b1 D1 + t∗ai Di + t∗bi Di + t∗−ai −bi Di . (4)
i=2

With these choices the divisor ∆ given by (3) has the required properties.
Essentially the same argument shows that |3D| also separates tangent vectors. Namely,
suppose P ∈ X and 0 *= τ ∈ TX,P . As the scheme-theoretic intersection ∩ni=1 Di reduces to the
single closed point 0, there is an index i such that t∗−P τ ∈ TX,0 does not lie in the subspace
TDi ,0 ⊂ TX,0 . Say this holds for i = 1. Take a1 = P , and take the remaining points ai and
bi such that P is not in the support of the divisor given by (4). This gives a divisor ∆ with
P ∈ Supp(∆) but τ not tangent to ∆. !
Later we shall prove that if D is an ample divisor on an abelian variety, then 3D is very
ample. In general 2D will not be very
! ample. "For an example, take an elliptic curve E and let
D = P , a point. Then L(2P ) = Γ E, O(2P ) has dimension 2, and |2P | defines a morphism
E → P1 of degree 2 with ramification divisor of degree 4. (In fact, if char(k) *= 2 this morphism
is ramified in 4 points.)

§ 3. Projective embeddings of abelian varieties.

Any smooth projective variety of dimension g can be embedded into P2g+1 , see [??]. We shall
now show that an abelian variety of dimension g cannot be embedded into P2g−1 and that an
embedding into P2g exists only for elliptic curves and for certain abelian surfaces. So in some
sense abelian varieties do not fit easily into projective space; this also helps to explain why it is
so difficult to write down explicit examples of abelian varieties.
In the proof of the next result we shall use the Chow ring CH(X) of X; we could also work
with a suitable cohomology theory (e.g., Betti cohomology or étale cohomology). In fact, all we
need are a couple of basic formulas which can be found in Fulton’s book [1]. The Chow ring of
an abelian variety is further studied in Chap. 13.

(2.27) Theorem. No abelian variety of dimension g can be embedded into P2g−1 . No abelian
variety of dimension g " 3 can be embedded into P2g .

– 26 –
Proof. Let X be an abelian variety, dim(X) = g, and suppose we have an embedding i: X "→
P = Pm . Consider the exact sequence of sheaves (“adjunction sequence”)

0 → TX → i ∗ T P → N → 0 , (5)

where N is the normal bundle of X in P and TX (resp. TP ) is the tangent bundle of X (resp. P).
Write h ∈ CH(X) for the class of a hyperplane section and ci = ci (N ) (for i = 1, . . . , g − 1)
for the ith Chern class of N . We know that the tangent bundle of X is trivial. Therefore, the
equality of total Chern classes resulting from (5) reads:
m−g
(
(1 + h)m+1 = 1 + ci .
i=1

(See Fulton [1], 3.2.12.) This implies immediately that hm−g+1 = 0 in CHg (X). But deg(hg )
equals the degree, say d, of X in Pm which is non-zero. We thus find m − g + 1 " g + 1,
i.e., m " 2g.
We now consider the case of an embedding into P2g . The previous argument gives
) *
2g + 1
cg = · hg .
g

Aplying the degree map we find

) * ) *
2g + 1 g 2g + 1
deg(cg ) = deg(h ) = d. (6)
g g

But since 2 dim(X) = dim(Pg ), the degree of the highest Chern class cg of the normal bundle N
on X is the self-intersection number of X in P2g , (see Fulton [1], §6.3), which is d2 . Together
with (6) this gives ) *
2g + 1
d= .
g
On the other hand, if we apply the Hirzebruch-Riemann-Roch theorem to the line bundle L =
O(1) and use that the Chern classes of X vanish we find that

χ(L) = c1 (L)g /g! ,

'g
where χ(L) = i=0 (−1)i dimk H i (X, L) is the Euler-Poincaré characteristic of L. Since χ(L) ∈
Z it follows that g! divides deg(hg ) = d. (For more details on Riemann-Roch see Chapter IX.)
But one easily checks that
) *
2g + 1
g! divides ⇒ g < 3.
g

This finishes the proof. !

The proof of the theorem shows that the possibilities for g = 1 and g = 2 are the cubic
curves in P2 and abelian surfaces of degree 10 in P4 . We have met the cubic curves in (1.7).
That there exist abelian surfaces of degree 10 in P4 was shown first by Comessatti in 1909.
He considered complex abelian surfaces C2 /Λ, where Λ ⊂ C2 is the lattice √ obtained from a
suitable embedding of OK ⊕ OK , with OK the ring of integers of K = Q( 5). Horrocks and

– 27 –
Mumford found abelian surfaces in P4 as zero sets of sections of the Horrocks-Mumford bundle,
an indecomposable rank two vector bundle on P2 . For further discussion we refer to Chap. ??.

Exercises.
(2.1) Let k ⊂ K be a field extension. Let X be a k-variety and F a sheaf of OX -modules.
Write XK for the K-variety obtained from X by extension of scalars, and let FK := (XK →
X)∗ F . Show that dimk H 0 (X, F ) = dimK H 0 (XK , FK ). Also show that F ∼
= OX if and only if
∼
FK = OXK .
(2.2) Show that the isomorphism in the Theorem of the Cube is canonical. By this we mean that
∼
to a given line bundle L on an abelian variety X we can associate an isomorphism τX,L : Θ(L) −→
OX×X×X in a functorial way, i.e, such that for every homomorphism f : Y → X we have
f ∗ (τX,L ) = τY,f ∗ L (via the canonical isomorphisms Θ(f ∗ L) ∼
= (f × f × f )∗ Θ(L) and OY ×Y ×Y ∼
=
∗
(f × f × f ) OX×X×X ).
(2.3) Let X be an abelian variety over an algebraically closed field. Show that every effective
divisor on X is linearly equivalent to an effective divisor without multiple components.
(2.4) Prove that no abelian variety of dimension g can be embedded into (P1 )2g−1 . Analyze
when an abelian variety of dimension g can be embedded into (P1 )2g .
(2.5) Let A and B be two abelian groups, written additively, and let n " 0 be an integer. If
f : A → B is a map (not necessarily a homomorphism), define a map θn (f ): An → B by
! " (
θn (f ) a1 , . . . , an = (−1)n+#I f (aI ) ,
I
'
where I runs over the non-empty subsets of {1, 2, . . . , n} and aI := i∈I ai . For instance,
θ0 (f ): {0} → B is the map with value 0 (by convention), θ1 (f ) = f , and
! "
θ2 (f ) a, a% = f (a + a% ) − f (a) − f (a% )
! "
θ3 (f ) a, a% , a%% = f (a + a% + a%% ) − f (a + a% ) − f (a + a%% ) − f (a% + a%% ) + f (a) + f (a% ) + f (a%% ) .
(i) Show that θn (f ): An → B is symmetric, i.e., invariant under the action of the group Sn
on An by permutation of the factors.
(ii) For n " 1, show that we have a relation
! "
θn+1 (f ) a1 , . . . , an , an+1 =
! " ! " ! "
θn (f ) a1 , . . . , an + an+1 − θn (f ) a1 , . . . , an − θn (f ) a1 , . . . , an+1 .
(iii) Use (i) and (ii) to show that θn+1 (f ) = 0 if and only if the map θn (f ): An → B is n-linear.
(iv) Let L be a line bundle on an abelian variety X over a field k. If T is a k-scheme, show
that the map X(T ) × X(T ) → Pic(T ) given by (x1 , x2 ) (→ (x1 + x2 )∗ L ⊗ x∗1 L−1 ⊗ x∗2 L−1 is
bilinear.

Notes. The Theorem of the Square and of the Cube are the pivotal theorems for divisors or line bundles on
abelian varieties. They are due to Weil [3]. Our discussion owes much to Mumford’s book MAV. Solomon
Lefschetz (1884–1972) gave a criterion for complex tori to be embeddable into projective space. This was re-
modelled by Weil to give the projectivity of abelian varieties; see Weil [5]. Our first proof of Theorem (2.25)
follows MAV; the argument given in (2.26) is the one found in Lang [1]. The definition of K(L) goes back to
Weil. Proposition (2.20) is due to M.V. Nori. Theorem (2.27) is due to Barth [1] and Van de Ven [1].

– 28 –
Chapter III. Basic theory of group schemes.

As we have seen in the previous chapter, group schemes come naturally into play in the study of
abelian varieties. For example, if we look at kernels of homomorphisms between abelian varieties
then in general this leads to group schemes that are not group varieties. In the next chapters
we shall have to deal with group schemes more often, so it is worthwile to set up some general
theory.
The present chapter mainly deals with some basic notions, covering most of what is needed
to develop the general theory of abelian varieties. We begin by introducing group schemes in
a relative setting, i.e., working over an arbitrary basis. After this, in order to avoid too many
technicalities, we shall focus on group schemes over a field and affine group schemes.

§ 1. Definitions and examples.

The definition of a group scheme is a variation on that of group variety, where we consider
arbitrary schemes rather than only varieties. This leads to the following, somewhat cumbersome,
definition.

(3.1) Definition. (i) Let S be a scheme. A group scheme over S, or an S-group scheme, is an
S-scheme π: G → S together with S-morphisms m: G ×S G → G (group law, or multiplication),
i: G → G (inverse), and e: S → G (identity section), such that the following identities of
morphisms hold:
m ◦ (m × idG ) = m ◦ (idG × m) : G ×S G ×S G → G ,
m ◦ (e × idG ) = j1 : S ×S G → G ,
m ◦ (idG × e) = j2 : G ×S S → G ,
and
e ◦ π = m ◦ (idG × i) ◦ ∆G/S = m ◦ (i × idG ) ◦ ∆G/S : G → G ,
∼ ∼
where j1 : S×S G −→ G and j2 : G×S S −→ G are the canonical isomorphisms. (Cf. the definitions
and diagrams in (1.1).)
(ii) A group scheme G over S is said to be commutative if, writing s: G ×S G → G ×S G
for the isomorphism switching the two factors, we have the identity m = m ◦ s: G ×S G → G.
(iii) Let (π1 : G1 → S, m1 , i1 , e1 ) and (π2 : G2 → S, m2 , i2 , e2 ) be two group schemes over S.
A homomorphism of S-group schemes from G1 to G2 is a morphism of schemes f : G1 → G2
over S such that f ◦ m1 = m2 ◦ (f × f ): G1 ×S G1 → G2 . (This condition implies that f ◦ e1 = e2
and f ◦ i1 = i2 ◦ f .)
In practice it will usually either be understood what m, i and e are, or it will be unnecessary
to make them explicit; in such case we will simply speak about “a group scheme G over S”
without further specification. (In fact, we already did so in parts (ii) and (iii) of the definition.)
If G is a group scheme over S and if S " → S is a morphism of schemes, then the pull-back
G" := G ×S S " inherits! the"structure of an S " -group scheme. In particular, if s ∈ S then the
fibre Gs := G ×S Spec k(s) is a group scheme over the residue field k(s).

BasGrSch, 15 september, 2011 (812)

– 29 –
 Given an S-group scheme G and an integer n, we define [n] = [n]G : G → G to be the
morphism which on sections—using multiplicative notation for the group law—is given by g %→
g n . If n ! 1 it factors as
∆n
G/S m(n)
[n] = (G −−−−→ GnS −−−→ G) ,

where m(n) is the “iterated multiplication map”, given on sections by (g1 , . . . , gn ) %→ g1 · · · gn .

For commutative group schemes [n] is usually called “multiplication by n”.

(3.2) The definitions given in (3.1) are sometimes not so practicable. For instance, to define a
group scheme one would have to give a scheme G, then one needs to define the morphisms m,
i and e, and finally one would have to verify that a number of morphisms agree. Would it not
be much simpler to describe a group as a scheme whose points form a group? Fortunately this
can be done; it provides a way of looking at group schemes that is often more natural than the
definition given above.
Suppose we have a scheme X over some base scheme S. For many purposes the underlying
point set |X| is not a good object to work with. For instance, if X is a group variety then
|X| will in general not inherit a group structure. However, there is another meaning of the
term “point of X”, and this notion is a very convenient one. Namely, recall that if T → S is
another S-scheme then by a T -valued point of X we mean a morphism of schemes x: T → X
over S. The set of such points is denoted X(T ). As a particular case, suppose S = Spec(k) and
T = Spec(K), where k ⊂ K is a field extension. Then one would also refer to a T -valued point
of X as a “K-rational point”, or in some contexts also as a “point of X with coordinates in K”.
It is useful to place our discussion in a more general context. For this, consider a category C.
The example to keep in mind is the category C = Sch/S of schemes over a base scheme S. Write
C# for the category of contravariant functors C → Sets with morphisms of functors as the
morphisms in C. # For X ∈ C, the functor hX = HomC (−, X) is an object of C. # Sending X
#
to hX gives a covariant functor h: C → C. The basic observation is that in this process we lose
no information, as made precise by the following fundamental lemma.

(3.3) Yoneda Lemma. The functor h: C → C # is fully faithful. That is, for all objects X and
" "
X of C, the natural map HomC (X, X ) → HomC#(hX , hX " ) is a bijection. More generally: for
# and X ∈ C, there is a canonical bijection F (X) → Hom (hX , F ).
every F ∈ C #
C

Proof. Suppose given F ∈ C # and X ∈ C. The identity morphism idX is an element of hX (X). If
α ∈ HomC#(hX , F ) then define ψ(α) := α(idX ) ∈ F (X). This gives a map ψ: HomC#(hX , F ) →
F (X). In the other direction, suppose we have β ∈ F (X). If x: T → X is an element of hX (T )
for some T ∈ C, define ϕ(β)(x) ∈ F (T ) to be the image of β under F (x): F (X) → F (T ). Now it
is straightforward to verify that this gives a map ϕ: F (X) → HomC#(hX , F ) which is an inverse
of ψ. "

(3.4) Definition. A functor F ∈ C# is said to be representable if it is isomorphic to a functor

hX for some X ∈ C. If this holds then it follows from the Yoneda lemma that X is uniquely
determined by F up to C-isomorphism, and any such X is said to represent the functor F .

(3.5) Continuing the discussion of (3.2), we define the notion of a group object in the category
C via the embedding into C. # Thus, if X is an object of C then we define a C-group law on X to
be a lifting of the functor hX : C → Sets to a group-valued functor h̃X : C → Gr. Concretely, to

– 30 –
give a group law on an object X means that for each object T in C we have to specify a group
law on the set hX (T ) = HomC (T, X), such that for every morphism f : T1 → T2 the induced
map hX (f ): hX (T2 ) → hX (T1 ) is a homomorphism of groups. An object of C together with a
C-group law on it is called a C-group, or a group object in C. In exactly the same way we can
define other algebraic structures in a category, such as the notion of a ring object in C.
Let us now suppose that C is a category with finite products. This means that C has a
final object (the empty product), which we shall call S, and that for any two objects X and Y
there exists a product X × Y . If G is a group object in C then the group structure on hG gives
a morphism of functors
m: hG×S G = hG × hG −→ hG .
The Yoneda lemma tells us that this morphism is induced by a unique morphism mG : G×S G →
G. In a similar way we obtain morphisms iG : G → G and eG : S → G, and these morphisms
satisfy the relations of (3.1)(i). Conversely, data (mG , iG , eG ) satisfying these relations define a
C-group structure on the object G.
Applying the preceding remarks to the category Sch/S of schemes over S, which is a category
with finite products and with S as final object, we see that a group scheme G over S is the
same as a representable group functor on Sch/S together with the choice of a representing object
(namely G). The conclusion of this discussion is so important that we state it as a proposition.

(3.6) Proposition. Let G be a scheme over a base scheme S. Then the following data are
equivalent:
(i) the structure of an S-group scheme on G, in the sense of Definition (3.1);
(ii) a group structure on the sets G(T ), functorial in T ∈ Sch/S .
For homomorphisms we have a similar assertion: if G1 and G2 are S-group schemes then the
following data are equivalent:
(i) a homomorphism of S-group schemes f : G1 → G2 , in the sense of Definition (3.1);
(ii) group homomorphisms f (T ): G1 (T ) → G2 (T ), functorial in T ∈ Sch/S .
In practise we often identify a group scheme G with the functor of points hG , and we use
the same notation G for both of them.
Already in the simplest examples we will see that this is useful, since it is often easier
to understand a group scheme in terms of its functor of points than by giving the structure
morphisms m, i and e. Before we turn to examples, let us use the functorial language to define
the notion of a subgroup scheme.

(3.7) Definition. Let G be a group scheme over S. A subscheme (resp. an open subscheme,
resp. a closed subscheme) H ⊂ G is called an S-subgroup scheme (resp. an open S-subgroup
scheme, resp. a closed S-subgroup scheme) of G if hH is a subgroup functor of hG , i.e., if
H(T ) ⊂ G(T ) is a subgroup for every S-scheme T . A subgroup scheme H ⊂ G is said to be
normal in G if H(T ) is a normal subgroup of G(T ) for every S-scheme T .
In what follows, if we speak about subgroup schemes it shall be understood that we give
H the structure of an S-group scheme induced by that on G. An alternative, but equivalent,
definition of the notion of a subgroup scheme is given in Exercise (3.1).

(3.8) Examples. 1. The additive group. Let S be a base scheme. The additive group over S,
denoted Ga,S , corresponds to the functor which associates to an S-scheme T the additive group
Γ(T, OT ). For simplicity, let us assume that S = Spec(R) is affine. Then Ga,S is represented by

– 31 –
 ! "
the affine S-scheme A1S = Spec R[x] . The structure of a group scheme is given, on rings, by
the following homomorphisms:

m̃: R[x] → R[x] ⊗R R[x] given by x %→ x ⊗ 1 + 1 ⊗ x , defining the group law;

ĩ: R[x] → R[x] given by x %→ −x , defining the inverse;
ẽ: R[x] → R given by x %→ 0 , defining the identity.
(See (3.9) below for further discussion of how to describe an affine group scheme in terms of a
Hopf algebra.)
2. The multiplicative group. This group scheme, denoted Gm,S , represents the functor
which associates to an S-scheme T the! multiplicative
" group Γ(T, OT )∗ of invertible elements of
−1
Γ(T, OT ). As a scheme, Gm = Spec OS [x, x ] . The structure of a group scheme is defined by
the homomorphisms given by
x %→ x ⊗ x defining the multiplication;
−1
x %→ x defining the inverse;
x %→ 1 defining the identity element.

3. n-th Roots of unity. Given a positive integer n, we have an S-group scheme µn,S which
associates to an S-scheme T the subgroup of Gm (T ) of elements whose order divides n. The
OS -algebra defining this group scheme is OS [x, x−1 ]/(xn − 1) with the group law given as in
Example 2. Put differently, µn,S is a closed subgroup scheme of Gm,S .
4. pn -th Roots of zero. Let p be a prime number and suppose that char(S) != p. Consider
n n "
the closed subscheme αpn ,S ⊂ Ga,S defined by the ideal (xp ); so αpn ,S := Spec OS [x]/(xp ) .
As is not hard to verify, this is in fact a closed subgroup scheme of Ga,S . If S = Spec(k) for a
field k of characteristic p then geometrically αpn ,k is just a “fat point” (a point together with its
(pn − 1)st infinitesimal neighbourhood); but as a group scheme it has an interesting structure.
n
If T is an S-scheme then αpn (T ) = {f ∈ Γ(T, OT ) | f p = 0}, with group structure given by
addition.
5. Constant group schemes. Let M be an arbitrary (abstract) group. Let MS := S (M ) ,
the direct sum of copies of S indexed by the set M . If T is an S-scheme then MS (T ) is the
set of locally constant functions of |T | to M . The group structure on M clearly induces the
structure of a group functor on MS (multiplication of functions), so that MS becomes a group
scheme. The terminology “constant group scheme” should not be taken to mean that the functor
T %→ MS (T ) has constant value M ; in fact, if M is non-trivial then MS (T ) = M only if T is
connected.

In Examples 1–3 and 5, the group schemes as described here are all defined over Spec(Z).
That is, in each case we have GS = GZ ×Spec(Z) S where GZ is “the same” example but now
over the basis Spec(Z). The group schemes αpn of Example 4 are defined over Spec(Fp ). The
subscript “S ” is sometimes omitted if the basis is Spec(Z) resp. Spec(Fp ), or if it is understood
over which basis we are working.
If G = Spec(A) is a finite k-group scheme then by the rank of G we mean the k-dimension of
its affine algebra A. Thus, for instance, the constant group scheme (Z/pZ)k , and (for char(k) =
p) the group schemes µp,k and αp,k all have rank p.

6. As is clear from the definitions, a group variety over a field k is the same as a geometrically
integral group scheme over k. In particular, abelian varieties are group schemes.

– 32 –
 7. Using the Yoneda lemma one easily sees that, for a group scheme G over a basis S, the
morphism i: G → G is a homomorphism of group schemes if and only if G is commutative.
8. Let S be a basis with char(S) = p. If G is an S-group scheme then G(p/S) naturally
inherits the structure of an S-group scheme (being the pull-back of G via the absolute Frobenius
morphism FrobS : S → S). The relative Frobenius morphism FG/S : G → G(p/S) is a homomor-
phism of S-group schemes.
9. Let V be a finite dimensional vector space over a field k. Then we can form the group
variety GL(V ) over k. If T = Spec(R) is an affine k-scheme then GL(V )(T ) is the group of
invertible R-linear transformations of V ⊗k R. If d = dimk (V ) then GL(V ) is non-canonically
(choice of a k-basis for V ) isomorphic to the group variety GLd,k of invertible d × d matrices; as
a scheme the latter is given by
$ %
GLd,k = Spec k[Tij , U ; 1 # i, j # d]/(det ·U − 1) ,

where det ∈ k[Tij ] is the determinant polynomial. (So “ U = det−1 ”.) We leave it to the reader
to write out the formulas for the group law.
More generally, if V is a vector bundle on a scheme S then we can form the group scheme
GL(V /S) whose T -valued points are the vector bundle automorphisms of VT over T . If V has
rank d then this group scheme is locally on S isomorphic to a group scheme GLd,S of invertible
d × d matrices.
10. As another illustration of the functorial point of view, let us define semi-direct prod-
ucts. Let N and Q be two group schemes over a basis S. Consider the contravariant functor
Aut(N ): Sch/S → Gr which associates to an S-scheme T the group of automorphisms of NT as
a T -group scheme. Suppose we are given an action of Q on N by group scheme automorphisms;
by this we mean that we are given a homomorphism of group functors
ρ: Q → Aut(N ) .
Then we can form the semi-direct product group scheme N !ρ Q. The underlying scheme is just
the product scheme N ×S Q. The group structure is defined on T -valued points by
! "
(n, q) · (n" , q " ) = n · ρ(q)(n" ), q · q " ,
as expected. By (3.6) this defines an S-group scheme N !ρ Q.
Here is an application. In ordinary group theory we know that every group of order p2 is
commutative. The analogue of this in the context of group schemes does not hold. Namely, if
k is a field of characteristic p > 0 then there exists a group scheme of rank p2 over k that is
not commutative. We construct it as a semi-direct product. First note that there is a natural
action of the group scheme Gm on the group scheme Ga ; on points it is given by the usual action
of Gm (T ) = Γ(T, OT )∗ on Ga (T ) = Γ(T, OT ). This action restricts to a (non-trivial) action of
µp,k ⊂ Gm,k on αp,k ⊂ Ga,k . Then the semi-direct product αp ! µp has rank p2 but is not
commutative.

(3.9) Affine group schemes. Let S = Spec(R) be an affine base scheme. Suppose G = Spec(A)
is an S-group scheme which is affine as a scheme. Then the morphisms m, i and e giving G its
structure of a group scheme correspond to R-linear homomorphisms
m̃: A → A ⊗R A called co-multiplication,
ĩ: A → A called antipode or co-inverse,
ẽ: A → R called augmentation or co-unit.

– 33 –
These homomorphisms satisfy a number of identities, corresponding to the identities in the
definition of a group scheme; see (3.1)(i). For instance, the associativity of the group law
corresponds to the identity

(m̃ ⊗ 1) ◦ m̃ = (1 ⊗ m̃) ◦ m̃: A → A ⊗R A ⊗R A .

We leave it to the reader to write out the other identities.

A unitary R-algebra equipped with maps m̃, ẽ and ĩ satisfying these identities is called a
Hopf algebra or a co-algebra over R. A Hopf algebra is said to be co-commutative if s ◦ m̃ =
m̃: A → A ⊗R A, where s: A ⊗R A → A ⊗R A is given by x ⊗ y %→ y ⊗ x. Thus, the category of
affine group schemes over R is anti-equivalent to the category of commutative R-Hopf algebras,
with commutative group schemes corresponding to Hopf algebras that are both commutative
and co-commutative. For general theory of Hopf algebras we refer to ??. Note that in the
literature Hopf algebras can be non-commutative algebras. In this chapter, Hopf algebras are
assumed to be commutative.
The ideal I := Ker(ẽ: A → R) is called the augmentation ideal. Note that A = R · 1 ⊕ I
as R-module, since the R-algebra structure map R → A is a section of the augmentation. Note
that the condition that e: S → G is a two-sided identity element is equivalent to the relation

m̃(α) = (α ⊗ 1) + (1 ⊗ α) mod I ⊗ I (1)

in the ring A ⊗R A. For the co-inverse we then easily find the relation

ĩ(α) = −α mod I 2 , if α ∈ I . (2)

(Exercise (3.3) asks you to prove this.)

The above has a natural generalization. Namely, suppose that G is a group scheme over
an arbitrary basis S such that the structural morphism π: G → S is affine. (In this situation
we say that G is an affine group scheme over S; cf. (3.10) below.) Let AG := π∗ OG , which is a
sheaf of OS -algebras. Then G ∼= Spec(AG ) as S-schemes, and the structure of a group scheme
is given by homomorphisms of (sheaves of) OS -algebras

m̃: AG → AG ⊗OS AG , ĩ: AG → AG , and ẽ: AG → OS

making AG into a sheaf of commutative Hopf algebras over OS . Note that the unit section
e: S → G gives an isomorphism between S and the closed subscheme of G defined by the
augmentation ideal I := Ker(ẽ).

§ 2. Elementary properties of group schemes.

(3.10) Let us set up some terminology for group schemes. As a general rule, if P is a property of
morphisms of schemes (or of schemes) then we say that a group scheme G over S with structural
morphism π: G → S has property P if π has this property as a morphism of schemes (or if G, as
a scheme, has this property). Thus, for example, we say that an S-group scheme G is noetherian,
or finite, if G is a noetherian scheme, resp. if π is a finite morphism. Other properties for which
the rule applies: the property of a morphism of schemes of being quasi-compact, quasi-separated,
(locally) of finite type, (locally) of finite presentation, finite and locally free, separated, proper,

– 34 –
flat, and unramified, smooth, or étale. Similarly, if the basis S is the spectrum of a field k
then we say that G is (geometrically) reduced, irreducible, connected or integral if G has this
property as a k-scheme.
Note that we call G an affine group scheme over S if π is an affine morphism; we do not
require that G is affine as a scheme. Also note that if G is a finite S-group scheme then this
does not say that G(T ) is finite for every S-scheme T . For instance, we have described the
group scheme αp (over a field k of characteristic p) as a “fat
! point”,
" so it should have a positive
dimensional tangent space. Indeed, αp (k) = {1} but αp k[ε] = {1 + aε ! | "a ∈ k}. We find
that the tangent space of αp at the origin has k-dimension 1 and that αp k[ε] is infinite if k is
infinite.
Let us also recall how the predicate “universal(ly)” is used. Here the general rule is the
following: we say that π: G → S universally has property P if for every morphism f : S " → S,
writing π " : G" → S " for the morphism obtained from π by base-change via f , property P holds
for G" over S " .
Let us now discuss some basic properties of group schemes. We begin with a general lemma.

(3.11) Lemma. (i) Let

i
X" −→ X
 
 g
g" ' '
j
Y" −→ Y

be a cartesian diagram in the category of schemes. If g is an immersion (resp. a closed immersion,

resp. an open immersion) then so is g " .
(ii) Let f : Y → X be a morphism of schemes. If s: X → Y is a section of f then s is an
immersion. If f is separated then s is a closed immersion.
(iii) If s: X → Y is a section of a morphism f , as in (ii), then s maps closed points of X to
closed points of Y .
Proof. (i) Suppose g is an immersion. This means we have a subscheme Z ⊂ Y such that g
∼
induces an isomorphism X −→ Z. If Z is an open subscheme (i.e., g an open immersion) then
Y " ×Y Z is naturally isomorphic to the open subscheme j −1 (Z) of Y " , and the claim follows. If
Z is a closed subscheme defined by some ideal I ⊂ OY (i.e, g a closed immersion) then Y " ×Y Z
is naturally isomorphic to the closed subscheme of Y " defined by the ideal generated by j −1 (I);
again the claim follows. The case of a general immersion follows by combining the two previous
cases.
(ii) By (i), it suffices to show that the commutative diagram
s
X −−−−−−−→ Y
 

s'
∆
' Y /X (3)
idY ×(s ◦ f )
Y −−−−−−−→ Y ×X Y

is cartesian. This can be done by working on affine open sets. Alternatively, if T is any scheme
then the corresponding diagram of T -valued points is a cartesian diagram of sets, as one easily
checks. It then follows from the Yoneda lemma that (3) is cartesian.
(iii) Let P ∈ X be a closed point. Choose an affine open U ⊂ Y containing s(P ). It suffices
to check that s(P ) is a closed point of U . (This is special about working with points, as opposed

– 35 –
to arbitrary subschemes.) But U → X is affine, hence separated, so (i) tells us that s(P ) is a
closed point of U . Alternatively, the assertion becomes obvious by working on rings. "

(3.12) Proposition. (i) An S-group scheme G is separated if and only if the unit section e is
a closed immersion.
(ii) If S is a discrete scheme (e.g., the spectrum of a field) then every S-group scheme is
separated.
Proof. (i) The “only if” follows from (ii) of the lemma. For the converse, consider the commu-
tative diagram π
G −−−−−−−→ S
 

∆G/S '
e
'
m ◦ (idG ×i)
G ×S G −−−−−−−→ G
For every S-scheme T it is clear that this diagram is cartesian on T -valued points. By the
Yoneda lemma it follows that the diagram is cartesian. Now apply (i) of the lemma.
(ii) Since separatedness is a local property on the basis, it suffices to consider the case that
S is a 1-point scheme. Then the unit section is closed, by (iii) of the lemma. Now apply (i). "
As the following example shows, the result of (ii) is in some sense the best possible. Namely,
suppose that S is a scheme which is not discrete. Then S has a non-isolated closed point s (i.e.,
a closed point s which is not open). Define G as the S-scheme obtained by gluing two copies of
S along S \ {s}. Then G is not separated over S, and one easily shows that G has a structure
of S-group scheme with Gs ∼ = (Z/2Z)k(s) . Notice that in this example G is even étale over S.

(Z/2Z) trivial fibres

↓ ↓ ↓
G

'π
S
s
Figure 3.

(3.13) Definition. (i) Let G be an S-group scheme with unit section e: S → G. Define
eG = e(S) ⊂ G (a subscheme of G) to be the image of the immersion e.
(ii) Let f : G → G" be a homomorphism of S-group schemes. Then we define the kernel of f
to be the subgroup scheme Ker(f ) := f −1 (eG" ) of G.
Note that the diagram
Ker(f ) (−→ G
 
 f
' '
e
S −−−→ G"
is cartesian. In particular, Ker(f ) represents the contravariant functor Sch/S → Gr given by
$ %
T %→ Ker f (T ): G(T ) −→ G" (T )

– 36 –
and is a normal subgroup scheme of G. If G" is separated over S then Ker(f ) ⊂ G is a closed
subgroup scheme.
As examples of kernels we have, taking S = Spec(Fp ) as our base scheme,
µp = Ker(F : Gm → Gm ) , αp = Ker(F : Ga → Ga ) ,
where in both cases F denotes the Frobenius endomorphism.

(3.14) Left and right translations; sheaves of differentials. Let G be a group scheme over a
basis S. Given an S-scheme T and a point g ∈ G(T ), the right translation tg : GT → GT and
the left translation t"g : GT → GT are defined just as in (1.4). Using the Yoneda lemma we can
also define tg and t"g by saying that for every T -scheme T " , the maps tg (T " ): G(T " ) → G(T " )
and t"g (T " ): G(T " ) → G(T " ) are given by γ %→ γg resp. γ %→ gγ. Here we view g as an element of
G(T " ) via the canonical homomorphism G(T ) → G(T " ).
If in the above we take T = G and g = idG ∈ G(G) then the resulting translations τ and
τ " : G ×S G → G ×S G are given by (g1 , g2 ) → (g1 g2 , g2 ), resp. (g1 , g2 ) → (g2 g1 , g2 ). Here we
view G ×S G as a scheme over G via the second projection. We call τ and τ " the universal right
(resp. left) translation. The point is that any other right translation tg : G ×S T → G ×S T as
above is the pull-back of τ via idG × g (i.e., the pull-back via g on the basis), and similarly for
left translations.
As we have seen in (1.5), the translations on G are important in the study of sheaves of
differentials. We will formulate everything using right translations. A 1-form α ∈ Γ(G, Ω1G/S )
is said to be (right) invariant if it is universally invariant under right translations; by this we
mean that for every T → S and g ∈ G(T ), writing αT ∈ Γ(T, Ω1GT /T ) for the pull-back of α
via GT → G, we have t∗g αT = αT . In fact, it suffices to check this in the universal case: α is
invariant if and only if p∗1 α ∈ Γ(G ×S G, p∗1 Ω1G/S ) is invariant under τ . The invariant differentials
form a subsheaf (π∗ Ω1G/S )G of π∗ Ω1G/S .
For the next result we need one more notation: if π: G → S is a group scheme with unit
section e: S → G, then we write
ωG/S := e∗ Ω1G/S ,
which is a sheaf of OS -modules. If S is the spectrum of a field then ωG/S is just cotangent space
of G at the origin.

(3.15) Proposition. Let π: G → S be a group scheme. Then there is a canonical isomorphism

∼
π ∗ ωG/S −→ Ω1G/S . The corresponding homomorphism ωG/S → π∗ Ω1G/S (by adjunction of the
∼
functors π ∗ and π∗ ) induces an isomorphism ωG/S −→ (π∗ Ω1G/S )G .
Proof. As in (1.5), the geometric idea is that an invariant 1-form on G can be reobtained from
its value along the zero section by using the translations, and that, by a similar proces, an
arbitrary 1-form can be written as a function on G times an invariant form. To turn this idea
into a formal proof we use the universal translation τ .
As above, we view G ×S G as a G-scheme via p2 . Then τ is an automorphism of G ×S G
over G, so we have a natural isomorphism
∼
τ ∗ Ω1G×S G/G −→ Ω1G×S G/G . (4)
We observe that G×S G/G is the pull back under p1 of G/S; this gives that Ω1G×S G/G = p∗1 Ω1G/S .
As τ = (m, p2 ): G ×S G → G ×S G, we find that (4) can be rewritten as
∼
m∗ Ω1G/S −→ p∗1 Ω1G/S .

– 37 –
Pulling back via (e ◦ π, idG ): G → G ×S G gives the isomorphism
∼
Ω1G/S −→ π ∗ e∗ Ω1G/S = π ∗ ωG/S . (5)

By adjunction, (5) gives rise to a homomorphism π ∗ : ωG/S → π∗ Ω1G/S associating to a

section β ∈ Γ(S, ωG/S ) the 1-form π ∗ β ∈ Γ(G, π ∗ ωG/S ) = Γ(G, Ω1G/S ). The isomorphism (5)
is constructed in such a way that π ∗ β is an invariant form.
! Clearly
" e∗ (π ∗ β) = β. Conversely,
if α ∈ Γ(G, ΩG/S ) is an invariant form then m (α) = τ p1 (α) = p∗1 (α). Pulling back (as in
1 ∗ ∗ ∗

the above argument) via (e ◦ π, idG ) then gives that α = π ∗ e∗ (α). This shows that the map
(π∗ Ω1G/S )G → ωG/S given on sections by α %→ e∗ α is an inverse of π ∗ . "

(3.16) The identity component of a group scheme over a field. Let G be a group scheme over
a field k. By (3.12), G is separated over k. The image of the identity section is a single closed
point e = eG of degree 1.
Assume in addition that G is locally of finite type over k. Then the scheme G is locally
noetherian, hence locally connected. If we write G0 for the connected component of G con-
taining e, it follows that G0 is an open subscheme of G. We call G0 the identity component
of G.
Geometrically, one expects that the existence of a group structure implies that G, as a
k-scheme, “looks everywhere the same”, so that certain properties need to be tested only at the
origin. The following proposition shows that for smoothness and reducedness this is indeed the
case. Note, however, that our intuition is a geometric one: in general we can only expect that
“G looks everywhere locally the same” if we work over k = k. In the following proposition it is
good to keep some simple examples in mind. For instance, let p be a prime number and consider
the group scheme µp over the field Q. The underlying topological space consists of two closed
points: the origin e = 1, and a point P corresponding to the non-trivial pth roots of unity. If
we extend scalars from Q to a field containing a pth root of unity then the identity component
(µp )0 = {e} stays connected but the other component {P } splits up into a disjoint union of p − 1
connected components.

(3.17) Proposition. Let G be a group scheme, locally of finite type over a field k.
(i) The identity component G0 is an open and closed subgroup scheme of G which is
geometrically irreducible. In particular: for any field extension k ⊂ K, we have (G0 )K = (GK )0 .
(ii) The following properties are equivalent:
(a1) G ⊗k K is reduced for some perfect field K containing k;
(a2) the ring OG,e ⊗k K is reduced for some perfect field K containing k;
(b1) G is smooth over k;
(b2) G0 is smooth over k;
(b3) G is smooth over k at the origin.
(iii) Every connected component of G is irreducible and of finite type over k.
Proof. (i) We first prove that G0 is geometrically connected; that it is even geometrically irre-
ducible will then follow from (iii). More generally, we show that if X is a connected k-scheme,
locally of finite type, that has a k-rational point x ∈ X(k) then X is geometrically connected.
(See EGA IV, 4.5.14 for a more general result.)
Let k be an algebraic closure of k. First we show that the projection p: Xk → X is open
and closed. Suppose {Vα }α∈I is an open covering of X. Then {Vα,k }α∈I is a covering of Xk . If
each Vα,k → Vα is open and closed then the same is true for p. Hence we may assume that X is

– 38 –
affine and of finite type over k. Let Z ⊂ Xk be closed. Then there is a finite extension k ⊂ K
inside k such that Z is defined over K; concretely this means that there is closed subscheme
ZK ⊂ XK with Z = ZK ⊗K k. Hence it suffices to show that the morphism pK : XK → X is
open and closed. But this is immediate from the fact that pK is finite and flat. (Use HAG,
Chap. III, Ex. 9.1 or EGA IV, Thm. 2.4.6.)
Now suppose we have two non-empty open and closed subsets U1 and U2 of Xk . Because
X is connected, it follows that p(U1 ) = p(U2 ) = X. The unique point x ∈ Xk lying over x is
therefore contained in U1 ∩ U2 ; hence U1 ∩ U2 is non-empty. This shows that Xk is connected.
(ii) The essential step is to prove that (a2) ⇒ (b1); all other implications are easy. (For
(b3) ⇒ (b1) use (3.15).) One easily reduces to the case that k = k and that G is reduced at
the origin. Using the translations on G it then follows that G is reduced. In this situation, the
same argument as in (1.5) applies, showing that G is smooth over k.
For (iii) one first shows that G0 is irreducible and quasi-compact. We have already shown
that (G0 )K = (GK )0 for any field extension k ⊂ K, so we may assume that k = k, in which case
we can pass to the reduced underlying group scheme G0red ; see Exercise (3.2). Note that G0red
has the same underlying topological space as G0 . By (ii), G0red is smooth over k. Every point of
G0red therefore has an open neighbourhood of the form U = Spec(A) with A a regular ring. As a
regular ring is a domain, such an affine scheme U is irreducible. Now suppose G0red is reducible.
Because it is connected, there exist two irreducible components C1 -= C2 with C1 ∩ C2 -= ∅. (See
EGA 0I , Cor. 2.1.10.) If y ∈ C1 ∩ C2 , let U = Spec(A) be an affine open neighbourhood of y in
G0red with A regular. Then one of C1 ∩ U and C2 ∩ U contains the other, say C2 ∩ U ⊆ C1 ∩ U .
But C2 ∩ U is dense in C2 , hence C2 ⊆ C1 . As C1 and C2 are irreducible components we must
have C2 = C1 , contradicting the assumption.
To prove quasi-compactness of G0 , take a non-empty affine open part U ⊂ G0 . Then U
is dense in G0 , as G0 is irreducible. Hence for every g ∈ G0 (k) the two sets g · U −1 and U
intersect. It follows that the map U × U → G0 given by multiplication is surjective. But U × U
is quasi-compact, hence so is G0 .
Now we look at the other connected components, working again over an arbitrary field k.
If H ⊂ G is a connected component, choose a closed point h ∈ H. Because G is locally of
finite type over k, there is a finite normal field extension k ⊂ L such that L contains the residue
field k(h). As in the proof of (i), the projection p: H ⊗k L → H is open and closed. One
easily shows that all points in p−1 (h) are rational over L. If h̃ ∈ p−1 (h) is one of these points
then using the translation th̃ one sees that the connected component C(h̃) of HL containing h̃
is isomorphic to G0L as an L-scheme. Then p(C(h̃)) ⊂ H is irreducible, closed and open. As
H is connected it follows that p(C(h̃)) = H and that H is irreducible. Finally, the preceding
arguments show that H ⊗k L is the union of the components C(h̃) for all h̃ in the finite set
p−1 (h). As each of these components is isomorphic to G0L , which is quasi-compact, it follows
that H is quasi-compact. "

(3.18) Remarks. (i) Let G be a k-group scheme as in the proposition. Suppose that G ⊗k K
is reduced (or that OG,e ⊗k K is reduced) for some non-perfect field K containing k. Then
it is not necessarily true that G is smooth over k. Here is an example: Suppose K = k is a
non-perfect! field of characteristic
" p. Choose an element
! α ∈" k not in k p . Let G be the k-scheme
p p 2
G = Spec k[X, Y ]/(X + αY ) . View Ak = Spec k[X, Y ] as a k-group scheme by identifying
it with Ga,k × Ga,k . Then G is a closed subgroup scheme of A2k . One easily checks that G is
√
reduced, but clearly it is not geometrically reduced (extend to the field k( p α)), and therefore
G is not a smooth group scheme over k.

– 39 –
 (ii) In (iii) of the proposition, let us note that the connected components of G are in general
not geometrically irreducible; see the example given before the proposition.

(3.19) Remark. Let G be a group scheme, locally of finite type over a field k. In case G is
affine, we have seen in (3.9) that we can study it through its Hopf algebra. For arbitrary G
there is no immediate substitute for this, not even if we are only interested in the local structure
of G at the origin. Note that the group law does not, in general, induce a co-multiplication
on the local ring OG,e . We do have a homomorphism OG,e → OG×k G,(e,e) but OG×k G,(e,e) is
in general of course not the same as OG,e ⊗k OG,e ; rather it is a localisation of it. In some
cases, however, something slightly weaker already suffices to obtain interesting conclusions. In
the proof of the next result we shall exploit the fact that, with m ⊂ OG,e the maximal ideal, we
do have a homomorphism m̃: OG,e → (OG,e /mq ) ⊗k (OG,e /mq ) for which the analogue of (1) in
section (3.9) holds.
Another possibility is to consider the completed local ring ÔG,e . The group law on G
ˆ k ÔG,e (completed tensor product). In this way we
induces a co-multiplication m̃: ÔG,e → ÔG,e ⊗ ! "
can associate to a group variety G a (smooth) formal group Ĝ = Spf ÔG,e . We shall further
go into this in ??.

(3.20) Theorem. (Cartier) Let G be a group scheme, locally of finite type over a field k of
characteristic zero. Then G is reduced, hence smooth over k.
Proof. We follow the elementary proof due to Oort [2]. Let A := OG,e be the local ring of G
at the identity element. Write m ⊂ A for the maximal ideal and nil(A) ⊂ A for the nilradical.
Since we are over a perfect field, the reduced scheme Gred underlying G is a subgroup scheme
(Exercise (3.2)), and by (ii) of Prop. (3.17) this implies that Ared := A/nil(A) is a regular local
ring. Writing mred := m/nil(A) ⊂ Ared , this gives
! "
dim(A) = dim(Ared ) = dimk (mred /m2red ) = dimk m/m2 + nil(A) .

In particular, we see that it suffices to show that nil(A) ⊂ m2 . Indeed, if this holds then
dim(A) = dim(m/m2 ), hence A is regular, hence nil(A) = 0.
Choose 0 -= x ∈ nil(A), and let n be the positive integer such that xn−1 -= 0 and xn = 0.
Because A is noetherian, we have ∩q!0 mq = (0), so there exists an integer q ! 2 with xn−1 ∈
/ mq .
Consider B := A/mq and m̄ := m/mq ⊂ B, and let x̄ ∈ B denote the class of x ∈ A modulo mq .
As remarked above, the group law on G induces a homomorphism m̃: A → B ⊗k B. Just as
in (3.9), the fact that e ∈ G(k) is a two-sided identity element implies that we have

m̃(x) = (x̄ ⊗ 1) + (1 ⊗ x̄) + y with y ∈ m̄ ⊗k m̄ . (6)

(See also Exercise (3.3).) This gives

0 = m̃(xn ) = m̃(x)n = ((x̄ ⊗ 1) + (1 ⊗ x̄) + y)n

(n ) *
n
= · (x̄ ⊗ 1)n−i · ((1 ⊗ x̄) + y)i .
i=0
i

From this we get the relation

$ %
n · (x̄n−1 ⊗ x̄) ∈ (x̄n−1 · m̄) ⊗k B + B ⊗k m̄2 ⊂ B ⊗k B .

– 40 –
But char(k) = 0, so that n is a unit, so that even (x̄n−1 ⊗ x̄) ∈ (x̄n−1 · m̄) ⊗k B + B ⊗k m̄2 .
Now remark that a relation of the form y1 ⊗ y2 ∈ J1 ⊗k B + B ⊗k J2 implies that either y1 ∈ J1
or y2 ∈ J2 . (To see this, simply view B, J1 and J2 as k-vector spaces.) But by the Nakayama
Lemma, x̄n−1 ∈ x̄n−1 · m̄ implies x̄n−1 = 0, which contradicts our choice of q. We conclude that
x̄ ∈ m̄2 ; hence x ∈ m2 , and we are done. "
The conclusion of this theorem does not hold over fields of positive characteristic. For
example, if char(k) = p > 0 then the group schemes µp,k and αp,k are not reduced, hence not
smooth over k. (The argument of the above proof breaks down if n is divisible by p.)

§ 3. Cartier duality.

(3.21) Cartier duality of finite commutative group schemes. We now discuss some aspects of
finite commutative group schemes that play an important role in the study of abelian varieties.
In particular, the Cartier duality that we shall discuss here comes naturally into play when we
discuss the dual of an abelian variety; see Chapter 7.
The Cartier dual of a group scheme can be defined in two ways: working functorially or
working with the underlying Hopf algebras. We first give two constructions of a dual group;
after that we prove that they actually describe the same object.
The functorial approach is based on the study of characters, by which we mean homo-
morphisms of the group scheme to the multiplicative group Gm . More precisely, suppose G is
any commutative group scheme over a basis S. Then we can define a new contravariant group
functor Hom(G, Gm,S ) on the category of S-schemes by

Hom(G, Gm,S ): T %→ Hom /T

(GT , Gm,T ) .

Next we define a dual object in terms of the Hopf algebra. For this we need to assume
that G is commutative and finite locally free over S. As in (3.9) above, write A := π∗ OG . This
A is a finite locally free sheaf of OS -modules which comes equipped with the structure of a
sheaf of co-commutative OS -Hopf algebras. (Recall that all our Hopf algebras are assumed to
be commutative.) Thus we have the following maps:

algebra structure map a: OS → A , augmentation ẽ: A → OS ,

ring multiplication µ: A ⊗OS A → A , co-multiplication m̃: A → A ⊗OS A ,
co-inverse ĩ: A → A .

We define a new sheaf of co-commutative OS -Hopf algebras AD as follows: first we set

D
A := Hom OS (A, OS ) as an OS -module. The above maps induce OS -linear maps

aD : AD → OS , ẽD : OS → AD ,
µD : AD → AD ⊗OS AD , m̃D : AD ⊗OS AD → AD ,
ĩD : AD → AD .

We give AD the structure of a sheaf of OS -algebras by defining m̃D to be the multiplication and
ẽD to be the algebra structure morphism. Next we define a Hopf algebra structure by using
µD as the co-multiplication, ĩD as the co-inverse, and aD as the co-unit. We leave it to the

– 41 –
reader (Exercise (3.8)) to verify that this gives AD a well-defined structure of a co-commutative
OS -Hopf algebra. Schematically, if we write the structure maps of a Hopf algebra in a diagram
..
.
..
multiplication . co-multiplication
antipode
..
algebra structure map . augmentation map
..
.

then the diagram corresponding to AD is obtained from that of A by first dualizing all maps
and then reflecting in the dotted line.
We write α: A → (AD )D for the OS -linear
! " a local section s ∈ A(U ) to the
map which sends
section evs = “evaluation at s” ∈ Hom OS Hom OS (A, OS ), OS (U ).

(3.22) Theorem. (Cartier Duality) Let π: G → S be a commutative S-group scheme which is

finite and locally free over S. Write A := π∗ OG , and define the sheaf of co-commutative Hopf
algebras AD over OS as above. Then GD := Spec(AD ) is a commutative, finite locally free
S-group scheme which represents the contravariant functor Hom(G, Gm,S ): Sch/S → Gr given
by
T %→ Hom /T
(GT , Gm,T ) .

The homomorphism (GD )D → G induced by the map α: A → (AD )D is an isomorphism.

Proof. That GD is indeed a commutative group scheme is equivalent to saying that AD is a
sheaf of co-commutative Hopf algebras, which we have left as an exercise to the reader. That
GD is again finite and locally free over S (of the same rank as G) is clear, and so is the claim
that (GD )D → G is an isomorphism.
Note that the functor G %→ GD is compatible with base-change: if T is an S-scheme and
G is a commutative, finite locally free S-group scheme then (GT )D ∼ = (GD )T canonically. In
particular, to prove that GD represents the functor Hom(G, Gm,S ) we may assume that the
basis is affine, say S = Spec(R), and it suffices to show that GD (S) is naturally isomorphic to
the group Hom /S
(G, Gm,S ). As S is affine we may view A simply as an R-Hopf algebra (i.e.,
replace the sheaf A by its R-algebra of global sections).
Among the identities that are satisfied by the structure homomorphisms we have that
(ẽ ⊗ id) ◦ m̃: A → R ⊗R A ∼ = A is the identity and that (ĩ, id) ◦ m̃: A → A is equal to the
composition a ◦ ẽ: A → R → A. In particular, if b ∈ A is an element with m̃(b) = b ⊗ b then it
follows that ẽ(b) · b = b and that ĩ(b) · b = ẽ(b). It follows that

{b ∈ A∗ | m̃(b) = b ⊗ b} = {b ∈ A | m̃(b) = b ⊗ b and ẽ(b) = 1} .

Write Agl for this set. (Its elements are sometimes referred to as the “group-like” elements of A.)
One easily checks that Agl is a subgroup of A∗ .
With these remarks in mind, let us compute Hom /S
(G, Gm,S ) and GD (S). The R-
algebra homomorphisms f : R[x, x ] → A are given by the elements b ∈ A∗ , via the corre-
−1

spondence b := f (x). The condition on b ∈ A∗ that the corresponding map f is a homo-

morphism of Hopf algebras is precisely that m̃(b) = b ⊗ b. Hence we find a natural bijection
∼
Hom S (G, Gm,S ) − → Agl , and one readily verifies this to be an isomorphism of groups.

– 42 –
 Every R-module homomorphism AD → R is of the form evb : λ %→ λ(b) for some b ∈ A.
Conversely, if b ∈ A then one verifies that

evb (1) = 1 ⇐⇒ ẽ(b) = 1

and
evb is a ring homomorphism ⇐⇒ m̃(b) = b ⊗ b .
∼
This gives a bijection GD (S) −→ Agl , and again one easily verifies this to be an isomorphism of
groups. "

(3.23) Definition. Let π: G → S be a commutative S-group scheme which is finite and

locally free over S. Then we call GD the Cartier dual of G. Similarly, if f : G1 → G2 is a
homomorphism between commutative, finite locally free S-group schemes then we obtain an
induced homomorphism f D : GD D
2 → G1 , called the Cartier dual of f .

(3.24) Examples. 1. Take G = (Z/nZ)S . Then it is clear from the functorial description of
the Cartier dual that GD = µn,S . Hence (Z/nZ) and µn are Cartier dual to each other. Note
that (Z/nZ)S and µn,S may well be isomorphic. For instance, if S = Spec(k) is the spectrum of
∼
a field and if ζ ∈ k is a primitive nth root of 1 then we obtain an isomorphism (Z/nZ)k −→ µn,k
sending 1̄ to ζ. In particular, if k = k and char(k) " n then (Z/nZ)k ∼ = µn,k . By contrast, if
char(k) = p > 0 and p divides n then (Z/nZ)k and µn,k are not isomorphic.
2. Let S be a scheme of characteristic p > 0. We claim that αp,S is its own Cartier dual.
Of course this can be shown at the level of Hopf algebras, but the functorial interpretation is
perhaps more instructive. As Cartier duality is compatible with base-change it suffices to do
the case S = Spec(Fp ).
Recall that if R is a ring of characteristic p then αp (R) = {r ∈ R | rp = 0} with its natural
structure of an additive group. If we want to make a homomorphism αp → Gm then the most
obvious guess is to look for an “exponential”. Indeed, if r ∈ αp (R) then

r2 rp−1
exp(r) = 1 + r + + ··· +
2! (p − 1)!

is a well-defined element of R∗ , and r %→ exp(r) defines a homomorphism αp (R) → Gm (R). Now

remark that αp (like Ga ) is not just a group scheme but has a natural structure of a functor
∼
in rings. The self-duality αp −→ αpD = Hom /Fp
(αp , Gm ) is obtained by sending a point
ξ ∈ αp (T ) (where T is an Fp -scheme) to the homomorphism of group schemes αp,T → Gm,T
given (on points with values in T -schemes) by x %→ exp(ξ · x).
3. After the previous example, one might guess that αpn is self-dual for all n. This is not
the case. Instead, (αpn )D can be described as the kernel of Frobenius on the group scheme Wn
of Witt vectors of length n. See Oort [3], § 10. For a special case of this, see also Exercise ??.

§ 4. The component group of a group scheme.

If X is a topological space then π0 (X) denotes the set of connected components of X. The
purpose of this section is to discuss a scheme-theoretic analogue of this for schemes that are

– 43 –
locally of finite type over a field k. To avoid confusion we shall use the notation π0 in the
topological context and /0 for the scheme-theoretic analogue.
If X/k is locally of finite type then /0 (X) will be an étale k-scheme, and X %→ /0 (X) is
a covariant functor. Furthermore, if G is a k-group scheme, locally of finite type over k, then
/0 (G) inherits a natural structure of a group scheme; it is called the component group (scheme)
of G.
We start with some generalities on étale group schemes. Let us recall here that, according
to our conventions, an étale morphism of schemes f : X → Y is only required to be locally of
finite type; see ??.

(3.25) Étale group schemes over a field. Let k be a field. Choose a separable algebraic closure ks
and write Γk := Gal(ks /k). Then Γk is a pro-finite group, (see Appendix ??) and Galois theory
tells us that L %→ Gal(ks /L) gives a bijection between the field extensions of k inside ks and the
closed subgroups of Γk . Finite extensions of k correspond to open subgroups of Γk . A reference
is Neukirch [1], Sect. 4.1.
By a Γk -set we mean a set Y equipped with a continuous left action of Γk . The continuity
assumption here means that for every y ∈ Y the stabilizer Γk,y ⊂ Γk is an open subgroup; this
implies that the Γk -orbits in Y are finite.
Let S := Spec(k). If X is a connected étale scheme over S, then X is of the form X =
Spec(L), with L a finite separable field extension of k. An arbitrary étale S-scheme can be written
as a disjoint union of its connected components, and is therefore of the form X = 2α∈I Spec(Lα ),
where I is some index set and where k ⊂ Lα is a finite separable extension of fields. Hence the
description of étale S-schemes is a matter of Galois theory. More precisely, if Et/k denotes the
category of étale k-schemes there is an equivalence of categories
eq ! "
Et/k −−→ Γk -sets .

associating to X ∈ Et/k the set X(ks ) with its natural Γk -action. To obtain a quasi-inverse,
write a Γk -set Y as a union of orbits, say Y = 2α∈I (Γk · yα ), let k ⊂ Lα be the finite field
extension (inside ks ) corresponding to the open subgroup Stab(yα ) ⊂ Γk , and associate to Y the
S-scheme 2α∈I Spec(Lα ). Up to isomorphism of S-schemes this does not depend on the chosen
base points of the Γk -orbits, and it gives a quasi-inverse to the functor X %→ X(ks ).
This equivalence of categories induces an equivalence between the corresponding categories
of group objects. This gives the following result.

(3.26) Proposition. Let k ⊂ ks and Γk = Gal(ks /k) be as above. Associating to an étale

k-group scheme G the group G(ks ) with its natural Γk -action gives an equivalence of categories
) *
étale eq ! "
−−→ Γk -groups ,
k-group schemes

where by a Γk -group we mean an (abstract) group equipped with a continuous left action of Γk
by group automorphisms.
The proposition tells us that every étale k-group scheme G is a k-form of a constant group
scheme. More precisely, consider the (abstract) group M = G(ks ). Then we can form the
constant group scheme Mk over k, and the proposition tells us that G ⊗ ks ∼ = Mk ⊗ ks . If G
is finite étale over k then we can even find a finite separable field extension k ⊂ K such that
GK ∼ = MK . So we can think of étale group schemes as “twisted constant group schemes”.

– 44 –
 For instance, if char(k) is prime to n then µn is a finite étale group scheme, and µn (ks ) is
(non-canonically) isomorphic to Z/nZ. The action of Γk on µn (ks ) is given by a homomorphism
χ: Γk → (Z/nZ)∗ ; here the rule is that if ζ ∈ ks ∗ is an n-th root of unity and σ ∈ Γk then
σ
ζ = ζ χ(σ) .
Now we turn to the scheme /0 (X) of connected components of X.

(3.27) Proposition. Let X be a scheme, locally of finite type over a field k. Then there is
an étale k-scheme /0 (X) and a morphism q: X → /0 (X) over k such that q is universal for k-
morphisms from X to an étale k-scheme. (By this we mean: for any k-morphism h: X → Y with
Y /k étale, there is a unique k-morphism g: /0 (X) → Y such that h = g ◦ q.) The morphism q
is faithfully flat, and its fibres are precisely the connected components of X.
Before we give the proof, let us make the last assertion more precise. If P is a point of
/0 (X) then {P } is a connected component of /0 (X), as the topological space of an étale scheme
−1
is discrete. The claim is then
+ that
+ q (P ), as an open subscheme of X, is a connected component
of X, for all points P ∈ +/0 (X)+.
! "
Proof. Consider the set π0geom (X) := π0 |X ⊗k ks | with its natural action of Γk . First we show
that the action of Γk is continuous. Let C ⊂ Xks be a connected component. Let D ⊂ X be the
connected component containing the image of C under the natural morphism Xks → X. Then
C is one of the connected components of D ⊗k ks . As D is locally of finite type over k, there
is a point x ∈ |D| such that k(x) is a finite extension of k. Let k " be the separable algebraic
closure of k inside k(x), and let k "" ⊂ ks denote the normal closure of k " . Then by EGA IV,
Prop. (4.5.15), all connected components of D ⊗k k "" are geometrically connected. Hence the
stabilizer of C contains the open subgroup Gal(ks /k "" ) ⊂ Γk , and is therefore itself open.
Define ,
/0geom (X) := Spec(ks )(α) ,
α∈π0geom (X)

the disjoint union of copies of Spec(ks ), one copy for each element of π0geom (X). Consider the
morphism q geom : Xks → /0geom (X) that on each connected component X (α) ⊂ Xks is given
by the structural morphism X (α) → Spec(ks )(α) . (So a point P ∈ Xks is sent to the copy of
Spec(ks ) labelled by the component of Xks that contains P .) Because the Γk -action on the
set π0geom (X)
! "is ∼
continuous, there is an étale k-scheme /0 (X) such that we have an isomorphism
β: /0 (X) ks −→ π0geom (X) of sets with Galois action. Up to isomorphism of k-schemes, this
∼
scheme is unique, and we have a unique isomorphism /0 (X) ⊗k ks −→ /0geom (X) that gives the
identity on ks -valued points. (Here we fix the identification β.) Then q geom can be viewed as a
morphism
q geom : X ⊗k ks → /0 (X) ⊗k ks ,
which is Γk -equivariant. By Galois descent this defines a morphism q: X → /0 (X) over k. (See
also Exercise (3.9).)
Next we show that the fibres of q are the connected components of X. Over ks this is clear
from the construction. Over k it suffices to show that distinct connected components of X are
mapped to distinct points of /0 (X). But the connected components of X correspond to the
Γk -orbits in π0geom (X), so the claim follows from the result over ks .
We claim that the morphism q: X → /0 (X) has the desired universal property. To see
this, suppose h: X → Y is a k-morphism with Y /k étale. Then Y ⊗k ks is a disjoint union
of copies of Spec(ks ). It readily follows from our construction of /0 (X) and q that there is

– 45 –
a unique morphism g geom : /0 (X) ⊗k ks → Y ⊗k ks such that hgeom : Xks → Yks factors as
hgeom = g geom ◦ q geom . Moreover, g geom is easily seen to be Galois-equivariant; hence we get the
desired morphism g: /0 (X) → Y with h = g ◦ q.
Finally we have to show that q is faithfully flat. But this can be checked after making a
base change to ks , and over ks it is clear from the construction. "

(3.28) In the situation of the proposition, we refer to /0 (X) as the scheme of connected com-
ponents of X. If f : X → Y is a morphism of schemes that are locally of finite type over k then
we write /0 (f ): /0 (X) → /0 (Y ) for the unique morphism such that qY ◦ f = /0 (f ) ◦ qX : X →
/0 (Y ).

(3.29) Let G be a k-group scheme, locally of finite type. The connected components ! of" Gks are
geom
geometrically
! " connected; see EGA IV, Prop. (4.5.21). Therefore π 0 (G) := π 0 |G ks | is equal
geom geom 0
to π0 |Gk | . The natural map q : G(k) → π0 (G) is surjective and has G (k) as its kernel.
As G0 (k) is normal in G(k), the set π0geom (G) inherits a group structure such that q geom is a
homomorphism. It is clear from the construction that Aut(k/k) acts on π0geom (G) through group
automorphisms. On the other hand, this action factors through Aut(k/k) → → Gal(ks /k) =: Γk ;
geom
hence we find that Γk acts on π0 (G) through group automorphisms.
We can view /0geom (G) as the constant group scheme associated to the abstract group
π0geom (G), and because Γk acts on π0geom (G) through group automorphisms, the étale scheme
/0 (G) over k inherits the structure of a k-group scheme. It is clear from the constructions that
q geom : Gks → /0geom (G) is a Γk -equivariant homomorphism of group schemes. It follows that
q: G → /0 (G) is a homomorphism of k-group schemes.
The conclusion of this discussion is that /0 (G) has a natural structure of an étale group
scheme over k, and that q: G → /0 (G) is a homomorphism. We refer to /0 (G) as the component
group scheme of G.
Another way to show that /0 (G), for G a k-group scheme, inherits the structure of a group
scheme is to use the fact that /0 (G ×k G) ∼ = /0 (G) ×k /0 (G); see Exercise 3.10. The group
law on /0 (G) is the map

/0 (m): /0 (G ×k G) ∼
= /0 (G) ×k /0 (G) −→ /0 (G)

induced by the group law m: G ×k G → G.

Exercises.
(3.1) Show that the following definition is equivalent to the one given in (3.7): If G is a group
scheme over a basis S then a subgroup scheme of G is a subscheme H ⊂ G such that (a) the
identity section e: S → G factors through H; (b) if j: H (→ G is the inclusion morphism then
the composition i ◦ j: H (→ G → G factors through H; (c) the composition m ◦ (j ×j): H ×S H →
G ×S G → G factors through H.
(3.2)
(i) Let G be a group scheme over a perfect field k. Prove that the reduced underlying scheme
Gred (→ G is a closed subgroup scheme. [Hint: you will need the fact that Gred ×k Gred is
again a reduced scheme; see EGA IV, § 4.6. This is where we need the assumption that k
is perfect.]

– 46 –
 (ii) Show, by means of an example, that Gred is in general not normal ! in G. p2 "
(iii) Let k be a field of characteristic p. Let a ∈ k,! and" set G := Spec k[x]/(x + axp ) . Show
that G is a subgroup scheme of Ga,k = Spec k[x] .
(iv) Assume that k is not perfect and that a ∈ k \ k p . Show that |G|, the topological space
underlying G, consists of p closed points, say |G| = {Q1 , Q2 , . . . , Qp }, where Q1 = e is
the origin. Show that G is reduced at the points Qi for i = 2, . . . , p but not geometrically
reduced. Finally show that the reduced underlying subscheme Gred (→ G is not a subgroup
scheme.
(3.3) Prove the relations (1) and (2) in (3.9). Also prove relation (6) in the proof of Theo-
rem (3.20).
! "
(3.4) Let G be a group scheme over a field k. Write TG,e = Ker G(k[ε]) → G(k) for the
tangent space of G at the identity element. Show that the map Te (m): TG,e × TG,e → TG,e
induced by the group law m: G ×k G → G on tangent spaces (the “derivative of m at e”) is
given by Te (m)(a, b) = a + b. Generalize this to group schemes over an arbitrary base.
(3.5) Let k be a field.
(i) If f : G1 → G2 is a homomorphism of k-group schemes, show that
! "
TKer(f ),e ∼
= Ker Te (f ): TG1 ,e → TG2 ,e .

(ii) If char(k) = p > 0, write G[F ] ⊂ G for the kernel of the relative Frobenius homomorphism
FG/k : G → G(p) . Show that TG[F ],e ∼
= TG,e .
(iii) If G is a finite k-group scheme and char(k) = p, show that G is étale over k if and only if
FG/k is an isomorphism. [Hint: in the “only if” direction, reduce to the case that k = k.]
(3.6) Let S = Spec(R) be an affine base scheme. Let G = Spec(A) be an affine S-group scheme
such that A is free of finite rank as an R-module. Choose an R-basis e1 , . . . , ed for A, and
-d
define elements aij ∈ A by m̃(ej ) = i=1 ei ⊗ aij . Let R[Tij , U ]/(det
! ·U" − 1) be the affine
algebra of GLd,R , where det ∈ k[Tij ] is the determinant of the matrix Tij . Show that there is
a well-defined homomorphism of R-algebras

ϕ: R[Tij , U ]/(det ·U − 1) −→ A

with Tij %→ aij . Show that the corresponding morphism G → GLd,R is a homomorphism and
gives an isomorphism of G with a closed subgroup scheme of GLd,R . [Hint: write Md,R for the
ring scheme over R of d × d matrices. First show that we get a morphism f : G → Md,R such
that f (g1 g2 ) = f (g1 )f (g2 ) for all g1 , g2 ∈ G. Next show that f (eG ) is the identity matrix, and
conclude that f factors through the open subscheme GLd,R ⊂ Md,R . Finally show that ϕ is
surjective. Use the relations between m̃, ẽ and ĩ.]
(3.7)! Let k be a field of characteristic
" p. Consider the group variety G := GLd,k . Let A =
Spec k[Tij , U ]/(det ·U − 1) be its affine algebra. Recall that we write [n]G : G → G for the
morphism given on points by g %→ g n .
(i) Let I ⊂ A be the augmentation ideal. Let [p]: A → A be the homomorphism of k-algebras
corresponding to [p]G . Show that [p](I) ⊆ I p .
(ii) Let H = Spec(B) be a finite k-group scheme. Let J ⊂ B be the augmentation ideal. Show
that [p](J) ⊆ J p . [Hint: use the previous exercise.] For an application of this result, see
Exercise (4.4).

– 47 –
(3.8) Let π: G → S be an affine S-group scheme. Set A := π∗ OG , so that G ∼
= Spec(A) as an
S-scheme. Let A := Hom OS (A, OS ). Show that with the definitions given in (3.21), AD is a
D

sheaf of co-commutative OS -Hopf algebras.

(3.9) Let k be a field, k ⊂ ks a separable algebraic closure, and write Γ := Gal(ks /k). Let X
be a scheme, locally of finite type over k, and let Y be an étale k-scheme. Note that Γ naturally
acts on the schemes Xks and Yks . If ϕ: Xks → Yks is a Γ-equivariant morphism of schemes
over ks , show that ϕ is defined over k, i.e., there is a (unique) morphism f : X → Y over k such
that fks = ϕ. [Hint: First reduce to the case that X is affine and that X and Y are connected.
Then work on rings.]
(3.10) Let X and Y be two schemes that are locally of finite type over a field k. Let qX : X →
/0 (X) and qY : Y → /0 (Y ) be the morphisms as in Prop. (3.27). By the universal property of
/0 (X ×k Y ), there is a unique morphism

ρ: /0 (X ×k Y ) → /0 (X) ×k /0 (Y )

such that ρ ◦ q(X×Y ) = (qX ◦ prX , qY ◦ prY ). Show that ρ is an isomorphism. In particular,
conclude that if k ⊂ K is a field extension then /0 (XK ) is naturally isomorphic to /0 (X)K .
[Hint: Reduce to the case k = ks . Use that if C and D are connected schemes over ks then C ×ks
D is again connected. See EGA IV, Cor. (4.5.8), taking into account loc. cit., Prop. (4.5.21).]

Notes. Proposition (3.17) is taken from SGA 3, Exp. VIA . The example following Proposition (3.12) is taken
from ibid., Exp. VIB , §5. A different proof of Prop. (3.27) can be found in the book of Demazure and Gabriel [1].

– 48 –
Chapter IV. Quotients by group schemes.

When we work with group schemes the question naturally arises if constructions from group
theory can also be carried out in the context of group schemes. For instance, we have seen
that if f : G → G! is a homomorphism then we can form the kernel group scheme, Ker(f ). In
this example the geometry and the group theory go hand in hand: there is an obvious scheme-
theoretic candidate for the kernel, namely the inverse image of the identity section of G! , and
this candidate also represents the kernel as a functor.
The present chapter is devoted to the formation of quotients, which is more delicate. (Nog
aanvullen)
The reader who wants to go on as quickly as possible with the general theory of abelian
varieties, may skip most of this chapter. The only results that are directly relevant for the
next chapters are the formation of quotients modulo finite group schemes, Thm. (4.16), Exam-
ple (4.40), and the material in § 4.

§ 1. Categorical quotients.

(4.1) Definition. (i) Let G be a group scheme over a basis S. A (left) action of G on an
S-scheme X is given by a morphism ρ: G ×S X → X such that the composition

∼ G e ×id ρ
X −→ S ×S X −−−−−X→ G ×S X −→ X

is the identity on X, and such that the diagram

idG ×ρ
G ×S G ×S X −−−−→ G ×S X
 

m×idX "
ρ
" (1)
ρ
G ×S X −−−−→ X

is commutative. In other words: for every S-scheme T , the morphism ρ induces a left action of
the group G(T ) on the set X(T ). We usually denote this action on points by (g, x) $→ g · x.
(ii) Given an action ρ as in (i), we define the “graph morphism”

Ψ = Ψρ := (ρ, pr2 ): G ×S X −→ X ×S X ;

on points this is given by (g, x) $→ (g · x, x). The action ρ is said to be free, or set-theoretically
free if Ψ is a monomorphism of schemes, and is said to be strictly free, or scheme-theoretically
free, if Ψ is an immersion.
(iii) If T is an S-scheme and x ∈ X(T ) then the stabilizer of x, notation Gx , is the subgroup
scheme of GT that represents the functor T ! $→ {g ∈ G(T ! ) | g · x = x} on T -schemes T ! . (See
also (4.2), (iii) below.)

QuoGrSch, 15 september, 2011 (812)

– 49 –
(4.2) Remarks. (i) In some literature the same terminology is used in a slightly different
meaning (cf. GIT, for example).
(ii) The condition that an action ρ is free means precisely that for all T and all x ∈ X(T )
the stabilizer Gx is trivial.
(iii) With notations as in the definition, we have a diagram with cartesian squares
id ×x
Gx "−→ GT = G × S T −−G−−→ G ×S X
  
 ax 
" " "Ψ
(x,idT ) id ×x
T −−−−−→ XT = X × S T −−X−−→ X ×S X ,

where the morphism ax is given by ax = (ρ ◦ (idG × x), pr2 ); on points: ax (g) = g · x. That the
functor T ! $→ {g ∈ G(T ! ) | g · x = x} is indeed representable by a subgroup scheme Gx ⊂ GT is
seen from this diagram, arguing as in (3.13).

(4.3) Examples. If G is a group scheme over S and H ⊂ G is a subgroup scheme then the
group law gives an action of H on G. The graph morphism Ψ: H ×S G → G ×S G is the
restriction to H ×S G of the universal right translation τ : G ×S G → G ×S G. Since τ is an
isomorphism, the action is strictly free.
More generally, if f : G → G! is a homomorphism of group schemes then we get a natural
action of G on G! , given on points by (g, g ! ) $→ f (g) · g ! . The action is free if and only if Ker(f )
is trivial, but if this holds the action need not be strictly free. For instance, with S = Spec(Q)
as a base scheme, take G = ZS to be the constant group scheme defined by the (abstract)
group Z, and take G! = Ga,S . We have a natural homomorphism f : ZS → Ga,S which, for
Q-schemes T , is given on points by the natural inclusion Z "→ Γ(T, OT ). This homomorphism f
is injective, hence it gives a free action of ZS on Ga,S . The graph morphism can be described
as the morphism
#
Ψ: A1 −→ A2
n∈Z

that maps the nth copy of A1 to the line L ⊂ A2 given by x − y = n. But this Ψ is not an
immersion (the image is not a subscheme of A2 ), so the action is not strictly free.

(4.4) The central issue of this chapter is the following question. Given a group scheme G acting
on a scheme X, does there exist a good notion of a quotient space G\X? As particular instances
of this question we have: given a homomorphism of group schemes f : G → G! , can we form
a cokernel of f ?, and if N ⊂ G is a normal subgroup scheme, can we define a quotient group
scheme G/N ?
Let us first look at an elementary example. Take an integer N ! 2, and consider the
endomorphism f : Gm → Gm over S = Spec(Z) given on points by q $→ q N . The kernel of f
is µN , by definition of the latter. As a morphism of schemes, f is faithfully flat, and if k
is any algebraically closed field then f is surjective on k-valued points. Therefore we would
expect that the cokernel of f is trivial, i.e., Coker(f ) = S. But clearly, the “cokernel functor”
C: T $→ Gm (T )/f (Gm (T )) is non-trivial. E.g., C(Q) is an infinite group. Moreover, from the
fact that C(Q) '= {1} but C(Q) = {1} it follows that C is not representable by a scheme. So,
in contrast with (3.13) where we defined kernels, the geometric and the functorial point of view
do not give to the same notion of a cokernel.

– 50 –
 The first notion of a quotient that we shall define is that of a categorical quotient. Though
we are mainly interested in working with schemes, it is useful to extend the discussion to a more
general setting.

(4.5) Definition. Let C be a category with finite products. Let G be a group object in C. Let
X be an object of C. Throughout, we simply write X(T ) for hX (T ) = HomC (T, X).
(i) A (left) action of G on X is a morphism ρ: G × X → X that induces, for every object T ,
a (left) action of the group G(T ) on the set X(T ).
(ii) Let an action of G on X be given. A morphism q: X → Y in C is said to be G-invariant
if q ◦ ρ = q ◦ prX : G × X → Y . By the Yoneda lemma this is equivalent to the requirement that
for every T ∈ C, if x1 , x2 ∈ X(T ) are two points in the same G(T )-orbit then q(x1 ) = q(x2 )
in Y (T ).
(iii) Let f, g: W ⇒ X be two morphisms in C. We say that a morphism h: X → Y is a
difference cokernel of the pair (f, g) if h ◦ f = h ◦ g and if h is universal for this property; by
this we mean that for any other morphism h! : X → Y ! with h! ◦ f = h! ◦ g there is a unique
α: Y → Y ! such that h! = α ◦ h.
(iv) Let ρ: G × X → X be a left action. A morphism q: X → Y is called a categorical
quotient of X by G if q is a difference cokernel for the pair (ρ, prX ): G×X ⇒ X. In other words, q
is a categorical quotient if q is G-invariant and if every G-invariant morphism q ! : X → Y ! factors
as q ! = α ◦ q for a unique α: Y → Y ! . The morphism q: X → Y is called a universal categorical
quotient of X by G if for every object S of C the morphism qS : XS → YS is a categorical
quotient of XS by GS in the category C/S .
In practice the morphism q is often not mentioned, and we simply say that an object Y is
the categorical quotient of X by G. Note that if a categorical quotient q: X → Y exists then it
is unique up to unique isomorphism.

(4.6) Examples. As in (4.4), let S = Spec(Z) and let G = Gm,S act on X = Gm,S by
ρ(g, x) = g N · x. If k = k then X(k) consists of a single orbit under G(k); this readily implies
that X → S is a categorical quotient of X by G. In fact, if we work a little harder we find that
X → S is even a universal categorical quotient; see Exercise (4.1).
As a second example, let k = k and consider the action of G = Gm,k on X = A1k given
on points by ρ(g, x) = g · x. There are two orbits in X(k), one given by the origin 0 ∈ X(k),
the other consisting of all points x '= 0. Suppose we have a G-invariant morphism q: X → Y
for some k-scheme Y . It maps X(k) \ {0} to a point y ∈ Y (k). Because X(k) \ {0} is Zariski
dense in X we find that q is the constant map with value y. This proves that the structural
morphism X → Spec(k) is a categorical quotient of X by G. We conclude that it is not possible
to construct a quotient scheme Y such that the two orbits {0} and A1 \ {0} are mapped to
different points of Y .

(4.7) Remark. Let G be an S-group scheme acting on an S-scheme X. Suppose there exists
a categorical quotient q: X → Y in Sch/S . To study q we can take Y to be our base scheme.
More precisely, GY := G ×S Y acts on X over Y and q is also a categorical quotient of X by GY
in the category Sch/Y . Taking Y to be the base scheme does not affect the (strict) freeness of
the action. To see this, note that the graph morphism Ψ: G ×S X → X ×S X factors through
the subscheme X ×Y X "→ X ×S X and that the resulting morphism

GY × Y X = G × S X → X × Y X

– 51 –
is none other than the graph morphism of GY acting on X over Y . Hence the action of G on X
over S is (strictly) free if and only if the action of GY on X over Y is (strictly) free.

§ 2. Geometric quotients, and quotients by finite group schemes.

We first give, in its simplest form, a result about the existence of quotients under finite groups.
This result will be generalized in (4.16) below. Here we consider an action of an abstract group Γ
on a scheme X; this means that for every element γ ∈ Γ we have a morphism ρ(γ): X → X,
satisfying the usual axioms for a group action. Such an action is the same as an action of the
constant group scheme Γ on X; hence we are in a special case of the situation considered in (4.1).

(4.8) Proposition. Let Γ be a finite (abstract) group acting on an affine scheme X = Spec(A).
Let B := AΓ ⊆ A be the subring of Γ-invariant elements, and set Y := Spec(B).
∼
(i) The natural morphism q: X → Y induces a homeomorphism Γ\|X| −→ |Y |, i.e., it
identifies the topological space |Y | with the quotient of |X| under the acion of Γ.
∼
(ii) The map q " : OY → q∗ OX induces an isomorphism OY −→ (q∗ OX )Γ , where the latter
denotes the sheaf of Γ-invariant sections of q∗ OX .
(iii) The ring A is integral over B; the morphism q: X → Y is quasi-finite, closed and
surjective.
Proof. Write Γ = {γ1 , . . . , γr }. Define the map N : A → AΓ = B by

N (a) = γ1 (a) · · · γr (a) .

If p and p! are prime ideals of A which lie in the same Γ-orbit then p ∩ AΓ = p! ∩ AΓ . Conversely,
if p ∩ AΓ = p! ∩ AΓ then N (x) ∈ p! for every x ∈ p, so p ⊂ γ1 (p! ) ∪ · · · ∪ γr (p! ). This implies (see
Atiyah-Macdonald [1], Prop. 1.11) that p ⊆ γi (p! ) for some i, and by symmetry we conclude
∼
that p and p! lie in the same Γ-orbit. Hence Γ\|X| −→ |Y | as sets, and q is quasi-finite.
$ %$ % $ %
For a ∈ A, let χa (T ) := T − γ1 (a) T − γ2 (a) · · · T − γr (a) ∈ A[T ]. Then it is clear that
χa (T ) is a monic polynomial in B[T ] and that χa (a) = 0. This shows that A is integral over B.
That the map q is closed and surjective then follows from Atiyah-Macdonald [1], Thm. 5.10; see
also (4.21) below.
∼
Finally we remark that for every f ∈ AΓ we have a natural isomorphism (AΓ )f −→ (Af )Γ .
As the special open subsets D(f ) := Y \ Z(f ) form a basis for the topology on Y , property (ii)
follows. #

(4.9) Remarks. (i) The morphism q: X → Y need not be finite. It may happen that A
is noetherian but that B := AΓ is not, and that A is not finitely generated as a B-module.
Examples of this kind can be found in Nagata [1], ??. However, if either the action on Γ on X is
free, or X is of finite type over a locally noetherian base scheme S and Γ acts by automorphisms
of X over S, then q is a finite morphism. See (4.16) below.
(ii) It is not hard to show that q: X → Y is a categorical quotient of X by G. (See
also Proposition (4.13) below.) More generally, if X → S is a morphism such that Γ acts
by automorphisms of X over S then also Y has a natural structure of an S-scheme, and q is
a categorical quotient in Sch/S . In general, q is not a universal
$ % categorical quotient. As an
example, let k be a field of characteristic p, take S = Spec k[ε] and X = A1S = Spec(A), with

– 52 –
A = k[x, ε]. We let the group Γ := Z/pZ act on X (over S); on rings we give the action of
n mod p by x $→ x + nε and ε $→ ε. The ring AΓ of invariants is generated as a k-algebra by ε,
xε, . . . , xp−1 ε and xp . But on the special fibre A1k the action is trivial. As

AΓ ⊗k[ε] k = k[ε, xε, . . . , xp−1 ε, xp ] ⊗k[ε] k = k[xp ]

is a proper subring of $ %Γ
A ⊗k[ε] k = k[x] ,
we see that Y := Spec(AΓ ) is not a universal categorical quotient of X in Sch/S .

(4.10) Suppose given an action of a group scheme G on a scheme X, over some basis S, say.
We should like to decide if there exists a categorical quotient of X by G in Sch/S , and if yes
then we should like to construct this quotient. Properties (a) and (b) in the above proposition
point to a general construction. Namely, if |X| is the topological space underlying X then we
could try to form a quotient of |X| modulo the action of G and equip this space with the sheaf
of G-invariant functions on X.
Another way to phrase this is the following. The category of schemes is a full subcategory of
the category LRS of locally ringed spaces, which in turn is a subcategory (not full) of the category
RS of all ringed spaces. If G is an S-group scheme acting on an S-scheme X then we shall show
that there exists a categorical quotient (G\X)rs in the category RS/S . It is constructed exactly
as just described: form the quotient “G\|X|” and equip this with the sheaf “(q∗ OX )G ”, where
q: |X| → G\|X| is the natural map. Then the question is whether (G\X)rs is a scheme and, if
so, if this scheme is a “good” scheme-theoretic quotient of X modulo G.
Before we give more details, let us note that in general (G\X)rs cannot be viewed as
a categorical quotient in the sense of Definition (4.5). Namely, because Sch/S is not a full
subcategory of RS/S , products in the two categories
$ may
% be different. Hence if G is an S-group
scheme then it is not clear if the ringed space |G|, OG inherits the structure of a group object
in RS/S . The assertion that (G\X)rs is a quotient of X by G will therefore be interpret as saying
that the morphism q is a difference cokernel of the pair of morphisms (ρ, prX ): G ×S X ⇒ X
in RS/S .

(4.11) Lemma. Let ρ: G ×S X → X be an action of an S-group scheme G on an S-scheme X.

Consider the continuous maps

|prX |: |G ×S X| −→ |X| and |ρ|: |G ×S X| −→ |X| .

Given P , Q ∈ |X|, write P ∼ Q if there exists a point R ∈ |G ×S X| with |prX |(R) = P and
|ρ|(R) = Q. Then ∼ is an equivalence relation on |X|.
Proof. See Exercise (4.2).
We refer to the equivalence classes under ∼ as the G-equivalence classes in |X|.

(4.12) Definition. Let ρ: G×S X → X be an action of an S-group scheme G on an S-scheme X.

Let |X|/∼ be the set of G-equivalence classes in |X|, equipped with the quotient topology. Write
q: |X| → |X|/∼ for the canonical map. Let U = q −1 (V ) for some open subset V ⊂ |X|/∼. If
f ∈ q∗ OX (V ) = OX (U ) then we can form the elements pr"X (f ) and ρ" (f ) in OG×S X (G ×S U ).
We say that f is G-invariant if pr"X (f ) = ρ" (f ). The G-invariant functions f form a subsheaf of
rings (q∗ OX )G ⊂ q∗ OX .

– 53 –
 We define $ %
(G\X)rs := |X|/∼, (q∗ OX )G ,

and write q: X → (G\X)rs for the natural morphism of ringed spaces.

If (G\X)rs is a scheme and q is a morphism of schemes then we say that it is a geometric
quotient of X by G. If moreover for every S-scheme T we have that (G\X)rs ×S T ∼
= (GT \XT )rs
then we say that (G\X)rs is a universal geometric quotient.
The phrase “if a geometric quotient of X by G exists” is used as a synonym for “if (G\X)rs
is a scheme and q: X → (G\X)rs is a morphism of schemes”.
The stalks of the sheaf (q∗ OX )G may not be local rings; for an example see ??. This is
the reason why we work in the category of ringed spaces rather than the category of locally
ringed spaces. Further we note that the formation of (G\X)rs does not, in general, commute
with base change; see (ii) of (4.9). However, if U ⊂ S is a Zariski open subset then (GU \XU )rs
is canonically isomorphic to the restriction of (G\X)rs to U .

(4.13) Proposition. In the situation of (4.12), q: X → (G\X)rs is a difference cokernel of the

pair of morphisms (ρ, prX ): G ×S X ⇒ X in the category RS/S . By consequence, if a geometric
quotient of X by G exists then it is also a categorical quotient in Sch/S .
Proof. The first assertion is an immediate consequence of how we constructed (G\X)rs . If
(G\X)rs is a geometric quotient then it is also a difference cokernel of (ρ, prX ) in the category
Sch/S because the latter is a subcategory of RS/S . This gives the second assertion. #

(4.14) Example. Let k be a field, and consider the k-scheme M2,k (=A4k ) of 2 × 2-matrices
over k. The linear algebraic group GL2,k acts on M2,k by conjugation: if g ∈ GL2 (T ) for some
k-scheme T then g acts on M2 (T ) by A $→ g · A · g −1 . Write ρ: GL2,k × M2,k → M2,k for the
morphism giving this GL2,k -action.
The trace and determinant give morphisms of schemes trace: M2,k → A1k and det: M2,k →
1
Ak . Now consider the morphism

p = (trace, det): M2,k → A2k .

Clearly p is a GL2 -invariant morphism, i.e., p ◦ pr2 : GL2,k × M2,k → M2,k → A2k is the same as
p ◦ ρ. It can be shown that the pair (A2k , p) is a (universal) categorical quotient of M2,k by GL2,k ,
see GIT, Chap. 1, § 2 and Appendix 1C.
On the other hand, it is quite easy to see that A2k is not a geometric quotient. Indeed, if
this were the case then on underlying topological spaces the map p should identify A2k as the set
of GL2,k -orbits in M2,k . But the trace and the determinant are not able to distinguish a matrix
& '
λ 1
Jλ :=
0 λ

from its semi-simple part & '

λ 0
.
0 λ

To give another explanation of what is going on, let us look at k-valued points, where k is
an algebraic closure of k. The theory of Jordan canonical forms tells us that the GL2 (k)-orbits in
M2 (k) are represented by the diagonal matrices diag(λ1 , λ2 ) together with the matrices Jλ . For

– 54 –
τ , δ ∈ k, write N (τ, δ) ⊂ M2,k for the 2-dimensional subvariety given by the conditions trace = τ
and det = δ. By direct computation one readily verifies that (i) the orbit of a diagonal matrix
A = diag(λ, λ) is the single closed point A; (ii) the orbit of a diagonal matrix diag(λ1 , λ2 ) with
λ1 '= λ2 equals N (λ1 + λ2 , λ1 λ2 ); (iii) the orbit of a matrix Jλ equals N (2λ, λ2 ) \ {diag(λ, λ)};
in particular, this orbit is not closed in M2,k .
From the observation that there are non-closed orbits in M2,k , it immediately follows that
there does not exist a geometric quotient. (Indeed, the orbits in M2 (k) would be the pre-
images of the k-valued points of the geometric quotient. Cf. the second example in (4.6).) This
suggests that the points in a subvariety of the form N (2λ, λ2 ) ⊂ M2,k are the “bad” points for
the given action of GL2 . Indeed, it can be shown that on the open complement U ⊂ M2,k given
by the condition 4 det −trace2 '= 0, the map p = (trace, det): U → D(4y − x2 ) ⊂ A2 (taking
coordinates x, y on A2 and writing D(f ) for the locus where a function f does not vanish) makes
D(4y − x2 ) ⊂ A2k a geometric quotient of U .
The notion of a geometric quotient plays a central role in geometric invariant theory. There,
as in the above simple example, one studies which points, or which orbits under a given group
action are so “unstable” that they obstruct the formation of a good quotient. (Which are the
“bad” points may depend on further data, such as the choice of an ample line bundle on the
scheme in question.) We refer the reader to the book GIT.
We now turn to the promised generalization of Proposition (4.8). First we need a lemma.

(4.15) Lemma. Let ϕ: A → C be a homomorphism of commutative rings that makes C a

projective A-module of rank r > 0. Let NormC/A : C → A be the norm map. Let ψ: Spec(C) →
Spec(A) be the morphism of affine schemes given by ϕ. If Z ⊂ Spec(C) is the zero locus of
f ∈ C then ψ(Z) ⊂ Spec(A) is the zero locus of NormC/A (f ).
Proof. The assumptions imply that ϕ is injective. As C is integral over A the map ψ is surjective;
see also (4.21) below. Let p ∈ Spec(A); write ψ −1 {p} = {q1 , . . . , qn }. By definition, N :=
NormC/A (f ) is the determinant of the endomorphism λf : c $→ f c of C as a module over A.
Write W ⊂ Spec(A) for the zero locus of N . Write ap for the image of an element a ∈ A
in Ap ; similar notation for elements of C. Then we have

/ W ⇐⇒ Np ∈ A∗p
p∈
⇐⇒ λf,p : Cp → Cp is an isomorphism
⇐⇒ fp ∈ Cp∗
⇐⇒ f ∈
/ qi for all i = 1, . . . , n
⇐⇒ qi ∈
/ Z for all i = 1, . . . , n ,

which proves the lemma. #

(4.16) Theorem. (Quotients by finite group schemes.) Let G be a finite locally free S-group
scheme acting on an S-scheme X. Assume that for every closed point P ∈ |X| the G-equivalence
class of P is contained in an affine open subset.
(i) The quotient Y := (G\X)rs is an S-scheme, which therefore is a geometric quotient of X
by G. The canonical morphism q: X → Y is quasi-finite, integral, closed and surjective. If S is
locally noetherian and X is of finite type over S then q is a finite morphism and Y is of finite
type over S, too.

– 55 –
 (ii) The formation of the quotient Y is compatible with flat base change (terminology: Y
is a uniform quotient). In other words, let h: S ! → S be a flat morphism. Let a prime ! denote
a base change via h, e.g., X ! := X ×S S ! . Then Y ! ∼= (G! \X ! )rs .
(iii) If G acts freely then q: X → Y is finite locally free and the morphism

G ×S X −→ X ×Y X

induced by Ψ = (ρ, prX ) is an isomorphism. Moreover, Y is in this case a universal geometric

quotient: for any morphism h: S ! → S, indicating base change via h by a prime ! , we have
Y! ∼
= (G! \X ! )rs .

(4.17) Remarks. (i) The condition that every G-equivalence class is contained in an affine open
subset is satisfied if X is quasi-projective over S. Indeed, given a ring R, a positive integer N ,
and a finite set V of closed points of PN N
R , we can find an affine open subscheme U ⊂ PR such
that V ⊂ U .
(ii) In the situation of the theorem we find that a free action is automatically strictly free.
Indeed, by (iii) the graph morphism Ψ gives an isomorphism of G ×S X with the subscheme
X ×Y X ⊂ X ×S X; hence Ψ is an immersion.
We break up the proof of the theorem into a couple of steps, (4.18)–(4.26).

(4.18) Reduction to the case that S is affine. Suppose S = ∪α Uα is a covering of S by Zariski

open subsets. As remarked earlier, the restriction of (G\X)rs to U = Uα is naturally isomorphic
to (GU \XU )rs . If we can prove the theorem over each of the open sets Uα then the result as stated
easily follows by gluing. In the rest of the proof we may therefore assume that S = Spec(Q) is
affine and that the affine algebra R of G is free of some rank r as a Q-module.

(4.19) Reduction to the case that X is affine. If P ∈ |X|, let us write G(P ) for its G-equivalence
class; note that this is a finite set. Note further that
$ % $ −1 %
G(P ) = ρ pr−1
X {P } = prX ρ {P } ,

by definition of G-equivalence. (Strictly speaking we should write |ρ| and |prX |.)
Say that a subset V ⊂ |X| is G-stable if it contains G(P ) whenever it contains P . If V is
open then there$ is a maximal
% open subset V ! ⊆ V which is G-stable. Namely, if Z := |X| \ V
then Z := prX ρ {Z} is closed (since prX : G ×S X → X is proper), and V ! := |X| \ Z ! has
! −1

the required property.

We claim that X can be covered by G-stable affine open subsets. It suffices to show that
every closed point P ∈ X has a G-stable affine open neighbourhood. By assumption there exists
an affine open V ⊂ X with G(P ) ⊂ V . Then also G(P ) ⊂ V ! . As G(P ) is finite there exists
an f ∈ Γ(V, OV ) such that, writing D(f ) ⊂ V for the open subset where f does not vanish,
G(P ) ⊂ D(f ) ⊆ V ! . In total this gives

G(P ) ⊂ D(f )! ⊆ D(f ) ⊆ V ! ⊆ V .

Our claim is proven if we can show that D(f )! is affine. Write f ! for the image of f
in Γ(V ! , OV " ), so that Z := V ! \ D(f ) is the zero locus of f ! . As V ! is G-stable we have
ρ−1 (V ! ) = G ×S V ! , which gives an element ρ" (f ! ) ∈ Γ(G ×S V ! , OG×S V " ). The zero locus of
ρ" (f ! ) is of course just ρ−1 (Z) ⊂ G ×S V ! . As G is finite locally free, the morphism prX makes

– 56 –
 $ % $ %
Γ G ×S V ! , OG×S V " into a projective module of finite rank over Γ V ! , OV " . This gives us a
norm map $ % $ %
Norm: Γ G ×S V ! , OG×S V " −→ Γ V ! , OV " .
$ %
Let F := Norm ρ" (f ! ) . By Lemma (4.15), the zero locus of F is the image of ρ−1 (Z) under the
projection to V !$. But the complement
% of this locus in V ! is precisely D(f )! . Hence if F ! is the
image of F in Γ D(f ), OD(f ) then D(f )! is the open subset of D(f ) where F ! does not vanish.
As this subset is affine open, our claim is proven.
Except for the last assertion of (i), the proof of the theorem now reduces to the case that
X is affine. Namely, by the previous we can cover X by G-stable affine open subsets, and if the
theorem is true for each of these then by gluing we obtain the result for X. The last assertion
of (i) will be dealt with in (4.23).

(4.20) From now on we assume that X = Spec(A) → S = Spec(Q). Further we assume that
G = Spec(R) for some Q-Hopf algebra R which is free of rank r as a module over Q. Much
of what we are going to do is a direct generalization of the arguments in (4.8); that proof may
therefore serve as a guide for the arguments to follow.
The action of G on X is given by a Q-algebra homomorphism σ: A → R ⊗Q A. Write
j: A → R ⊗Q A for the map given by a $→ 1 ⊗ a. (In other words, we write σ for ρ" and j
for pr"X .) Define a subring B := AG ⊂ A of G-invariants by

B := {a ∈ A | σ(a) = j(a)} .

We are going to prove that Y := Spec(B) is the geometric quotient of X under the given action
of G.
As a first step, let us show that A is integral over B. For a ∈ A, multiplication by σ(a) is
an endomorphism of R ⊗Q A, and we can form its characteristic polynomial

χ(t) = tr + cr−1 tr−1 + · · · + c1 t + c0 ∈ A[t] .

We have cartesian squares

m̃⊗id id ⊗σ
R ⊗Q A −−−−−A→ R ⊗Q R ⊗Q A R ⊗Q A −−R−−→ R ⊗Q R ⊗Q A
( ( ( (

j
j  j
 2,3 and j  2,3 (2)
j σ
A −−−−−→ R ⊗Q A A −−−−→ R ⊗Q A ,
where the map j2,3 is given by r ⊗ a $→ 1 ⊗ r ⊗ a. We view R $ ⊗Q%R ⊗Q A as a module over
R ⊗Q A via j2,3 . It follows from the left-hand diagram that j χ(t) , the polynomial $ obtained
%
from χ(t) by applying j to its coefficients,
$ % is the characteristic polynomial of m̃⊗id A $σ(a) %. The
right-hand diagram tells us that σ χ(t) is the characteristic polynomial of idR ⊗ $ σ σ(a)
% . But
the
$ commutativity
% $ of diagram
% $ (1)%in Definition (4.1) gives the identity m̃ ⊗ idA σ(a) = idR ⊗
σ σ(a) . Hence j χ(t) = σ χ(t) , which means that χ(t) is a polynomial with coefficients ci
in the ring B of G-invariants.
The Cayley-Hamilton theorem tells us that

σ(a)r + j(cr−1 )σ(a)r−1 + · · · + j(c1 )σ(a) + j(c0 ) = 0 .

As j(ci ) = σ(ci ) for all i we can rewrite this as

$ %
σ χ(a) = σ(a)r + σ(cr−1 )σ(a)r−1 + · · · + σ(c1 )σ(a) + σ(c0 ) = 0 . (3)

– 57 –
But σ is an injective map, because we have the relation (ẽ ⊗ idA ) ◦ σ = idA , which translates the
fact that the identity element of G acts as the identity on X. Hence (3) implies that χ(a) = 0.
This proves that A is integral over B.

(4.21) The fact that A is integral over B has the following consequences.
(i) If p1 ⊆ p2 are prime ideals of A with p1 ∩ B = p2 ∩ B then p1 = p2 . Geometrically this
means that all fibres of Spec(A) → Spec(B) have dimension 0.
(ii) The natural map q: X = Spec(A) → Y = Spec(B) is surjective.
(iii) The map q is closed, i.e., if C ⊂ X is closed then q(C) ⊂ Y is closed too.
Properties (i) and (ii) can be found in many textbooks on commutative algebra, see for instance
Atiyah-Macdonald [1], Cor. 5.9 and Thm. 5.10. For (iii), suppose C ⊂ X is the closed subset
defined by an ideal a ⊂ A. We may identify C with Spec(A/a). The composite map C =
Spec(A/a) "→ Spec(A) → Spec(B) factors through the closed subset Spec(B/b) ⊂ Spec(B),
where b = a ∩ B. Note that A/a is again integral over its subring B/b. Applying (ii) with A
and B replaced by A/a and B/b, we find that C = Spec(A/a) → Spec(B/b) is surjective. Hence
q(C) is the closed subset of B defined by b.
Define a map N : A → B by
$ %
N (a) = NormR⊗Q A/A σ(a) .

Note that N (a) = (−1)n c0 , where c0 is the constant coefficient of the characteristic polyno-
mial χ(t) considered in (4.20); hence N (a) is indeed an element of B. The relation χ(a) = 0
gives $ %
N (a) = (−1)n+1 · a · an−1 + cn−1 an−2 + · · · + c1 .
In particular, if a ∈ a for some ideal a ⊂ A then N (a) ∈ a ∩ B.

(4.22) Recall that Y := Spec(B). We are going to prove that Y = (G\X)rs . Note that the
natural map |X| → |Y | is surjective, by (ii) in (4.21).
By definition, two prime ideals p and p! of A are in the same G-equivalence class if there
exists a prime ideal Q of R ⊗Q A with σ −1 (Q) = p and j −1 (Q) = p! . If such a prime ideal Q
exists then it is immediate that p ∩ B = p! ∩ B, so G-equivalent points of X are mapped to the
same point of Y .
Conversely, suppose p ∩ B = p! ∩ B. There are finitely many prime ideals Q1 , . . . , Qn of
R⊗Q A with the property that j −1 (Qi ) = p! . (The morphism prX : G×S X → X is finite because
G is finite.) Set qi = σ −1 (Qi ). Note that qi ∩ B = p ∩ B. Our goal is to prove that p = qi for
some i. By property (i) above it suffices to show that p ⊆ qi for some i. Suppose this is not the
case. Then there exists an element a ∈ p that is not contained in q1 ∪ · · · ∪ qn . (Use Atiyah-
Macdonald [1], Prop. 1.11, and cf. the proof of Prop. (4.8) above.) Lemma (4.15), applied with
f = σ(a) ∈ R ⊗Q A, tells us that the prime ideals of A containing N (a) are all of the form j −1 (r)
with r a prime ideal of R ⊗Q A that contains σ(a). But a ∈ p, hence N (a) ∈ p ∩ B = p! ∩ B.
Hence one of the prime ideals Qi contains σ(a), contradicting our choice of a.
We have now proven that the map X → Y identifies |Y | with the set |X|/∼ of G-equivalence
classes in X. Further, by (iii) in (4.21) the quotient map |X| → |Y | is closed, so the topology
on |Y | is the quotient topology. If V = DY (f ) ⊂ Y is the fundamental open subset given by
f ∈ B then q −1 (V ) = DX (f ), and we find

∼ $ %G $ %G
OY (V ) = Bf = (AG )f −→ (Af )G = OX (q −1 (V )) = (q∗ OX )(V ) .

– 58 –
As the fundamental open subsets form a basis for the topology on Y , it follows that q " : OY →
∼
q∗ OX induces an isomorphism OY −→ (q∗ OX )G .

(4.23) Let us now prove the last assertion of part (i) of the theorem. As before we may assume
that S = Spec(Q) is affine. Let q: X → Y := (G\X)rs be the quotient morphism, which we
have already shown to exist. Let U = Spec(A) be a G-stable affine open subset $ of% X, and let
B = AG . By construction, q(U ) = Spec(B) is an open subset of Y , and q −1 q(U ) = U . If X
is locally of finite type over S then A is a finitely generated Q-algebra, a fortiori also of finite
type as a B-algebra. But A is also integral over B. It follows that A is finitely generated as a
B-module (see e.g. Atiyah-Macdonald [1], Cor. 5.2). Hence q is a finite morphism.
If S is locally noetherian then we may assume, arguing as in (4.18), that Q is a noetherian
ring. Choose generators a1 , . . . , an for A as a Q-algebra. We have seen that for each i we can
find a monic polynomial fi ∈ B[T ] with fi (ai ) = 0. Let B ! ⊂ B be the Q-subalgebra generated
by the coefficients of the polynomials fi . Then A is integral over B ! , and by the same argument
as above it follows that A is finitely generated as a B ! -module. Because B ! is finitely generated
over Q it is a noetherian ring. But then B ⊂ A is also finitely generated as a B ! -module, hence
finitely generated as a Q-algebra. This shows that Y is locally of finite type over S.
So far we have used only that X is locally of finite type over S. Assume, in addition, that
the morphism f : X → S is quasi-compact. Let g: Y → S be the structural morphism of Y . It
remains to be shown that g is$ quasi-compact.
% But this is clear, for if V ⊂ S is a quasi-compact
open subset then g (V ) = q f (V ) , which is quasi-compact because f −1 (V ) is.
−1 −1

(4.24) Proof of (ii) of the theorem. Let S ! → S be a flat morphism. We want to show that
Y ! := Y ×S S ! is a geometric quotient of X ! by G! . Arguing as in (4.18) one reduces to the case
that S ! → S is given by a flat homomorphism of rings Q → Q! . Note that every G! -equivalence
class in X ! is again contained in an affine open subset. As in (4.19) one further reduces to
the case that X, X ! , Y and Y ! are all affine. With notations as above we have Y = Spec(B),
where B = Ker(j − σ). We want to show, writing a prime ! for extension of scalars to Q! , that
B ⊗Q Q! = Ker(j ! − σ ! : A! ⊗Q" R! → A! ⊗Q" A! ). But this is obvious from the assumption that
Q → Q! is flat.

(4.25) We now turn to part (iii) of the theorem. As before, everything reduces to the situation
where S, G, X and Y are all affine, with algebras Q, R, A and B = AG , respectively, and that
R is free of rank r as a module over Q. We view R ⊗Q A as an A-module via j. Let

ϕ: A ⊗B A → R ⊗Q A

be the homomorphism given by ϕ(a1 ⊗ a2 ) = σ(a1 ) · j(a2 ) = σ(a1 ) · (1 ⊗ a2 ).

Assume that G acts freely on X. This means that the morphism Ψ: G ×S X → X ×S X
is a monomorphism in the category of schemes. The corresponding map on rings is given by
Ψ" = ϕ ◦ q, where q: A ⊗Q A →→ A ⊗B A is the natural map. Since a morphism of affine schemes
is a monomorphism if and only if the corresponding map on rings is surjective, it follows that ϕ
is surjective.
Let q be a prime ideal of B and write Aq = (B − q)−1 A ∼ = A ⊗B Bq . Note that Aq is a
semi-local ring, because X → Y is quasi-finite. Let r ⊂ Aq be its radical. We claim that Aq
is free of rank r = rank(G) as a module over Bq . If this holds for all q then A is a projective
B-module of rank r; use Bourbaki [2], Chap. II, § 5, Thm. 2. Furthermore, ϕ is then a surjective
map between projective A-modules of the same rank and is therefore an isomorphism.

– 59 –
 We first prove that Aq is Bq -free of rank r in the case where the residue field k of Bq is
infinite. Consider the Bq -submodule

N := {σ(a) | a ∈ Aq } ⊂ M := R ⊗Q Aq .

Because ϕq : Aq ⊗Bq Aq → R ⊗Q Aq is surjective, N spans M as an Aq -module. Therefore

N/rN spans M/rM ∼ = (Aq /r)r as a module over Aq /r, which is a product of fields. Using that
k is an infinite subfield of Aq /r it follows that N/rN contains a basis of M/rM over Aq /r; see
Exercise (4.3). Applying the Nakayama lemma, it follows that N contains a basis of M over Aq ,
i.e., we have elements a1 , . . . ar ∈ Aq such that the elements ϕq (ai ⊗ 1) = σ(ai ) form an Aq -basis
of R ⊗Q Aq . Hence for every a ∈ Aq there are unique coordinates x1 , . . . , xr ∈ Aq such that

σ(a) = x1 · σ(a1 ) + · · · + xr · σ(ar )

(4)
= (1 ⊗ x1 ) · σ(a1 ) + · · · + (1 ⊗ xr ) · σ(ar ) .

We view R!! := R ⊗Q R ⊗Q Aq as a module over R ⊗Q Aq via the homomorphism j2,3 given by

r ⊗ a $→ 1 ⊗ r ⊗ a. The diagrams (2) tell us that the elements
$ % $ %
γi := (m̃ ⊗ idA ) σ(ai ) = (idR ⊗ σ) σ(ai )

form an R ⊗Q Aq -basis of R!! . Applying m̃ ⊗ idA and idR ⊗ σ to (4) gives

$ %
(m̃ ⊗ idA ) σ(a) = (1 ⊗ 1 ⊗ x1 ) · γ1 + · · · + (1 ⊗ 1 ⊗ xr ) · γr
)
)
)
$ % $ % $ %
(idR ⊗ σ) σ(a) = 1 ⊗ σ(x1 ) · γ1 + · · · + 1 ⊗ σ(xr ) · γr .

Hence the coordinates xi lie in B, and (4) becomes σ(a) = σ(x1 a1 + · · · + xr ar ). But we have
seen in (4.20) that σ is injective, hence a = x1 a1 + · · · + xr ar . This proves that the elements
a1 , . . . , ar span Aq as a Bq -module. On the other hand, since the map a $→ σ(a) is Bq -linear,
the elements a1 , . . . , ar are linearly independent over Bq . Hence Aq is free of rank r over Bq .
Finally we consider the case that Bq has a finite residue field. By what was explained in
Remark (4.7) we may assume that S = Y . Because B → Bq is flat we may, by (ii) of the theorem,
further reduce to the case where B = Bq . Let h: B → B ! be a faithfully flat homomorphism,
where B ! is a local ring with infinite residue field; for instance we could take B ! to be a strict
henselization of B = Bq . In order to show that A = Aq is free of rank r over B, it suffices
to show that A! := A ⊗B B ! is free of rank r over B ! , see EGA IV, 2.5.2. But, again by (ii),
Spec(B ! ) is the quotient of Spec(A! ) under the G-action obtained by base-change. Hence we are
reduced to the case treated above.

(4.26) As the final step in the proof we show that if G acts freely, Y is a universal geometric
quotient. Consider a morphism h: S ! → S. Let us indicate base change via h by a ! , so
X ! := X ×S S ! , etc. Then G! acts again freely on X ! , and it is easy to see that every G! -
equivalence class of closed points in |X ! | is contained in an affine open subset. (Since this
statement only involves the fibres of X ! we may assume that S ! is affine, in which case the
morphism X ! → X is affine.) Hence there exists a geometric quotient, say qZ : X ! → Z. As
Z is a categorical quotient of X ! by G! , the morphism q ! : X ! → Y ! factors as q ! = f ◦ qZ with
f : Z → Y ! . We want to show that f is an isomorphism.

– 60 –
 As before we may assume that G is free of rank r over S. Then X ! is free of rank r over Z
but at the same time it is free of the same rank r over Y ! . But then Z has to be locally free of
∼
rank 1 over Y ! , so f : Z −→ Y ! . This completes the proof of Theorem (4.16). #

§ 3. FPPF quotients.

Consider an action of an S-group scheme G on an S-scheme X. In general there is not a simple

procedure to construct a “good” quotient of X by G in the category Sch/S . Of course we have
the notion of a categorical quotient, but this is only a “best possible approximation in the given
category”, and its definition gives no clues about whether there exists a categorical quotient
and, if so, how to describe it.
Most approaches to the formation of quotients follow the same pattern:
(a) replace the category Sch/S of S-schemes by some “bigger” category, in which the formation
of quotients is easier;
(b) form the quotient Y := G\X in this bigger category;
(c) study under which assumptions the quotient Y is (representable by) a scheme.
Thus, for instance, in our discussion of geometric quotients the “bigger” category that we used
was the category of ringed spaces over S.
The approach usually taken in the theory of group schemes is explained with great clarity
in Raynaud [2]. The idea is that one chooses a Grothendieck topology on the category of S-
schemes and that all objects in question are viewed as sheaves on the resulting site. The quotient
spaces that we are interested in exist as sheaves—this usually involves a sheafification—and their
construction has good functorial properties. Then it remains to be investigated under what
conditions the quotient sheaf is representable by a scheme. For the choice of the topology, a
couple of remarks have to be taken into account. First, we want our original objects, schemes,
to be sheaves rather than presheaves; this means that the topology should be no finer than the
canonical topology (see Appendix ??). On the other hand, the finer the topology, the weaker
the condition that a sheaf is representable. Finally the topology has to be accessible by the
methods of algebraic geometry. In practice one usually works with the étale topology, the fppf
topology or the fpqc topology. We shall mostly work with the fppf topology. See (4.36) below
for further discussion.
From a modern perspective, perhaps the most natural choice for the “bigger category” in
which to work, is the category of algebraic stacks. An excellent reference for the foundations of
this theory is the book by Laumon and Moret-Bailly [1]. For general results about the formation
of quotients as algebraic spaces we recommend the papers by Keel and Mori [1] and Kollár [1].
However, at this stage in our book we shall not assume any knowledge of algebraic spaces or
stacks (though algebraic spaces will be briefly mentioned in our discussion of Picard functors in
Chap. 6).
Finally let us remark that we shall almost exclusively deal with quotients modulo a group
action, and not with more general equivalence relations or groupoids. It should be noted that
even if one is interested only in group quotients, the proofs often involve more general groupoids.

(4.27) We shall use some notions that are explained in more detail in Appendix ??.
Let S be a scheme. We write (S)FPPF for the big fppf site of S, i.e., the category Sch/S of
S-schemes equipped with the fppf topology. We write FPPF(S) for the category of sheaves on

– 61 –
(S)FPPF .
The fppf topology is coarser than the canonical topology; this means that for every S-
scheme X the presheaf hX = HomS (−, X) is a sheaf on (S)FPPF . As explained in A?? this is
essentially a reformulation of results in descent theory. Via X $→ hX we can identify Sch/S with
a full subcategory of FPPF(S). We shall usually simply write X for hX .
Denote by ShGr/S and ShAb/S the categories of sheaves of groups, respectively sheaves of
abelian groups, on (S)FPPF . The category ShAb/S is abelian; ShGr/S is not abelian (excluding
S = ∅) but we can still speak about exact sequences. Unless specified otherwise, we shall from
now on view the category of S-group schemes as a full subcategory of ShGr/S . For example, we
shall say that a sequence of S-group schemes

ϕ ψ
G! −→ G −→ G!!

is exact if it is exact as a$ sequence in ShGr/S , %i.e., if Ker(ψ) represents the fppf sheaf associated
to the presheaf T $→ Im ϕ(T ): G! (T ) → G(T ) .

(4.28) Definition. Let G be an S-group scheme acting, by ρ: G×S X → X, on an S-scheme X.

We write (G\X)fppf , or simply G\X, for the fppf sheaf associated to the presheaf

T $→ G(T )\X(T ) .

If G\X is representable by a scheme Y then we refer to Y (or to the quotient morphism q: X →

Y ) as the fppf quotient of X by G.
We often say that “an fppf quotient exists” if (G\X)fppf is representable by a scheme. Note
that the sheaf G\X is a categorical quotient of X by G in FPPF(S), so we are indeed forming the
quotient in a “bigger” category. Note further that if (G\X)fppf is representable by a scheme Y
then by the Yoneda lemma we have a morphism of schemes q: X → Y .
As we are mainly interested in the formation of quotients of a group scheme by a subgroup
scheme, we shall mostly restrict our discussion of fppf quotients to the case that the action is
free.

(4.29) Example. Consider the situation as in (iii) of Theorem (4.16). So, G is finite locally free
over S, acting freely on X, and every orbit is contained in an affine open set. Let qY : X → Y
be the universal geometric quotient, as we have proven to exist. We claim that Y is also an fppf
quotient. To see this, write Z := (G\X)fppf and write qZ : X → Z for the quotient map. As Z
is a categorical quotient in FPPF(S), the morphism qY , viewed as a morphism of fppf sheaves,
factors as qY = r ◦ qZ for some r: Z → Y . To prove that r is an isomorphism it suffices to show
that it is both a monomorphism and an epimorphism.
By (iii) of (4.16), the morphism qY is fppf. By A?? this implies it is an epimorphism of
sheaves. But then r is an epimorphism too. On the other hand, suppose T is an S-scheme and
suppose a, b ∈ Z(T ) map to the same point in Y (T ). There exists an fppf covering T ! → T such
that a and b come from points a! , b! ∈ X(T ! ). But we know that Ψ = (ρ, prX ): G×S X → X×Y X
is an isomorphism, so there is a point c ∈ G ×S X(T ! ) with ρ(c) = a! and prX (c) = b! . By
construction of Z := (G\X)fppf this implies that a = b. Hence r is a monomorphism.

(4.30) The formation of fppf quotients is compatible with base change. To explain this in more
detail, suppose j: S ! → S is a morphism of schemes. Then j gives rise to an inverse image

– 62 –
functor j ∗ : FPPF(S) → FPPF(S ! ) which is exact. Concretely, if f : T → S ! is an S ! -scheme
then j ◦ f : T → S is an S-scheme, and if F is an fppf sheaf on S then we have j ∗ F (f : T →
S ! ) = F (j ◦ f : T → S). In particular, on representable sheaves j ∗ is simply given by base-change:
j ∗ X = X ×S S ! . Writing X ! = X ×S S ! and G! = G ×S S ! , we conclude that j ∗ (G\X) = (G! \X ! )
as sheaves on (S ! )FPPF . Hence if q: X → Y is an fppf quotient over S then Y ! := Y ×S S ! is
an fppf quotient of X ! by G! . Put differently: An fppf quotient, if it exists, is automatically a
universal fppf quotient.

(4.31) Proposition. Let G be an S-group scheme acting freely on an S-scheme X. Suppose

the fppf sheaf (G\X)fppf is representable by a scheme Y . Write q: X → Y for the canonical mor-
phism. Then q is an fppf covering and the morphism Ψ: G ×S X → X ×Y X is an isomorphism.
This gives a commutative diagram with cartesian squares
∼ pr1
G ×S X −−−→ X ×Y X −−→ X
  

pr2 "

pr2 "
q
"
q
X === X −→ Y .

In particular, X is a G-torsor over Y in the fppf topology which becomes trivial over the covering
q: X → Y .
Proof. By construction, the projection X → Y is an epimorphism of fppf sheaves. This implies
that it is an fppf covering; see A??. Further, Ψ: G ×S X → X ×Y X is an isomorphism of fppf
sheaves, again by construction of Y = G\X. By the Yoneda lemma (3.3), Ψ is then also an
isomorphism of schemes. #

(4.32) In the situation of the proposition, a necessary condition for (G\X)fppf to be repre-
sentable by a scheme is that the action of G on X is strictly free. Indeed, this is immediate
from the fact that X ×Y X is a subscheme of X ×S X. But the good news contained in (4.31)
is that if an fppf quotient exists, it has very good functorial properties. Let us explain this in
some more detail.
We say that a property P of morphisms of schemes is fppf local on the target if the following
two conditions hold:
(a) given a cartesian diagram
h
X ! −→ X
 
 f
f "" "
S! −→ S
g

we have P (f ) ⇒ P (f ! ) (we say: “P is stable under base change”);

(b) if furthermore g: S ! → S is an fppf covering then P (f ) ⇔ P (f ! ).
Many properties that play a role in algebraic geometry are fppf local on the target. More
precisely, it follows from the results in EGA IV, § 2 that this holds for the property P of a
morphism of schemes of being flat, smooth, unramified, étale, (locally) of finite type or finite
presentation, (quasi-) separated, (quasi-) finite, (quasi-) affine, or integral.

(4.33) Corollary. Let P be a property of morphisms of schemes which is local on the target
for the fppf topology. If q: X → Y is an fppf quotient of X under the free action of an S-group

– 63 –
scheme G, then
q: X → Y pr2 : G ×S X → X π: G → S
⇐⇒ ⇐=
has property P has property P has property P
where moreover the last implication is an equivalence if X → S is an fppf covering.
∼
Proof. Clear, as q: X → Y is an fppf covering and G ×S X −→ X ×Y X. #
In the applications we shall see that this is a most useful result. After all, it tells us that
an fppf quotient morphism q: X → Y inherits many properties from the structural morphism
π: G → S. To study π we can use the techniques discussed in Chapter 3. To give but one
example, suppose S = Spec(k) is the spectrum of a field and that G and X are of finite type
over k. As before we assume that G acts freely on X. Then the conclusion is that an fppf
quotient morphism q: X → Y is smooth if and only if G is a smooth k-group scheme. By (3.17)
it suffices to test this at the origin of G, and if moreover char(k) = 0 then by (3.20) G is
automatically smooth over k.

(4.34) At this point, let us take a little step back and compare the various notions of a quotient
that we have encountered.
Consider a base scheme S, an S-group scheme G acting on an S-scheme X, and suppose
q: X → Y is a morphism of S-schemes. Then q realizes Y as
—a categorical quotient of X by G if q is universal for G-equivariant morphisms from X to
an S-scheme with trivial G-action;
—a geometric quotient of X by G if |Y | = |X|/ ∼ and OY = (q∗ OX )G , i.e., Y represents
the quotient of X by G formed in the category of ringed spaces;
—an fppf quotient of X by G if Y represents the fppf sheaf associated to the presheaf
T $→ G(T )\X(T ), i.e., Y represents the quotient of X by G formed in the category of fppf
sheaves.
Further we have defined what it means for Y to be a universal categorical or geometric
quotient. As remarked earlier, an fppf quotient is automatically universal.
The following result is due to Raynaud [1] and gives a comparison between fppf and geo-
metric quotients.

(4.35) Theorem. Let G be an S-group scheme acting on an S-scheme X.

(i) Suppose there exists an fppf quotient Y of X by G. Then Y is also a geometric quotient.
(ii) Assume that X is locally of finite type over S, and that G is flat and locally of finite
presentation over S. Assume further that the action of G on X is strictly free. If there exists
a geometric quotient Y of X by G then Y is also an fppf quotient. In particular, the quotient
morphism q: X → Y is an fppf morphism and Y is a universal geometric quotient.
Proof. For the proof of (ii) we refer to Anantharaman [1], Appendix I. Let us prove (i). Suppose
that q: X → Y is an fppf quotient. Write r: X → Z := (G\X)rs for the quotient of X by G
in the category of ringed spaces over S. Since r is a categorical quotient in RS/S we have a
unique morphism of ringed spaces s: Z → Y such that q = s ◦ r. Our goal is to prove that s is
an isomorphism. First note that q, being an fppf covering, is open and surjective. Since also r
is surjective, this implies that the map s is open and surjective.
Next we show that s is injective. Suppose A and B are points of |X| that map to the same
point C in |Y |. We have to show that ρ−1 {A} ∩ pr−1 X {B} is non-empty, for then A and B map
to the same point of Z, and the injectivity of s follows. Choose a field extension κ(C) ⊂ K

– 64 –
and K-valued points a ∈ X(K) and b ∈ X(K) with support in A and B, respectively, such
that q(a) = q(b). By construction of the fppf quotient, there exists a K-algebra L of finite type
and an L-valued point d ∈ G ×S X(L) with ρ(d) = a and prX (d) = b. But then the image of
d: Spec(L) → G ×S X is contained in ρ−1 {A} ∩ pr−1 X {B}.
Finally, let U an open part of Y . There is a natural bijection between Γ(U, OY ) and the
morphisms U → A1S over S. Write V := q −1 (U ) and W := ρ−1 (V ) = pr−1 X (V ). By the Yoneda
1
lemma the morphisms U → AS as schemes are the same as the morphisms as fppf sheaves. By
construction of the fppf quotient we therefore find that Γ(U, OY ) is in bijection with the set of
morphisms f : V → A1S over S such that f ◦ ρ = f ◦ prX : W → A1S . Writing f := q ◦ ρ = q ◦ prX ,
this shows that OY is the kernel of q∗ OX ⇒ f∗ OG×S X , which, by definition, is the subsheaf of
G-invariant sections in q∗ OX . This proves that s is an isomorphism of ringed spaces, so that Y
is also a geometric quotient of X by G. #

To summarize, we have the following relations between the various notions:*

universal universal
fppf quotient =⇒ =⇒
geometric quotient categorical quotient
* *
* *
1 + +
geometric quotient =⇒ categorical quotient

where the implication “geometric ⇒ fppf” is valid under the assumptions as in (ii) of the theorem.

(4.36) The sheaf-theoretic approach that we are discussing here of course also makes sense for
other Grothendieck topologies on Sch/S , such as the étale topology. Thus, for instance, suppose
q: X → Y is an fppf quotient of X by the action of an S-group scheme G. One may ask if q is
also an étale quotient. But for this to be the case, q has to be an epimorphism of étale sheaves,
which means that étale-locally on Y it admits a section. If this is not the case then q will not
be a quotient morphism for the étale topology.
To give a simple geometric example, suppose q: X → Y is a finite morpism of complete
non-singular curves over a field such that the extension k(Y ) ⊂ k(X) on function fields is Galois
with group G. Then q is an fppf quotient of X by G, but it is an étale quotient only if there is
no ramification, i.e., if q is étale.
Conversely, if étale-locally on Y the morphism q has a section then q is an epimorphism
of étale sheaves and one shows without difficulty that q is an étale quotient of X by G. (Note
that q is assumed to be an fppf quotient morphism, so we already know it is faithfully flat, and
in particular also surjective.) But as the simple example just given demonstrates, for a general
theory of quotients we obtain better results if we use a finer topology, such as the fppf topology.

(4.37) Working with sheaves of groups has the advantage that many familiar results from
ordinary group theory readily generalize. For instance, if H is a normal subgroup scheme
of G then the fppf quotient sheaf G/H is naturally a sheaf of groups, and the canonical map
q: G → G/H is a homomorphism. Hence if G/H is representable then it is a group scheme
and the sequence 0 −→ H −→ G −→ G/H −→ 0 is exact. In this case, if f : G → G! is a
homomorphism of S-group schemes such that f|H is trivial then f factors uniquely as f = f ! ◦ q,
where f ! : G/H → G! is again a homomorphism of group schemes.

* schuine pijl moet gestreepte pijl worden

– 65 –
 To conclude our general discussion of fppf quotients, let us now state two existence results.
For some finer results see Raynaud [1] and [2], SGA 3, Exp. V and VI, and Anantharaman [1].

(4.38) Theorem. Let G be a proper and flat group scheme of finite type over a locally
noetherian basis S. Let ρ: G ×S X → X define a strictly free action of G on a quasi-projective
S-scheme X. Then the fppf quotient G\X is representable by a scheme.
A proof of this result can be found in SGA 3, Exp. V, § 7.

(4.39) Theorem. Let G be a flat group scheme of finite type over a locally noetherian base
scheme S. Let H ⊂ G be a closed subgroup scheme which is flat over S. Suppose that we are
in one of the following cases:
(a) dim(S) $ 1;
(b) G is quasi-projective over S and H is proper over S;
(c) H is finite locally free over S such that every fibre Hs ⊂ Gs is contained in an affine
open subset of G.
Then the fppf quotient sheaf G/H is representable by an S-scheme. If H is normal in G then
G/H has the structure of an S-group scheme such that the natural map q: G → G/H is a
homomorphism.
For the proof of this result in case (a) see Anantharaman [1], § 4. In case (b) the assertion
follows from (4.38), and case (c) is an application of Thm. (4.16); cf. Example (4.29).

(4.40) Example. Let X be an abelian variety over a field k. If H ⊂ X is a closed subgroup

scheme then by Thm. (4.38) there exists an fppf quotient q: X → Y := X/H. By Thm. (4.35) q
is also a geometric quotient, and from this it readily follows that Y is again an abelian variety.

§ 4. Finite group schemes over a field.

Now that we have some further techniques at our disposal, let us return to the study of group
schemes. As an application of the above, we sketch the proof of a useful general result.

(4.41) Theorem. If k is a field then the category of commutative group schemes of finite type
over k is abelian.
Proof (sketch). Write C for the category of commutative group schemes of finite type over k.
We view C as a full subcategory of the category ShAb/k of fppf sheaves of abelian groups on
Spec(k), which is an abelian category. Clearly C is an additive subcategory, and by (3.13) it is
stable under the formation of kernels.
Let f : G1 → G2 be a morphism in C. In the category ShAb/k we can form the quotients
∼
q1 : G1 → G1 /Ker(f ) and q2 : G2 → G2 /G1 , and we have an isomorphism α: G1 /Ker(f ) −→
Ker(q2 ). First one shows that the quotient morphism q1 exists as a homomorphism of group
schemes; see also (4.39) below. Let Ḡ1 := G1 /Ker(f ), and let f¯: Ḡ1 → G2 be the homomorphism
induced by f . Note that f¯ is a monomorphism. Now one proves that the quotient sheaf G2 /Ḡ1
is also representable by a k-scheme of finite type; for the details of this see SGA 3, Exp VIA ,
Thm. 3.2. But the natural map of sheaves G2 /G1 → G2 /Ḡ1 is an isomorphism, so it follows that
G2 /G1 is a group scheme. In particular, C is stable under the formation of cokernels, and since

– 66 –
 ∼
C is a full subcategory of ShAb/k we have an isomorphism α: G1 /Ker(f ) −→ Ker(q2 ) in C. #
We now focus on finite group schemes.

(4.42) Definition. Let G be a finite group scheme over a field k. We say that G is
— étale if the structural morphism G → Spec(k) is étale;
— local if G is connected.
Next suppose that G is commutative. Recall that we write GD for the Cartier dual of G. We
say that G is
— étale-étale if G and GD are both étale;
— étale-local if G is étale and GD is local;
— local-étale if G is local and GD is étale;
— local-local if G and GD are both local.
Let us note that if k ⊂ K is a field extension and if G is étale (resp. local) then GK is
étale (resp. local), too. For étaleness this is clear; for the property of being local this is just
Prop. (3.17), part (i).

(4.43) Examples. If char(k) = 0 then it follows from Thm. (3.20) that every finite commutative
k-group scheme is étale-étale. If char(k) = p > 0 then all four types occur:
type: étale-étale étale-local local-étale local-local
n
example: (Z/mZ) with p ! m (Z/p Z) µp n αp n

(4.44) Lemma. Let G1 and G2 be finite group schemes over a field k, with G1 étale and G2
local. Then the only homomorphisms G1 → G2 and G2 → G1 are the trivial ones.
Proof. Without loss of generality we may assume that k = k. Then G2,red ⊂ G2 is a connected
étale subgroup scheme; hence G2,red ∼= Spec(k). Now note that any homomorphism G1 →
G2 factors through G2,red . Similarly, any homomorphism G2 → G1 factors through G01 ∼ =
Spec(k). #
Note that the assertion about homomorphisms from an étale to a local group
$ scheme
% does
not generalize to arbitrary base
$ schemes.% For instance, if we take S = Spec k[ε] as a base
scheme then the group HomS (Z/pZ), µp is isomorphic to the additive group k, letting a ∈ k
correspond to the homomorphism (Z/pZ)S → µp,S given on points by (n mod p) $→ (1 + aε)n .

(4.45) Proposition. Let G be a finite group scheme over a field k. Then G is an extension
of an étale k-group scheme Gét = /0 (G) by the local group scheme G0 ; so we have an exact
sequence
1 −→ G0 −→ G −→ Gét −→ 1 . (5)

If k is perfect then this sequence splits (i.e., we have a homomorphic section G ← Gét ) and G is
isomorphic to a semi-direct product G0 " Gét . In particular, if k is perfect and G is commutative
then G ∼ = G0 × Gét .
Note that the étale quotient Gét is nothing but the group scheme /0 (G) of connected
components introduced in (3.28). In the present context it is customary to think of Gét as a
“building block” for G, and it is more customary to use a notation like Gét .
Proof. Define Gét := /0 (G), and consider the homomorphism q: G → Gét as in Prop. (3.27).

– 67 –
As shown there, q is faithfully flat, and the kernel of q is precisely the identity component G0 .
Hence we have the exact sequence (5).
Let us now assume that k is perfect. Then Gred ⊂ G is a closed subgroup scheme
(Exercise 3.2) which by (ii) of Prop. (3.17) is étale over k. We claim that the composition
Gred "→ G →→ Gét is an isomorphism. To see this we may assume that k = k. But then G, as a
scheme, is a finite disjoint union of copies of G0 . If there are n components then Gred and Gét
are both isomorphic to the disjoint union of n copies of Spec(k), and it is clear that Gred → Gét
is an isomorphism of group schemes. The inverse of this isomorphism gives a splitting of (5). #
Combining this with Lemma (4.44) we find that the category C of finite commutative group
schemes over a perfect field k decomposes as a product of categories:

C = Cét,ét × Cét,loc × Cloc,ét × Cloc,loc .

As remarked above, C = Cét,ét if char(k) = 0.

(4.46) Lemma. Let S be a connected base scheme. If 0 −→ G1 −→ G2 −→ G3 −→ 0 is an

exact sequence of finite locally free S-group schemes then rank(G2 ) = rank(G1 ) · rank(G3 ).
Proof. Immediate from the fact that G2 is a G1 -torsor over G3 for the fppf topology, as this
implies that OG2 is locally free as an OG3 -module, of rank equal to rank(G1 /S). #

(4.47) Proposition. Let k be a field of characteristic p > 0. Let G be a finite connected

k-group scheme. Then the rank of G is a power of p.
Proof. Let FG/k : G → G(p) be the relative Frobenius homomorphism. Write G[F ] := Ker(FG/k ).
The strategy of the proof is to use the short exact sequence 1 −→ G[F ] −→ G −→ G/G[F ] −→ 1
and induction on the rank of G. The main point is then to show that the affine algebra of G[F ]
is of the form k[X1 , . . . , Xd ]/(X1p , . . . , Xdp ) with d = dimk (TG,e ). To prove this we use certain
differential operators.
Write G = Spec(A), and let I ⊂ A be the augmentation ideal. We have an isomorphism
∼
I/I 2 −→ ΩA/k ⊗A k, sending the class of ξ ∈ I to dξ ⊗ 1. Further, (3.15) tells us that ΩA/k ∼ =
(ΩA/k ⊗A k) ⊗k A. In total we find

Derk (A) = HomA (ΩA/k , A) ∼

= Homk (I/I 2 , A) ,

where the derivation Dϕ : A → A corresponding to ϕ: I/I 2 → A satisfies Dϕ (ξ) = ϕ(ξ) mod I

for all ξ ∈ I.
Choose elements x1 , . . . , xd ∈ I whose classes form a k-basis for I/I 2 . By the previous
remarks, there exist k-derivations Di : A → A such that Di (xj ) = δi,j mod I for all i and j.
We claim that for all non-negative numbers m1 , . . . , md and n1 , . . . , nd with m1 + · · · + md =
n1 + · · · + nd we have
,
md md−1 m1
$ n1 nd
% n1 ! n2 ! · · · nd ! mod I, if mi = ni for all i;
Dd Dd−1 · · · D1 x1 · · · xd ≡ (6)
0 mod I, otherwise.

To see this, note that for every D ∈ Derk (A) the product rule implies that D(I r ) ⊆ I r−1 . With
this remark, (6) follows by induction on the number m1 + · · · + md .
By Nakayama’s lemma the xi generate I, so we have

A∼
= k[X1 , . . . , Xd ]/(f1 , . . . , fq )

– 68 –
via xi ←$ Xi . Let J = (f1 , . . . , fq ) ⊂ A. We claim that J ⊆ (X1p , . . . , Xdp ). To see this, suppose
we have a polynomial relation between the xi such that there are no terms xai with a ! p. Write
this relation as
0 = h0 + h1 (x1 , . . . , xd ) + · · · + hr (x1 , . . . , xd ) ,

where hj is a homogeneous polynomial of degree j. Let j be the smallest integer such that
hj '= 0. Suppose xn1 1 · · · xnd d (with n1 + · · · + nd = j) is a monomial occurring with non-zero
nd−1
coefficient. Applying the differential operator Ddnd Dd−1 · · · D1n1 and using (6) we obtain the
relation n1 !n2 ! · · · nd ! ∈ I. This contradicts the fact that k is a field of characteristic p and that
all ni are < p. Hence J ⊆ (X1p , . . . , Xdp ), as claimed.
Let FG/k : G → G(p) be the relative Frobenius homomorphism. On rings it is given by

(p)
k[X1 , . . . , Xd ]/(f1 , . . . , fq(p) ) −→ k[X1 , . . . , Xd ]/(f1 , . . . , fq ) , Xi $→ Xip .

As the zero section of G(p) is given by sending all Xi to 0 we find that the affine algebra of
G[F ] := Ker(FG/k ) is

AG[F ] = k[X1 , . . . , Xd ]/(X1p , . . . , Xdp , f1 , . . . , fq ) = k[X1 , . . . , Xd ]/(X1p , . . . , Xdp ) .

d
$ % $ %
In particular,
$ % G[F ] has rank p . Further, rank(G) = rank G[F ] · rank G/G[F ] = pd ·
rank G/G[F ] by Lemma (4.46). As G = G0 we have d > 0 if G '= {1}; now the proposi-
tion follows by induction on rank(G). #

(4.48) Corollary. If char(k) = p then a finite commutative k-group scheme is étale-étale if and
only if p ! rank(G).
Proof. In the “if” direction this is a direct consequence of the proposition combined with (4.45)
and Lemma (4.46). Conversely, suppose G is étale-étale. We may assume that k = k, in which
case G is a constant group scheme. If p | rank(G) then G has a direct factor (Z/pn Z). But then
GD has a factor µpn and is therefore not étale. #

Exercises.

(4.1) Let S be a base scheme. Fix an integer N ! 2. Take G = X = Gm,S , and let g ∈ G act
on X as multiplication by g N .
(i) Let T be an S-scheme. Let x1 and x2 be T -valued points of X; they correspond to elements
γ1 , γ$2 ∈ Γ(T, OT )∗ . % Let c := γ1 /γ2 , and define a scheme T ! , affine over T , by T ! :=
Spec OT [t]/(tN − c) . Show that the images of x1 and x2 in X(T ! ) lie in the same orbit
under G(T ! ).
(ii) Show that T ! → T is an epimorphism of schemes over S. (By definition this means that for
every S-scheme Z the induced map Z(T ) → Z(T ! ) is injective.)
(iii) Suppose that q: X → Y is a G-invariant morphism of S-schemes. Show that for every
S-scheme T the image of q(T ): X(T ) → Y (T ) consists of a single point. Conclude that
X → S is a universal categorical quotient of X by G.
(iv) Show that the endomorphism Gm → Gm given by g $→ g N is faithfully flat and of finite
presentation. Use this to show that the fppf sheaf G\X is represented by the scheme S.

– 69 –
(4.2) Let ρ: G ×S X → X be an action of an S-group scheme G on an S-scheme X. Define the
relation P ∼ Q on |X| as in (4.11). The goal of this exercise is to show that ∼ is an equivalence
relation.
(i) Let Ψ = Ψρ be the graph morphism, as defined in (4.1). Write

- |G ×S X| → |X| ×|S| |X|

Ψ:

for the composition of the map |Ψ|: |G ×S X| → |X ×S X| and the canonical (surjective)
map |X ×S X| → |X| ×|S| |X|. Show that P ∼ Q precisely if (P, Q) ∈ ImΨ. -
(ii) Write e(S) ⊂ G for the image of the identity section. Show that the projection e(S)×S X →
X is an isomorphism. Conclude that ∼ is reflexive.
(iii) Let s: X ×S X −→ X ×S X be the morphism reversing the factors. Find a morphism
f : G ×S X → G ×S X such that s ◦ Ψ = Ψ ◦ f . Conclude that ∼ is symmetric.
(iv) Show that ∼ is transitive. [Hint: use that the natural map
. .
. .
.(G ×S X) × (G ×S X). −→ |G ×S X| × |G ×S X|
. ρ,X,pr
. |ρ|,|X|,|pr
X X|

is surjective.]
(4.3) Let k be an infinite field. Let Λ be a k-algebra which is a product of fields. Suppose
M is a free Λ-module of finite rank. Let N ⊂ M be a k-submodule such that N spans M as
a Λ-module. Show that N contains a Λ-basis for M . Show by means of an example that the
condition that k is infinite is essential.
(4.4) Let π: G → S be a locally free group scheme of rank r over a reduced, irreducible base
scheme S. The goal of this exercise is to show that G is annihilated by its rank, i.e., the morphism
[r]G : G → G given on points by g $→ g r equals the zero morphism [0]G = e ◦ π: G → S → G.
(i) Suppose S is the spectrum of a field k. Reduce the problem to the case that G = G0 . [Hint:
Use (4.45) and Lemma (4.46). For étale group schemes reduce the problem to Lagrange’s
theorem in group theory.]
(ii) Suppose S = Spec(k) with char(k) = p. Suppose further that G = G0 = Spec(A). By
(4.47) we have rank(G) = pn for some n. If I ⊂ A is the augmentation ideal, show that
n
I p = (0). Now use the result of Exercise (3.7) to derive that [pn ](I) = (0). Conclude that
[pn ]G = [0]G .
(iii) Prove the stated result over an arbitrary reduced and irreducible basis. [Hint: use that the
generic fibre of G is Zariski dense in G.]
[Remark: for commutative finite locally free group schemes the result holds without any restric-
tion on the basis. This was proven by Deligne; see Tate-Oort [1]. It is an open problem if the
result is also valid over arbitrary base schemes for non-commutative G.]
(4.5) Let S be a locally noetherian scheme. Let G be a finite locally free S-group scheme acting
on an S-scheme X of finite type. Assume that for every closed point P ∈ |X| the G-equivalence
class of P is contained in an affine open subset. Write q: X → Y for the quotient morphism. If
x ∈ |X| then we write ÔX,x for the completed local ring of X at the point x; likewise for other
schemes.
/
(i) Let y ∈ |Y |. Show that the scheme F̂y := x∈q−1 (y) Spec(ÔX,x ) inherits a G-action, and
that Spec(ÔY,y ) is the quotient of F̂y modulo G. [Hint: First reduce to the case that S = Y ;
then apply a flat base change.]

– 70 –
(ii) Suppose S = Spec(k) is the spectrum of a field. Let x ∈ X(k) be a k-rational point with
∼
image y ∈ Y (k) under q. Show that q induces an isomorphism ÔY,y −→ (ÔX,x )Gx .
(4.6) Let X → S be a morphism of schemes. Let G be a finite group that acts on X over S.
(i) For g ∈ G, define a scheme X g and a morphism ig : X g → X by the fibre product square
ig
Xg −→ X
 
 ∆
" " X
X −−→ X ×S X
ψg

where the morphism ψg : X → X ×S X is given by x $→ (g · x, x). Show that ig is an

immersion and that it is a closed immersion if X/S is separated.
(ii) Define X G "→ X as the scheme-theoretic intersection of the subschemes X g , for g ∈ G. (In
other words, if G = {g1 , . . . , gn } then X G := X g1 ×X X g2 ×X · · · ×X X gn .) Show that X G
is a subscheme of X, and that it is a closed subscheme if X/S is separated. Further show
that for any S-scheme T we have X G (T ) = X(T )G . The subscheme X G "→ X is called the
fixed point subscheme of the given action of G.

– 71 –
Chapter V. Isogenies.

In this chapter we define the notion of an isogeny, and we discuss some basic examples, in-
cluding the multiplication by an integer n != 0 and the relative Frobenius homomorphism in
characteristic p. As applications we obtain results about the group of n-torsion points on an
abelian variety. If the ground field has positive characteristic p this leads to the introduction of
an invariant, the p-rank of the abelian variety.

§ 1. Definition of an isogeny, and basic properties.

(5.1) Lemma. (i) Let X and Y be irreducible noetherian schemes which are both regular and
with dim(X) = dim(Y ). Let f : X → Y be a quasi-finite morphism. Then f is flat.
(ii) Let f : X → Y be a morphism of finite type between noetherian schemes, with Y reduced
and irreducible. Then there is a non-empty open subset U ⊆ Y such that either f −1 (U ) = ∅ or
the restricted morphism f : f −1 (U ) → U is flat.
A proof of (i) can be found in Altman-Kleiman [1], Chap. V, Cor. 3.6 or Matsumura [1],
Thm. 23.1. For (ii) we refer to Mumford [2], Lecture 8.

(5.2) Proposition. Let f : X → Y be a homomorphism of abelian varieties. Then the following

conditions are equivalent:
(a) f is surjective and dim(X) = dim(Y );
(b) Ker(f ) is a finite group scheme and dim(X) = dim(Y );
(c) f is a finite, flat and surjective morphism.
Proof. We shall use that if h: Z1 → Z2 is a flat morphism of k-varieties and F ⊂ Z1 is the fibre
of h over a closed point of Z2 then F is equidimensional and

dim(Z1 ) = dim(Z2 ) + dim(F ) . (1)

(This is a special case of HAG, Chap. III, Prop. 9.5.)

Let us first assume that (b) holds. As f is proper and all fibres are translates of Ker(f ) it
follows that f is finite. Hence f (X) is closed in Y , of dimension equal to dim(X) = dim(Y ).
Hence f is surjective. Further, by (i) of the lemma, f is flat. This shows that (a) and (c) hold.
Next suppose that (a) holds. By (ii) of the lemma, f is flat over a non-empty open subset
U ⊆ Y . As all fibres of f are translates of Ker(f ), (b) follows from (1). That (c) implies (b)
again readily follows from (1). !
By making use of the results about quotients that were discussed in the previous chapter,
we could do without Lemma (5.1). We leave such an alternative proof of the proposition to the
reader.

(5.3) Definition. A homomorphism f : X → Y of abelian varieties is called an isogeny if f

satisfies the three equivalent conditions (a), (b) and (c) in (5.2). The degree of an isogeny f is

Isogs, 15 september, 2011 (812)

– 72 –
the degree of the function field extension [k(X): k(Y )]. (Note that we have a homomorphism
k(Y ) → k(X), since an isogeny is surjective.)
∼
If f : X → Y is an isogeny then f induces an isomorphism X/Ker(f ) −→ Y . Because all
fibres of f are translates of Ker(f ) the sheaf f∗ OX is a locally free OY -module of finite rank.
Computing this rank at the generic point of Y , respectively the closed point 0 ∈ Y , gives
! "
deg(f ) = rankOY (f∗ OX ) = rank Ker(f ) .
! "
(Here rank Ker(f ) denotes the rank of the finite group scheme Ker(f ).) If f : X → Y and
g: Y → Z are isogenies then so is g ◦ f , and deg(g ◦ f ) = deg(g) · deg(f ).

(5.4) Lemma. Let f : W → X and h: Y → Z be isogenies of abelian varieties over k. If g1 ,

g2 : X → Y are homomorphisms such that h ◦ g1 ◦ f = h ◦ g2 ◦ f then g1 = g2 .
Proof. We may assume that k = k. Suppose h ◦ g1 ◦ f = h ◦ g2 ◦ f . Because f is faithfully flat, it
is an epimorphism of schemes, so it follows that h ◦ g1 = h ◦ g2 . Hence g1 − g2 maps X into the
finite group scheme Ker(h). As X is connected and reduced, g1 − g2 factors through Ker(h)0red ,
which is trivial. !

(5.5) We recall the notion of a purely inseparable morphism (French: morphisme radiciel).
In EGA Inew , Prop. 3.7.1 it is shown that the following conditions on a morphism of schemes
f : X → Y are equivalent:
(a) f is universally injective; this means that for every Y $ → Y the morphism f $ : X $ → Y $
obtained from f by base change is injective;
!(b) f"is injective and for every x ∈ X the residue field k(x) is a purely inseparable extension
of k f (x) ;
(c) for every field K, the map X(K) → Y (K) induced by f is injective.
A morphism that satisfies these conditions is called a purely inseparable morphism.

(5.6) Proposition. Let f : X → Y be an isogeny.

(i) The following conditions are equivalent.
(a) The function field k(X) is a separable field extension of k(Y );
(b) f is an étale morphism;
(c) Ker(f ) is an étale group scheme.
(ii) The following conditions are equivalent.
(a) The function field k(X) is a purely inseparable field extension of k(Y );
(b) f is a purely inseparable morphism;
(c) Ker(f ) is a connected group scheme.
Proof. (i) That (b) and (c) are equivalent is clear from (4.33). If f is étale then for every x ∈ X,
writing y = f (x) ∈ Y , the residue field k(x) is a finite separable extension of k(y). If we apply
this with x the generic point of X, we see that (b) implies (a).
Now assume that k(X) is a finite separable extension of k(Y ). As f is a finite flat morphism,
it is étale at a point x ∈ X if and only if (Ω1X/Y )x = 0. But Ω1X/Y is a coherent OX -module,
hence its support is closed, and it follows that the locus where f is étale is an open subset
U ⊂ X. The assumption that k(X) is finite separable over k(Y ) means that the generic point
of X is in U , so U is non-empty. As f is proper it follows that there is an open subset V ⊂ Y
such that f −1 (V ) is étale over V . But V is the quotient of f −1 (V ) under Ker(f ), so it follows
from (4.33) that Ker(f ) is étale.

– 73 –
 (ii) We can factor f as a composition of two isogenies: X → X/Ker(f )0 → Y . The kernel
of the second isogeny is Ker(f )/Ker(f )0 , which is étale. (See also Prop. (4.45).) Using (i) it
follows that (a) implies (c).
That (b) implies (a) is immediate from property (b) in (5.5), applied to the generic point
of X.
Finally suppose that N := Ker(f ) is a connected group scheme. Let k ⊂ K be a field
extension. Let A be the affine algebra of N and write AK = A ⊗k K. If y: Spec(K) → Y is
a K-valued point then the scheme-theoretic fibre f −1 (y) := X ×Y,y Spec(K) is isomorphic to
NK = Spec(AK ). As AK has finite K-dimension it is an artinian ring. Any artinian ring is
a product of artinian local rings; this corresponds to the decomposition of f −1 (y) as a union
of connected components. But we know from (i) of (3.17) that NK is a connected scheme.
Hence AK is artinian local and |f −1 (y)| consists of a single point. This shows that f satisfies
condition (c) of (5.5) and is therefore purely inseparable. !

(5.7) Definition. An isogeny f : X → Y is called separable if it satisfies the three equivalent

conditions in (5.6)(i). It is called a (purely) inseparable isogeny if it satisfies the equivalent
conditions of (5.6)(ii).

(5.8) Corollary. Every isogeny f : X → Y can be factorized as f = h ◦ g, where g: X → Z is

an inseparable isogeny and h: Z → Y is a separable isogeny. This factorization is unique up to
isomorphism, in the sense that if f = h$ ◦ g $ : X → Z $ → Y is a second such factorization then
∼
there is an isomorphism α: Z −→ Z $ with g $ = α ◦ g and h = h$ ◦ α.
Proof. Immediate from the above and Prop. (4.45). !
An important example of an isogeny is the multiplication [n]X : X → X by an integer n != 0.
We write X[n] := Ker([n]X ) ⊂ X.

(5.9) Proposition. For n != 0, the morphism [n]X is an isogeny. If g = dim(X), we have

deg([n]X ) = n2g . If (char(k), n) = 1 then [n]X is separable.
Proof. Choose an ample and symmetric line bundle L on X. (Recall that L is said to be
symmetric if (−1)∗ L ∼ = L, and note that if L is ample then L ⊗ (−1)∗ L is ample and symmetric.)
By (2.12) we know that n∗X L ∼
2
= L⊗n . The restiction of n∗X L to Ker(f ) is a trivial bundle which
is ample. (Here we use that n != 0.) This implies that Ker(f ) is finite, hence [n]X is an isogeny.
To compute the degree we use intersection theory on smooth varieties. Choose an ample
symmetric divisor D. Then deg([n]X ) · (D)g = ([n]∗X D)g . But [n]∗X D is linearly equivalent to
n2 · D, so ([n]∗X D)g = n2g · (D)g , and we find that deg([n]X ) = n2g .
If char(k) = 0 then the last assertion is trivial. If char(k) = p > 0 with p ! n then also
2g
p ! n #= rank(X[n]), $and the result follows from Cor. (4.48). Alternatively, as p does not divide
n2g = k(X1 ) : k(X2 ) , the field extension k(X2 ) ⊂ k(X1 ) given by f is separable. !

(5.10) Corollary. If X is an abelian variety over an algebraically closed field k then X(k) is a
divisible group. That is, for every P ∈ X(k) and n ∈ Z\{0} there exists a point Q ∈ X(k) with
n · Q = P.
Note that if the ground field k is only assumed to be separably closed then it is not true in
general that X(k) is a divisible group. See ?? for an example.

– 74 –
(5.11) Corollary. If (char(k), n) = 1 then X[n](ks ) = X[n](k) ∼
= (Z/nZ)2g .
Proof. We know that X[n] is an étale group scheme of rank n2g . Hence X[n](ks ) = X[n](k) is
an abelian group of order n2g , killed by n. Further, for every divisor d of n the subgroup of
elements killed by d is just X[d](ks ) and has order d2g . It now readily follows from the structure
theorem for finite abelian groups that we must have X[n](ks ) ∼ = (Z/nZ)2g . !

(5.12) Proposition. If f : X → Y is an isogeny of degree d then there exists an isogeny

g: Y → X with g ◦ f = [d]X and f ◦ g = [d]Y .
Proof. If deg(f ) = d then Ker(f ) is a finite group scheme of rank d and is therefore annihilated
by multiplication by d; see Exercise (4.4). It follows that [d]X factors as

f g
[d]X = (X −→ Y −→ X)

for some isogeny g: Y → X. Then g ◦ [d]Y = [d]X ◦ g = (g ◦ f ) ◦ g = g ◦ (f ◦ g), and by Lemma (5.4)
it follows that f ◦ g = [d]Y . !

(5.13) Corollary. The relation

def
X ∼k Y = there exists an isogeny f : X → Y

is an equivalence relation on the set of abelian varieties over k.

If there is no risk of confusion we shall use the notation X ∼ Y instead of X ∼k Y . Note,
however, that the ground field plays a role: if k ⊂ K is a field extension then X ∼k Y implies
that XK ∼K YK , but the converse does not hold in general.
If there exists an isogeny f : X → Y then we say that X and Y are isogenous. Again this
notion is relative to a given ground field; if necessary we may specify that X and Y are isogenous
over the given field k.

(5.14) Example. Suppose we work over the field C of complex numbers. If X is an abelian
variety over C, the associated analytic manifold X an is a complex torus; see also (1.10). So
we can write X an = V /L, where V is a complex vector space and L ⊂ V is a lattice. More
intrinsically, V can be identified with the tangent space of X an at the origin, and the projection
map V → X is then the exponential map in the sense of Lie theory. We shall come back to this
in more detail in Chapter ??.
Let X1 and X2 be complex abelian varieties; write Xian = Vi /Li . Let f : X1 → X2
be a homomorphism. It follows from the previous remarks that the associated analytic map
f an : X1an → X2an is given by a C-linear map ϕ: V1 → V2 such that ϕ(L1 ) ⊆ L2 . Conversely, any
such ϕ gives an analytic map ϕ̄: X1an → X2an , and it can be shown (using a result of Chow, see
HAG, Appendix B, Thm. 2.2) that there exists a unique algebraic homomorphism f : X1 → X2
with ϕ̄ = f an .
As an example, multiplication by n on X corresponds to ϕ = n·idV , which obviously maps L
into itself. We find that the group of n-torsion points X[n](C) is isomorphic to n−1 L/L ⊂ V /L,
and if g = dim(X) then indeed n−1 L/L ∼ = (Z/nZ)2g .
As an application we find that X1 ∼ X2 if and only if there exists a C-linear isomorphism
∼
α: V1 −→ V2 such that α(L1 ⊗ Q) = L2 ⊗ Q; in other words, there should exist positive integers
m and n with m · L2 ⊆ α(L1 ) ⊆ n−1 · L2 .

– 75 –
§ 2. Frobenius and Verschiebung.

As the next example of an isogeny, we look at Frobenius in characteristic p > 0.

(5.15) Proposition. Let X be a g-dimensional abelian variety over a field k with char(k) =
p > 0. Then the relative Frobenius homomorphism FX/k : X → X (p) is a purely inseparable
isogeny of degree pg .
Proof. Write X[F ] := Ker(FX/k ). On underlying topological spaces, the absolute Frobenius
FrobX : X → X is the identity. It follows that the topological space underlying X[F ] is the
singleton {e}. Let now U = Spec(A), with A = k[x1 , . . . , xr ]/(f1 , . . . , fn ), be an affine open
neigbourhood of e in X such that e corresponds to the maximal ideal m = (x1 , . . . , xr ) ⊂ A.
(p)
Write fi ∈ k[x1 , . . . , xr ] for the polynomial obtained from fi by raising all coefficients to the pth
(p) (p)
power. Then U (p) = Spec(A(p) ), with A(p) = k[x1 , . . . , xr ]/(f1 , . . . , fn ), and FU/k : U → U (p) ,
the restriction of FX/k to U , is given on rings by
(p)
A = k[x1 , . . . , xr ]/(f1 , . . . , fn ) ←− A(p) = k[x1 , . . . , xr ]/(f1 , . . . , fn(p) )
xpi ←−, xi .

It follows that X[F ] = Spec(B), with B = k[x1 , . . . , xr ]/(xp1 , . . . , xpr , f1 , . . . , fn ). In particular,

X[F ] is finite, hence FX/k is an isogeny.
Write  for the m-adic completion of A. Without loss of generality we may assume that
x1 , . . . , xg form a basis of m/m2 = TX,e
∨
. The structure theory for complete regular local rings
tells us that there is an isomorphism
∼
k[[t1 , . . . , tg ]] −→ Â

sending ti to xi . (See Bourbaki [2], Chap. VIII, § 5, no 2.) Since (xp1 , . . . , xpr ) ⊂ m, we find that

B = A/(xp1 , . . . , xpr )A ∼ p
= Â/(x1 , . . . , xpr )Â
∼ Â/(xp , . . . , xp )Â
= 1 g
∼ p
= k[[t1 , . . . , tg ]]/(t1 , . . . , tpg )
∼ p
= k[t1 , . . . , tg ]/(t , . . . , tp ) .
1 g

In particular this shows that deg(FX/k ) = rank(X[F ]) = pg and that X[F ] is a connected group
scheme. !
Our next goal is to define the Verschiebung isogeny for abelian varieties in characteristic p.
In fact, under a suitable flatness assumption the Verschiebung can be defined for arbitrary
commutative group schemes over a basis S with char(S) = p; we shall give the construction in
this generality. First we need some preparations.

(5.16) Let R be a ring with char(R) = p > 0. Let A be an R-algebra. Write T p (A) :=
A ⊗R ⊗R · · · ⊗R A for the p-fold tensor product of A over R. The symmetric group Sp on p
letters naturally acts on T p (A) by ring automorphisms. Write S p (A) ⊂ T p (A) for the subalgebra
of Sp -invariants, i.e., the subalgebra of symmetric tensors.
Let N : T p (A) → S p (A) be the “symmetrizer” map, i.e., the map given by
%
N (a1 ⊗ · · · ⊗ ap ) = aσ(1) ⊗ · · · ⊗ aσ(p) .
σ∈Sp

– 76 –
If !s ∈ S p "(A) is a symmetric tensor and t ∈ T p (A) then N (st) = sN (t). It follows that J :=
N T p (A) is an ideal of S p (A).
Write U := Spec(A) → T := Spec(R). Applying Thm. (4.8) we find that the quotient S p (U )
of UTp := U ×T!U ×T "· · · ×T U (p factors) under the natural action of Sp exists and is given by
S p (U ) = Spec S p (A) . The scheme S p (U ) is called the p-th symmetric power of U over T . Note
that S p (U/T ) would be a better notation, as the base scheme is important in the construction.
We trust, however, that the simpler notation S p (U ) will not cause any confusion. Let U [p/T ] #→
S p (U ) be the closed subscheme defined by the ideal! J. If η: T p (A)" → A is the multiplication
map, given by a1 ⊗ · · · ⊗ ap ,→ a1 · · · ap , then η N (a1 ⊗ · · · ⊗ ap ) = p! · (a1 · · · ap ) = 0. This
means that the morphism
∆p
U −−−−→ UTp −→ S p (U )
U/T

factors through U [p/T ] ⊂ S p (U ). Write FU/T

$
: U → U [p/T ] for the morphism thus obtained.
Write A(p/R) := A ⊗R,F R, where F = FrobR : R → R is the Frobenius homomorphism,
given by r ,→ rp . We view A(p/R) as an R-algebra via r ,→ 1 ⊗ r; so for a ∈ A !and r ∈" R we
have the relations rp · (a ⊗ 1) = a ⊗ rp = (ra) ⊗ 1. By definition, U (p/T ) = Spec A(p/R) . Now
observe that we have a well-defined map

ϕA/R : A(p/R) → S p (A)/J

sending a ⊗ r ∈ A(p/R) to (ra ⊗ a ⊗ · · · ⊗ a) mod J. Note that (ra ⊗ a ⊗ · · · ⊗ a) is an element

of S p (A) because all tensors are taken over the ring R. Also note that ϕA/R is well-defined
precisely because we use p-tensors. (Check this yourself!) Write ϕU/T : U [p/T ] → U (p/T ) for the
$
morphism of schemes induced by ϕA/R . It is clear from the definitions that FU/T = ϕU/T ◦ FU/T .
We now globalize these constructions. For this, consider a base scheme S of characteristic p
and an S-scheme π: X → S. Define S p (X), the pth symmetric power of X over S, to be the
quotient of XSp under the natural action of Sp . If U ⊂ X and T ⊂ S are affine open subsets with
π(U ) ⊆ T then S p (U ) is an affine open subset of S p (X). The closed subschemes U [p/T ] #→ S p (U )
glue to a locally closed subscheme X [p/S] #→ S p (X). Also, the morphisms FU/T $
and ϕU/T glue
and give a factorization of the relative Frobenius morphism FX/S as
"
FX/S ϕX/S
FX/S = (X −−−−→ X [p/S] −−−−→ X (p/S) ) .
$
By construction, the composition of FX/S and the inclusion X [p/S] #→ S p (X) is the same as the
composition of the diagonal ∆X/S : X → XSp and the natural projection XSp → S p (X). Summing
p

up, we have a commutative diagram

∆X/S
p
X −−−−→ XS
 
F " 
' X/S '
p
FX/S X [p/S] #−→ S (X)

ϕX/S
'

X (p/S)

(5.17) Lemma. (i) The construction of X [p/S] , as well as the formation of FX/S
$
and ϕX/S , is
functorial in X and compatible with flat base change T → S.

– 77 –
 (ii) If X is flat over S then ϕX/S : X [p/S] → X (p/S) is an isomorphism of S-schemes.
Proof. Part (i) of the lemma is a straightforward verification. For (ii), it suffices to treat the
case that X = U = Spec(A) and S = T = Spec(R). Let M be an R-module. Just as before we
can form the p-fold tensor product T p (M ) of M over R and the submodule S p (M ) ⊂ T p (M ) of
symmetric tensors, and there is a symmetrizer map N : T p (M ) → S p (M ). We have a well-defined
map
! "
ϕM/R : M (p/R) −→ S p (M )/N T p (M ) given by m ⊗ r ,→ [rm ⊗ m ⊗ · · · ⊗ m] .

Suppose M is a free R-module with a basis {ei }i∈I . The tensors ei := ei1 ⊗ ei2 ⊗ · · · ⊗ eip with
i = (i1 , . . . , ip ) ∈ I p , form a basis of T p (M ). Such a tensor ei can be symmetrized in a minimal
way. Namely, if H ⊂ Sp is the stabilizer of (i1 , . . . , ip ) in the natural action of Sp on I p then
for σ̄ ∈ H\Sp the element eiσ̄(1) ⊗ eiσ̄(2) ⊗ · · · ⊗ eiσ̄(p) is well-defined; now set
%
si := eiσ̄(1) ⊗ eiσ̄(2) ⊗ · · · ⊗ eiσ̄(p) .
σ̄∈H\Sp

The symmetric tensors si obtained in this way span S p (M ); note however that different se-
quences i may give the same tensor si . If i1 = i2 = · · · = ip then N (ei ) = p! · si = 0; if not all
ij are equal then N (ei ) is a unit times si . (Recall that! R is an" Fp -algebra.) We conclude that
the tensors ei ⊗ ei ⊗ · · · ⊗ ei form a basis of S p (M )/N T p (M ) , and it follows that ϕM/R is an
isomorphism if M is free over R.
Now we use a non-trivial result from commutative algebra. Namely, if M is flat over R
then it can be written as a filtered direct limit, say M = lim−→
Mα , of free R-modules. For a proof
see [??]. Since lim
−→
is right exact and commutes with tensor products, ϕM/R can be identified
with lim ϕ
−→ Mα /R
and is therefore again an isomorphism. Applying this to M = A the lemma
follows. !
We now consider a commutative S-group scheme G. The morphism m(p) : GpS → G given
on sections by (g1 , g2 , . . . , gp ) ,→ g1 g2 · · · gp factors through S p (G), say via m̄(p) : S p (G) → G. It
follows that [p]: G → G, which is equal to m(p) ◦ ∆pG/S , factors as

! FG/S
"
m̄(p)
[p] = G −−−−→ G[p/S] #−→ S p (G) −−−→ G) . (2)

(5.18) Definition. If G is a commutative flat group scheme over a basis S of characteristic p

then we define the Verschiebung homomorphism

VG/S : G(p/S) −→ G

to be the composition

! ϕ−1
G/S m̄(p) "
VG/S = G(p/S) −−−−→ G[p/S] #−→ S p (G) −−−→ G .

That VG/S is indeed a homomorphism of group schemes follows from (i) of the lemma.

(5.19) Proposition. Let S be a scheme with char(S) = p > 0. Let G be a flat S-group scheme.

– 78 –
 (i) We have VG/S ◦ FG/S = [p]G : G −→ G.
(ii) If G is finite locally free over S then the Verschiebung is Cartier dual to the Frobenius
homomorphism; more precisely, we have (VG/S )D = FGD /S and VG/S = (FGD /S )D .
Proof. Statement (i) follows from the definitions; indeed, if we write j: G[p/S] #−→ S p (G) for the
inclusion morphism then

VG/S ◦ FG/S = (m̄(p) ◦ j ◦ ϕ−1 $

G/S ) (ϕG/S FG/S ) = m̄
◦ ◦
(p) ◦ ◦ $
j FG/S = [p]G

by (2).
For (ii), suppose G is finite locally free over S. Without loss of generality we may assume
that S = Spec(R) is affine, so that G is given by an R-algebra A. Possibly after further
localization on S we may assume that A is free as a module over R, say with basis {e1 , . . . , en }.
Recall from the proof of Lemma (5.17) that given a sequence i = (i1 , i2 , . . . , ip ) ∈ {1, 2, . . . , n}p ,
we can symmetrize the tensor ei1 ⊗ ei2 ⊗ · · · ⊗ eip in a minimal way. The resulting collection of
tensors
{si }1!i1 !i2 !···!ip !n
is a basis of S p (A). It follows from the proof of Lemma (5.17) that the Verschiebung VG/S is
given on rings by the composition

m̄(p)
A −−−→ S p (A) −→ A(p/R) ,

where m̄(p) is the homomorphism that corresponds to the morphism m̄(p) : S p (G) → G, and
where the homomorphism S p (A) → A(p/R) is given by
(
0, if ij < ij+1 for some j;
si ,→ .
ei ⊗ 1 if i = (i, i, . . . , i).

Now we apply the functor ( )D = HomR (−, R). We have an isomorphism

! D "(p/R) ∼ ! (p/R) "D
A −→ A

by sending ϕ !⊗ ρ ∈ "AD ⊗R,F R to the map a ⊗ r ,→ rρϕ(a)p . Further there is a canonical

isomorphism S p (A) D ∼ = Symp (AD ); here we note that by our general conventions in (??),
Symp (AD ) is a quotient of the p-fold tensor product T p (AD ), whereas S p (A) is a sub-algebra
of T p (A). Using these identifications, and writing {ε1 , . . . , εn } for the R-basis of AD dual to
{e1 , . . . , en }, the dual of the map S p (A) → A(p/R) is the map
! D "(p/R)
A −→ Symp (AD ) given by εi ⊗ ρ ,→ [ρεi ⊗ εi ⊗ · · · ⊗ εi ] .

Furthermore, by definition of the ring structure on AD , the dual of the map m̄(p) : A → S p (A)
is the multiplication map Symp (AD ) → AD given by [ϕ1 ⊗ · · · ⊗ ϕp ] ,→ ϕ1 · · · ϕp . Combining
this we see that the Cartier dual of VG/S is given on rings by the map
! "(p/R)
AD → A sending ϕ⊗r to r · ϕp .

This shows that (VG/S )D = FGD /S . By Cartier duality then also VG/S = (FGD /S )D . !
Now we apply this to abelian varieties.

– 79 –
(5.20) Proposition. Let X be an abelian variety over a field k with char(k) = p. Then the
Verschiebung homomorphism VX/k : X (p) → X is an isogeny of degree pg . We have VX/k ◦ FX/k =
[p]X and FX/k ◦ VX/k = [p]X (p) .
Proof. Write F = FX/k and V = VX/k . We have already seen that V ◦ F = [p]X . It follows that
V satisfies (a) of Proposition
! " (5.2); hence it is an isogeny. That V has degree pg follows from
the relation p2g = deg [p] = deg(V ) · deg(F ) = deg(V ) · pg . Finally, F ◦ V ◦ F = F ◦ [p] = [p] ◦ F ,
and because F is an epimorphism this implies that F ◦ V = [p]. !
m
(5.21) Let X be a k-scheme, where k is a field of characteristic p. For m " 1 we write X (p )
for the base change of X over the mth power Frobenius homomorphism Frobm k : k → k. By a
slight abuse of notation we write
2 m
m
FX/k = FX (pm−1 ) /k ◦ · · · ◦ FX (p) /k ◦ FX/k : X → X (p) → X (p )
→ · · · → X (p )

for the “mth power” of Frobenius, or “iterated Frobenius”. Similarly, we can define an “mth
m
m
iterated Verschiebung” VX/k : X (p ) → X by

m
VX/k = VX/k ◦ VX (p) /k ◦ · · · ◦ VX (pm−1 ) /k .

By an easy induction on m we find that [pm ]X = VX/k m ◦ m

FX/k and [pm ]X (pm ) = FX/k
m ◦ m
VX/k .
Indeed, for m = 1 this is just Proposition (5.20), and to make the induction we note that
m+1 ◦ m+1
VX/k FX/k = VX/k ◦ VXm(p) /k ◦ FX
m
(p) /k ◦ FX/k

= VX/k ◦ [pm ]X (p) ◦ FX/k

= [pm ]X ◦ VX/k ◦ FX/k = [pm+1 ]X .

(Likewise for the relation [pm ]X (pm ) = FX/k

m ◦ m
VX/k .)

Let us now look what is the analogue of (5.11) in case char(k) | n. In fact, since all X[n](k)
are finite abelian, it suffices to consider the case that n = pm , where p = char(k) > 0.

(5.22) Proposition. Suppose char(k) = p > 0. There is an integer f = f (X), with 0 # f #

g = dim(X), such that X[pm ](k) ∼
= (Z/pm Z)f for all m " 0. If Y is isogenous to X then
f (Y ) = f (X).
Proof. We can factor pm : X → X as
) m
*
FX/k
m (pm ) h1 h2
[p ]X = X −−−−→ X −−→ Y −−→ X ,

m
where h1 ◦ FX/k is purely inseparable and h2 is a separable isogeny. Looking at the degrees we
find that X[p ](k) is an abelian group of rank deg(h2 ) = pd(m) , where d(m) is an integer with
m

0 # d(m) # gm. Write f = d(1), so that X[p](k) ∼ = (Z/pZ)f . It follows from Corollary (5.10)
that we have exact sequences of (abstract) groups

pm−1
0 −→ X[pm−1 ](k) −→ X[pm ](k) −−−−→ X[p](k) −→ 0 .

The claim that X[pm ](k) ∼

= (Z/pm Z)f for all m " 0 follows by induction on m.

– 80 –
 Finally, suppose h: X → Y is an isogeny, say of degree d. Then X[pm ](k) maps to Y [pm ](k),
and the kernel has order at most d. Taking m large enough, it follows that f (Y ) " f (X). As
X ∼ Y is a symmetric relation, we conclude that f (X) = f (Y ). !

(5.23) Definition. The integer f = f (X), which lies in the range 0 # f # g := dim(X), is
called the p-rank of X.

(5.24) Caution. Let X be an abelian variety of p-rank f > 0 over a non-perfect field k, and
let k ⊂ ks ⊂ k be respectively a separable closure and an algebraic closure of k. Then we have
natural injective maps X[pm ](ks ) → X[pm ](k), but these are not, in general, isomorphisms. In
other words, in order to see all pmf distinct physical points of order pm , in general we need an
inseparable extension of the ground field.
At first sight this may seem to contradict the fact that an étale k-group scheme becomes
constant over ks . For instance, taking m = 1 we have a short exact sequence of k-group schemes

1 −→ X[p]loc −→ X[p] −→ X[p]ét −→ 1 ,

(see Prop. (4.45)) and X[p]ét ⊗k ks is isomorphic to (Z/pZ)f . However, in order to split the
exact sequence, and hence to be able to lift the points of X[p]ét to points of X[p], we in general
need to pass to an inseparable extension. See also the examples in (5.26) and (5.27) below for
a concrete illustration of this point.

(5.25) Remarks. (i) The p-rank does not depend on the ground field. More precisely, if k ⊂ K
is a field extension and X is an abelian variety over k then f (X) = f (XK ). To see this we may
assume that k and K are both algebraically closed. By (4.45) the group scheme X[p] is a product
of its local and étale parts, i.e., X[p] ∼
= X[p]loc × X[p]ét . Over k = k the étale part becomes
a constant group scheme, i.e., X[p]ét = Γk with Γ = X[p](k). But after extension of scalars
to K the local and étale parts of X[p] remain local and étale, respectively; see ??. Therefore
X[p](K) = Γk (K) = Γ, so indeed f (X) = f (XK ).
(ii) Later we shall prove that the p-rank may take any value between 0 and dim(X): given
a field k with char(k) = p > 0 and integers 0 # f # g, there exists an abelian variety X over k
with dim(X) = g and f (X) = f . In fact, as clearly f (X1 × X2 ) = f (X1 ) + f (X2 ), it suffices to
show that there exist elliptic curves X0 and X1 over k with f (Xi ) = i.
(iii) An elliptic curve X is said to be ordinary if f (X) = 1 and supersingular if f (X) = 0.
In the examples below we shall use this terminology. In Chapter ??, we shall define the notions
“ordinary” and “supersingular” for abelian varieties of arbitrary dimension. It should be noted
that for dim(X) > 2, “supersingular” is not equivalent to “p-rank = 0”.

(5.26) Example. Let X be an elliptic curve over a field k with char(k) = 2. Then X can be
given by a Weierstrass equation

y 2 + a1 xy + a3 y = x3 + a2 x2 + a4 x + a6 , (3)

such that the origin is the “point at infinity” ∞ = (0 : 1 : 0). A point P ∈ X(k) with affine
coordinates (ξ, η) is a 2-torsion point precisely if the tangent line at P passes through ∞. An
easy calculation shows that this happens if and only if a1 ξ + a3 = 0. We cannot have a1 = a3 ,
because X then would be singular. We conclude:
(
0 if a1 = 0;
f (X) =
1 if a1 != 0.

– 81 –
 It should be noted that if a1 = 0 and k = k then there is a linear change of coordinates
such that the equation for X becomes y 2 + y = x3 . So, up to isomorphism this is the only
supersingular elliptic curve in characteristic 2 (over k = k).
In the ordinary case, a1 != 0, we find that the non-trivial point of order 2 in X(k) is the
point with affine coordinates (a3 /a1 , η), where η ∈ k satisfies

η 2 = (a3 /a1 )3 + a2 (a3 /a1 )2 + a4 (a3 /a1 ) + a6 .

In particular, we see illustrated here the point made in (5.24) that in general we need to pass
to an inseparable extension of the ground field in order to have all p-torsion points rational.

(5.27) Example. Let X be an elliptic curve given by a Weierstrass equation (3), this time over
a field k with char(k) = 3. Then P ∈ X(k) \ {0} is a 3-torsion point if and only if P is a flex
point, i.e., a point at which the tangent line TX,P intersects X with multiplicity 3. (As X is
a nonsingular cubic curve the intersection multiplicity cannot be bigger.) Again this allows to
compute the p-rank by hand. To simplify, let us assume that a1 = a3 = 0; this is achieved after
a linear change of variables. Then P = (ξ, η) ∈ X(k) is a flex point if and only if

4a2 η 2 = 4a22 ξ 2 + 4a2 a4 ξ + a24 . (4)

Combined with the equation for X this is equivalent to

4a2 ξ 3 + (4a2 a6 − a24 ) = 0 . (5)

As X is nonsingular we cannot have a2 = a4 = 0. Hence

def
X is ordinary ⇐⇒ X[3](k) ∼
= Z/3Z ⇐⇒ a2 != 0 .

Note that if a2 != 0 then (5) has a unique solution for ξ ∈ k, and if ±η are the corresponding
solutions of (5.27.1) then (ξ, ±η) are the only two non-trivial 3-torsion points in X(k). So
indeed X[3](k) ∼= Z/3Z and f = 1. Further note that solving (4) in general requires passing to
an inseparable extension of k.

(5.28) Example. Let k be a field of characteristic 2. Consider the elliptic curve X ⊂ P2k given
by the homogeneous equation x21 x2 + x1 x22 = x30 , with ∞ = (0 : 1 : 0) as origin. As we have seen
above, X is supersingular, which for an elliptic curve is the same as saying that X has p-rank
zero. ! "
Recall that the group scheme α2 = α2,k is given by α2 = Spec k[ε]/(ε2 ) , with co-
multiplication ε ,→ ε ⊗ 1 + 1 ⊗ ε. We are going to give an action ρ: α2 × X → X of α2
on X. For this, write X as the union of two affine open subsets: X = U1 ∪ U2 , with
! "
U1 = X \ {(0 : 1 : 0)} = Spec k[x, y]/(x3 − y 2 − y)

and ! "
U2 = X \ {(0 : 0 : 1)} = Spec k[x, z]/(x3 − z 2 − z) .
Now we can give the action ρ on rings: let ρ1 : α2 × U1 → U1 be given by the homomorphism

k[x, y]/(x3 − y 2 − y) −→ k[x, y, ε]/(x3 − y 2 − y, ε2 ) with x ,→ x + ε , y ,→ y + εx2 ,

– 82 –
and, similarly, let ρ2 : α2 × U2 → U2 be given on rings by x ,→ x + ε and z ,→ z + εx2 . It is not
hard to verify that these homomorphisms are well-defined, that ρ1 and ρ2 agree on U1 ∩ U2 , and
that the resulting morphism ρ is indeed a group scheme action. Note that the points (0 : 1 : 0)
and (0 : 0 : 1) are α2 -stable when viewed as points in the underlying topological space |X|, but
that they are not fixed points of the action. In fact, the action is strictly free.
On U1 the functions ξ := x2 and η := y 2 are α2 -invariant. They generate a subring of
O(U1 ) of index 2; as the functions x and y themselves are clearly not invariant we conclude that

O(U1 )α2 ∼
= k[ξ, η]/(ξ 3 − η 2 − η) #−→ O(U1 ) = k[x, y]/(x3 − y 2 − y) .

Similarly, the algebra of α2 -invariants in O(U2 ) is generated by x2 and z 2 . We find that the
quotient α2 \X is isomorphic to X itself, where the quotient map X → X is just the Frobenius
endomorpism, given on points by (x, y) ,→ (x2 , y 2 ).
It can be shown that there is an isomorphism X[F ] ∼ = α2 such that the action ρ described
above becomes precisely the action of X[F ] on X by translations. As Exercise (??) shows,
this does not immediately follow from the fact that the quotient map for the α2 -action is the
Frobenius morphism. Note that from the given definition of the action ρ it is not clear that this
is an action of a subgroup scheme by translations. We shall return to this later; see (??).

(5.29) Example. Let X be an elliptic curve over a field k with char(k) = p, such that X[F ] ∼
=
∼
αp,k . It is not hard to verify! that k −
"→ End k (α p,k ), where the map sends λ ∈ k to the
endomorphism of αp,k = Spec k[t]/(tp ) given on rings by t ,→ λ · t. For (λ, µ) ∈ A2 (k) we
obtain an embedding ϕ(λ,µ) : αp,k #→ X × X by taking the composition

(λ,µ)
αp,k −−−→ αp,k × αp,k ∼
= X[F ] × X[F ] ⊂ X × X .

The image of ϕ(λ,µ) only depends on (λ : µ) ∈ P1 (k).

Figure ??.

We get a family of abelian surfaces over P1 by considering Y(λ:µ) := (X × X)/ϕ(λ,µ) (αp ). It can
be shown that given (λ0 : µ0 ) ∈ P1 (k), there are only finitely many (λ : µ) with Y(λ:µ) ∼
= Y(λ0 :µ0 ) .
The conclusion is that we have a non-trivial “continuous” family of isogenies X × X → Y(λ:µ) .
As we shall see later, such examples only exist in characteristic p > 0.

§ 3. Density of torsion points.

(5.30) Theorem. Let X be an abelian variety over a field k and let p be a prime number.
Then the collection of subschemes X[pm ] for m " 0 is scheme-theoretically dense in X.

– 83 –
 Let im : X[pm ] #→ X be the inclusion homomorphism. By definition, saying that the col-
lection of subschemes X[pm ] ⊂ X is scheme-theoretically dense in X means that there does not
exist a proper closed subscheme Y " X such that all im factor through Y . If p != char(k) we can
express the density of the torsion points of p-power order in a more elementary way. Namely, in
that case the following statements hold, as we shall see
+ inmthe+ proof.
(1) Topological density: the union of the subspaces X[p ]+ ⊂ |X| is dense in |X|;
+
,
(2) Function-theoretic density: the homomorphism of sheaves OX → m"0 OX[pm ] that is
induced by the homomorphisms im is injective.
Because X is reduced, properties (1) and (2) are equivalent, and (1) immediately implies that
the collection of subschemes X[pn ] is scheme-theoretically dense in X.
By contrast, if p = char(k) then (1) and (2) do not hold, in general. Indeed, if the p-
rank of X is zero then the group schemes X[pm ] are local, which means that the underlying
topological space is reduced to the single point 0. So in this case we can only interprete the
density statement scheme-theoretically.
Proof. We give separate proofs for the cases p = char(k) and p != char(k).
First assume that p != char(k). It suffices to prove the assertion for k = k, which from now
on we assume. In this case we know that X[pm ] is étale and consists of p2gm distinct closed
points. Let T ⊂ X(k) be the union of all X[pm ](k), and let Y ⊂ X be the smallest closed
subscheme such that all im factor through Y . Note that Y is reduced; it is in fact just the
reduced closed subscheme of X whose underlying space is the Zariski closure of T . We shall first
prove that Y is a subgroup scheme of X.
If x ∈ T then the translation tx : X → X maps T into itself; hence tx (Y ) ⊆ Y . This holds
for all x ∈ T , so it follows that for all y ∈ Y (k) also the translation ty maps T into itself, and
hence ty (Y ) ⊂ Y . Because Y and Y ×k Y are reduced, this implies that under the group law
m: X × X → X we have m(Y × Y ) ⊂ Y . As further it is clear that also Y is mapped into itself
under the inverse ι: X → X, we conclude that Y is indeed a subgroup scheme of X.
The identity component Y 0 is an abelian subvariety of X. Let N be the number of con-
nected components of Y . Further, let g = dim(X) and h = dim(Y 0 ). By Prop. (5.9) we have
#Y 0 [pm ](k) = p2mh for all m " 0, and it follows that #Y [pm ](k) # N · p2mh . (If W ⊂ Y is a
connected component that contains a torsion point w with pm ·w = 0 then translation by w gives
∼
an isomorphism Y 0 [pm ] −→ W ∩ X[pm ].) But by construction, Y contains all torsion points
of X of p-power order; so #Y [pm ](k) = p2mg . Taking m very large we see that we must have
h = g, which gives that Y 0 = X.
m
Next we deal with the case p = char(k). Let F m = FX/k m
: X → X (p ) be the mth power
of the Frobenius homomorphism, and let X[F m ] ⊂ X be the kernel. Because [pm ] = V m ◦ F m
(with V m = VX/k m
the iterated Verschiebung; see (5.21)) we have X[F m ] ⊂ X[pm ]. So we are
done if we can prove that the collection of group schemes X[F m ] is scheme-theoretically dense
in X. As in the proof of Prop. (5.15), let U = Spec(A) with A = k[x1 , . . . , xr ]/(f1 , . . . , fn ) be an
affine open neigbourhood ofmthe origin e in X such that e corresponds to the maximal ideal m =
(p )
(x1 , . . . , xr ) ⊂ A. Write fi ∈ k[x1 , . . . , xr ] for the polynomial obtained from fi by raising all
m (pm ) (pm )
coefficients to the power p , and write A(p ) = k[x1 , . . . , xr ]/(f1 , . . . , fn ). The restriction
m
m m
of F m to U is given on rings by the homomorphism A(p ) → A that sends xj to xpj . It follows
m m
that X[F m ] is the closed subscheme of U defined by the ideal (xp1 , . . . , xpr , f1 , . . . , fn ) ⊂ A.
Suppose Y ⊂ X is a closed subscheme such that all inclusion homomorphisms X[F m ] #→ X
factor through Y . Let J ⊂ A be the ideal of Y ∩ U . As in the proof of Prop. (5.15), let Â
be the m-adic completion of A and choose the coordinates xi in such a way that x1 , . . . , xg

– 84 –
 ∼
(with g = dim(X)) form a basis of m/m2 . We then have an isomorphism k[[t1 , . . . , tg ]] −→ Â
via ti ,→ xi , and we shall identify  with k[[t1 , . . . , tg ]] via this isomorphism. The assumption
m m
that X[F m ] is a subscheme of Y means that J Â is contained in the ideal (tp1 , . . . , tpg ). The
m m
intersection of the ideals (tp1 , . . . , tpg ) ⊂ Â for all m " 0 is the zero ideal, so we conclude
that J Â = (0). But then the complete local ring ÔY,e = Â/J Â of Y at the origin has Krull
dimension g, and consequently Y = X. !
We now prove the fact stated in Remark (2.14) that the results in (2.13) are true over an
arbitrary, not necessarily perfect, ground field.

(5.31) Proposition. Let X be an abelian variety over a field k. If Y #→ X is a closed subgroup

scheme then the connected component Y 0 ⊂ Y that contains the origin is an open and closed
0
subgroup scheme of Y that is geometrically irreducible. The reduced underlying scheme Yred #→
X is an abelian subvariety of X.
Proof. The assertion that Y 0 is open and closed in Y and is geometrically irreducible, was proven
0
in Prop. (3.17). To prove that Yred is an abelian subvariety of X we may assume, to simplify
0
notation, that Y = Y . We are going to prove that Yred is geometrically reduced. Before we
give the argument, let us explain how the desired conclusion follows. If Yred is geometrically
reduced then we have, with k ⊂ k an algebraic closure, that Yred,k = (Yk )red is a closed subgroup
scheme of Yk ; see Exercise (3.2). But then also Yred is a closed subgroup scheme of Y . Indeed,
the assertion that Yred is a subgroup scheme just means that the morphism Yred × Yred → Y
given on points by (y1 , y2 ) ,→ y1 − y2 factors through Yred ⊂ Y . If this holds after extension of
scalars to k then it also holds over k. So the conclusion is that Yred is a subgroup scheme of X
that is geometrically integral; hence it is an abelian subvariety.
We now prove that Yred is geometrically reduced. If char(k) = 0 then Y = Yred by
Thm. (3.20) and we are done by Prop. (3.17). Assume then that char(k) = p > 0. For all positive
integers n with p ! n the subgroup scheme Y [n] ⊂ Y is étale; hence we have Y [n] ⊂ Yred ⊂ Y .
This gives us a homomorphism of sheaves hn : OYred → OY [n] on |Yred | = |Y |, and we define
-
h: OYred → OY [n]
p!n

,
by h(f ) = n hn (f ). Further we know that (Yk )red ⊂ Xk is an abelian subvariety. By
Thm. (5.30) the collection of Y [n]k , for n " 1 with p ! n, is topologically dense in |Yk | = |(Yk )red |.
This implies that also the collection of all Y [n] is topologically dense in |Y | = |Yred |, and because
Yred is reduced, the homomorphism h is injective.
Suppose that Yred is not geometrically reduced. Then there is a finite, purely inseparable
field extension k ⊂ K such that (Yred )K is not reduced. (See EGA IV, Prop. 4.6.1.) As
k ⊂ K is purely inseparable, we have |(Yred )K | = |Yred | and |Y [n]K | = |Y [n]| for all n. The
structure sheaves of (Yred )K and Y [n]K are just OYred ⊗k K and OY [n] ⊗k K, respectively, and
the homomorphism
!- "
h ⊗ id: OYred ⊗k K → OY [n] ⊗k K
p!n

can be identified with the map

-
hK : O(Yred )K → OY [n]K
p!n

– 85 –
induced by the inclusions Y [n]K #→ (Yred )K . By our assumptions, (Yred )K is not reduced,
whereas all Y [n]K are reduced schemes. Hence h ⊗ id = hK must have a non-trivial kernel. But
then also h has a non-trivial kernel (k ⊂ K being faithfully flat), which contradicts our earlier
conclusion that it is injective. !

Exercises.
(5.1) Let f : X → Y be a surjective homomorphism of abelian varieties. Show that f is flat.
(5.2) Let k = Fp . By definition, αp is a subgroup scheme of Ga , so that we get a natural
action ρ: αp × Ga → Ga . Similarly, µp is a subgroup scheme of Gm , which gives an action
σ: µp × Gm → Gm .
(i) Identify Gm with the open subscheme of Ga given by x != 0. Show that the action ρ restricts
to a free action ρ$ of αp on Gm , and that the Frobenius endomorphism F : Gm → Gm , given
on points by x ,→ xp , is a quotient morphism for ρ$ .
(ii) Conclude that σ and ρ$ give rise to the same quotient morphism, even though αp ∼ ! µp .
=

– 86 –
Chapter VI. The Picard scheme of an abelian variety.

§ 1. Relative Picard functors.

To place the notion of a dual abelian variety in its context, we start with a short discussion of
relative Picard functors. Our goal is to sketch some general facts, without much discussion of
proofs.
Given a scheme X we write

Pic(X) = H 1 (X, OX
∗
) = {isomorphism classes of line bundles on X} ,

which has a natural group structure. (If τ is either the Zariski, or the étale, or the fppf topology
on Sch/X then we can also write Pic(X) = Hτ1 (X, Gm ), viewing the group scheme Gm = Gm,X
as a τ -sheaf on Sch/X ; see Exercise ??.)
If C is a complete non-singular curve over an algebraically closed field k then its Jacobian
Jac(C) is an abelian variety parametrizing the degree zero divisor classes on C or, what is the
same, the degree zero line bundles on C. (We refer to Chapter 14 for further discussion of
Jacobians.) Thus, for every k ⊂ K the degree map gives a homomorphism Pic(CK ) → Z, and
we have an exact sequence

0 −→ Jac(C)(K) −→ Pic(CK ) −→ Z −→ 0 .

In view of the importance of the Jacobian in the theory of curves one may ask if, more generally,
the line bundles on a variety X are parametrized by a scheme which is an extension of a discrete
part by a connected group variety.
If we want to study this in the general setting of a scheme f : X → S over some basis S, we
are led to consider the contravariant functor PX/S : (Sch/S )0 → Ab given by

PX/S : T $→ Pic(XT ) = H 1 (X ×S T, Gm ) .

However, one easily finds that this functor is not representable (unless X = ∅.). The reason
for this is the following. Suppose {Uα }α∈A is a Zariski covering of S and L is a line bundle
on X such that the restrictions L|X×S Uα are trivial. Then it is not necessarily the case that
L is trivial. This means that PX/S is not a sheaf for the Zariski topology on Sch/S , hence not
representable. (See also Exercise (6.1).)
The previous arguments suggest that in order to arrive at a functor that could be repre-
sentable we should first sheafify (or “localize”) PX/S with respect to some topology.

(6.1) Definition. The relative Picard functor PicX/S : (Sch/S )0 → Ab is defined to be the fppf
sheaf (on (S)FPPF ) associated to the presheaf PX/S . An S-scheme representing PicX/S (if such
a scheme exists) is called the relative Picard scheme of X over S.
Concretely, if T is an S-scheme then we can describe an element of PicX/S (T ) by giving an
fppf covering T $ → T and a line bundle L on XT ×T T $ such that the two pull-backs of L to

DualAV1, 15 september, 2011 (812)

– 87 –
XT ×T (T $ ×T T $ ) are isomorphic. Now suppose we have a second datum of this type, say an fppf
covering U $ → T and a line bundle M on XT ×T U $ whose two pull-backs to XT ×T (U $ ×T U $ )
are isomorphic. Then (T $ → T, L) and (U $ → T, M ) define the same element of PicX/S (T )
if there is a common refinement of the coverings T $ and U $ over which the bundles L and M
become isomorphic.
As usual, if PicX/S is representable then the representing scheme is unique up to S-
isomorphism; this justifies calling it the Picard scheme.

(6.2) Let us study PicX/S in some more detail in the situation that

the structure morphism f : X → S is quasi-compact and quasi-separated,
(∗) f (O ) = OT for all S-schemes T ,
 ∗ X×S T
f has a section ε: S → X.

For instance, this holds if S is the spectrum of a field k and X is a complete k-variety with
X(k) (= ∅ (see also Exercise ??); this is the case we shall mostly be interested in.
Rather than sheafifying PX/S we may also rigidify the objects we are trying to classify.
This is done as follows. If L is a line bundle on XT for some S-scheme T then, writing εT : T →
XT for the section induced by ε, by a rigidification of L along εT we mean an isomorphism
∼
α: OT −→ ε∗T L. (In the sequel we shall usually simply write ε for εT .)
Let (L1 , α1 ) and (L2 , α2 ) be line bundles on XT with rigidification along ε. By a homomor-
phism h: (L1 , α1 ) → (L2 , α2 ) we mean a homomorphism of line bundles h: L1 → L2 with the
property that (ε∗ h) ◦ α
$ 1 = α2 . In %particular, an endomorphism of (L, α) is given by an element
∗
h ∈ Γ(XT , OXT ) = Γ T, f∗ (OXT ) with ε (h) = 1. By the assumption that f∗ (OXT ) = OT we
therefore find that rigidified line bundles on XT have no nontrivial automorphisms.
Now define the functor PX/S,ε : (Sch/S )0 → Ab by
& '
isomorphism classes of rigidified
PX/S,ε : T $→ ,
line bundles (L, α) on X ×S T

with group structure given by

(L, α) · (M, β) = (L ⊗ M, γ) ,
γ = α ⊗ β: OT = OT ⊗ OT → ε∗ L ⊗ ε∗ M = ε∗ (L ⊗ M ) .
OT OT

If h: T $ → T is a morphism of S-schemes and (L, α) is a rigidified line bundle on X ×S T then

$ $ $ $ ∗
PX/S,ε (h): PX/S,ε (T ) →$ PX/S,ε
% (T ) sends (L, α) to (L , α ), where L = (idX × h) L and where
$ ∼ ∗ $ ∗ ∗
α : OT " −→ εT " L = h εT L is the pull-back of α under h.
Suppose PX/S,ε is representable by an S-scheme. On X ×S PX/S,ε we then have a universal
rigidified line bundle (P, ν); it is called the Poincaré bundle. The universal property of (P, ν)
is the following: if (L, α) is a line bundle on X ×S T with rigidification along the section ε then
there exists a unique morphism g: T → PX/S,ε such that (L, α) ∼ = (idX × g)∗ (P, ν) as rigidified
bundles on XT .
Under the assumptions (∗) on f it is not so difficult to prove the following facts. (See for
example BLR, § 8.1 for details.)
(i) For every S-scheme T there is a short exact sequence

pr∗
T
0 −→ Pic(T ) −−−
→ Pic(XT ) −→ PicX/S (T ) . (1)

– 88 –
This can be viewed as a short exact sequence obtained from a Leray spectral sequence. The
existence of a section is not needed for this.
(ii) For every S-scheme T , we have an isomorphism
∼
Pic(XT )/pr∗T Pic(T ) −→ PX/S,ε (T )

obtained by sending the class of a line bundle L on XT to the bundle L ⊗ f ∗ ε∗ L−1 with its
canonical rigidification.
(iii) The functor PX/S,ε is an fppf sheaf. (Descent theory for line bundles.)
Combining these facts we find that PX/S,ε ∼ = PicX/S and that these functors are given by

Pic(XT ) {line bundles on XT }

T $→ = .
∗
prT Pic(T ) {line bundles of the form f ∗ L, with L a line bundle on T }

In particular, the exact sequence (1) extends to an exact sequence

0 −→ Pic(T ) −→ Pic(XT ) −→ PicX/S (T ) −→ 0 . (2)

It also follows that PicX/S equals the Zariski sheaf associated to PX/S .

(6.3) Returning to the general case (i.e., no longer assuming that f satisfies the conditions (∗)
in (6.2)), one finds that PicX/S cannot be expected to be representable unless we impose further
conditions on X/S. (See Exercise ?? for an example.) The most important general results
about representability all work under the assumption that f : X → S is proper, flat and of finite
presentation. We quote some results:
(i) If f is flat and projective with geometrically integral fibres then PicX/S is representable
by a scheme, locally of finite presentation and separated over S. (Grothendieck, FGA, Exp. 232.)
(ii) If f is flat and projective with geometrically reduced fibres, such that all irreducible
components of the fibres of f are geometrically irreducible then PicX/S is representable by
a scheme, locally of finite presentation (but not necessarily separated) over S. (Mumford,
unpublished.)
(iii) If S = Spec(k) is the spectrum of a field and f is proper then PicX/S is representable
by a scheme that is separated and locally of finite type over k. (Murre [1], using a theorem of
Oort [1] to reduce to the case that X is reduced.)
If we further weaken the assumptions on f , e.g., if in (ii) we omit the condition that the
irreducible components of the fibres are geometrically irreducible, then we may in general only
hope for PicX/S to be representable by an algebraic space over S. Also if we only assume X/S
to be proper, not necessarily projective, then in general PicX/S will be an algebraic space rather
than a scheme. For instance, in Grothendieck’s FGA, Exp. 236 we find the following criterion.
(iv) If f : X → S is proper and locally of finite presentation with geometrically integral
fibres then PicX/S is a separated algebraic space over S.
We refer to ??, ?? for further discussion.

(6.4) Remark. Let X be a complete variety over a field k, let Y be a k-scheme and let L be
a line bundle on X × Y . The existence of maximal closed subscheme Y0 '→ Y over which L is
trivial, as claimed in (2.4), is an immediate consequence of the existence of PicX/k . Namely, the
line bundle L gives a morphism Y → PicX/k and Y0 is simply the fibre over the zero section of
PicX/k under this morphism. (We use the exact sequence (1); as remarked earlier this does not
require the existence of a rational point on X.)

– 89 –
 Let us now turn to some basic properties of PicX/S in case it is representable. Note that
PicX/S comes with the structure of an S-group scheme, so that the results and definitions of
Chapter 3 apply.

(6.5) Proposition. Assume that f : X → S is proper, flat and of finite presentation, with
geometrically integral fibres. As discussed above, PicX/S is a separated algebraic space over S.
(Those who wish to avoid algebraic spaces might add the hypothesis that f is projective, as in
that case PicX/S is a scheme.)
(i) Write T for the relative tangent sheaf of PicX/S over S. Then the sheaf e∗ T (“the
tangent space of PicX/S along the zero section”) is canonically isomorphic to R1 f∗ OX .
(ii) Assume moreover that f is smooth. Then every closed subscheme Z '→ PicX/S which
is of finite type over S is proper over S.
For a proof of this result we refer to BLR, Chap. 8.

(6.6) Corollary. Let X be a proper variety over a field k.

(i) The tangent space of PicX/S at the identity element is isomorphic to H 1 (X, OX ). Fur-
ther, Pic0X/S is smooth over k if and only if dim Pic0X/S = dim H 1 (X, OX ), and this always holds
if char(k) = 0.
(ii) If X is smooth over k then all connected components of PicX/k are complete.
Proof. This is immediate from (6.5) and the results discussed in Chapter 3 (notably (3.17)
and (3.20)). As we did not prove (6.5), let us here give a direct explanation of why the tangent
space of PicX/S at the identity element is isomorphic to H 1 (X, OX ), and why the components
of PicX/k are complete.
$ %
Let S = Spec k[ε] , where k[ε] is the ring of dual numbers over k. Note that X and XS
have the same underlying topological space. On this space we have a short exact sequence of
sheaves
h ∗ res ∗
0 −→ OX −→ OX S
−−→ OX −→ 1
where h is given on sections by f $→ exp(εf ) = 1 + εf and where res is the natural restriction
map. On cohomology in degree zero this gives the exact sequence

0 −→ k −→ k[ε]∗ −→ k ∗ −→ 1

where the maps are given by f $→ 1 + εf and a + εb $→ a. On cohomology in degree 1 we then

find an exact sequence
h res
0 −→ H 1 (X, OX ) −→ Pic(XS ) −−→ Pic(X) . (3)

Concretely, if γ ∈ H 1 (X, OX ) is represented, on some open covering U = {Uα }α∈A , by a Čech

1-cocyle {fαβ ∈ OX (Uα ∩ Uβ )} then h(γ) is the class of the line bundle on XS which is trivial
on each Uα (now to be viewed as an open subset of XS ) and with transition functions 1 + εfαβ .
Write T for the tangent space of PicX/k at the identity element. We can descibe T as the
kernel of the restriction map PicX/k (S) → PicX/k (k); see Exercise 1.3. If γ ∈ H 1 (X, OX ) then
h(γ) restricts to the trivial class on X. Hence γ defines an element of T , and this gives a linear
map ξ: H 1 (X, OX ) → T . As Pic(S) = {1} it follows from the exact sequences (1) and (3) that
ξ is injective.
So far we have
$ not used % anything about X. To prove that ξ is also surjective it suffices to
show that dim H 1 (X, OX ) = dim(T ). Both numbers do not change if we extend the ground

– 90 –
field. Without loss of generality we may therefore assume that X(k) is non-empty, so that
assumptions (∗) in (6.2) are satisfied. Then the surjectivity of the map ξ follows from the exact
∼
sequence (2). This proves that H 1 (X, OX ) −→ T .
Next let us explain why the components of PicX/S are complete. We already know that
PicX/S is a group scheme, locally of finite type over k. By Propositions (3.12) and (3.17), all
connected components are separated and of finite type over k. To show that they are complete,
we may extend the ground field; hence we can again assume that the assumptions (∗) in (6.2) are
satisfied. In this situation we apply the valuative criterion for properness. Let R be a k-algebra
which is a dvr. Let K be its fraction field, and suppose we have a K-valued point of PicX/k ,
say represented by a line bundle L on XK . We want to show that L extends to a line bundle
on XR . Since X/k is smooth, L is represented by a Weil divisor. But if P ⊂ XK is any prime
divisor then the closure of P inside XR is a prime divisor of XR . It follows that L extends to a
line bundle on XR . !

(6.7) Remark. If char(k) = p > 0 then PicX/k is in general not reduced, even if X is smooth
and proper over k. An example illustrating this will be given in (7.31) below.

(6.8) Let C be a complete curve over a field k. Then PicC/k is a group scheme, locally of finite
type over k; see (6.3). We claim that PicC/k is smooth over k. To see this we may extend
the ground field and assume that C(k) (= ∅, so that the assumptions (∗) in (6.2) are satisfied.
Because PicC/k is locally of finite type over k, it suffices to show that any point of PicC/k with
values in R0 := k[t]/(tn ) can be lifted to a point with values in R := k[t]/(tn+1 ). But if we have
a line bundle L0 on C ⊗k R0 then the obstruction for extending L0 to a line bundle on C ⊗k R
lies in H 2 (C, OC ), which is zero because C is a curve.
In particular, we find that the identity component Pic0C/k is a group variety over k. If in
addition we assume that C is smooth then by Cor. (6.6) Pic0C/k is complete, and is therefore an
abelian variety. In this case we usually write Jac(C) for Pic0C/k ; it is called the Jacobian of C.
Jacobians will be further discussed in Chapter 14. We remark that the term “Jacobian of C”,
for a complete and smooth curve C/k, usually refers to the abelian variety Jac(C) := Pic0C/k
together with its natural principal polarisation.

(6.9) Remark. Suppose X is a smooth proper variety over an algebraically closed field k. Recall
that two divisors D1 and D2 are said to be algebraically equivalent (notation D1 ∼alg D2 ) if
there exist (i) a smooth k-variety T , (ii) codimension 1 subvarieties Z1 , . . . , Zn of X ×k T which
(n
are flat over T , and (iii) points t1 , t2 ∈ T (k), such that D1 − D2 = i=1 (Zi )t1 − (Zi )t2 as
divisors on X; here (Zi )t := Zi ∩ (X × {t}), viewed as a divisor on X. Translating this to line
bundles we find that D1 ∼alg D2 precisely if the classes of L1 = OX (D1 ) and L2 = OX (D2 ) lie
in the same connected component of PicX/k . (Note that the components of the reduced scheme
underlying PicX/k are smooth k-varieties.) The discrete group π0 (PicX/k ) = PicX/k /Pic0X/k is
therefore naturally isomorphic to the Néron-Severi group NS(X) := Div(X)/ ∼alg . For a more
precise treatment, see section (7.24).

§ 2. Digression on graded bialgebras.

In our study of duality, we shall make use of a structure result for certain graded bialgebras.

– 91 –
Before we can state this result we need to set up some definitions.
Let k be a field. (Most of what follows can be done over more general ground rings; for our
purposes the case of a field suffices.) Consider a graded k-module H • = ⊕n!0 H n . An element
x ∈ H • is said to be homogeneous if it lies in H n for some n, in which case we write deg(x) = n.
By a graded k-algebra we shall mean a graded k-module H • together with a unit element 1 ∈ H 0
and an algebra structure map (multiplication) γ: H • ⊗k H • → H • such that
(i) the element 1 is a left and right unit for the multiplication;
(ii) the multiplication γ is associative, i.e., γ(x, γ(y, z)) = γ(γ(x, y), z) for all x, y and z;
(iii) the map γ is a morphism of graded k-modules, i.e., it is k-linear and for all homogeneous
elements x and y we have that γ(x, y) is homogeneous of degree deg(x) + deg(y).
If no confusion arises we shall simply write xy for γ(x, y).
A homomorphism between graded k-algebras H1• and H2• is a k-linear map f : H1• → H2•
which preserves the gradings, with f (1) = 1 and such that f (xy) = f (x)f (y) for all x and y
in H1• .
We say that the graded algebra H • is graded-commutative if

xy = (−1)deg(x)deg(y) yx

for all homogeneous x, y ∈ H • . (In some literature this is called anti-commutativity, or some-
times even commutativity.) The algebra H • is said to be connected if H 0 = k · 1; it is said
to be of finite type over k if dimk (H n ) < ∞ for all n (which is weaker than saying that H • is
finite-dimensional).
If H1• and H2• are graded k-algebras then the graded k-module H1• ⊗k H2• inherits the
structure of a graded k-algebra: for homogeneous elements x, ξ ∈ H1• and y, η ∈ H2• one sets
(x ⊗ y) · (ξ ⊗ η) = (−1)deg(y)deg(ξ) · (xξ ⊗ yη). As an exercise the reader may check that H • is
graded-commutative if and only if the map γ: H • ⊗ H • → H • is a homomorphism of graded
k-algebras. The field k itself shall be viewed as a graded k-algebras with all elements of degree
zero.

(6.10) Definition. A graded bialgebra over k is a graded k-algebra H • together with two
homomorphisms of k-algebras

• • •
µ: H → H ⊗k H called co-multiplication,
•
ε: H → k the identity section,

such that
• • • •
(µ ⊗ id) ◦ µ = (id ⊗ µ) ◦ µ: H → H ⊗k H ⊗k H

and
• •
(ε ⊗ id) ◦ µ = (id ⊗ ε) ◦ µ: H → H

(using the natural identifications H • ⊗k k = H • = k ⊗k H • ).

(6.11) Examples. (i) If all elements of H • have degree zero, i.e., H • = H 0 , then we can ignore
the grading and we “almost” find back the definition of a Hopf algebra as in (3.9). The main
distinction between Hopf algebras and bialgebras is that for the latter we do not require an
antipode.

– 92 –
 (ii) If V is a vector space over k then we can form the exterior algebra ∧• V = ⊕n!0 ∧n V .
The multiplication is given by the “exterior product”, i.e.,

(x1 ∧ · · · ∧ xr ) · (y1 ∧ · · · ∧ ys ) = x1 ∧ · · · ∧ xr ∧ y1 ∧ · · · ∧ ys .

By definition we have ∧0 V = k.
A k-linear map V1 → V2 induces a homomorphism of graded algebras ∧• V1 → ∧• V2 . Fur-
∼
thermore, there is a natural isomorphism ∧• (V ⊕ V ) −→ (∧• V ) ⊗ (∧• V ). Therefore, the di-
agonal map V → V ⊕ V induces a homomorphism µ: ∧• V → ∧• V ⊗ ∧• V . Taking this as
co-multiplication, and defining ε: ∧• V → k to be the projection onto the degree zero component
we obtain the structure of a graded bialgebra on ∧• V .
(iii) If H1• and H2• are two graded bialgebras over k then H1• ⊗k H2• naturally inherits
(
the structure of a graded bialgebra; if a ∈ H1• with µ1 (a) = xi ⊗ ξi and b ∈ H2• with
(
µ2 (b) = yj ⊗ ηj then the co-multiplication µ = µ1 ⊗ µ2 is described by
)
µ(a ⊗ b) = (−1)deg(yj )deg(ξi ) (xi ⊗ yj ) ⊗ (ξi ⊗ ηj ) .
i,j

(iv) Let x1 , x2 , . . . be indeterminates. We give each of them a degree di = deg(xi ) " 1 and we
choose si ∈ Z!2 ∪ {∞}. Then we can define a graded-commutative k-algebra H • = k1x1 , x2 , . . .2
generated by the xi , subject to the conditions xsi i = 0. Namely, we take the monomials

m = xr11 xr22 · · · (ri (= 0 for finitely many i)

as a k-basis, with deg(m) = r1 d1 + r2 d2 + · · ·, and where we set xsi i = 0. Then there is a unique
graded-commutative multiplication law such that γ(xi , xj ) = xi xj for i # j, and with this
multiplication k1x1 , x2 , . . .2 becomes a graded k-algebra. Note that k1x1 , x2 , . . . xN 2 is naturally
isomorphic to k1x1 2 ⊗k · · · ⊗ k1xN 2.
It is an interesting question whether k1x1 , x2 , . . .2 can have the structure of a bialgebra. It
turns out that the existence of such a structure imposes conditions on the numbers di and si .
Let us first do the case of one generator; the case of finitely many generators will follow from
this together with Borel’s theorem to be discussed next. So, we consider a graded k-algebra
H • = k1x | xs = 02 with deg(x) = d > 0. Suppose that H • has the structure of a bialgebra.
Then:
conditions on s:
char(k) = 0, d odd s=2
char(k) = 0, d even s=∞
char(k) = 2 either s = ∞ or s = 2n for some n
char(k) = p > 2, d odd s=2
char(k) = p > 2, d even either s = ∞ or s = pn for some n
For a proof of this result (in fact a more general version of it) we refer to Milnor and Moore [1],
§ 7. Note that the example given in (ii) is of the form k1x1 , x2 , . . .2 where all xi have di = 1 and
si = 2.

(6.12) Theorem. (Borel-Hopf structure theorem) Let H • be a connected, graded-commutative

bialgebra of finite type over a perfect field k. Then there exist graded bialgebras Hi• (i = 1, . . . , r)
and an isomorphism of bialgebras

H ∼
• • •
= H1 ⊗ k · · · ⊗ k Hr

– 93 –
such that the algebra underlying Hi• is generated by one element, i.e., the algebras Hi• are of
the form k1xi | xsi i = 02, with deg(xi ) = di > 0.
For a proof of this result, which is due to A. Borel, we refer to Borel [1] or Milnor and
Moore [1].

(6.13) Corollary. Let H • be as in (6.12). Assume there is an integer g such that H n = (0) for
all n > g. Then dimk (H 1 ) # g. If dimk (H 1 ) = g then H • ∼
= ∧ H 1 as graded bialgebras.
•

Proof. Decompose H • = H1• ⊗k · · · ⊗k Hr• as in (6.12). Note that dimk (H 1 ) equals the number
of generators xi such that di = 1. Now x1 · · · xr (:= x1 ⊗ · · · ⊗ xr ) is a nonzero element of
H • of degree d1 + · · · + dr . Therefore d1 + · · · + dr # g, which implies dimk (H 1 ) # g. Next
suppose dimk (H 1 ) = g, so that all generators xi have degree 1. If x2i (= 0 for some i then
x1 · · · xi−1 x2i xi+1 · · · xg is a nonzero element of degree g + 1, contradicting our assumptions.
Hence x2i = 0 for all i which means that H • ∼ = ∧ H 1.
•
!

(6.14) Let us now turn to the application of the above results to our study of abelian varieties.
Given a g-dimensional variety X over a field k, consider the graded k-module

g
*
H n (X, OX ) .
• •
H = H (X, OX ) :=
n=0

Cup-product makes H • into a graded-commutative k-algebra, which is connected since X is

connected.
In case X is a group variety the group law induces on H • the structure of a graded bialgebra.
Namely, via the Künneth formula H • (X ×k X, OX×X ) ∼ • •
= H (X, OX ) ⊗k H (X, OX ) (which is an
isomorphism of graded k-algebras), the group law m: X ×k X → X induces a co-multiplication

• • •
µ: H → H ⊗k H .

For the identity section ε: H • → k we take the projection onto the degree zero component, which
can also be described as the map induced on cohomology by the unit section e: Spec(k) → X.
Now the statement that these µ and e make H • into a graded bialgebra over k becomes a simple
translation of the axioms in (1.2) satisfied by m and e.
As a first application of the above we thus find the estimate dimk (H 1 (X, OX )) # g for a
g-dimensional group variety X over a field k. (Note that dimk (H 1 (X, OX )) does not change
if we pass from k to an algebraic closure; we therefore need not assume k to be perfect.) For
abelian varieties we shall prove in (6.18) below that we in fact have equality.
We summarize what we have found.

(6.15) Proposition. Let X be a group variety over a field k. Then H • (X, OX ) has a natural
structure of a graded k-bialgebra. We have dimk (H 1 (X, OX )) # dim(X).

To conclude this digression on bialgebras, let us introduce one further notion that will be
useful later.

(6.16) Definition. Let H • be a graded bialgebra with comultiplication µ: H • → H • ⊗k H • .

Then an element h ∈ H • is called a primitive element if µ(h) = h ⊗ 1 + 1 ⊗ h.

– 94 –
(6.17) Lemma. Let V be a finite dimensional k vector space, and consider the exterior algebra
∧• V as in (6.11). Then V = ∧1 V ⊂ ∧• V is the set of primitive elements in ∧• V .
Proof. We follow Serre [1]. Since the co-multiplication µ is degree-preserving, an element of a
graded bialgebra H • is primitive if and only if all its homogeneous components are primitive.
Thus we may restrict our attention to homogeneous elements of ∧• V .
It is clear that the non-zero elements of ∧0 V = k are not primitive. Further we see directly
from the definitions that the elements of ∧1 V = V are primitive. Let now y ∈ ∧n V with n " 2.
Write *
[(∧ V ) ⊗ (∧ V )]n = ∧p V ⊗ ∧q V ,
• •

p+q=n
(
and write µ(y) = µ(y)p,q with µ(y)p,q ∈ ∧p V ⊗ ∧q V . For instance, one easily finds that
µ(y)n,0 = y = µ(y)0,n via the natural identifications ∧n V ⊗ k = ∧n V = k ⊗ ∧n V . Similarly, we
find that the map y $→ µ(y)1,n−1 is given (on decomposable tensors) by

n
)
v1 ∧ · · · ∧ vn $→ (−1)i+1 vi ⊗ (v1 ∧ · · · ∧ v+i ∧ · · · ∧ vn ) .
i=1

It follows that for λ ∈ V ∗ the composition ∧n V → V ⊗ ∧n−1 V → ∧n−1 V given by y $→

(λ ⊗ id)(µ(y)1,n−1 ) is just the interior contraction y $→ y $ λ. The assumption that y is primitive
and n " 2 implies that µ(y)1,n−1 = 0 so we find y $ λ = 0 for all λ ∈ V ∗ . This only holds for
y = 0. !

§ 3. The dual of an abelian variety.

From now on, let π: X → S = Spec(k) be an abelian variety over a field k. We shall admit from
the general theory that PicX/k is a group scheme over k with projective connected components.
One of the main results of this section is that Pic0X/k is reduced, and is therefore again an abelian
variety.
Note that PicX/k also represents the functor PX/k,0 of line bundles with rigidification along
the zero section. As above, the identification between the two functors is given by sending the
class of a line bundle L on X ×k T to the class of L ⊗ pr∗T e∗ L−1 with its canonical rigidification
along {0} × T . (In order to avoid the notation 0∗ L we write e for the zero section of XT .)
In particular, we have a Poincaré bundle P on X ×k PicX/k together with a rigidification
∼
α: OPicX/k −→ P|{0}×PicX/k .
If L is a line bundle on X we have the associated Mumford bundle Λ(L) on X × X. In
order to distinguish the two factors X, write X (1) = X × {0} and X (2) = {0} × X. Viewing
Λ(L) as a family of line bundles on X (1) parametrised by X (2) we obtain a morphism

ϕL : X = X (2) −→ PicX/k

which is the unique morphism with the property that (idX × ϕL )∗ P ∼ = Λ(L). On points, the
∗ −1
morphism ϕL is of course given by x $→ [tx L ⊗ L ], just as in (2.10). We have seen in (2.10),
as a consequence of the Theorem of the Square, that ϕL is a homomorphism. Further we note
that ϕL factors through Pic0X/k , as X is connected and ϕL (0) = 0.

– 95 –
(6.18) Theorem. Let X be an abelian variety over a field k. Then Pic0X/k is reduced, hence it
is an abelian variety. For every ample line bundle L the homomorphism ϕL : X → Pic0X/k is an
isogeny with kernel K(L). We have dim(Pic0X/k ) = dim(X) = dimk H 1 (X, OX ).
Proof. Let L be an ample line bundle on X. By Lemma (2.17), ϕL has kernel K(L). Since
K(L) is a finite group scheme it follows that dim(Pic0X/k ) " dim(X). Combining this with (6.6)
and (6.15) we find that dim(Pic0X/k ) = dim(X) = dimk H 1 (X, OX ) and that Pic0X/k is re-
duced. !

(6.19) Definition and Notation. The abelian variety X t := Pic0X/k is called the dual of X.
We write P, or PX , for the Poincaré bundle on X × X t (i.e., the restriction of the Poincaré
bundle on X × PicX/k to X × X t ). If f : X → Y is a homomorphism of abelian varieties over k
then we write f t : Y t → X t for the induced homomorphism, called the dual of f or the transpose
of f . Thus, f t is the unique homomorphism such that

(id × f t )∗ PX ∼
= (f × id)∗ PY

as line bundles on X × Y t with rigidification along {0} × Y t .

(6.20) Remark. We do not yet know whether f $→ f t is additive; in other words: if we have
two homomorphisms f , g: X → Y , is then (f + g)t equal to f t + g t ? Similarly, is (nX )t equal to
nX t ? We shall later prove that the answer to both questions is “yes”; see (7.17). Note however
that such relations certainly do not hold on all of PicX/k ; for instance, we know that if L is a
line bundle with (−1)∗ L ∼= L then n∗ L ∼
2
= Ln which is in general not isomorphic to Ln .

Exercises.
(6.1) Show that the functor PX/S defined in §1 is never representable, at least if we assume X
to be a non-empty scheme.
(6.2) Let X and Y be two abelian varieties over a field k.
(i) Write iX : X → X × Y and iY : Y → X × Y for the maps given by x $→ (x, 0) and y $→ (0, y),
respectively. Show
$ that the map (i%tX , itY ): (X × Y )t → X t × Y t that sends a class [L] ∈
0
Pic(X×Y )/k to [L|X×{0} ], [L|{0}×Y ] , is an isomorphism. [Note: in general it is certainly
not true that the full Picard scheme PicX×Y /k is isomorphic to PicX/k × PicY /k .]
(ii) Write

p: X × Y × X t × Y t −→ X × X t and q: X × Y × X t × Y t −→ Y × Y t

for the projection maps. Show that the Poincaré bundle of X × Y is isomorphic to p∗ PX ⊗
q ∗ PY .
(6.3) Let L be a line bundle on an abelian variety X. Consider the homomorphism (1, ϕL ): X →
X × X t . Show that (1, ϕL )∗ PX ∼
= L ⊗ (−1)∗ L.
(6.4) The goal of this exercise is to prove the restrictions listed in (iv) of (6.11). We consider a
graded bialgebra H • over a field k, with co-multiplication µ. We define the height of an element
x ∈ H • to be the smallest positive integer n such that xn = 0, if such an n exists, and to be ∞
if x is not nilpotent.

– 96 –
 (i) If y ∈ H • is an element of odd degree, and char(k)$ % (= 2, show that y 2 = 0.
( n
(ii) If x ∈ H • is primitive, show that µ(xn ) = i=0 ni xi ⊗ xn−i . Conclude that if x has height
n < ∞ then char(k) = p > 0 and n is a power of p.
(iii) If H • = k1x | xs = 02 with deg(x) = d, show that x is a primitive element. Deduce the
restrictions on the height of x listed in (iv) of (6.11).

– 97 –
Chapter VII. Duality.

§ 1. Formation of quotients and the descent of coherent sheaves.

(7.1) Definition. Let S be a base scheme. Let ρ: G ×S X → X be an action (from the left) of
an S-group scheme G on an S-scheme X. Let F be a coherent sheaf of OX -modules. Then an
∼
action of G on F , compatible with the action ρ, is an isomorphism λ: pr∗2 F −→ ρ∗ F of sheaves
on G ×S X, such that on G ×S G ×S X we have a commutative diagram
pr∗ (λ)
pr∗3 F 23
−−−−−→ pr∗23 ρ∗ F
 

(m×idX )∗ (λ)"
(id ×ρ)∗ (λ)
" G
(m × idX )∗ ρ∗ F === (idG × ρ)∗ ρ∗ F .

Here is a more concrete explanation of what this means. If T is an S-scheme and g ∈ G(T ),
write ρg : XT → XT for the action of the element g. Then to have an action of G on F
that is compatible with ρ means that for every g ∈ G(T ) we have an isomorphism of sheaves
∼
λg : FT −→ ρ∗g FT such that λgh = ρ∗h (λg ) ◦ λh for all g, h ∈ G(T ).
If F is a locally free OX -module we can take a more geometric point of view. First recall
that a locally free OX -module is “the same” as a geometric vector bundle over X. Namely,
V := V(F ∨ ) is a geometric vector bundle over X, and F is the sheaf of sections of the structure
morphism π: V → X. Then a ρ-compatible G-action on F corresponds to an action ρ̃: G×S V →
V such that (i) the structure morphism π: V → X is G-equivariant, and (ii) the action ρ̃ is
“fibrewise linear”, meaning that for every S-scheme T and every g ∈ G(T ), x ∈ X(T ), the
isomorphism ρ̃(g): Vx → Vgx is OT -linear. We refer to such an action ρ̃ as a lifting of ρ.
With this notion of a G-action on a sheaf, we can formulate a useful result on the descent
of modules.

(7.2) Proposition. Let ρ: G ×S X → X be an action of an S-group scheme G on an S-

scheme X. Suppose there exists an fppf quotient p: X → Y of X by G. If F is a coherent
∼
sheaf of OY -modules then the canonical isomorphism λcan : pr∗2 (p∗ F ) −→ ρ∗ (p∗ F ) defines a ρ-
compatible G-action on p∗ F . The functor F %→ (p∗ F, λcan ) gives an equivalence between the
category of coherent OY -modules and the category of coherent OX -modules with (ρ-compatible)
G-action. This restricts to an equivalence between the category of finite locally free OY -modules
and the category of finite locally free OX -modules with G-action.
This proposition should be seen as a statement in (faithfully flat) descent theory; it follows
for instance from the results of SGA 1, Exp. VIII, § 1. (See also [BLR], § 6.1, Thm. 4.) Given
such results in descent theory, the only point here is that a ρ-compatible G-action on a coherent
OX -module is the same as a descent datum on this module. (Recall that we have an isomorphism
∼
(ρ, pr2 ): G ×S X −→ X ×Y X.) The assertion that finite locally free OX -modules with G-action
give rise to finite locally free OY -modules follows from EGA IV, Prop. 2.5.2.

DualAV2, 15 september, 2011 (812)

– 98 –
(7.3) Example. We consider the situation of the proposition. The geometric vector bundle
corresponding to the structure sheaf OX is just the affine line A1X over X.
On OX (geometrically: on A1X ) we have a “trivial” action ρ̃triv , given by
ρ̃triv = ρ × idA1S : G ×S A1X = G ×S X ×S A1S −→ X ×S A1S = A1X .
The OY -module corresponding to (OX , ρ̃triv ) is just OY itself.
Let ρ̃ be some other lifting of ρ to a G-action on A1X . Let T be an S-scheme and g ∈
G(T ). The automorphism ρ̃(g) · ρ̃triv (g)−1 of A1X ×S T = A1XT is given on every fibre A1x
by some (invertible) scalar multiplication. This means that ρ̃(g) · ρ̃triv (g)−1 is given by an
∗
element ν(g) ∈ Γ(XT , OX T
). We find that an action ρ̃ gives rise to a morphism of functors
ν: G → ResX/S Gm,X on the category#Sch $ /S . The#condition that # $ ρ̃ is a group action means that
ν satisfies
# $ a cocycle condition ν(g g
1 2 ) x = ν(g 1 ) g 2 x) · ν(g 2 x , where we simply write g2 x for
)
ρ(g2 ) x . Conversely, given a morphism ν: G → ResX/S Gm,X that satisfies this condition, one
finds back a G-action ρ̃ by ρ̃(g) = ν(g) · ρ̃triv (g).
Now suppose that the structure morphism f : X → S satisfies f∗ (OXT ) = OT for all S-
schemes T . This holds for instance if X is a proper variety over a field. Then ResX/S Gm,X ∼ =
Gm,S as functors on Sch/S . In particular, any morphism ν: G → #ResX/S Gm,X #is $G-invariant,
in the sense that for all g1 , g2 ∈ G(T ) and x ∈ X(T ) we have ν(g1 ) g2 x) = ν(g1 ) x . Hence the
cocycle condition in this case just says that ν is a homomorphism. So the conclusion is that the
liftings ρ̃ of ρ to a G-action on A1X are in bijective correspondence with Hom /S
(G, Gm ). In
case G is a commutative, finite locally free S-group scheme this is just the Cartier dual GD (S).
Via Proposition (7.2), we can use this to obtain a description of the line bundles L on Y
such that p∗ L ∼
= OX . The result is as follows.

(7.4) Proposition. Let G be a commutative, finite locally free S-group scheme. Let ρ: G ×S
X → X be a free action of G on an S-scheme X. Let p: X → Y be the quotient of X by G.
Suppose that f∗ (OXT ) = OT for all S-schemes T . Then for any S-scheme T there is a canonical
isomorphism of groups
% &
isomorphism classes of line bundles ∼
δT : −→ GD (T ) ,
L on YT with p∗ L ∼= O XT
and this isomorphism is compatible with base change T & → T .
Proof. To define δT for arbitrary S-schemes T we may replace S by T and p: X → Y by
pT : XT → YT . Note that by Theorem (4.16) and what was explained in Example (4.29), pT is
again the quotient morphism of XT by the action of GT , and of course also the assumption that
f∗ (OXT ) = OT for all S-schemes T is preserved under base change. Hence it suffices to define
the isomorphism δS .
∼
Let L be a line bundle on Y with p∗ L ∼ = OX . Via the choice of an isomorphism α: p∗ L −→
∼
OX (or, more geometrically, the isomorphism α: p∗ V(L−1 ) −→ A1X over X) the canonical G-
action on p∗ L translates into a G-action ρ̃ on A1X , and as explained above this gives us a character
ν: G → Gm,S . We claim that this character is independent of the choice of α. In general, any
∼
other isomorphism p∗ L −→ OX is of the form α& = γ ◦ α for some γ ∈ Γ(X, OX ∗
). Write ρ̃ and ρ̃&
for the G-actions on A1X obtained using α and α& , respectively, and let ν and ν & be the associated
characters. If g ∈ G(T ) and y is a T -valued point of p∗ V(L−1 ) lying over x ∈ X(T ) then we
have the relations
# $ # $ # $ # $
ρ̃triv g, α& (y) = γ(x) · ρ̃triv g, α(y) and ρ̃& g, α& (y) = γ(gx) · ρ̃ g, α(y) ,

– 99 –
 ∗
where γ(x) is the image of γ under the homomorphism Γ(X, #OX ) → $Γ(T, OT∗ ) induced by x: T →
X, and similarly for γ(gx). (Note that elements such as ρ̃ g, α(y) are T -valued points of A1X
lying over the point gx ∈ X(T ), and on such elements we have the “fibrewise” multiplication by
functions on T .) But now our assumption that f∗ (OX ) = OS implies that γ is the pull-back of
an element in Γ(S, OS∗ ), so γ(x) = γ(gx). Hence ν = ν & , as claimed.
Now we can simply apply the conclusion from (7.3), and define δS as the map that sends the
isomorphism class of L to the character ν: G → Gm,S given on points by ν(g) = ρ̃(g) · ρ̃triv (g)−1 .
By Proposition (7.2), together with what was explained in Example (7.3), the map δS thus
obtained is indeed an isomorphism.
Finally we note that the maps δT are indeed compatible with base change, as is immediate
from the construction. !

§ 2. Two duality theorems.

(7.5) Theorem. Let f : X → Y be an isogeny of abelian varieties. Then f t : Y t → X t is again

an isogeny and there is a canonical isomorphism of group schemes
∼
Ker(f )D −→ Ker(f t ) .

Proof. If T is a k-scheme, any class in Ker(f t )(T ) is uniquely represented by a line bundle L
on YT such that f ∗ L ∼= OXT . Indeed, if L& represents a class in Ker(f t )(T ) then there is a line
bundle M on T such that f ∗ L& ∼ = pr∗T M . Then L := L& ⊗ pr∗T M −1 represents the same class
& ∗ ∼
as L and satisfies f L = OXT . Conversely, if L1 and L2 represent the same class then they
differ by a line bundle of the form pr∗T M ; hence f ∗ L1 ∼
= f ∗ L2 implies L1 ∼
= L2 .
∼
Applying Proposition (7.4) we obtain the desired isomorphism Ker(f t ) −→ Ker(f )D . In
particular this shows that f t has a finite kernel and therefore is again an isogeny. !

(7.6) Proposition. Let f : X → Y be a homomorphism. Let M be a line bundle on Y and

write L = f ∗ M . Then ϕL : X → X t equals the composition

f ϕM ft
X −→ Y −−→ Y t −−→ X t .

If f is an isogeny and M is non-degenerate then L is non-degenerate too, and rank(K(L)) =

deg(f )2 · rank(K(M )).
Proof. That ϕL = f t ◦ ϕM ◦ f is clear from the formula t∗x f ∗ M = f ∗ t∗f (x) M . For the second
assertion recall that a line bundle L is non-degenerate precisely if ϕL is an isogeny, in which
case rank(K(L)) = deg(ϕL ). Now use (7.5). !

(7.7) The Poincaré bundle on X × X t comes equipped with a rigidification along {0} × X t . As
P|X×{0} ∼= OX we can also choose a rigidification of P along X × {0}. Such a rigidification is
∗
unique up to an element of Γ(X, OX ) = k ∗ . Hence there is a unique rigidification along X × {0}
such that the two rigidifications agree at the origin (0, 0).
Now we view P as a family of line bundles on X t parametrised by X. This gives a morphism

κX : X −→ X tt .

– 100 –
As κX (0) = 0 it follows from Prop. (1.13) that κX is a homomorphism.

(7.8) Lemma. Let L be a line bundle on X. Then ϕL = ϕtL ◦ κX : X → X t .

Proof. Let s: X × X → X × X and s: X × X t → X t × X be the morphisms switching the two
factors; on points: s(x, y) = (y, x). We have a canonical isomorphism s∗ Λ(L) ∼
= Λ(L). Let T be
a k-scheme and x ∈ X(T ). Writing [M ] for the class of a bundle M on X × T in Pic0X/k (T ) we
have
' (
id×x
ϕL (x) = (X × T −−−→ X × X)∗ Λ(L)
' (
id×x s
= (X × T −−−→ X × X −→ X × X)∗ Λ(L)
' id×ϕL
(
id×x s t ∗
= (X × T −−−→ X × X −→ X × X −−−−→ X × X ) P
' ϕL ×id
(
id×x s
= (X × T −−−−→ X t × T −−−→ X t × X −→ X × X t )∗ P = ϕtL ◦ κX (x) .

As this holds for all T and x the lemma is proven. !

(7.9) Theorem. Let X be an abelian variety over a field. Then the homomorphism κX : X −→
X tt is an isomorphism.
Proof. Choose an ample line bundle L on X. The formula ϕL = ϕtL ◦ κX shows that Ker(κX ) is
finite; hence κX is an isogeny. Furthermore,
# $ # $
rank K(L) = deg(ϕL ) = deg(ϕtL ) · deg(κX ) = rank K(L)D · deg(κX ) ,
# $ # $
using (7.5). But rank K(L)D = rank K(L) , so κX has degree 1. !

(7.10) Corollary. If L is a non-degenerate line bundle on X then K(L) ∼

= K(L)D .
Proof. Apply (7.5) to ϕL and use (7.8) and (7.9). !

§ 3. Further properties of Pic0X/k .

Let X be an abelian variety over a field k. A line bundle L on X gives rise to a homo-
morphism ϕL : X → X t . We are going to extend this construction to a more general situation.
Namely, let T be a k-scheme, and suppose L is a line bundle on XT := X ×k T . We are going
to associate to L a homomorphism ϕL : XT → XTt .
As usual we write Λ(L) := m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1 for the Mumford bundle on XT ×T XT
associated to L. (Note that we are working in the relative setting, viewing T as the base scheme.
If we rewrite XT ×T XT as X ×k X ×k T then Λ(L) becomes (m × idT )∗ L ⊗ p∗13 L−1 ⊗ p∗23 L−1 .) In
(1) (2)
order to distinguish the two factors XT , let us write XT = XT ×T e(T ) and XT = e(T ) ×T XT
where e(T ) ⊂ XT is the image of the zero section e: T → XT . Viewing Λ(L) as a family of line
(1) (2)
bundles on XT parametrized by XT we obtain a morphism
(2)
ϕL : XT = XT −→ PicXT /T = PicX/k ×k T .

As ϕL (0) = 0 and the fibres Xt are connected, ϕL factors through XTt = Pic0X/k ×k T .

– 101 –
(7.11) Lemma. (i) The morphism ϕL only depends on the class of L in PicX/k (T ).
(ii) Let f : T → S be a morphism of k-schemes. If M is a line bundle on XS and L =
(idX × f )∗ M on XT , then ϕL : XT → XTt is the morphism obtained from ϕM : XS → XSt by
pulling back via f on the basis.
(iii) The morphism ϕL : XT → XTt is a homomorphism.
Part (i) of the lemma will be sharpened in (7.15) below. As a particular case of (ii), note
that the fibre of ϕL above a point t ∈ T is just ϕLt , where we write Lt for the restriction of L
to X × {t}.
Proof. (i) If L1 and L2 have the same class then they differ by a factor pr∗T M . But then Λ(L1 )
and Λ(L2 ) differ by a factor π ∗ M −1 , where π: XT ×T XT → T is the structural morphism. This
implies that ϕL1 = ϕL2 , as claimed.
(ii) This readily follows from the definitions.
(iii) The assertion that ϕL is a homomorphism means that we have an equality of two
morphisms
ϕL ◦ m = m ◦ (ϕL × ϕL ): XT ×T XT −→ XTt .

For every t ∈ T we already know that the two morphisms agree on the fibres above t. Hence
the lemma is true if T is reduced. In particular, the lemma is true in the “universal” case that
T = PicX/k and L is the Poincaré bundle on X ×k PicX/k . In the general case, consider the
morphism f : T → PicX/k associated to the line bundle L. This morphism is characterized by the
property that L and (id × f )∗ P have the same class in PicX/k (T ). Now apply (i) and (ii). !
In the above we allow L—to be thought of as a family of line bundles on X parametrized
by T —to be non-constant. But the abelian variety we work on is a constant one. We can go
one step further by also letting the abelian varieties Xt “vary with t”. This generalization will
be discussed in Chapter ??; see in particular (?.?).
We write K(L) := Ker(ϕL ) ⊂ XT . It is the maximal subscheme of XT over which Λ(L) is
trivial, viewing XT ×T XT as a scheme over XT via the second projection. In particular, ϕL = 0
if and only if Λ(L) is trivial over XT , meaning that Λ(L) = pr∗2 M for some line bundle M on XT .
Using (2.17) we can make this a little more precise.

(7.12) Lemma. Let T be a locally noetherian k-scheme. Write π: XT ×T XT → T for the

structural morphism. For a line bundle L on XT , consider the following conditions.
(a) ϕL = 0.
(b) Λ(L) ∼
= pr∗2 M for some line bundle M on XT .
(c) Λ(L) ∼
= π ∗ N for some line bundle N on T .
(d) ϕLt = 0 for some t ∈ T .
Then (a) ⇔ (b) ⇔ (c) ⇒ (d), and if T is connected then all four conditions are equivalent. If
these equivalent conditions are satisfied then N ∼
= e∗ L−1 and M = pr∗T N .
Proof. The implications (d) ⇐ (a) ⇔ (b) ⇐ (c) are clear. Let us write XT ×T XT as X ×k X ×k T .
In this notation we have Λ(L) = (m × idT )∗ L ⊗ p∗13 L−1 ⊗ p∗23 L−1 and π becomes the projection
onto the third factor. Set N := e∗ L−1 . We find that

Λ(L)|{0}×X×T ∼
= pr∗T N ∼
= Λ(L)|X×{0}×T

as line bundles on X × T .

– 102 –
 Suppose T is connected and ϕLt = 0 for some t ∈ T . Then

Λ(L)|X×X×{t} ∼
= OX×X×{t}

by (iii) of (2.17). By Thm. (2.5) the line bundle Λ(L) ⊗ p∗3 N −1 on X × X × T is trivial, i.e.,
Λ(L) ∼= π ∗ N . This shows that (d) ⇒ (a) for connected T . For arbitrary T we get the implication
(a) ⇒ (c) by applying the previous to each of its connected components.
The last assertion of the lemma is obtained by restricting Λ(L) to {0} × {0} × T and to
{0} × X × T . !

(7.13) Fact. Let X and Y be two projective varieties over a field k. Then the contravariant
functor
Hom (X, Y ): (Sch/k ) → Sets given by T %→ Hom /T (XT , YT )
is representable by a k-scheme, locally of finite type.
This fact is a consequence of the theory of Hilbert schemes. A reference is ??. Note that in
this proof the projectivity of X and Y is used in an essential way. See also Matsumura-Oort [1]
for related results for non-projective varieties.

(7.14) Proposition. Let X and Y be two abelian varieties over a field k. Then the functor

Hom (X, Y ): (Sch/k ) → Ab given by T %→ Hom /T

(XT , YT )

is representable by an étale commutative k-group scheme.

Proof. Let H = Hom (X, Y ) and H & = Hom (X × X, Y ). Let f : XH → YH be the universal
morphism. Consider the morphism g: (X × X)H → YH given on points by g(x1 , x2 ) = f (x1 +
x2 ) − f (x1 ) − f (x2 ). Consider also the “trivial” morphism e: (X × X)H → YH given on points
by e(x1 , x2 ) = eY . Then g and e are H-valued points of H & ; in other words, they correspond to
morphisms ψg , ψe : H → H & . The functor Hom (X, Y ) is represented by the subscheme of H
given by the condition that ψg = ψe ; in other words, it is given by the cartesian diagram

Hom (X, Y ) −−−−−→ H&

 
 ∆ #
" " H /k
(ψg ,ψe )
H −−−−−→ H & ×k H & .

To get a group scheme structure on Hom (X, Y ) we just note that Hom (X, Y ) is natu-
rally a group functor; now apply (3.6).
It remains to be shown that Hom (X, Y ) is an étale group scheme. We already know it is
locally of finite type over k, so it suffices to show that its tangent space at the origin is trivial.
It suffices to prove this in the special case that Y = X, for Hom (X, Y ) embeds as a closed
subgroup scheme of E nd# (X ×$ Y ) := Hom (X × Y, X × Y ) by sending f : X → Y to the
endomorphism (x, y) %→ 0, f (x) of X × Y .
A tangent vector
# $ of E nd (X) at the point idX is the same as a homomorphism ξ: Xk[ε] →
Xk[ε] over Spec k[ε] that reduces to the identity modulo ε. Note that ξ is necessarily an
automorphism. (It is the identity on underlying topological spaces, and it is an easy exercise to
show that ξ gives an automorphism of the structure sheaf.) Hence by the results in Exercise (1.3),
ξ corresponds to a global vector field Ξ on X. As we know, the global vector fields on X are

– 103 –
precisely the translation-invariant vector fields. On the other hand, a necessary condition for ξ
to be an endomorphism is that it maps the identity section of Xk[ε] to itself. This just means
that Ξ(eX ) = 0. Hence Ξ is the trivial vector field. This shows that idX has non non-trivial
first order deformations. !
In line with the notational conventions introduced in (1.16), we shall usually simply write
Hom(X, Y ) for the group scheme of homomorphisms from X to Y . If we wish to refer to the
bigger scheme of arbitrary scheme morphisms from X to Y , or if there is a risk of confusion, we
shall use a subscript “AV” or “Sch” to indicate which of the two we mean.
By (i) and (ii) of Lemma (7.11), L %→ ϕL gives rise to a morphism of functors ϕ: PicX/k →
Hom(X, X t ). If L and M are line bundles on XT then Λ(L ⊗ M ) ∼ = Λ(L) ⊗ Λ(M ) and we find
that ϕL⊗M = ϕL + ϕM . Summing up, we obtain a homomorphism of k-group schemes

ϕ: PicX/k → Hom(X, X t ) .

(7.15) Lemma. Let T be a connected k-scheme. Let L be a line bundle on XT . Write Lt

for L|X×{t} . Then for any two k-valued points s, t ∈ T (k) we have ϕLs = ϕLt . In particular,
Pic0X/k ⊂ Ker(ϕ).
Proof. By (d) ⇒ (a) of (7.12), applied with T = X t and with L = P the Poincaré bundle, we
find that X t = Pic0X/k ⊂ Ker(ϕ). As ϕ is a homomorphism, it is constant on the connected
components of PicX/k .
Let f : T → PicX/k be the morphism corresponding to L; it factors through some connected
component C ⊂ PicX/k . Let M := P|X×C be the restriction of the Poincaré bundle to X × C.
Using (i) and (ii) of (7.11) we find that ϕL : XT → XTt is obtained from ϕM : XC → XCt by
pulling back via f on the basis. But by the above, ϕMf (s) = ϕMf (t) . !

(7.16) Lemma. Let X be an abelian variety over k. Let T be a k-scheme and let L be a line
bundle on XT such that ϕL = 0.
(i)
) If Y is a* T -scheme
) ∗ then* for any two morphisms f , g: Y → XT of schemes over T we
have (f + g) L = f L ⊗ g ∗ L in PicY /T (T ).
∗

(ii) For n ∈ Z we have [n∗ L] = [Ln ] in PicX/k (T ).

Proof. If ϕL = 0 then Λ(L) = π ∗ N for some line bundle N on T . Pulling back via (f, g): Y →
XT ×T XT gives (f + g)∗ L = f ∗ L ⊗ g ∗ L ⊗ π ∗ N , where π: Y → T is the structural morphism.
But π ∗ N is trivial in PicY /T (T ), so we get (i). Applying this with f = idXT and g = nXT gives
the relation [(n + 1)∗ L] = [L ⊗ n∗ L]. By double induction on n, starting with the cases n = 0
and n = 1, we obtain (ii). !
Using that Pic0X/k ⊂ Ker(ϕ) we obtain a positive answer to the questions posed in (6.20).

(7.17) Corollary. Let X and Y be abelian varieties over k. Then the map Hom(X, Y ) →
Hom(Y t , X t ) given on points by f %→ f t is a homomorphism of k-group schemes. For all n ∈ Z
we have (nX )t = nX t .
Combining this last result with (7.5) we find that X t [n] is canonically isomorphic to the
Cartier dual of X[n], for every n ∈ Z>0 .

(7.18) Let X be an abelian variety. We call a homomorphism f : X → X t symmetric if f = f t ,

∼
taking the isomorphism κX : X −→ X tt of (7.9) as an identification. It follows from the previous

– 104 –
corollary that the functor of symmetric homomorphisms X → X t is represented by a closed
subgroup scheme
Hom symm (X, X t ) ⊂ Hom(X, X t ) .
In fact, Hom symm (X, X t ) is just the kernel of the endomorphism of Hom(X, X t ) given by f %→
f − f t.
By Lemma (7.8), the homomorphism ϕ: PicX/k → Hom(X, X t ) factors through the sub-
group Hom symm (X, X t ). (Because Hom(X, X t ) is étale, it suffices to know that ϕ maps into
Hom symm for points with values in a field.)
Our next goal is to show that not only Pic0X/k ⊂ Ker(ϕ) but that the two are in fact equal.
First we prove a lemma about the cohomology of line bundles L with ϕL = 0. Note that we are
here again working over a field; this lemma has no straightforward generalization to the relative
setting.

(7.19) Lemma. Let L be a line bundle on X with ϕL = 0. If L ,∼

= OX then H i (X, L) = 0 for
all i.
Proof. First we treat the group H 0 (X, L). If there is a non-trivial section s then (−1)∗ s is a
non-trivial section of (−1)∗ L ∼= L−1 ; so both L and L−1 have a non-trivial section, and this
implies that L is trivial. Since we have assumed this is not the case, H 0 (X, L) = {0}.
Let now i " 1 be the smallest positive integer such that H i (X, L) ,= 0. Consider the
composition
m
X → X × X −→ X , given by x %→ (x, 0) %→ x .
On cohomology this induces the maps

H i (X, L) → H i (X × X, m∗ L) → H i (X, L) ,

the composition of which is the identity. But since m∗ L ∼= p∗1 L ⊗ p∗2 L, the Künneth formula
gives +
H i (X × X, m∗ L) ∼
= H i (X × X, p∗1 L ⊗ p∗2 L) ∼
= H a (X, L) ⊗ H b (X, L) .
a+b=i
0
Since H (X, L) = {0} we may consider only those terms in the RHS where a " 1 and b " 1.
But then a < i which by our choice of i implies that H a (X, L) = 0. This shows that the identity
map on H i (X, L) factors via zero. !
In the proof of the next proposition we need some facts about cohomology and base change.
Here is what we need.

(7.20) Fact. Let f : X → Y be a proper morphism of noetherian schemes, with Y reduced and
connected. Let F be a coherent sheaf of OX -modules on X.
(i) If y %→ dimk(y) H q (Xy , Fy ) is a constant function on Y then Rq f∗ (F ) is a locally free sheaf
on Y , and for all y ∈ Y the natural map Rq f∗ (F ) ⊗OY k(y) → H q (Xy , Fy ) is an isomorphism.
(ii) If Rq f∗ (F ) = 0 for all q " q0 then H q (Xy , Fy ) = 0 for all y ∈ Y and q " q0 .
A proof of this result can be found in [MAV], § 5.

(7.21) Proposition. Let X be an abelian variety over an algebraically closed field k. Let L
be an ample line bundle on X and M a line bundle with ϕM = 0. Then there exists a point
x ∈ X(k) with M ∼= t∗x L ⊗ L−1 .

– 105 –
Proof. We follow Mumford’s beautiful proof. The idea is to look at the cohomology on X × X
of the line bundle
K := Λ(L) ⊗ p∗2 M −1 .
The projections p1 , p2 : X × X → X give rise to two Leray spectral sequences

E2p,q = H p (X, Rq p1,∗ (K)) ⇒ H p+q (X × X, K)

and
E2&p,q = H p (X, Rq p2,∗ (K)) ⇒ H p+q (X × X, K) .

The restrictions of K to the horizontal and vertical fibres are given by

K|{x}×X ∼
= t∗x L ⊗ L−1 ⊗ M −1 ,
K|X×{x} ∼
= t∗ L ⊗ L−1 .
x

Assume that there is no x ∈ X(k) such that t∗x L⊗L−1 ∼ = M . It then follows that K|{x}×X is
a non-trivial bundle in Ker(ϕ) for every x. (Note that [t∗x L ⊗ L−1 ] = ϕL (x) ∈ Pic0X/k ⊂ Ker(ϕ).)
By Lemma (7.19) and (7.20) this gives Rq p1,∗ (K) = (0) for all q, and from the first spectral
sequence we find that H n (X × X, K) = 0 for all n.
Now use the second spectral sequence. For x ∈ / K(L) the bundle t∗x L ⊗ L−1 is a non-trivial
bundle in Ker(ϕ). Again by Lemma (7.15) we find that supp(Rq p2,∗ K) ⊂ K(L). Since K(L) is
a finite subscheme of X (the bundle L being ample) we find
, -
&p,q
Rq p2,∗ (K)x if p = 0;
E2 = x∈K(L)
0 otherwise.

As we only have non-zero terms for p = 0, the spectral sequence degenerates at level E2& . This
gives H n (X × X, K) = ⊕x∈K(L) Rn p2,∗ (K)x .
Comparing the two answers for H n (X × X, K) we find that Rn p2,∗ (K) = 0 for all n. By
(7.20) this implies that H n (X, K|X×{x} ) = 0 for all x. But K|X×{0} is the trivial bundle, so
taking n = 0 and x = 0 gives a contradiction. !

0
#
(7.22) Corollary.
$ Let X be an abelian variety over a field k. Then Pic X/k = Ker ϕ: PicX/k →
Hom(X, X t ) .
Proof. We already know that Ker(ϕ) is a subgroup scheme of PicX/k that contains Pic0X/k .
Hence Ker(ϕ) is the union of a number of connected components of PicX/k . By the proposition,
every k-valued point of Ker(ϕ) lies in Pic0 . The claim follows. !

(7.23) Corollary. Let X be an abelian variety over a field k. Let L be a line bundle on X.
(i) If [Ln ] ∈ Pic0X/k for some n ,= 0 then [L] ∈ Pic0X/k . In particular, if L has finite order,
i.e., Ln ∼= OX for some n ∈ Z!1 , then [L] ∈ Pic0X/k .
(ii) We have [L ⊗ (−1)∗ L−1 ] ∈ Pic0X/k .
(iii) We have

[L] ∈ Pic0X/k ⇐⇒ n∗ L ∼
= Ln for all n ∈ Z
⇐⇒ n∗ L ∼
= Ln for some n ∈ Z \ {0, 1}.

– 106 –
Proof. (i) Since ϕ is a homomorphism we have ϕLn (x) = n · ϕL (x) = ϕL (n · x). Hence if
[Ln ] ∈ Pic0 (X) then ϕL is trivial on all points in the image of nX . But nX is surjective, so ϕL
is trivial.
(ii) Direct computation shows that ϕ(−1)∗ L (x) = −ϕL (x) for all L and x. Since also
ϕL−1 (x) = −ϕL (x), we find that [L ⊗ (−1)∗ L−1 ] ∈ Ker(ϕ).
(iii) The first implication “⇒” was proven in (7.16) above; the second is trivial. Suppose that
∗ ∼ n
/ {0, 1}. Since n∗ L ∼
2
n L = L for some n ∈ = Ln ⊗ [L ⊗ (−1)∗ L](n −n)/2 it follows that L ⊗ (−1)∗ L
has finite order, hence its class lies in Pic0X/k . By (ii) we also have [L ⊗ (−1)∗ L−1 ] ∈ Pic0X/k .
Hence [L2 ] ∈ Pic0X/k and by (i) then also [L] ∈ Pic0X/k . !

(7.24) In (3.29) we have associated to any group scheme G locally of finite type over a field k
an étale group scheme of connected components, denoted by -0 (G). We now apply this with
G = PicX/k for X/k an abelian variety. The associated component group scheme

NSX/k := -0 (PicX/k )

is called the Néron-Severi group scheme of X over k. The natural homomorphism q: PicX/k →
NSX/k realizes NSX/k as the fppf quotient of PicX/k modulo Pic0X/k ; hence we could also write

NSX/k = PicX/k /Pic0X/k .

We refer to the group

NS(X) := NSX/k (k)
as the Néron-Severi group of X. Note that NS(X) equals the subgroup of Gal(ks /k)-invariants
in NS(Xks ).
We say that two line bundles L and M are algebraically equivalent, notation L ∼alg M , if
[L] and [M ] have the same image in NS(X). As NS(X) naturally injects into NS(Xk ), algebraic
equivalence of line bundles (or divisors) can be tested over k, and there it coincides with the
notion defined in Remark (6.9). Hence we can think of the Néron-Severi group scheme as being
given by the classical, geometric Néron-Severi group NS(Xks ) = NS(Xk ) of line bundles (or
divisors) modulo algebraic equivalence, together with its natural action of Gal(ks /k). Note,
however, that a k-rational class ξ ∈ NS(X) may not always be representable by a line bundle
on X over the ground field k.
Let us rephrase some of the results that we have obtained in terms of the Néron-Severi
group.

(7.25) Corollary. The Néron-Severi group NS(X) is torsion-free. If n ∈ Z and L is a line

2
bundle on X then n∗ L is algebraically equivalent to Ln ; in other words, n∗ : NS(X) → NS(X)
is multiplication by n2 .
Proof. The first assertion is just (i) of Corollary (7.23). The second assertion follows from (ii)
of that Corollary together with Corollary (2.12). !

(7.26) Corollary (7.22) can be restated by saying that the natural homomorphism ϕ: PicX/k →
Hom symm (X, X t ) ⊂ Hom(X, X t ) factors as

q ψ
PicX/k −→ NSX/k .−→ Hom symm (X, X t )

– 107 –
for some injective homomorphism ψ: NSX/k .→ Hom symm (X, X t ). This says that the homomor-
phism ϕL associated to a line bundle L only depends on the algebraic equivalence class of L,
and that ϕL = ϕM only if L ∼alg M . We shall later show that ψ is actually an isomorphism;
see Corollary (11.3).

§ 4. Applications to cohomology.

(7.27) Proposition. Let X be an abelian variety with dim(X) = g. Cup-product gives an

∼
isomorphism ∧• H 1 (X, OX ) −→ H • (X, OX ). For every p and q we have a natural isomorphism
H q (X, ΩpX/k ) ∼
= (∧q TX t ,0 ) ⊗ (∧p TX,0
∨
). The Hodge numbers hp,q = dim H q (X, ΩpX/k ) are given
# $# $
by hp,q = gp gq .
Proof. Use (6.13) and the isomorphisms ΩpX/k ∼
= (∧p TX,0
∨
) ⊗k OX . !

(7.28) Corollary. Multiplication by an integer n on X induces multiplication by np+q on

H q (X, ΩpX ).
Proof. Immediate from the fact that nX induces multiplication by n on TX,0 , applied to both
X and X t . !
Before we state the next corollary, let us recall that the algebraic de Rham cohomology of
a smooth proper algebraic variety X over a field k is defined to be the hypercohomology of the
de Rham complex
d d d
ΩX/k = (OX −→ Ω1X/k −→ Ω2X/k −→ · · ·) ,
•

with OX in degree zero. We have the so-called “stupid filtration” of this complex, by the
subcomplexes σ!p Ω•X/k given by
,
0 for i < p
• i
[σ!p ΩX/k ] =
ΩiX/k for i " p.

This gives rise to a spectral sequence

E1p,q = H q (X, ΩpX ) ⇒ HdR

p+q
(X/k)

called the “Hodge–de Rham” spectral sequence.

If k has characteristic zero then it follows from Hodge theory that this spectral sequence
degenerates at the E1 -level, see Deligne [1], section 5. If k has characteristic p > 0 then this is
no longer true in general. For examples and further results we refer to Deligne-Illusie [1] and
Oesterlé [1].
As we shall now show, for abelian varieties the degeneration of the Hodge-de Rham spectral
sequence at level E1 follows from (6.12) without any restrictions on the field k.

(7.29) Corollary. Let X be an abelian variety over a field k. Then the “Hodge-de Rham”
spectral sequence of X degenerates at level E1 .
Proof. We follow the proof given by Oda [1]. We have to show that the differentials dr : Erp,q →
Erp+r,q−r+1 are zero for all r " 1. By induction we may assume that this holds at all levels < r.
(The empty assumption if r = 1.)

– 108 –
 Write Er∗ (X) = ⊕Erp,q , graded by total degree. Cup-product makes Er∗ (X) into a connected,
graded-commutative k-algebra. By our induction assumption and the Künneth formula there is
a canonical isomorphism
Er∗ (X × X) ∼= Er∗ (X) ⊗k Er∗ (X) .
Write µ: Er∗ (X) → Er∗ (X) ⊗k Er∗ (X) for the map induced by the multiplication law on X, and
write ε: Er∗ (X) → Ek0 (X) = k for the projection onto the degree zero component. One checks
that µ and ε give Er∗ (X) the structure of a graded bialgebra over k.
Let g = dim(X). By what was shown above, Er1 (X) = H 1 (X, OX ) ⊕ H 0 (X, Ω1X/k ) has
dimension 2g. Also, Eri (X) = 0 for i > 2g. The Borel-Hopf structure theorem (6.12) then gives

Er∗ (X) ∼
= ∧∗ Er1 (X) .

Since dr is compatible with the product structure (cup-product) on Er∗ (X), it suffices to
show that dr is zero on Er1 (X), which is just the space of primitive elements of Er∗ (X). (See
6.17.) By functoriality of the Hodge–de Rham spectral sequence we have µ ◦ dr = (dr ⊗ dr ) ◦ µ.
Therefore, for ξ ∈ Er1 (X) we have µ(dr (ξ)) = dr (ξ) ⊗ 1 + 1 ⊗ dr (ξ). This shows that dr (ξ) is
again a primitive element. But dr (ξ) ∈ Er2 (X) which, by (6.17), contains no non-zero primitive
elements. This shows that dr = 0. !

(7.30) Corollary. There is an exact sequence

0 −→ Fil1 HdR
1 1
(X/k) −→ HdR (X/k) −→ H 1 (X, OX ) −→ 0 ,

where Fil1 HdR

1
(X/k) := H 0 (X, Ω1X/k ) ∼ ∨
= TX,0 .
To close this section let us fulfil an earlier promise and give an example of a smooth projec-
tive variety with non-reduced Picard scheme. We refer to Katsura-Ueno [1] for similar examples.

(7.31) Example. Let k be an algebraically closed field of characteristic 3. Let E1 be the elliptic
curve over k given by the Weierstrass equation y 2 = x3 − x. From (5.27) we know that E1 is
supersingular. Let σ be the automorphism of E1 given by (x, y) %→ (x + 1, y). Then σ has
order 3, so that we get an action of G := Z/3Z on E1 . The quotient of E1 by G is isomorphic
to P1k ; in affine coordinates te quotient map is just (x, y) %→ y.
Let E2 be an ordinary elliptic curve over k. Let τ be the translation over a point of (exact)
order 3 on E2 . Then (σ, τ ) is an automorphism of order 3 of the abelian surface X := E1 ×E2 ; this
gives a strictly free action of G := Z/3Z on X, and we can form the quotient π: X → Y := G\X.
By (??) π is an étale morphism, so Y is again a non-singular algebraic surface. We have a
natural morphism Y → (G\E1 ) ∼ = P1 ; this exhibits Y as an elliptic surface over P1 . In fact, for
1
all P ∈ P (k) with P ,= ∞ the fibre YP above P is isomorphic to E2 .
We compute h1 (Y, OY ) using Hirzebruch-Riemann-Roch and Chern numbers for algebraic
surfaces. (A reference is ??.) The Euler number c2 of Y is a multiple of the Euler number of X,
and this is 0. By the Hirzebruch-Riemann-Roch formula we have

1 − h1 (Y, OY ) + h2 (Y, OY ) = (c21 + c2 )/12 = 0 ,

since c21 = 0 for every elliptic surface. By Serre duality, h2 (Y, OY ) = h0 (Y, Ω2Y /k ). Now we use
that H 0 (Y, Ω2Y /k ) is isomorphic to the space of G-invariants in H 0 (X, Ω2X/k ). If ωi is a basis for
H 0 (Ei , Ω1Ei /k ) then ω1 ∧ ω2 is a basis for H 0 (X, Ω2X/k ). But ω1 is a multiple of dy, which is

– 109 –
invariant under σ, and ω2 is translation invariant, in particular invariant under τ . In sum, we
find that h2 (Y, OY ) = 1 and h1 (Y, OY ) = 2.
On the other hand, π: X → Y induces a homomorphism π ∗ : Pic0Y /k → X t = Pic0X/k . The
same arguments as in the proof of Theorem (7.5) show that Ker(π ∗ ) ∼ = µ3 . On the other hand,
π ∗ factors via the subscheme of G-invariants in X t . (See Exercise ?? for the existence of such
a subscheme of G-invariants.) The point here is that we are describing line bundles on Y as
coming from line bundles L on X together with an action of G. But such an action is given by
∼
an isomorphism ρ∗ L −→ pr∗X L of line bundles on G ×k X. The existence of such an isomorphism
says precisely that L corresponds to a G-invariant point of X t .
∼
By Exercises ?? and ??, X t −→ X. The induced action of G on X t is given by the
automorphism (σ, id). (Cf. Exercise ??) Therefore, the subscheme of G-invariants in X t is
)σ*
E1 × E2 . The only geometric point of E1 fixed under σ is the origin. A computation in local
)σ*
coordinates reveals that E1 is in fact the Frobenius kernel E1 [F ] ⊂ E1 which can be shown to
be isomorphic to α3 . In any case, we find that Pic0Y /k is 1-dimensional, whereas we have shown
its tangent space at the identity, isomorphic to H 1 (Y, OY ), to be 2-dimensional. Hence Pic0Y /k
is non-reduced.

§ 5. The duality between Frobenius and Verschiebung.

(7.32) Let S be a scheme of characteristic p. Recall that for any S-scheme aX : X → S we have
a commutative diagram with Cartesian square

WX/S
X (p/S) −−−−→ X
 
 (p) a
"aX " X
Frob
S
S −−−−→ S

If there is no risk of confusion we simply write X (p) for X (p/S) . Note that if aT : T → S
is an S-scheme then we have aT ◦ FrobT = FrobS ◦ aT and this gives a natural identification
(p)
(XT )(p/T ) = (X (p/S) )T . We denote this scheme simply by XT .
Let us write T(p) for the scheme T viewed as an S-scheme via the morphism aT(p) :=
FrobS ◦ aT = aT ◦ FrobT : T → S. The morphism FrobT : T → T is not, in general, a morphism
of S-schemes, but if we view it as a morphism T(p) → T then it is a morphism over S. To avoid
confusion, let us write FrT : T(p) → T for the morphism of S-schemes given by FrobT .
Let Y be an S-scheme. Recall that we write Y (T ) for the T -valued points of Y . It is
understood here (though not expressed in the notation) that all schemes and morphisms of
schemes are over a fixed base scheme S; so Y (T ) is the set of morphisms T → Y over S. There
is a natural bijection
∼
wY,T : Y (p) (T ) −→ Y (T(p) ) ,

sending a point η: T → Y (p) to WY /S ◦ η, which is a T(p) -valued point of Y . The composition

wY,T ◦ FY /S (T ): Y (T ) → Y (T(p) )

– 110 –
is the map that sends y ∈ Y (T ) to y ◦ FrT : T(p) → Y , which is the same as y ◦ FrobT : T → Y
viewed as a morphism T(p) → Y .

(7.33) Consider an abelian variety X over a field k of characteristic p. Take S := Spec(k). If

T is any S-scheme then X ×S T(p) is the same as X (p) ×S T , and we find that
. /
(p) ∼ isomorphism classes of rigidified
PicX/k (T ) −−−−−−−→ PicX/k (T(p) ) =
wPicX/k ,T line bundles (L, α) on X ×S T(p)
. /
isomorphism classes of rigidified
= = PicX (p) /S (T ) .
line bundles (L, α) on X (p) ×S T

(p) ∼
In this way we obtain an isomorphism PicX/S −→ PicX (p) /S , which we take as an identification.
Applying (7.32) with Y = PicX/k we find that the relative Frobenius of PicX/k over k is the
homomorphism that sends a point y ∈ PicX/k (T ) to y ◦ FrobT , viewed as a morphism T(p) →
PicX/k . Because the diagram
(p) WXT /T
XT −−−−−→ XT
 
(p)  a
aX " " X
Frob
T −−−−T→ T
is Cartesian this just means that FPic/k : PicX/k → PicX (p) /k sends the class of a rigidified line
# $ (p)
bundle (L, α) on XT to the class of L(p) , α(p) on XT , where L(p) := WX ∗
T /T
L, and where
∼
α(p) : OT −→ e∗ L(p) = Frob∗T (e∗ L) is the rigidification of L(p) along the zero section obtained by
pulling back α via FrobT .

(7.34) Proposition. Let X be an abelian variety over a field k of characteristic p. We identify

(X t )(p) = (X (p) )t as in (7.33), and we denote this abelian variety by X t,(p) . Then we have the
identities

t
FX/k = VX t /k : X t,(p) → X t and t
VX/k = FX t /k : X t → X t,(p) .

t t t
Proof. It suffices to prove that FX/k X /k : X → X equals [p]X t , because if this holds then
◦F t

t
together with Proposition (5.20) and the fact that FX t /k is an isogeny it follows that FX/k =
VX t /k . The other assertion follows by duality.
Let T be a k-scheme. Consider a rigidified line bundle (L, α) on
# (p) $ XT that gives a point
t (p) (p) ∗
of X (T ). As explained in (7.33) FX t /k sends (L, α) to L , α with L = WX T /T
L.
∗
Because WXT /T ◦ FXT /T = FrobXT , pull-back via FXT /T gives the line bundle FrobXT L on XT .
But if Y is any scheme of characteristic p and M is a line bundle on Y then Frob∗Y (M ) ∼ = M p ; this
follows for instance by taking a trivialization of M and remarking that FrobY raises all transition
functions to the power p. The rigidification we have on FX ∗
T /T
WX∗
T /T
L = Frob∗XT L = Lp is the
isomorphism

∼
OT = Frob∗T OT −→ e∗XT FX
∗
T /T
∗
WX T /T
L = e∗X (p) WX
∗
T /T
L = Frob∗T e∗XT L = (e∗XT L)p
T

that is obtained from α by pulling back via FrobT , which just means it is αp . In sum, FX/k
t ◦F t
X /k
p p
sends (L, α) to (L , α ), which is what we wanted to prove. !

– 111 –
Exercises.
(7.1) Let X be an abelian variety. Let mX : X × X → X be the group law, and let ∆X : X →
X × X be the diagonal morphism. Show that (mX )t = ∆X t : X t × X t → X t , and that (∆X )t =
mX t : X t × X t → X t .
(7.2) Let L be a line bundle on an abelian variety X.
(i) Show that, for n ∈ Z,
n∗ L ∼
= OX ⇐⇒ Ln ∼ = OX .
(ii) Show that, for n ∈ Z \ {−1, 0, 1},

n∗ L ∼
=L ⇐⇒ Ln−1 ∼
= OX .

(7.3) Let X be an abelian variety over an algebraically closed field k. Show that every line
bundle L on X can be written as L = L1 ⊗ L2 , where L1 is symmetric and [L2 ] ∈ Pic0X/k . [Hint:
By (7.23), the class of the line bundle (−1)∗ L ⊗ L−1 is in Pic0X/k . As Pic0 is an abelian variety
and k = k, there exists a line bundle M on X with [M ] ∈ Pic0 and M 2 ∼ = (−1)∗ L ⊗ L−1 . Now
show that L ⊗ M is symmetric.]
(7.4) Let P be the Poincaré bundle on X × X t . For m, n ∈ Z, consider the endomorphism
(m, n) of X × X t . Show that (m, n)∗ P ∼
= P mn .
(7.5) Let P be the Poincaré bundle on X × X t . Show that the# associated
$ homomorphism
t t tt
ϕP : X ×X → X ×X is the homomorphism given by ϕP (x, ξ) = ξ, κX (x) . [Hint: Compute
the restrictions of t( x, ξ)∗ P ⊗ P −1 to X × {0} and {0} × X t .]
(7.6) If τ is a translation on an abelian variety, then what is the induced automorphism τ t of
the dual abelian variety?
(7.7) Let X be an abelian variety over a field k. Let i: Y .→ X be an abelian subvariety. Write
q: X → Z := X/Y for the fppf quotient morphism, which exists by Thm. (4.38). Note that Z
is an abelian variety; see Example (4.40).
(i) Show that for any k-scheme T we have q∗ (OXT ) = OZT .
(ii) Prove that q t : Z t → X t is injective and gives an isomorphism between Z t and Ker(it : X t →
Y t )0red .
(7.8) Let L be a line bundle on an abelian variety X. For a symmetric m × m-matrix S with
integer coefficients sij we define a line bundle L"S on X m by
0m 1 0 1
L"S := ⊗ p∗i Lsii ⊗ ⊗ p∗ij Λ(L)sij .
i=1 1#i<j#m

# $
If α = aij is an integer valued matrix of size m × n we define a homomorphism of abelian
2n
varieties [α]X : X n → X m by α(x1 , . . . , xn ) = (y1 , . . . , ym ) with yi = j=1 aij xj .
# $ t
(i) Prove that [α]∗X L"S is algebraically equivalent to L"( αSα) .
# $
(ii) Assume that L is a symmetric line bundle. Prove that [α]∗X L"S ∼
t
= L"( αSα) .

Notes. (nog aanvullen)

– 112 –
Chapter VIII. The Theta group of a line bundle.

To a line bundle L on an abelian variety X we shall associate a group scheme, the theta group
G (L) of L. If the class of L is in Pic0X/k = X t then G (L) is an extension of X by the multiplicative
group Gm and is commutative. If [L] !∈ Pic0X/k then G (L) is much smaller and in general not
commutative. The theta group is a convenient tool for studying when a line bundle descends
over an isogeny. Further we study the structure of so-called non-degenerate theta groups, and
their representations.

§ 1. The theta group G (L).

(8.1) Let X be an abelian variety over a field k. Let L be a line bundle on X. Write L = V(L∨ )
for the corresponding geometric line bundle over X.
For a k-scheme T define G (L)(T ) to be the set of pairs (x, ϕ) where x ∈ X(T ) and where
∼
ϕ: LT → t∗x LT is an isomorphism. Geometrically this means that we have ϕL : LT −→ LT ,
fibrewise linear, fitting in a commutative diagram
ϕL
LT −−→ LT
 
 
" "
XT −−→ XT .
tx

Note that x is uniquely determined by ϕ, so that G (L)(T ) is in natural bijection with the set of
∼
ϕL : LT −→ LT lying over a translation on XT .
The set G (L)(T ) carries a natural group structure, with multiplication given by (x1 , ϕ1 ) ·
(x2 , ϕ2 ) = (x1 + x2 , t∗x2 ϕ1 ◦ ϕ2 ). Since the association T %→ G (L)(T ) is functorial in T we obtain
a group functor G (L): (Sch/k )0 → Gr.
The (fibrewise linear) automorphisms of LT lying over the identity on XT are just the
multiplications by elements of Γ(XT , OXT )∗ = Γ(T, OT )∗ . This gives an identification of Gm,k
with the kernel of the natural homomorphism (of group functors) G (L) → K(L) ⊂ X. Notice
that Gm,k is central in G (L).

(8.2) Lemma. The group functor G (L) is representable. There is an exact sequence of group
schemes
0 −→ Gm,k −→ G (L) −→ K(L) −→ 0 , (1)

where the map G (L) → K(L) is given on points by (x, ϕ) %→ x.

Proof. Since the functor K(L) is representable, it suffices to show that the homomorphism
π: G (L) → K(L) is (relatively) representable by a Gm -torsor. So, let T be a k-scheme and
x ∈ K(L)(T ). Write
M := prT,∗ (L−1 ∗
T ⊗ tx L T ) ,

ThetaGr, 15 september, 2011 (812)

– 113 –
 ∼
which is a line bundle on T . If T % is a T -scheme then the ϕ: LT " −→ LT " such that (x, ϕ) ∈
G (L)(T % ) are precisely the nowhere vanishing sections of MT " . Thus, writing M∗ for the Gm,T -
torsor corresponding to M (i.e., the T -scheme obtained from the geometric line bundle M :=
V(M ∨ ) by removing the zero section), we find that the fibre π −1 (x) is representable by the
T -scheme M∗ . That the sequence (1) is exact (even as a sequence of Zariski sheaves on Sch/k )
is clear from the remarks preceding the lemma and the definition of K(L). !

(8.3) We indicate another proof of (8.2). For this, consider the Gm -torsor L∗ over X associated
to L. Write ξ: L∗ → X for the structure morphism. Let Y = ξ −1 (K(L)) = K(L) ×X L∗ , the
scheme obtained by pulling back L∗ via the inclusion map K(L) $→ X. Choose a k-rational point
P ∈ L∗ (0). (This gives a trivialization L∗ (0) ∼
= Gm,k .) We obtain a morphism rP : G (L) → Y by
sending a point (x, ϕ) ∈ G (L)(T ) to the image point ϕL (P ) ∈ Y (T ) ⊂ L∗ (T ). It is not difficult
to see that rP gives an isomorphism of (set-valued) functors. So G (L) is represented by the
scheme Y = ξ −1 (K(L)).

(8.4) Definition. Let L be a line bundle on an abelian variety. The group scheme G (L) is
called the theta group of L.

(8.5) Consider the morphism [ , ]: G (L)×G (L) → G (L) given on points by (g1 , g2 ) %→ [g1 , g2 ] =
g1 g2 g1−1 g2−1 . Since K(L) is commutative this morphism factors through Gm ⊂ G (L). The fact
that Gm is central in G (L) then implies that [ , ] factors modulo Gm × Gm . We thus obtain a
pairing
eL : K(L) × K(L) → Gm ,
called the commutator pairing of the theta group. Note that eL is alternating, meaning that
eL (x, x) = 1 for every x ∈ K(L). For fixed x ∈ K(L)(T ) the morphisms K(L)T → Gm,T given
by y %→ eL (x, y) resp. y %→ eL (y, x) are homomorphisms. In sum we find that the theta group
G (L) gives rise to an alternating bilinear form eL .
The alternating form eL has the following properties.

(8.6) Proposition. (i) If f : X → Y is a homomorphism of abelian varieties and L is a line

bundle on Y then
∗
ef (L)
= eL ◦ (f, f ) on f −1 (K(L)) × f −1 (K(L)) .

(ii) If L and M are line bundles on X then eL⊗M = eL · eM on K(L) ∩ K(M ).

(iii) If [L] ∈ Pic0X/k then eL = 1.
Ln
(iv) For x ∈ K(L) and y ∈ n−1 n
X (K(L)) = K(L ) we have e (x, y) = eL (x, ny).
Proof. (i) Note that f −1 (K(L)) ⊆ K(f ∗ L), for if x ∈ X then

t∗x f ∗ L = f ∗ (t∗f (x) L) . (2)

Now suppose T is a k-scheme and x1 , x2 ∈ f −1 (K(L))(T ). We can cover T by Zariski-open

subsets U such that there exist automorphisms ϕ1,L and ϕ2,L of the geometric line bundle LU ,
∗
lying over the translations tf (x1 ) and tf (x2 ) , respectively. As it suffices to show that ef (L) =
eL ◦ (f, f ) on [f −1 (K(L)) × f −1 (K(L))](U ) for every such U , we may replace T by U .
By construction, the automorphism [ϕ1,L , ϕ2,L ] of L is the (fibrewise) multiplication by
eL (f (x1 ), f (x2 )). Then f ∗ ϕ1,L and f ∗ ϕ2,L are automorphisms of f ∗ L which, by formula (2), lie

– 114 –
over the translations tx1 resp. tx2 on X. Since clearly [f ∗ ϕ1,L , f ∗ ϕ2,L ] = f ∗ [ϕ1,L , ϕ2,L ], we find
∗
that ef (L) (x1 , x2 ) = eL (f (x1 ), f (x2 )).
(ii) If we have elements (ϕ1 , x), (ϕ2 , y) ∈ G (L)(T ) and (ψ1 , x), (ψ2 , y) ∈ G (M )(T ) then
(ϕ1,L ⊗ ψ1,M ) ◦ (ϕ2,L ⊗ ψ2,M ) = (ϕ1,L ◦ ϕ2,L ) ⊗ (ψ1,M ◦ ψ2,M ) as (fibrewise linear) automorphisms of
L ⊗ M. The claim readily follows from this.
(iii) If the class of L is in Pic0 then K(L) = X, and since X is complete the pairing
eL : X × X → Gm must be constant.
(iv) For n " 0 this follows by induction from (ii) and the bilinearity of the pairing eL . The
case n # 0 then follows from (ii) and (iii). !
Let k be a field. As we have seen in (4.41), the category C of commutative group schemes
of finite type over k is abelian. In particular, given objects A and B of C we can form the
groups ExtnC (A, B) of n-extensions of A by B. If there is no risk of confusion we shall simply
write Ext(A, B) for Ext1C (A, B). Thus, the elements of Ext(A, B) are equivalence classes of
exact sequences
0 −→ B −→ E −→ A −→ 0
where E is again an object of C.
It can be shown (but this requires some work) that C does not contain any injective or
projective objects. In particular, the computation of Ext-groups by homological methods cannot
be done “directly” in C. To repair this, one may work in an Ind- or Pro-category, cf. ?? and ??.
We shall further discuss extensions of group schemes in Chapter ??. In this chapter we only
need the following two facts.

(8.7) Fact. Let k be an algebraically closed field.

(i) Write C for the category of commutative group schemes of finite type over k. If G is a
finite commutative k-group scheme then Ext1C (G, Gm ) = 0. In other words, for every extension
0 → Gm,k → G → G → 0 with G commutative, there exists a section s: G → G which is a
homomorphism of group schemes.
(ii) Let G be a finite k-group scheme of prime order. If 0 → Gm,k → G → G → 0 is an
exact sequence of k-group schemes then G is commutative.

(8.8) We shall use the notion of a theta group to obtain an interpretation of X t = Pic0X/k as
being ExtC (X, Gm ), where C is the category of commutative k-group schemes of finite type.
In one direction this is quite easy. Namely, suppose that L is a line bundle on X which
gives a class in Pic0 . Then K(L) = X and the pairing eL is trivial. This means that G = G (L)
is a commutative group scheme fitting in an exact sequence
0 −→ Gm −→ G −→ X −→ 0. (3)
Thus, if [L] ∈ Pic0X/k then G (L) gives an element of Ext(X, Gm ).
Conversely, suppose G is a commutative k-group scheme for which we have an exact se-
quence (3). Then G can be viewed as a Gm -torsor over X. Write LG for the corresponding line
bundle on X. (See Appendix ??) We claim that LG is a line bundle in Pic0X/k with theta group
isomorphic to G. To see this, suppose that G1 and G2 are commutative k-group schemes and
that we have a commutative diagram
j1 π1
0 −→ Gm −→ G1 −−→ X −→ 0
  

γ"
ϕ f
" "
j2 π 2
0 −→ Gm −→ G2 −−→ X −→ 0

– 115 –
with exact rows. Here f is only required to be a morphism of schemes and ϕ is required to
be “fibrewise linear” (meaning that ϕ(j1 (c) · g) = j2 γ(c) · ϕ(g) for all c ∈ Gm and g ∈ G1 .)
∼
Then ϕ gives an isomorphism of Gm -torsors G1 −→ f ∗ G2 , hence it induces a homomorphism
∼
ϕ: LG1 −→ f ∗ LG2 . Now take G1 = G2 = G and let ϕ = tg be the translation over an element
g ∈ G. If x ∈ X is the image of g then we obtain a pair (x, ϕ) ∈ G (LG ). Since this construction is
obviously functorial, it gives a homomorphism h: G → G (LG ), compatible with the projections
to X. In particular this shows that K(LG ) = X, so that the class of LG is in Pic0X/k . Furthermore
it is clear that h is injective, and it follows that h is an isomorphism.
In sum, we can pass from line bundles L on X with [L] ∈ Pic0 to commutative group
schemes G as in (3) and vice versa.

(8.9) Theorem. Let X be an abelian variety over a field k. Write C for the (abelian) category
of commutative group schemes of finite type over k. Associating G (L) to a line bundle L with
∼
[L] ∈ Pic0X/k gives an isomorphism X t (k) −→ Ext1C (X, Gm ).

Proof. All that remains to be shown is that L ∼ = LG (L) as line bundles on X. This follows
from the construction in (8.3), as it shows that G (L) is (non-canonically) isomorphic to L∗ as a
Gm -torsor. !
∼
We shall later extend this result, obtaining an isomorphism of group schemes X t −→
E xt(X, Gm ). The main problem here is to set up a framework in which we can define E xt(X, Gm )
correctly.

§ 2. Descent of line bundles over homomorphisms.

Theta groups are a useful tool in studying when a line bundle on an abelian variety descends
over an isogeny. The basic result is in fact just a reformulation of what we have seen in (7.2).

(8.10) Theorem. Let f : X → Y be a surjective homomorphism of abelian varieties. Let L

be a line bundle on X. Then there is a bijective correspondence between the M ∈ Pic(Y ) with
f ∗M ∼= L and the homomorphisms Ker(f ) → G (L) lying over the natural inclusion Ker(f ) $→ X.

Note that such homomorphisms Ker(f ) → G (L) can only exist if Ker(f ) ⊆ K(L) and
Ker(f ) is totally isotropic for the pairing eL .

Proof. Write V1 for the set of isomorphism classes of pairs (M, α) where M is a line bundle
∼
on Y and α: f ∗ M −→ L. Write V2 for the set of isomorphism classes of line bundles M on Y
such that f ∗ M ∼= L. Using that Aut(M ) = k ∗ = Aut(L) we see that the natural map V1 → V2
(forgetting α) is a bijection.
Write H = Ker(f ). Then Y represents the fppf quotient of X by H. We have seen in (7.2)
that the pairs (M, α) ∈ V1 correspond to the H-actions on L compatible with the natural action
of H on X. It is an immediate translation of the definitions that such H-actions correspond to
homomorphisms H → G (L) lifting the inclusion H $→ X. !

For isogenies over an algebraically closed field this leads to a handy criterion for when a
line bundle descends. To prove it we shall make use of a result about extensions that we stated
above.

– 116 –
(8.11) Corollary. Let X and Y be abelian varieties over an algebraically closed field k. Let
f : X → Y be an isogeny. Then a line bundle L on X is the pull-back of a line bundle on Y if
and only if Ker(f ) is a subgroup scheme of K(L) which is totally isotropic with respect to the
pairing eL .
Proof. According to the preceding theorem one must check whether Ker(f ) $→ X can be lifted
to a homomorphism Ker(f ) → G (L). If it can then Ker(f ) is a subgroup scheme of K(L) and
eL is trivial when restricted to Ker(f ) × Ker(f ).
Conversely, if Ker(f ) is a totally isotropic subgroup scheme of K(L) then we consider the
extension
0 −→ Gm −→ π −1 (Ker(f ))) −→ Ker(f ) −→ 0 ,

where π: G (L) → K(L) is the projection. Since we assume Ker(f ) to be totally isotropic, the
group scheme G := π −1 (Ker(f ))) is commutative. By (8.7), the extension splits, i.e., there exists
a (homomorphic) section Ker(f ) → G. !

(8.12) Remarks. (i) In the “if” statement of the theorem we really need the assumption that
k = k̄: if k is an arbitrary field and Ker(f ) is a totally isotropic subgroup scheme of K(L) then
in general L descends to a line bundle on Y only after we pass to a finite extension of k.
(ii) The condition in (8.11) that the kernel of f is finite is necessary. If K(L) is not finite (i.e.,
L is degenerate) then Y := K(L)0red is a nonzero abelian subvariety of X (assuming the ground
field is perfect), and the quotient Z = X/Y exists as an abelian variety; see Example (4.40). In
this situation L is not, in general, the pullback of a line bundle on Z, even though Y ⊂ K(L) is
totally isotropic with respect to eL . For example, if the class of L is in Pic0X/k then Y = X, so
if L is non-trivial it is not a pullback from Z = {0}.
If q: X → Z is the quotient map then possibly after replacing the ground field by a finite
separable extension it is still true that there exists a line bundle M on Z such that L ⊗ q ∗ M −1
is in Pic0X/k ; see Exercise (11.3) below.

(8.13) Definition. A level subgroup of the theta group G (L) is a subgroup scheme H̃ ⊂ G (L)
such that Gm ∩ H̃ = {1}, i.e., H̃ maps isomorphically to its image H ⊂ K(L) under π.
With this notion of a level subgroup we have the following corollary to the theorem.

(8.14) Corollary. Let L be a line bundle on an abelian variety X over a field k. Then there
is a bijective correspondence between the set of level subgroups H̃ ⊂ G (L) and the set of
isomorphism classes of pairs (f, M ) where f : X → Y is a surjective homomorphisms and# M$ is
a line bundle on Y with f ∗ M ∼
= L. If H̃ corresponds to the pair (f, M ) then Ker(f ) = π H̃ .
∼
Proof. Given a level subgroup H̃ ⊂ G (L), set H := π(H̃) ⊂ K(L) and write ξ: H −→ H̃ ⊂ G (L)
for the inverse of π|H̃ . The projection f : X → X/H =: Y is a surjective homomorphism and
Theorem (8.10) shows that ξ corresponds to a line bundle M on Y with f ∗ M ∼ = L.
Conversely, if f : X → Y is a surjective homomorphism and M is a line bundle on Y
with f ∗ M ∼
= L then the image of the corresponding homomorphism Ker(f ) → G (L) is a level
subgroup. One now easily verifies that these two constructions give the desired bijection. !
Given a pair (f, M ) as in the corollary, we can describe the theta group G (M ) in terms
of G (L) and the level subgroup H̃.

(8.15) Proposition. Let f : X → Y be a surjective homomorphism of abelian varieties. Let L

– 117 –
 ∼ L. Write H̃ ⊂ G (L) for the
be a line bundle on X and let M be a line bundle on Y with# f ∗ M =
−1
$
level subgroup corresponding to the pair (f, M ). Then f K(M ) ⊆ K(L), the centralizer CH̃
of H̃ inside G (L) is given by

CH̃ = {g ∈ G (L) | π(g) ∈ f −1 (K(M ))} ,

and G (M ) ∼
= CH̃ /H̃.
# $
Proof. As already remarked in the proof of (8.6), we have f −1 K(M ) ⊆ K(L). Write H =
Ker(f ) = π(H̃) ⊂ K(L). Let ξ: H → G (L) be the homomorphism giving the canonical H-action
on L. By construction, H̃ is the image of ξ. As remarked after (7.2), such an H-action on L
(compatible with the H-action on X by translations) is nothing but a descent datum on L with
respect to the morphism f .
Let T be a k-scheme and (x, ϕ) ∈ G (L)(T ). Write y = f (x) ∈ Y (T ). Then t∗x ξ: H → G (t∗x L)
gives a descent datum on t∗x L. This descent datum corresponds to the line bundle t∗y M on Y , and
∼
we have a natural identification f ∗ (t∗y M ) = t∗x L. Now the isomorphism ϕ: L −→ t∗x L descends
∼
to an isomorphism ψ: M −→ t∗y M if and only if ϕ is equivariant with respect to the descent
data ξ and t∗x ξ. This last condition precisely means that (x, ϕ) · (h, ξ(h)) = (h, ξ(h)) · (x, ϕ) for
all (h, ξ(h)) ∈ H̃, i.e., (x, ϕ) ∈ CH̃ . Thus we obtain a homomorphism γ: CH̃ → G (M ).
By construction, if (x, ϕ) maps to (y, ψ) then f ∗ ψ = ϕ as homomorphisms from L to t∗x L.
∼
Thus, if (x, ϕ) ∈ Ker(γ) then x ∈ H = Ker(f ) and ϕ = ξ(x): L −→ t∗x L. This means precisely
that (x, ϕ) ∈ H̃ ⊂ CH̃ . (Note that H̃ is commutative, being isomorphic to H, so that H̃ is
indeed contained in CH̃ .)
Conversely, if (ψ, y) ∈ G (M )(T ), then there is an fppf cover T % → T and an x ∈ X(T % ) with
f (x) = y. Then (f ∗ ψ, x) is an element of CH̃ (T % ) with γ(f ∗ ψ, x) = (ψ, y). Thus γ is surjective
and G (M ) ∼ = CH̃ /H̃.
Finally, it is clear from the above that CH̃ ⊆ {g ∈ G (L) | π(g) ∈ f −1 (K(M ))}. Conversely,
if g = (ϕ, x) ∈ G (L) with f (x) ∈ K(M ) then we have shown that there exists an element of the
form (ϕ% , x) in CH̃ . As CH̃ clearly contains the central subgroup Gm,k ⊂ G (L), it follows that
also g ∈ CH̃ . !

If x is a T -valued point of K(L) for some k-scheme T then y %→ eL (x, y) defines a homo-
morphism
# K(L)$T → Gm,T . The bilinearity of the pairing eL implies that the map K(L) →
Hom K(L), Gm given on points by x %→ eL (x, −) is a homomorphism of group schemes.

(8.16) Corollary. In the situation of the proposition we have f −1 (K(M )) = H ⊥ := {k ∈

K(L) | eL (k, h) = 1 for every h ∈ H}. We have K(M ) ∼
= H ⊥ /H.
Proof. It easily follows from the definition of eL that CH̃ = {(x, ϕ) ∈ G (L) | x ∈ H ⊥ }. With
this remark the corollary directly follows from the proposition. !

§ 3. Theta groups of non-degenerate line bundles.

It will be helpful to reformulate some of the notions we have encountered without reference to
line bundles.

(8.17) Definition. Let k be a field. A theta-group over k is an exact sequence of k-group

– 118 –
schemes
i π
0 −→ Gm,k −→ G −→ K −→ 0

such that i(Gm,k ) is contained in the center of G and K is commutative. The commutator
pairing e: K × K → Gm,k of the theta group is the alternating bilinear pairing induced by
the commutator [ , ]: G × G → G . We say that two theta groups are isomorphic if they are
isomorphic as extensions of a group scheme K by Gm,k .
Suppose that we have a theta group as above such that K is finite. If T is a k-scheme and
x ∈ K(T ) then y %→ e(x, y) defines a homomorphism KT → Gm,T , i.e., an element of K D (T ).
In this way the pairing e gives a homomorphism ν: K → K D . The relation e(x, y) = e(y, x)−1
gives that ν D = ν −1 .

(8.18) Definition. A theta group G as above is said to be non-degenerate if K is finite and if

ν: K → K D is an isomorphism.
Notice that the non-degeneracy condition can also be expressed by saying that i(Gm,k ) is
the center of G .
As the terminology suggests, the theta group of a non-degenerate line bundle is non-
degenerate. This is a consequence of the following result.

(8.19) Proposition. Let L be a non-degenerate line bundle on an abelian variety X. If H ⊂

K(L) is a subgroup scheme which is maximal totally isotropic with respect to the pairing eL
then H = H ⊥ and rank(H)2 = rank(K(L)).
Proof. It suffices to prove this over an algebraically closed field. Write f : X → X/H =: Y for
the projection. By (8.11) there is a line bundle M on Y with f ∗ M ∼ = L.
We claim that K(M ) = {1}. Suppose not. Then there is a subgroup scheme K % ⊂ K(M ) of
prime order. By (8.7) this K % is totally isotropic for eM . Then (i) of (8.6) shows that f −1 (K % ) is
totally isotropic for eL . As H ! f −1 (K % ) this contradicts our choice of H. So indeed K(M ) =
{1}. It then follows from (8.16) that H ⊥ = H and by (7.6) we have rank(K(L)) = rank(H)2 . !

(8.20) Corollary. If L is a non-degenerate line bundle on an abelian variety then the theta
group G (L) is non-degenerate.
Proof. Choose H as above. Remark that ν: K(L) → K(L)D fits in a commutative diagram with
exact rows

0 −→ H⊥ −→ K(L) −→ K(L)/H ⊥ −→ 0
  
 " ν 
"ν " "ν̄
0 −→ [K(L)/H]D −→ K(L)D −→ HD −→ 0 .

By definition of H ⊥ the homomorphism ν̄ is injective. Now ν D = ν −1 so that ν̄ D : H →

[K(L)/H ⊥ ]D is the map obtained by restricting ν −1 to H. But H = H ⊥ , so we find that
ν % = (ν̄ D )−1 is surjective. By rank considerations it follows that ν % and ν̄ are isomorphisms.
Hence ν is an isomorphism. !

(8.21) Heisenberg groups. We now discuss an important example of non-degenerate theta-

groups, the so-called Heisenberg groups.

– 119 –
 We work over a field k. Let H be a finite abelian group; we shall view it as a (constant)
k-group scheme. Write H D := Hom(H, Gm,k ) for its Cartier dual. (If k = k one would also refer
to H D as the character group of H.) So

H∼
= (Z/d1 Z) × · · · × (Z/dn Z) , HD ∼
= µd1 ,k × · · · × µdn ,k ,

with d1 |d2 | · · · |dn . To the pair (H, H D ) we associate a Heisenberg group H ; it is defined by
H = Gm,k × H × H D as a k-scheme, with multiplication given by

(λ, x, χ) · (λ% , x% , χ% ) = (λλ% χ% (x), x + x% , χχ% ) . (4)

Then H is a theta-group: we have an exact sequence

0 −→ Gm,k −→ H −→ H × H D −→ 0 .

The commutator pairing e: (H × H D ) × (H × H D ) → Gm,k is given by e((x, χ), (x% , χ% )) =

χ% (x)χ(x% )−1 . As this is clearly a perfect pairing, H is non-degenerate.
The construction clearly generalizes to the case where we start with an arbitrary finite
commutative k-group scheme H. For H we now take Gm,k × H × H D , and the group structure
is again given on points by (4). We refer to the resulting theta group as the Heisenberg group
associated to the group scheme H (or to the pair (H, H D )).
Our next goal is to show that under suitable assumptions the theta group of a line bundle
can be described as a Heisenberg group.
π
(8.22) Lemma. Let 0 −→ Gm,k −→ G −→ K −→ 0 be a non-degenerate theta group over a
field k. Let H ⊂ K be a subgroup scheme. Consider the following conditions.
(i) H is maximal totally isotropic w.r.t. the commutator pairing e: K × K → Gm,k ,
(ii) H is totally isotropic and rank(H)2 = rank(K),
(iii) H = H ⊥ .
Then (iii) ⇔ (ii) ⇒ (i). If k is algebraically closed the three conditions are equivalent.
∼ ∼
Proof. The isomorphism ν: K −→ K D induces an isomorphism K/H ⊥ −→ H D . In particular,
rank(K) = rank(H) · rank(H ⊥ ). Now H is totally isotropic precisely if H ⊆ H ⊥ . This readily
gives (iii) ⇔ (ii) ⇒ (i).
To see that (i) ⇒ (iii) if k = k, let H be maximal totally isotropic and assume that
π
H ! H ⊥ . By (8.7) the extension 0 −→ Gm,k −→ π −1 (H) −→ H −→ 0 splits, so there exists
a level subgroup H̃ ⊂ G with π(H̃) = H. Writing G % := π −1 (H ⊥ )/H̃ we obtain a new theta
π"
group 0 −→ Gm,k −→ G % −−→ H ⊥ /H −→ 0. As H != H % and k = k we can choose a subgroup
−1
scheme Γ ⊂ H ⊥ /H of prime order. By (8.7) π % (Γ) is commutative. It follows that the inverse
image of Γ under H ⊥ → H ⊥ /H is totally isotropic. This contradicts the assumption that H is
maximal totally isotropic. !

(8.23) Definition. Let G be a non-degenerate theta group over a field k.

(i) A k-subgroup scheme H ⊂ K satisfying (ii) and (iii) in (8.22) is called a Lagrangian
subgroup. If H̃ ⊂ G is a level subgroup then we say that H̃ is a Lagrangian level subgroup if
π(H̃) ⊂ K is Lagrangian.
∼ ∼
(ii) A Lagrangian decomposition of K is an isomorphism K −→ H1 ×H2 such that ν: K −→
∼
K D induces an isomorphism ν̄: H1 −→ H2D .

– 120 –
 Condition (i) in (8.22) shows that for every non-degenerate theta group over k = k there
exist Lagrangian subgroups H ⊂ K. By (8.7) every such H can be lifted (still with k = k) to a
Lagrangian level subgroup of G .
If H is a Heisenberg group then, with the notations of (8.21), H and H D are Lagrangian
subgroups. So, a necessary condition for a theta group G to be a Heisenberg group is that there
exists a Lagrangian decomposition. This is not always the case. For instance, suppose that E is
a supersingular elliptic curve over a field k of characteristic p > 0 and that G is a theta group
with finite quotient equal to E[p]. One can show that E[p] has a unique non-trivial subgroup
scheme, isomorphic to αp . It follows that G is not a Heisenberg group H as in (8.21).
If the ground field is algebraically closed and K admits a Lagrangian decomposition then
we can describe G as a Heisenberg group.

(8.24) Lemma. Suppose G is a non-degenerate theta group over an algebraically closed field k.
(i) Assume that K = G /Gm,k admits a Lagrangian decomposition, say K ∼ = H1 × H2 . Then
G is isomorphic, as a theta group, to the Heisenberg group associated to the pair (H1 , H1D ).
(ii) If rank(K) is prime to char(k) then K admits a Lagrangian decomposition.
∼
Proof. (i) Lift Hi (i = 1, 2) to a Lagrangian level subgroup H̃i ⊂ G and write ξi : Hi −→ H̃i for
the inverse of the projection. Write H = Gm,k × H1 × H1D for the Heisenberg group associated
∼ ∼
to the pair (H1 , H1D ). If α: H1D −→ H2 is the inverse of ν̄ D : H2 −→ H1D then the map H → G
given by
(λ, x, χ) %→ i(λ) · ξ2 (α(χ)) · ξ1 (x)

gives the desired isomorphism of theta groups.

The proof of (ii) is done by the usual procedure of putting a symplectic pairing in canonical
form. For details we refer to Exercise ??. !
We apply this to non-degenerate line bundles L such that K(L) is finite and prime to char(k).
This last condition is equivalent to saying that the isogeny ϕL : X → X t is separable (see Exercise
??), hence we say that L is a non-degenerate line bundle of separable type.

(8.25) Corollary. Let k be algebraically closed field. Let X be an abelian variety over k
and let L be a non-degenerate line bundle on X of separable type. Then there is a sequence
of integers d1 |d2 | · · · |dn , called the type of L, such that G (L) is isomorphic to the Heisenberg
group associated to the group H = (Z/d1 Z) × · · · × (Z/dn Z).
Here is another case where a theta group can be descibed as a Heisenberg group.

(8.26) Example. Let X be an abelian variety. If P = PX is the Poincaré bundle on X × X t

then we know from Exercise (7.5) that ϕP : X × X t → X t × X is given by (x, y) %→ (y, x). Hence
K(P) = {0} and G (P) = Gm is the trivial theta group.
Next consider an isogeny h: X → X t , and let M := (1 × h)∗ PX on X × X. (If h = ϕL for
some non-degenerate line bundle L then M = Λ(L).) Note that also M = (ht × 1)∗ s∗ P, where
∼
s: X × X t −→ X t × X is the isomorphism switching the two factors. Identifying Ker(ht ) =
Ker(h)D as in Thm. (7.5) we find that {0} × Ker(h) ⊂ K(M ) and Ker(h)D × {0} ⊂ K(M ), and
by comparing ranks it follows that in fact K(M ) = Ker(h)D × Ker(h).
We claim that the #theta group G (M$ ) is naturally isomorphic to the Heisenberg group H
associated to the pair Ker(h)D , Ker(h) . We already have natural actions of Ker(h)D × {0}
and {0} × Ker(h) on M , compatible with the actions on the basis by translations; this realizes

– 121 –
 ∼
Ker(h)D and Ker(h) as Lagrangian level subgroups of G (M ). An isomorphism H −→ G (M ) is
then obtained in the same manner as in the proof of (8.24).

§ 4. Representation theory of non-degenerate theta groups.

As we have seen in (2.27), an abelian variety X of dimension g can only be embedded in

projective spaces of dimension at least 2g. Hence we will need, at least for large g, a rather large
number of equations to describe X. In a beautiful series of papers, Mumford [1] showed how one
can nevertheless set up a systematic study of the equations defining an abelian variety. Theta
groups play a crucial role in this. To explain why, suppose we choose an ample line bundle L
on X. To find the equations for X in the projective embedding defined by (some power of) L,
we must try to describe the kernel of the map
%
Sym H 0 (X, L) −→ H 0 (X, Ln )
•

n!0

given by cup-product. The key observation is that H 0 (X, L) has a natural action of G (L).
Under suitable assumptions we can identify G (L) with a Heisenberg group, in which case the
representation H 0 (X, L) can be described very precisely. This then allows to choose a basis
for H 0 (X, L) (the elements of which are referred to as theta functions) that has particular
properties.
What is sketched here is discussed in much greater detail in Chapter ?? below. First,
however, we shall study representations of non-degenerate theta groups.

(8.27) Definition. Let G be a theta group over a field k. If ρ: G → GL(V ) is a representation

of G then we say that ρ is a representation of weight n (n ∈ Z) if ρ ◦ i: Gm,k → GL(V ) is given
by z %→ z n · idV .
We shall mainly be interested in representations of weight 1.

(8.28) Theorem. Let k be an algebraically closed field. Let 0 → Gm,k → G → K → 0 be a

non-degenerate theta group over k such that rank(K) is prime to char(k). Then G has a unique
irreducible representation ρ = ρG : G → GL(V ) of weight 1 (up to isomorphism). We have
dim(V )2 = rank(K).
If W is any representation of G of weight 1 then W is isomorphic to a direct sum of copies
of ρG . More precisely, if H̃ ⊂ G is a maximal level subgroup then W ∼
= V ⊕a with a = dimk (W H̃ )
equal to the dimension of the subspace of H̃-invariants in W .
Proof. Choose a maximal level subgroup H̃ ⊂ G . (By (8.22) it is Lagrangian.) As rank(K) is
prime to char(k) and k = k we can view K and H̃ as constant groups.
Let τ : G → GL(W ) be a representation of weight 1. Viewing W as a module under H̃
(which is abelian) it decomposes as a direct sum of character spaces:
%
W = Wχ , with Wχ := {w ∈ W | τ (h)(w) = χ(h) · w for all h ∈ H̃} .
χ∈Hom(H̃,k∗ )

An element g ∈ G defines a character χg ∈ Hom(H̃, k ∗ ) by g −1 · h · g = χg (h) · h. (That is,

χg (h) = [g −1 , h] ∈ Gm .) Then τ (g)Wχ = Wχ·χg .

– 122 –
 As H̃ is Lagrangian, its centralizer CH̃ ⊂ G equals Gm · H̃ (cf. the proof of (8.16)) and
g %→ χg gives an isomorphism
∼
γ: CH̃ \G −→ Hom(H̃, k ∗ ) .
As furthermore the elements of CH̃ = Gm · H̃ act on each Wχ through scalar multiplications,
it follows that: (a) all Wχ have the same dimension, and (b) if 0 != w ∈ Wχ then the elements
τ (g)(w) span a G -submodule V ⊂ W with dim(V ∩ Wχ ) = 1 for all χ.
Choose a section σ: Hom(H̃, k ∗ ) ∼= CH̃ \G → G of the projection G → → CH̃ \G . Suppose
∗
that W is irreducible. Choose 0 != w1 ∈ W1 , where 1 ∈ Hom(H̃, k ) is the trivial character.
For χ ∈ Hom(H̃, k ∗ ) set wχ := τ (σ(χ))(w1 ) ∈ Wχ . Then {wχ } is a k-basis of W . If g ∈ G has
image η in Hom(H̃, k ∗ ) then there is a unique c = c(g, χ) ∈ CH̃ such that g · σ(χ) = c · σ(η · χ).
Then the representation τ is completely described by τ (g)(wχ ) = τ (c(g, χ))(wηχ ). (Note that
c(g, χ) ∈ Gm · H̃, so we know how it acts on the spaces Wψ .) As the elements c(g, χ) only depend
on the structure of G and the chosen section σ, it follows that there is at most one irreducible
representation of weight 1, up to isomorphism. Conversely, our description gives a simple recipe
of how to construct one. (See also (8.29) below.) This shows that there is a unique irreducible
representation ρ: G → GL(V ) of weight 1.
To prove the last assertions, write r = rank(K) = rank(H̃)2 and consider the subgroup
G [r] ⊂ G of elements g with g r = 1. As G is generated by Gm and G [r], a weight 1 representation
W of G is completely reducible (i.e., a direct sum of irreducible representations) if and only if it is
completely reducible as a representation of G [r]. But G [r] is a finite group of order not divisible
by char(k). Therefore all k-representations of G [r] are completely reducible. If W ∼ = V ⊕a then
H̃
a = dim(W1 ) = dim(W ). !

(8.29) The standard representation of a Heisenberg group. By (8.24), the theta group G in the
theorem is isomorphic to a Heisenberg group H = Gm,k × H × H D with

H∼
= (Z/d1 Z) × · · · × (Z/dn Z) , H D = Hom(H, k ∗ ) ∼
= µd1 (k) × · · · × µdn (k) .

We can take H̃ := {(1, h, 1) | h ∈ H} as a Lagrangian level subgroup. In the proof of the

theorem we have seen how to construct an irreducible weight 1 representation. The result can
be described as follows.
Let V be the space of functions on H with values in k. Then we have a representation
ρ: H → GL(V ) given by [ρ(λ, x, χ)(f )](h) = λ · χ(h) · f (x + h) for f ∈ V and h ∈ H. One easily
checks that this indeed gives an irreducible representation of weight 1.
More generally, let H be an arbitrary finite commutative group scheme over a field k. Write
AH := Γ(H, OH ) for its affine algebra and let H = Gm,k ×H ×H D be the associated Heisenberg
group, as defined in (8.21). Then we have a representation

ρ: H −→ GL(AH )

by letting (λ, x, χ) ∈ H act on AH by

f %→ λ · χ · t∗x (f ) ,

where we view χ as an invertible element of AH . More precisely, we should write (λ, x, χ) ∈

H (T ), where T is a k-scheme. For simplicity, assume that T = Spec(R) is affine. Then ρ(λ, x, χ)
is an R-linear automorphism of AH ⊗k R. Now notice that χ is given by an invertible element
of AH ⊗k R.

– 123 –
 Again this resentation ρ is irreducible of weight 1. We shall refer to it as ??.

(8.30) We now wish to lift the restrictions on the characteristic of k. In the case considered
above the desired representation was realized as a representation on the space of functions on a
Lagrangian level subgroup. Therefore is natural to consider representations of G on spaces of
functions on G .
We work over an algebraically closed field k. As K is a semi-local scheme, we can trivialize
∼
G as a Gm -torsor over K. So, we can choose an isomorphism G −→ Gm × K of K-schemes via
which the Gm -action on G corresponds to multiplication in Gm on the right hand term. Writing
K = Spec(A0 ) this gives G = Spec(B), with
%
B = A0 [t, t−1 ] = Ai , where we set Ai := A0 · ti .
i∈Z

We can view Ai as the vector space of those functions f : G → A1 such that f (λx) = λi · f (x)
for all λ ∈ Gm . (This has to be read functorially: if R is a k-algebra then Ai ⊗k R = {f ∈
HomR (GR , A1R ) | f (λx) = λi · f (x) for all λ ∈ Gm }.)
As will become clear in the proof of (8.32), the space A1 is the most interesting for us.
(That is, if we want to study representations of weight 1.) Note that dim(A1 ) is the square
of the dimension of the irreducible G -representation that we are looking for. We consider the
action of G × G on A1 given by

[(g, h) · f ](x) = f (h−1 xg) for f ∈ A1 , (g, h) ∈ G × G and x∈G.

(Again this has to be read functorially.)

(8.31) Lemma. With this action A1 is an irreducible G × G -module.

Proof. First we look at the diagonal G -action. If g ∈ G and f ∈ A1 then (g, g) · f is the function
given by

x %→ f (g −1 xg) = f (g −1 xgx−1 · x)
= f ([g −1 , x] · x)
= [g −1 , x] · f (x) (because f ∈ A1 and [g −1 , x] ∈ Gm )
= e(π(g)−1 , π(x)) · f (x) .

Each γ ∈ K defines a character e(γ, −): K → Gm . If χ is any such character then χ ◦ π: G →

∆
Gm ⊂ A1 can be viewed as an element of A0 . The previous calculations show that G −→ G ×G →
GL(A1 ) factors through G →
→ K and that the resulting action of K on A1 is given by

γ · f = [π ◦ e(γ −1 , −)] · f for γ ∈ K , f ∈ A1 .

Suppose that E ⊂ A1 is a G × G -submodule. The non-degeneracy of our theta group means

that every character χ: K → Gm is of the form e(γ −1 , −) for some γ ∈ K. Furthermore, the
k-characters χ form a k-basis of the k-algebra A0 ; see Exercise ??. It thus follows from the above
that E is an A0 -submodule of A1 . But A1 = A0 · t, so E is of the form IA1 with I an ideal
of A0 .
So far we have only used the diagonal action of G . Using the full action of G × G again we
see that E is stable under all translations by elements of G . Combined with the previous it now
readily follows that E = (0) or E = A1 !

– 124 –
 We can now generalize Theorem (8.28).

(8.32) Theorem. Let G be a non-degenerate theta group over an algebraically closed field k.
Then G has a unique irreducible representation ρ = ρG : G → GL(V ) of weight 1 (up to isomor-
phism). We have dim(V )2 = rank(K). If W is any representation of G of weight 1 then W is
isomorphic to a direct sum of copies of ρG .
Proof. Let τ : G → GL(W ) be a representation of G of weight 1. Then τ gives rise to a homo-
morphism of G × G -modules r: W ∗ ⊗ W → A1 , by

r(ϕ ⊗ w)(g) = ϕ(τ (g)(w)) for ϕ ∈ W ∗ , w ∈ W and g∈G.

Suppose that W is irreducible. Then W ∗ ⊗k W is an irreducible G × G -module. (Here we

need that k = k ! The point is that EndG ×G (W ∗ ⊗ W ) = EndG (W ∗ ) ⊗k EndG (W ). As k = k
the irreducibility of W implies that EndG (W ) = k = EndG (W ∗ ).) As r is obviously not the zero
map it follows from the lemma that r is an isomorphism. We conclude that A1 ∼ = W ⊕···⊕W
(dim(W ) factors) as a G -module, that there is a unique irreducible G -module of weight 1, and
that rank(K) = dim(A1 ) is the square of its dimension. We also see that A1 is completely
reducible as a G -module.
Now let W be an arbitrary G -module of weight 1 again. Then r gives a k-linear map
r% : W → Hom(W ∗ , A1 ) = W ⊗ A1 , sending w ∈ W to ϕ %→ r(ϕ ⊗ w). From r(ϕ ⊗ w)(1) = ϕ(w)
we see that r% is injective. Moreover, if we let G act on W ⊗ A1 through its action on A1 then
dim(W )
r% is G -equivariant. We conclude that W is isomorphic to a G -submodule of A1 . As A1 is
a completely reducible G -module, W is also completely reducible. !

Exercises.
(8.1) Let k be an algebraically closed field. Let K be a finite commutative group scheme of
order prime to char(k). Let e: K × K → Gm be a non-degenerate alternating bilinear pairing,
i.e., e is a morphism of k-schemes such that (a) e(x, y) = e(y, x)−1 , (b) for fixed x ∈ K the maps
y %→ e(x, y) and y %→ e(y, x) are homomorphisms, and (c) the homomorphism K → K D given
by x %→ e(x, −) is an isomorphism.
(i) Show that K is isomorphic to a constant group scheme of the form

K∼
= (Z/d1 Z) × (Z/d2 Z) × · · · × (Z/dn Z) ,

where we may require that d1 |d2 | · · · |dn . (And if char(k) = p > 0 then p " dn .)
(ii) Choose an element a ∈ K such that K is a product K = /a0 × K % . Let d be the order of
a and let ζd ∈ k be a primitive dth root of unity. Show that there is a unique b ∈ K with
e(a, b) = ζd and e(k, b) = 1 for all k ∈ K % .
(iii) Let K %% := {k ∈ K % | e(a, k) = 1}. Show that K decomposes as a product of groups
K = /a0 × /b0 × K %% . Also show that the restriction of e to K %% × K %% is again non-degenerate.
∼
(iv) Prove that there exists a finite commutative k-group scheme H and an isomorphism K −→
H × H D via which the pairing e corresponds to the pairing on H × H D given by
# $
(x, χ), (x% , χ% ) %→ χ% (x) · χ(x% )−1 .

– 125 –
Chapter IX. The cohomology of line bundles.

In this chapter we study the cohomology of line bundles on abelian varieties. The main results
are the Riemann-Roch Theorem (9.11) and the Vanishing Theorem for non-degenerate line
bundles (9.14). The key step in deriving these results is the computation of the cohomology of
the Poincaré bundle on X × X t .

(9.1) Theorem. Let X be a g-dimensional abelian variety over a field k. Let P be the Poincaré
bundle on X × X t and write p2 : X × X t → X t for the second projection. Then the sheaves
Rn p2,∗ P and the cohomology of P are given by
!
n 0 if n #= g;
R p2,∗ P =
i0 (k) if n = g,
and !
n t 0 if n #= g;
H (X × X , P) =
k if n = g.

Here i0 (k) denotes the skyscraper sheaf at 0 ∈ X t with stalk k.

Proof. As the proof is a somewhat long we divide it into steps, (9.2)–(9.9).

(9.2) We look at the higher direct image sheaves Rn p2,∗ P on X t . If y ∈ X t \ {0} then the
restriction of P to X × {y} is a non-trivial line bundle on X with class in Pic0 . As was proven
in (7.19) such sheaves have zero cohomology. Applying (i) of (7.20), it follows that Rn p2,∗ P
has support only at 0 ∈ X t , for all n. As the closed point 0 is a zero-dimensional subscheme
of X t we have H i (X t , Rn p2,∗ P) = 0 for all i ! 1. (Use HAG, III, Thm. 2.7 and Lemma 2.10.)
Applying the Leray spectral sequence

E2p,q = H p (X t , Rq p2,∗ P) ⇒ H p+q (X × X t , P)

we find that
H n (X × X t , P) ∼
= H 0 (X t , Rn p2,∗ P) . (1)

As p2 is projective of relative dimension g we have (HAG, III, Cor. 11.2) Rn p2,∗ P = 0 for all
n > g. Hence also H n (X × X t , P) = 0 for n > g.
Next we apply Serre duality to the Poincaré bundle. We have P −1 ∼ = (−1, 1)∗ P ∼ =
∗ −1
(1, −1) P; see Exercise (7.4). In particular the cohomology of P is the same as that of P.
As X × X t is an abelian variety its dualizing sheaf is trivial, and Serre duality (in the form given
by HAG, III, Cor. 7.7) gives

H n (X × X t , P) ∼
= H 2g−n (X × X t , P −1 )∨ ∼
= H 2g−n (X × X t , P)∨ .

Hence H n (X × X t , P) = 0 for all n < g too. By (1) and the fact that the Rn p2,∗ P are
supported at 0 we also have Rn p2,∗ P = 0 for n #= g.

CohomLB, 15 september, 2011 (812)

– 126 –
(9.3) Let A := OX t ,0 be the local ring of X t at 0. Let m ⊂ A be the maximal ideal. It follows
from (1) that (Rg p2,∗ P)0 is an A-module of finite length. By (??) the natural map
" #
(Rg p2,∗ P) ⊗OX t κ(0) −→ H g X × {0}, P|X×{0} = H g (X, OX ) ∼ =k

is an isomorphism. Using the Nakayama lemma, we find that (Rg p2,∗ P)0 ∼ = A/a for some
m-primary ideal a ⊂ A.
To complete the proof of (9.1) it remains to be shown that a = m. This is the hardest
part of the proof. We need to exploit the fact that P is the universal line bundle on X × X t ;
thus far we have not made full use of this. In particular, we know that P is trivial over
X × {0} = X × Spec(A/m), but not over any “thickening” X × Spec(A/J) for J ! m. The
problem is how to translate this into information about Rg p2,∗ P.
We shall give two proofs of the fact that a = m. The first proof uses Grothendieck duality
and is fairly short; the second relies on essentially the same ideas but is more elementary.

(9.4) Let Z be a scheme. Write Mod(Z) for the category of OZ -modules and D(Z) for its
derived category. If F is a sheaf of OZ -modules and n ∈ Z, write F [n] for the object of D(Z)
represented by the complex whose only non-zero term is the sheaf F , sitting in degree −n. The
functor Mod(Z) → D(Z) given by F *→ F [0] realizes Mod(Z) as a full subcategory of D(Z).
If C • is a complex of OZ -modules with the property that H i (C • ) = 0 for all i #= n, for some
integer n, then C • ∼
= H n (C )[−n] in D(Z).
•

To simplify notation we write Y := X × X t . A corollary of Grothendieck duality, applied to

the morphism p2 : Y → X t , is that for quasi-coherent OX t -modules G we have an isomorphism
∼
HomOY (P, p∗2 G) −→ HomOX t (Rg p2,∗ P, G) , (2)

which is functorial in G. Before we start exploiting this, let us indicate how this is obtained
from the general machinery of Grothendieck duality.
We already know that p2 is a smooth morphism of relative dimension g and that ΩgY /X t ∼ =
OY . Consider a bounded complex F • of quasi-coherent OY -modules and a bounded complex G•
of quasi-coherent OX t -modules. Then a consequence of Grothendieck duality is that we have a
canonical isomorphism
" • # ∼
HomD(Y ) F , p∗2 G [g] −→ HomD(X t ) (Rp2,∗ F , G ) .
• • •
(3)

See Hartshorne [1], Chap. III, § 11, and use that the functor p!2 is given by G• *→ p∗2 G• [g]; see
op. cit., Chap. III, § 2. We apply this with F • = P. We already know that Rp2,∗ F • only has
cohomology in degree g. As explained above, this implies that Rp2,∗ F • is isomorphic, in D(X t ),
to Rg p2,∗ P[−g]. If we now apply (3) with G• = G[−g] for some quasi-coherent OX t -module G
then we obtain (2).

(9.5) Let J ⊂ A be a proper ideal (with A = OX t ,0 , as above). Write Z(J) := Spec(A/J),

t
# Z(J) "→ X be the natural closed immersion. Write Y (J) := X × Spec(A/J) =
and"let i(J):
−1
p2 Z(J) ⊂ Y . In particular, Y (m) = X ×"{0}. If we write # P(J) for the restriction of P
to Y (J) then HomOY (P, OY (J) ) = HomOY (J) P(J), OY (J) .
Suppose J is an m-primary ideal. Via the natural map OX t → i(J)∗ OZ(J) , the structure
sheaf OZ(J) is then just the skyscraper sheaf i0 (A/J) at 0 ∈ X t with stalk A/J. Further Y (J) is
the closed subscheme of Y = X × X t with underlying topological space |X × {0}| and structure
sheaf p∗2 OZ(J) = OX ⊗k A/J.

– 127 –
 As explained in (9.3), we have Rg p2,∗ P = i0 (A/a) for some m-primary ideal a ⊂ A. Now
consider the commutative diagram
∼ " #
HomOY (P, OY (a) ) −→ HomOX t i0 (A/a), i0 (A/a) ∼
= A/a
  
% % %
∼ " #
HomOY (P, OY (m) ) −→ HomOX t i0 (A/a), i0 (k) ∼
= k ,

where the horizontal arrows are given by (2) and the vertical arrows are induced by the quotient
map A/a → A/m = k.
∼
We have a natural isomorphism h: P(m) = P|X×{0} −→ OX . This gives us an element
" # " #
h ∈ HomOY P, OY (m) = HomOX×{0} P|X×{0} , OX×{0} .

From the diagram we see that h can be lifted to an element

" #
h̃ ∈ HomOY (P, OY (a) ) = HomOY (a) P(a), OY (a) .

Then h̃: P(a) → OY (a) is a homomorphism of line bundles on Y (a) which is an isomorphism
modulo m. It follows that h̃ is an isomorphism, too. This shows that the pull-back of P under
idX × i(a): X × Spec(A/a) "→ X × X t is trivial. By the universal property of P this implies
that i(a): Spec(A/a) "→ X t factors through the closed point {0} = Spec(k) ⊂ X t . Hence a = m
and Rg p2,∗ P = i0 (k). This finishes our (first) proof of Theorem (9.1). "

(9.6) Our second proof that (Rg p2,∗ P)0 = ∼ k is not very different from the first, but it replaces
Grothendieck duality by more explicit arguments.
We use the notation introduced in (9.5). In particular, if J ⊂ A is a proper ideal, the second
projection p2 : X × X t → X t restricts to a morphism p2 : Y (J) → Z(J). We " shall
# systematically
g t
g
" # of global sections. Note that Z (0) g = Spec(A) →
confuse R p2,∗ P(J) with its A/J-module " X#
is a flat morphism; hence R p2,∗ P (0) is the same as the restriction of "R p#2,∗ P to Z (0) .
(See HAG, Chap. III, Prop. 9.3.) Thus, our goal is to prove that Rg p2,∗ P (0) ∼ = k.
We apply the results about cohomology and base-change explained in (??). This gives us a
length g complex (with g = dim(X) = dim(A)) of finitely generated free A-modules

d0 d1 dg−1
0 −→ K 0 −−→ K 1 −−→ · · · −→ K g−1 −−−→ K g −→ 0
•
K : (4)

with the property that for all ideals J ⊂ A and all n we have

Rn p2,∗ P(J) ∼
= H n (K ⊗A A/J) ,
•

functorially in A/J. (In fact, a similar statement holds with A/J replaced by an arbitrary A-
algebra, but we will not need this.) In particular H n (K • ) ∼
= Rn p2,∗ P. But as shown in (9.2),
R p2,∗ P = 0 for n < g; so K is a resolution of H := H g (K • ). We want to show that
n • g

Hg∼ = A/m = k.
Consider the “dual” complex

δ0 δ1 δ g−1
0 −→ L0 −−→ L1 −−→ · · · −→ Lg−1 −−−→ Lg −→ 0
•
L :

– 128 –
where Lj := HomA (K g−j , A), and where δ j is the map induced by dg−1−j . Set

Q := H g (L ) = Coker(δ g−1 : Lg−1 → Lg ) .

•

The next lemma (taken from MAV, p. 127) tells us that L• is a free resolution of Q. (Note
that all H n (L• ) are artinian A-modules, as easily follows from the corresponding fact for the
complex K • .)

(9.7) Lemma. Let A be a g-dimensional regular local ring. Let

0 −→ C 0 −→ C 1 −→ · · · −→ C g −→ 0
•
C :

be a complex of finitely generated free A-modules such that all cohomology groups H j (C • ) are
artinian A-modules. Then H j (C • ) = 0 for all j < g.
Proof. We use induction on g. For g = 0 there is nothing to prove, so we may assume that g > 0
and that the lemma holds in smaller dimensions. Choose x ∈ m \ m2 , so that A/(x) is regular
•
of dimension g − 1. Put C := C • /(x), so that we have an exact sequence of complexes

• ·x • •
0 −→ C −→ C −→ C −→ 0 .

In cohomology this gives the long exact sequence

·x ·x
· · · −→ H i (C ) −→ H i (C ) −→ H i (C ) −→ H i+1 (C ) −→ H i+1 (C ) −→ · · ·
• • • • •
.

• •
We see from this that the H i (C ) are artinian modules, and by induction H i (C ) = 0 for all
i < g − 1. Hence multiplication by x is injective on H j (C • ) for all j < g. But H j (C • ) is
artinian, so it is killed by xN for N + 0. This proves the induction step. "

(9.8) From (7.27) we know the cohomology of the complex K • ⊗A k = [0 → K 0 /mK 0 →

K 1 /mK 1 → · · ·]. In particular we have H 0 (K • ⊗A k) = H 0 (X, OX ) = k and H g (K • ⊗A k) =
H g (X, OX ) = k. This gives us that H g /mH g ∼ = k and Q/mQ ∼ = k. By Nakayama’s Lemma
g
it follows that the A-modules H and Q are both generated by a single element, so there exist
ideals a and b of A with
Hg∼ = A/a and Q∼= A/b .

(For H g this repeats what was explained in (9.3).)

Let J ⊂ A be an ideal. Put
" #
HJ0 := Ker(K 0 /JK 0 → K 1 /JK 1 ) = H 0 Y (J), P(J) .

Applying HomA (−, A/J) to the exact sequence Lg−1 /JLg−1 → Lg /JLg → Q/JQ → 0 gives
the exact sequence

d¯0
0 −→ HomA (Q/JQ, A/J) −→ K 0 /JK 0 −−→ K 1 /JK 1 ,

which shows that

HJ0 ∼
= HomA (A/b + J, A/J) . (5)

– 129 –
The isomorphism (5) is functorial in the ideal J, in the sense that for J1 ⊆ J2 the natural
reduction map HJ01 → HJ02 corresponds to the natural map

HomA (A/b + J1 , A/J1 ) → HomA (A/b + J1 , A/J2 ) = HomA (A/b + J2 , A/J2 ) .

We now use that, by definition of X t , the closed point 0 ∈ X t is the maximal closed
subscheme over which P is trivial, in the sense of (2.4). " Taking J1 #= b and J2 = m in the
above we find that the section 1 ∈ k = H 0 (X, OX ) = H 0 Y (m), P(m) lifts to a global section
of P(b). With the same arguments as in (9.5) it follows that P(b) ∼
= OY (b) , and by the universal
property of P this is possible only if b = m.

(9.9) We have shown that L• is a free resolution of the A-module A/m = k. Another way to
obtain such a resolution is to use a Koszul complex. This works as follows. Choose a regular
system of parameters x1 , x2 , . . . , xg ∈ m, i.e., a sequence of elements which generate m and which
give a k-basis for m/m2 . Consider the complex

0 −→ F 0 −→ F 1 −→ · · · −→ F g −→ 0
•
F :

where F j = ∧jA (Ag ) and where, writing x = (x1 , . . . , xg ) ∈ Ag , the differential

dj : ∧jA (Ag ) −→ ∧j+1 g

A (A )

is given by v *→ x ∧ v. Then F • , the so-called Koszul complex associated to the sequence x, is

also a free resolution of k.
By (??) the complexes L• and F • are homotopy equivalent. “Dualizing back” we then find
that the complex K • is homotopy equivalent to the dual of the Koszul complex F • . The first
terms of F • are given by

d0
0 −→ A −−→ Ag −→ · · · with d0 : a *→ (x1 a, x2 a, . . . , xg a) .

The last non-zero terms of the dual complex are therefore given by

(d0 )∗
· · · −→ Ag −−−−→ A −→ 0 with (d0 )∗ : (a1 , a2 , . . . , ag ) *→ x1 a1 + x2 a2 + · · · + xg ag .

With this we can finally compute:

" • # " #
(Rg p2,∗ P)0 ∼
= H g (K ) ∼
= H g (F )∗ = Coker (d0 )∗ = A/m = k
•

and this finishes the (second) proof of Theorem (9.1). "

(9.10) The following result we want to prove is the Riemann-Roch theorem for abelian varieties.
Let X be a proper scheme of finite type over a field k. If F is a quasi-coherent OX -module
then its Euler characteristic is defined to be the integer
&
χ(L) := (−1)i · dimk H i (X, F ) .
i!0

Suppose X is projective and H is a very ample line bundle on X. Then n *→ χ(F ⊗ H n )

is a polynomial function of n. More precisely, there is a polynomial with rational coefficients

– 130 –
Φ = ΦF,H ∈ Q[t], called the Hilbert polynomial of F (with respect to H), such that Φ(n) =
χ(F ⊗ H n ) for all n ∈ Z. Note that there is a natural number n0 such that H i (X, F ⊗ H n ) = 0
for all i > 0 and all n ! n0 ; hence Φ(n) = dimk H 0 (X, F ⊗ H n ) for all n ! n0 .
This “polynomial behaviour” of the Euler characteristic with respect to its entries is a
much more general phenomenon. For instance, suppose X is a smooth proper variety over k
and F1 , . . . , Fr are vector bundles on X (or, more generally, coherent OX -modules). Then the
function (n1 , . . . , nr ) *→ χ(F1n1 ⊗ · · · ⊗ Frnr ) is polynomial in the r-tuple of integers (n1 , . . . , nr ).
This is a consequence of the Hirzebruch-Riemann-Roch theorem. When X is an abelian variety
the Riemann-Roch formula takes a particularly simple form and the polynomial dependence of
χ(F1n1 ⊗ · · · ⊗ Frnr ) on the exponents ni becomes obvious; cf. (9.13).

(9.11) Riemann-Roch Theorem. Let L be a line bundle on a g-dimensional abelian vari-

ety X. Then
χ(L) = c1 (L)g /g! and χ(L)2 = deg(ϕL ) .

Thus, if L ∼= OX (D) for some divisor D then the first equation says that χ(L) equals
(D )/g!, where (Dg ) is g-fold self-intersection
g
" # ' number of D. Notice that, by slight abuse of
g g g
notation, we write c1 (L) for deg c1 (L) = X c1 (L) .
We shall prove the theorem together with the following corollary.

(9.12) Corollary. Let f : Y → X be an isogeny. If L is a line bundle on X then χ(f ∗ L) =

deg(f ) · χ(L).
Proof of (9.11) and (9.12). First we show that χ(L) = c1 (L)g /g!. For this we use the Hirzebruch-
Riemann-Roch formula, which says that
(
χ(L) = ch(L) · td(TX ) . (6)
X

Here ch(L), the Chern character of L, is the power series

" #
ch(L) = exp c1 (L) = 1 + c1 (L) + c1 (L)2 /2 + · · ·

which should be thought of as a formal expression. Similiarly, td(TX ), the Todd class of the
tangent bundle TX , is a formal power series in the Chern classes of TX . As TX is trivial we
have ci (T
' X ) = g0 for all i ! 1 and td(TX ) = 1. This reduces (6) to the desired equality
χ(L) = X c1 (L) /g!. Notice that, in particular,

χ(Lm ) = mg · χ(L) (7)

for all m ∈ Z.
To prove (9.12)
∗ g ∗
" we
g
# may assume that k = k. Let f : Y → X be an isogeny of degree d. Then
c1 (f L) = f c1 (L) in the Chow ring of Y . (Alternatively we may use any Weil cohomology,
such as &-adic cohomology for some & #= char(k), or Betti cohomology in case the ground field
g
is C.) But c1 (L)
' ∗is represented by a 0-cycle (a formal sum of points), so all that remains−1 to be
shown is that Y f [P ] = d for every point P ∈ X. This is clear if f is separable, for then f (P )
consists of d distinct points, each with multiplicity 1. It is also clear if f is purely inseparable,
because then f −1 (P ) consists of one single point, say Q, and OY,Q is free of rank d over OX,P .
The general result follows by combining these two cases, using (5.8). This proves Cor. (9.12).

– 131 –
 2
Next we show that χ(L)
" =
# deg(ϕL ). We first do this for non-degenerate line bundles L.
The idea is to compute χ Λ(L) in two different ways.
So, assume that L is non-degenerate. As usual we write Λ(L) for the associated Mumford
bundle on X × X. We have a cartesian diagram
idX ×ϕL
X ×X −−−−−−→ X × Xt
 

p2 %
p $
% 2
X −−→ Xt .
ϕL

Further we know that Λ(L) = (idX × ϕL )∗ P, and ϕL is an isogeny with kernel {1} × K(L). By
(9.1) and flat base change,
!
n
" # 0 if n #= g;
R p2,∗ Λ(L) = ϕ∗L Rn p&2,∗ P =
i∗ OK(L) if n = g,

where i: K(L) "→ X is the inclusion. Using a Leray spectral sequence, as in (9.2), we find
!
n
" # 0 if n #= g;
h X × X, Λ(L) = (8)
deg(ϕL ) if n = g.

Here, as usual, we write hn (−) := dim H n (−). In particular,

" #
χ Λ(L) = (−1)g · deg(ϕL ) . (9)

(A quicker proof of (9) is to use (9.12),

" but #we shall need (8) later.) ∗
For the second computation of χ Λ(L) , recall that Λ(L) := m L ⊗ p∗1 L−1 ⊗ p∗2 L−1 . The
projection formula therefore gives
" #
Rn p2,∗ Λ(L) = Rn p2,∗ m∗ L ⊗ p∗1 L−1 ⊗ L−1 .

We know that Rn p2,∗ Λ(L) is supported on the finite subscheme K(L) ⊂ X. As L can be
trivialized over K(L) we find that
" # " #
Rn p2,∗ m∗ L ⊗ p∗1 L−1 ⊗ L−1 = Rn p2,∗ m∗ L ⊗ p∗1 L−1 .

Once again computing cohomology via a Leray spectral sequence we conclude that
" #
H n X × X, Λ(L) ∼= H n (X × X, m∗ L ⊗ p∗1 L−1 ) for all n. (10)
" #
Now remark that (m × p1 ): X × X → X × X is an isomorphism with (m × p1 )∗ p∗1 L ⊗ p∗2 L−1 =
m∗ L ⊗ p∗1 L−1 . By the Künneth formula it follows that
)
H n (X × X, m∗ L ⊗ p∗1 L−1 ) ∼
= H n (X × X, p∗1 L ⊗ p∗2 L−1 ) ∼
= H p (X, L) ⊗ H q (X, L−1 ) . (11)
p+q=n

Combining (10) and (11) we find

" # " #
χ Λ(L) = χ p∗1 L ⊗ p∗2 L−1 = χ(L) · χ(L−1 ) = (−1)g · χ(L)2 , (12)

– 132 –
where the last equality follows from (7). Comparing the two answers (9) and (12) proves that
χ(L)2 = deg(ϕL ) for non-degenerate L.
Now suppose that L is degenerate. Then ϕL is not finite and, by convention, deg(ϕL ) = 0.
We want to show that χ(L) = 0 too. It is still true that Λ(L) = (idX × ϕL )∗ P. We rewrite this
as
" #
m∗ L ⊗ p∗2 L−1 = (idX × ϕL )∗ P ⊗ p∗1 L .
" #
The same argument as above gives that χ m∗ L ⊗ p∗2 L−1 = (−1)g · χ(L)2 . (Notice that this part
of the above argument works without the assumption that L is non-degenerate.) If H ⊂ K(L) is
" idX × ϕL factors
a subgroup scheme of order r then # through the projection X × X → X × X/H,
and by (9.12) it follows that χ m∗ L ⊗ p∗2 L−1 is divisible by r. But K(L) contains subgroup
schemes of arbitrarily large order (in fact, K(L)0red is an abelian subvariety of X of positive
dimension), and we conclude that χ(L) = 0. This finishes the proof of the theorem. "

(9.13) Remark. If F is a coherent

' sheaf on a g-dimensional abelian variety X then Hirzebruch-
Riemann-Roch gives χ(F ) = X chg (F ) where chg is a certain polynomial in the Chern classes
of F . See Fulton [1], Example 3.2.3.
Looking at the proof of (9.11) we see that for non-degenerate bundles we can draw one
further conclusion.

(9.14) Vanishing Theorem. If L is a non-degenerate line bundle then there is a unique

integer i (necessarily with 0 # i # g) such that H i (X, L) #= 0.
Proof. Combining (8), (10) and (11) we have shown that

& !
p q −1 0 if n #= g;
h (L) · h (L )=
p+q=n
deg(ϕL ) if n = g.

As all hi (L) and hj (L−1 ) are in Z!0 this is possible only if there are unique p and q (with
p + q = g) such that hp (L) #= 0 and hq (L−1 ) #= 0. "

(9.15) Definition. If L is a non-degenerate line bundle then the unique index i = i(L) such
that hi (L) #= 0 is called the index of L.
Note that i(L) = 0 just means that L is effective.

(9.16) Example.
" Let# D be a divisor of degree d on an elliptic curve E. Riemann-Roch for
curves gives χ OE (D) = d. It follows that

D is degenerate ⇐⇒ d=0
D is non-degenerate of index 0 ⇐⇒ d>0
D is non-degenerate of index 1 ⇐⇒ d<0

(9.17) Corollary. Let X be an abelian variety over an algebraically closed field k. Let L be a
non-degenerate line bundle on X with index i = i(L). Then H i (X, L) is the unique irreducible
weight 1 representation of the theta group G (L).

– 133 –
Proof. That H i (X, L) is a G (L)-representation of weight 1 is clear, for instance, using Čech
cohomology. The corollary thus follows from (8.32) by a dimension count. Indeed, we have
" #2 " #
dim H i (X, L) = χ(L)2 = deg(ϕL ) = rank K(L) ,

as required. "
If L is a non-degenerate line bundle with index i then χ(L) = (−1)i · hi (L). In particular,
χ(L) has sign equal to (−1)i(L) . We shall later see how the index can be read off from the
Hilbert polynomial of L. As a preparation for this we collect some properties of the index as a
function on the set of non-degenerate bundles.

(9.18) Proposition. (i) Let L be a non-degenerate line bundle on a g-dimensional abelian

variety X. Then i(L−1 ) = g − i(L).
(ii) “The index is (locally) constant in algebraic families”: If T is a locally noetherian k-
scheme and M is a line bundle on X × T such that all Mt := M|X×{t} are non-degenerate then
the function t *→ i(Mt ) is locally constant on T . In particular, if L is as in (i) and L& is a line
bundle on X with [L& ] ∈ Pic0X/k then i(L) = i(L ⊗ L& ).
(iii) Let f : X → Y be an isogeny of degree prime to char(k). If M is a non-degenerate line
bundle on Y then f ∗ M is non-degenerate too and i(f ∗ M ) = i(M ).
(iv) If L is non-degenerate and m #= 0 then Lm is non-degenerate too. Furthermore, if
m > 0 and char(k) " m then i(Lm ) = i(L).
(v) If L1 , L2 and L1 ⊗ L2 are all non-degenerate then i(L1 ⊗ L2 ) # i(L1 ) + i(L2 ).
(vi) If H is ample and L and L ⊗ H are both non-degenerate then i(L ⊗ H) # i(L).
Notes: In (9.23) below we shall sharpen (iv), showing that i(Lm ) = i(L) for all m > 0. In
(9.26) we shall show that (iii) holds without the assumption that deg(f ) is prime to char(k). If in
(ii) the scheme T is geometrically connected then it suffices to require that Mt is non-degenerate
for some t ∈ T (as K(Mt ) does not jump in such families), and the conclusion is that t *→ i(Mt )
is constant on T . The requirement that T is locally noetherian is in fact superfluous, as we can
reduce to the “universal” case T = PicX/k .
Proof. Statement (i) was already found in the proof of (9.14). Alternatively, it follows from
Serre duality.
The first statement of (ii) follows from the fact (HAG, III, Thm. 12.8) that for all j the
function t *→ dimk(t) H j (X ⊗ k(t), Mt ) is upper semi-continuous. The second statement follows
by applying this to the Poincaré bundle over X × PicX/k . Alternatively, passing to an algebraic
closure of k the bundles L ⊗ L& with [L& ] ∈ Pic0X/k are precisely the line bundles of the form t∗x L.
In cohomology the translation tx induces an isomorphism between H j (X, L) and H j (X, t∗x L).
(iii) As shown in (7.6), f ∗ M is again non-degenerate. We have f∗ (f ∗ M ) = M ⊗OY f∗ OX .
We claim that the sheaf OY is a direct summand of f∗ OX , hence M is a direct summand of
f∗ f ∗ M . Indeed, if r = deg(f ) then f∗ OX is locally free of rank r over OY and by assumption r
is invertible in OY . If trace: f∗ OX → OY is the trace map then (1/r) · trace is a section of the
natural map OY → f∗ OX , so f∗ OX = OY ⊕ Ker(trace).
Since f is finite, a Leray spectral sequence shows that H i (X, f ∗ M ) ∼
= H i (Y, f∗ f ∗ M ) for
i
all i (see also HAG, III, Exercise 4.1), and we conclude that H (Y, M ) is isomorphic to a direct
summand of H i (X, f ∗ M ). This proves
" #(iii).
m −1
(iv) We have K(L ) = m K(L) . Hence Lm is non-degenerate for m #= 0. Now assume
that m > 0 is relatively prime with char(k). We use the notation and the results of Exercise (7.8).

– 134 –
Consider the line bundle L"4 on X 4 given by L"4 = L"id4 = ⊗4i=1 p∗i L. (Here id4 denotes the
identity
" "4 # matrix of size 4 × 4.) It is readily seen that L"4
" is #again non-degenerate (in fact,
K L = K(L) ), and by the Künneth formula we have i L"4 = 4 · i(L).
4

We write m > 0 as a sum of four squares, say m = a2 + b2 + c2 + d2 . Consider the matrix

 
a −b −c −d
 b a −d c 
A= c d

a −b 
d −c b a

which should be thought of as representing the quaternion a+bi+cj+dk. We have At ·A = m·id4 .

Now consider the homomorphism" α =# αA : X 4 → X 4 associated to A, and apply part (i) of
Exercise (7.8). This gives that α∗ L"4 and (Lm )"4 differ by something in Pic0X/k ; hence by (ii)
" #
they have the same index. But by (iii) the index of α∗ L"4 equals that of L"4 . Putting
everything together we find that
" # " #
i(L) = 1/4 · i L"4 = 1/4 · i (Lm )"4 = i(Lm ) ,

as claimed.
(v) Let i1 , i2 and ι be the indices of L1 , L2 and L1 ⊗L2 , respectively. Consider the line bundle
N := p∗1 L1 ⊗ p∗2 L2 on X × X, and let ν: X × X → X be given by ν(x, y) = x − y. The fibre of ν
over 0 is the diagonal X ∼ = ∆(X) ⊂ X × X, over which N restricts to the bundle L1 ⊗ L2 . By (ii)
it follows that all fibres of N have index ι, so that Rj ν∗ N = 0 for all j < ι. By a Leray spectral
sequence this implies that H j (X × X, N ) = 0 for all j < ι. But the Vanishing Theorem together
with the Künneth decomposition show that H i1 (X, L1 ) ⊗k H i2 (X, L2 ) ∼ = H i1 +i2 (X × X, N ).
Finally, (vi) follows from (v), as it follows from (iv) that ample bundles have index 0. "

(9.19) Remark. The fact used in the proof of (iii) that OY is a direct summand of f∗ OX is
not necessarily true if the degree of f is divisible by char(k). For instance, suppose X is an
abelian variety over a field k of characteristic p > 0, such that X is not ordinary, i.e., f (X) < g.
Then the relative Frobenius map FX/k : X → X (p) is an isogeny of abelian varieties, but it
can be shown that OX (p) is in this case not a direct summand of FX/k,∗ OX . In the literature
one finds this as the statement that a non-ordinary abelian variety is not Frobenius split; see
Mehta-Srinivas [??].
For the proof of the following proposition we need a somewhat technical, but important
lemma.

(9.20) Lemma. Let Y be a d-dimensional projective scheme over a field. Let L1 , · · · , Lr be line
bundles on Y . For a = (a1 , . . . , ar ) ∈ Zr , set |a| := |a1 | + · · · + |ar | and La := La1 1 ⊗ · · · ⊗ Lar r .
Then there is a constant C > 0, only depending on Y and the bundles Lj , such that

hi (Y, La ) # C · (1 + |a|r )

for all i and all a ∈ Zr .

Proof. If all Li are trivial then the assertion is clear; this covers the cases d = 0 and r = 0. Next
we reduce to the case when all Lj are very ample. For this, choose a very ample bundle M such
that each of the
Mj := Lj ⊗ M (1 # j # r)

– 135 –
is very ample, too. Suppose we know the lemma for the line bundles M1 , . . . , Mr , Mr+1 . If C is
the resulting constant then for all a ∈ Zr , putting σ(a) := a1 + · · · + ar ,

−σ(a) " #
hi (Y, La ) = hi (Y, M1a1 ⊗ · · · ⊗ Mrar ⊗ Mr+1 ) # C · 1 + {|a| + |σ(a)|}r
" #
# C · 1 + {2|a|}r # (3r+1 C) · (1 + |a|r ) .

From now on we may therefore assume all Lj to be very ample.

We proceed by induction on the integer d + r. The case d + r = 0 is already dealt with.
Assume the lemma is true whenever d + r # ν. As the lemma is true when r = 0, it suffices to
do the case when we have r + 1 very ample bundles, say L1 , . . . , Lr and M , on a d-dimensional
projective scheme Y , such that d + r + 1 = ν + 1.
Let Z ⊂ Y be a hyperplane section for the projective embedding given by M . For every
a ∈ Zr and b ∈ Z we have an exact sequence

0 −→ La ⊗ M b−1 −→ La ⊗ M b −→ (La ⊗ M b )|Z −→ 0 .

In cohomology this gives an exact sequence

H i−1 (Z, La ⊗ M b ) −→ H i (Y, La ⊗ M b−1 ) −→ H i (Y, La ⊗ M b ) −→ H i (Z, La ⊗ M b )

which gives
hi (Y, La ⊗ M b ) # hi (Y, La ⊗ M b−1 ) + hi (Z, La ⊗ M b ) (13)

and
hi (Y, La ⊗ M b−1 ) # hi (Y, La ⊗ M b ) + hi−1 (Z, La ⊗ M b ) . (14)

By induction hypothesis we have estimates for hi (Y, La ⊗ M b ) when b = 0 and for the terms
hi (Z, La ⊗ M b ). For b > 0 we get the desired estimates by iterated application of (13); for b < 0
we do the same using (14). "

(9.21) To obtain further results on the index function, we investigate in more detail what
happens in the situation of (vi) in (9.18). We fix a non-degenerate bundle L and an ample
bundle H. As remarked above, ample bundles have index 0; in other words: they are effective.
Set l = c1 (L) and h = c1 (H). Consider the homogeneous polynomial of degree g

P (s, t) := (sl + th)g ∈ Z[s, t] ,

whose coefficients are intersection numbers. Notice that P (m, n) = g! · χ(Lm ⊗ H n ) for all
integral m and n. Further note that P is homogeneous of degree g, so P (m, n) = mg P (1, n/m) =
g!mg ΦL,H (n/m) where ΦL,H is the Hilbert polynomial of L with respect to H. In other words:
P is “the Hilbert polynomial made homogeneous of degree g”. If we want to indicate which
bundles L and H we are working with then we use the notation PL,H . For later use let us remark
that
PLm ,H (s, t) = PL,H (ms, t) = mg · PL,H (s, t/m) (15)

for all integers m #= 0.

(9.22) Proposition. Suppose that both L and L ⊗ H are non-degenerate, and that i(L) #=
i(L ⊗ H). Then P (1, t) has a root in the interval [0, 1] ⊂ R.

– 136 –
Proof. Let M be a square not divisible by char(k). By (iv) of (9.18) we have

i(LM ) = i(L) #= i(L ⊗ H) = i(LM ⊗ H M ) .

Assume that P (1, t) does not vanish on [0, 1], so that there exists a constant C > 0 with
|P (1, t)| > C for all t ∈ [0, 1]. As degenerate line bundles have zero Euler characteristic this
implies that all line bundles LM ⊗ H n with 0 # n # M are non-degenerate. Let n be the
smallest positive integer such that i(LM ⊗ H n−1 ) #= i(LM ⊗ H n ). Set

i1 = i(L) = i(LM ) = · · · = i(LM ⊗ H n−1 )

i2 = i(LM ⊗ H n ) ,

and observe that i2 < i1 by (vi) of (9.18).

Choose an effective divisor D ∈ |H| and consider the short exact sequence

0 −→ LM ⊗ H n−1 −→ LM ⊗ H n −→ (LM ⊗ H n )|D −→ 0 .

Looking at the associated long exact cohomology sequence and using that i1 > i2 we find that

H i2 (X, LM ⊗ H n ) "−→ H i2 (D, LM ⊗ H n ) .

In particular, hi2 (D, LM ⊗ H n ) ! M g · |P (1, n/M )|, which by our choice of C is at least
M g · C. Since this holds with arbitrarily large M , and since D has dimension g − 1, we obtain
a contradiction with (9.20). "

(9.23) Corollary. If L is non-degenerate then i(Lm ) = i(L) for all m > 0.

Proof. Write L = H1 ⊗ H2−1 as the difference of two ample bundles. Choose M ! 2 big enough
such that both polynomials PL,H1 (1, t) and PL,H2 (1, t) have no zeroes in the interval [0, 1/M ],
which is possible since PL,Hj (1, 0) = g! · χ(L) #= 0. By (15) it follows that for m ! M both
PLm ,H1 (1, t) and PLm ,H2 (1, t) have no zeroes in the interval [0, 1]. By the proposition this implies

i(Lm+1 ) = i(Lm+1 ⊗ H2 ) = i(H1m+1 ⊗ H2−m ) = i(Lm ⊗ H1 ) = i(Lm ) .

Hence for large enough m the index of Lm is independent of m. Using properties (i) and (iv)
in (9.18) the corollary follows. "

(9.24) Lemma. Let L be non-degenerate, H ample, and let P (s, t) := PL,H (s, t) be the poly-
nomial defined above. Suppose P (1, t) has a unique root τ ∈ [0, 1], of multiplicity µ and with
τ #= 1. Then i(L) # i(L ⊗ H) + µ.
Proof. As PLm ,H m (s, t) = m2g · PL,H (s, t) we may assume, using (9.23), that H is very ample.
Also we may assume that i(L) #= i(L ⊗ H), so that also i(Lm ) #= i(Lm ⊗ H m ) for all m #= 0.
Let m ∈ Z>0 and n ∈ Z with n < m. In the rest of the proof we shall only consider
integers m which are coprime with all denominators of rational roots of P (1, t). This ensures
that Lm ⊗ H n is non-degenerate; indeed, if Lm ⊗ H n is degenerate then P (1, n/m) = 0.
With m and n as above, suppose that i(Lm ⊗ H n−1 ) #= i(Lm ⊗ H n ). Note that

n−1+t
PLm ⊗H n−1 ,H (1, t) = mg · PL,H (1, ),
m
– 137 –
so it follows from (9.22) that P (1, t) has a root in the interval [(n − 1)/m, n/m]. By the
assumptions of the lemma we conclude that for given m > 0 there is a unique n with 1 < n # m
(depending on m) such that

i(Lm ) = · · · = i(Lm ⊗ H n−1 ) > i(Lm ⊗ H n ) = · · · = i(Lm ⊗ H m ) . (16)

Let X =: Z0 ⊃ Z1 ⊃ Z2 ⊃ · · · be obtained by taking hyperplane sections for the projective

embedding given by H. So, Z1 ⊂ X is a hyperplane section, Z2 is a hyperplane section of Z1 ,
etc. We have exact sequences

0 −→ (Lm ⊗ H q−1 )|Zr −→ (Lm ⊗ H q )|Zr −→ (Lm ⊗ H q )|Zr+1 −→ 0 . (17)

Fix m > 0 and let n = n(m) < m be determined by (16). Set i1 := i(L) = i(Lm ) and
i2 := i(L ⊗ H) = i(Lm ⊗ H m ). Note that i1 > i2 and i(Lm ⊗ H q ) ! i1 for all q # n − 1. Similar
to what we did in the proofs of (9.20) and (9.22), we shall use the exact sequences (17) to obtain
dimension estimates for cohomology groups. As a first step, take r = 0 in (17). Since i2 < i1
we find that H i2 (X, Lm ⊗ H n ) injects into H i2 (Z1 , Lm ⊗ H n ) and that H j (Z1 , Lm ⊗ H q ) = 0
for all j < i1 − 1 and q # (n − 1). Next we want to take r = 1, in which case we have the exact
sequence

H i2 (Z1 , Lm ⊗ H n−1 ) −→ H i2 (Z1 , Lm ⊗ H n ) −→ H i2 (Z2 , Lm ⊗ H n ) .

Applying the previous conclusions we see that the first term vanishes if i2 < i1 − 1. If this holds
then H i2 (Z1 , Lm ⊗ H n ) injects into H i2 (Z2 , Lm ⊗ H n ); further we then find that H j (Z2 , Lm ⊗
H q ) = 0 for all j < i1 − 2 and q # (n − 1).
Proceeding by induction we find that if r < i1 − i2 then

H i2 (Zr−1 , Lm ⊗ H n ) "−→ H i2 (Zr , Lm ⊗ H n )

and
H j (Zr , Lm ⊗ H q ) = 0 for all j < i1 − r and q # (n − 1).
(The induction breaks down for r ! i1 − i2 .) The conclusion of this (terminating) induction is
that H i2 (X, Lm ⊗ H n ) maps injectively to H i2 (Zi1 −i2 , Lm ⊗ H n ). Comparing dimensions and
using (9.20) we find that there exists a constant C such that
0 g 0 0 0
0m · P (1, n/m)0 # C · 0mg−(i1 −i2 ) 0 (18)

for all sufficiently large m. Here n = n(m) < m is a function of m.

Next we write P (1, t) = (t − τ )µ · R(t) where R(t) does not have roots in [0, 1]. Choose a
constant C & > 0 with |R(t)| > C & for all t ∈ [0, 1]. Combined with (18) this gives
0 & 0 0 " #0 0 0
0C · (n/m − τ )µ 0 # 0P 1, n/m 0 # C · 0m−(i1 −i2 ) 0 (19)

for all sufficiently large m.

To finish the argument we distinguish two cases. First assume that τ ∈ Q. Let f be its
denominator. Recall that we only consider integers m that are coprime with f . For all such m
and all 1 # n # m we have |n/m − τ | ! 1/f m. Using this in (19) and letting m get large
we find the desired estimate i1 # i2 + µ. Similarly, if τ is irrational then it suffices to show
that there is an infinite sequence of values for m, say m1 , m2 , . . ., and a constant C && such that

– 138 –
|nj /mj − τ | ! C && /mj for all j. (Note that the n’s are still a function of the m’s, determined by
the rule that τ lies in the interval [(n − 1)/m, n/m].) This is achieved by a theorem of Kronecker
which says that the fractional parts of the numbers m · τ , for m ∈ N, lie dense in the interval
]0, 1[; see Hardy and Wright [1], Chap. 23. "
After all these preparations we are now ready for the main result about the relation between
the index and the Hilbert polynomial of L.

(9.25) Theorem. (Kempf-Mumford-Ramanujam) Let L be a non-degenerate line bundle on

an abelian variety X. Let H be an ample line bundle on X and write Φ(t) ∈ Z[t] for the Hilbert
polynomial of L with respect to H. (So Φ(n) = χ(L ⊗ H n ) for all n.) Then all complex roots
of Φ are real, and the index i(L) equals the number of positive roots, counted with multiplicities.
Proof. Writing P (s, t) = (sl + th)g for the 2-variable polynomial as introduced before (9.22), we
have Φ(t) = P (1, t)/g!. For the rest of the proof we may therefore work with P (1, t). Notice
that this is a polynomial of degree g.
Let τ1 , . . . , τh be the real roots of P (1, t), say with multiplicities µ1 , . . . , µh , respectively.
(It will be clear from the arguments below that h > 0.) Choose m ∈ Z>0 and n1 , . . . , nh ∈ Z
such that τj lies in the interval [(nj − 1)/m, nj /m]. We can make these choices such that P (1, t)
has no roots of the form n/m, so that all bundles Lm ⊗ H n are non-degenerate.
For n >> 0, say n ! N2 , the bundle Lm ⊗ H n is ample, so that i(Lm ⊗ H n ) = 0. Similarly,
for n # N1 the bundle Lm ⊗ H n is anti-ample, in which case i(Lm ⊗ H n ) = g. (That h > 0 is
now clear from (9.22).)
Applying Proposition (9.22) and Lemma (9.24) we find that for every n ∈ Z,
either: P (1, t) has no root in the interval [(n − 1)/m, n/m] and i(Lm ⊗ H n−1 ) = i(Lm ⊗ H n ),
or: n = nj (for some j), and P (1, t) has a unique root in [(n − 1)/m, n/m], of multiplicity
µj ; in this case i(Lm ⊗ H n−1 ) # i(Lm ⊗ H n ) + µj .

index increases index stays

i(Lm ⊗H N1 )=g by at most µj constant i(Lm ⊗H N2 )=0
↓ ←− ←− ↓
··· · ··· ···
N1 nj −1 ↑ nj 1 23 4 N2
m·τj no roots

Starting at n = N2 and descending in steps of length 1 we find

&
g = i(N1 ) − i(N2 ) # µj .
j

5
On the other hand, as P (1, t) has degree g we have j µj # g. The conclusion is that we have
equality everywhere: P (1, t) has all its roots real and i(Lm ⊗ H nj −1 ) = i(Lm ⊗ H nj ) + µj for
all j. This also gives that
&
i(L) = i(Lm ) = i(Lm ⊗ H 0 ) = µj ,
j;τj >0

and the theorem is proven. "

– 139 –
(9.26) Corollary. Let f : X → Y be an isogeny. If L is a non-degenerate line bundle on Y
then i(L) = i(f ∗ L).
Proof. Choose and ample line bundle H on Y . By (9.12), the Hilbert polynomial of f ∗ L with
repsect to the ample bundle f ∗ H is just deg(f ) times the Hilbert polynomial of L with respect
to H. Now apply the theorem. "
The reason that in (9.25) we restrict ourselves to non-degenerate bundles is that only for
such bundles the index is well-defined. Without this restriction we still have a quantative result,
though.

(9.27) Theorem. Let L be a line bundle on an abelian variety X over a field k. Let H be an
ample line bundle on X and write Φ(t) ∈ Z[t] for the Hilbert polynomial of L with respect to H.
Then the multiplicity of 0 as a root of Φ equals the dimension of K(L).
Proof. Write Y := K(L)0red , which is an abelian subvariety of X. There exists an abelian
subvariety Z ⊂ X such that the homomorphism ν: Y × Z → X given by (y, z) *→ y + z is an
isogeny; see Exercise ?? or Theorem (12.2) below. Let M := (ν ∗ L)|{0}×Z . Note that M is a
non-degenerate bundle on Z. We claim that ν ∗ L differs from p∗Z M by an element in Pic0(Y ×Z)/k .
Indeed, if we let N := ν ∗ L ⊗ p∗Z M −1 then K(N ) contains both {0} × Z (because N|{0}×Z is
trivial) and Y × {0} (because N|Y ×{0} = L|Y and Y ⊂ K(L)); hence K(N ) = Y × Z, which
by Cor. (7.22) means that the class of N lies in Pic0(Y ×Z)/k . Writing l = c1 (L) and m = c1 (M )
we therefore have ν ∗ l = p∗Z m. Let g = dim(X) and s = dim(Z), and write h = c1 (H). Using
Corollary (9.12) we find
" #g " #g
deg(ν) · Φ(t) = deg(ν) · (l + t · h)g = ν ∗ (l + th) = p∗Z m + t · ν ∗ h
&s 6 7
g " ∗ #
= (pZ m)j · (ν ∗ h)g−j · tg−j ,
j=0
j

since mj = 0 if j > s = dim(Z). Moreover, ms #= 0 because M is non-degenerate; and because

ν ∗ h is an ample class then also (p∗Z m)s · (ν ∗ h)g−s #= 0. This shows that Φ(t) is exactly divisible
by tg−s . "

– 140 –
Chapter X. Tate modules, p-divisible groups, and the fundamental group.

If X is an abelian variety over the complex numbers, the associated analytic manifold can be
described as a complex torus V /Λ, with V a C-vector space and Λ ⊂ V a lattice. Topologically
this is a product of spheres, and the fundamental group can be identified with Λ ∼ = Z2g (with
g = dim(X)). Many properties of X can be expressed in terms of this lattice, and in fact we see
that Λ together with the complex structure on V = Λ ⊗Z R completely determines X.
Over an arbitrary ground field, we can no longer naturally associate a lattice of rank 2g
to a g-dimensional abelian variety (see ??), and we have to look for a substitute for Λ. The
starting point is the remark that, over C, the fundamental group is also the group of covering
transformations of the universal covering of X, and its pro-finite completion classifies the finite
coverings of X. Analytically, such finite coverings can be described as V /Λ! → V /Λ where
Λ! ⊂ Λ is a subgroup of finite index; the covering group is then Λ/Λ! . In particular, one finds
that any finite covering is dominated by a covering of the form [n]X : X → X (which corresponds
to taking Λ! = nΛ), which has covering group isomorphic to the n-torsion subgroup X[n] ⊂ X.
This leads to a description of the pro-finite completion of π1 (X, 0) as the projective limit of the
finite groups X[n]. (Cf. Cor. (10.37).)
Finite coverings of X, as well as torsion points of X, can be studied over arbitrary ground
fields. Restricting to the "-primary part, for a prime number ", we are led to consider the so-
called Tate-"-module T! X of X, which is a good "-adic analogue of the fundamental group, and
which can be defined in elementary terms. These Tate modules turn out to be very useful, and
will play an important role in the study of endomorphisms.
If the ground field has positive characteristic p then the Tate-p-module of X has a somewhat
different structure than the T! X for " %= p, and there is another object, called the p-divisible
group, that contains finer information. This p-divisble group, denoted X[p∞ ], will be introduced
in the second paragraph.
In the second half of the chapter we give a brief introduction to Grothendieck’s theory of
the algebraic fundamental group. We then compute the (algebraic) π1 of an abelian variety, and
show that it can indeed be expressed—as already suggested above—in terms of Tate modules.
Throughout this chapter we work over a field k. We let ks denote a separable closure of k and
k an algebraic closure. The letter " is reserved for a prime number different from char(k).

§ 1. Tate-"-modules.

(10.1) Let X be a g-dimensional abelian variety over a field k. Let " be a prime number different
from char(k). As we have seen in (5.9) the group scheme X["n ] has rank "2ng , and since this is
not divisible by char(k), Cor. (4.48) shows that X["n ] is étale-étale.
In (3.26) we have seen that a finite étale group scheme is fully described by its group of
ks -valued points equipped with its !natural
" action of Gal(ks /k). In the case of X["n ] this means
we have to look at the group X[" ] ks of "n -torsion points in X(ks ), equipped with its natural
n

TateBT, 15 september, 2011 (812)

– 141 –
Galois action.
Multiplication by " on X induces a homomorphism of group schemes ": X["n+1 ] → X["n ].
Under the correspondence of (3.26) it corresponds to the homomorphism of abstract groups
! " ! "
": X["n+1 ] ks → X["n ] ks , (1)
# ! "$
which is Gal(ks /k)-equivariant. For varying n these maps make the collection X["n ] ks n∈Z
!0
into a projective system of abelian groups with Gal(ks /k)-action.

(10.2) Definition. Let X be an abelian variety over a field k, and let " be a prime number dif-
ferent from char(k).#Then !we "$
define the Tate-"-module of X, notation T! X, to be the projective
n
limit of the system X[" ] ks n∈Z with respect to the transition maps (1). In other words,
!0

% ! " ! ! " ! ! " ! &

!
T! X := lim {0} ←− X["] ks ←− X["2 ] ks ←− X["3 ] ks ←− · · · .

If char(k) = p > 0 then we define

% p ! " p ! " p ! " p &
Tp,ét X := lim {0} ←− X[p] k ←− X[p2 ] k ←− X[p3 ] k ←− · · · .

In concrete terms this means that an element of T! X is a sequence x = (0, x1 , x2 , . . .) with

xn ∈ X(ks ) an "n -torsion point, and with " · xn+1 = xn for all n. The addition on T! X is
done coordinatewise, and if we have an "-adic number a = (a0 , a1 , a2 , . . .) with ai ∈ Z/"i Z and
ai+1 mod "i = ai , then a · x = (0, a1 x1 , a2 x2 , . . .).
In practice we often simply call T! X the Tate module of X, especially when the choice of "
plays no particular role.

n
! " that forn "! %=
Note " char(k) we get the same module T! X if in the definition we replace
X[" ] ks by X[" ] k ; see Prop. (5.11). In fact, we prefer to state the definition using the
separable closure ks , as we usually want to consider T! X with its natural action of Gal(ks /k);
see below. For the definition of Tp,ét X, it does make a difference that we work with torsion
points over k (and not ks ); see (5.24).
Though the definition of Tp,ét X is perfectly analogous to that of T! X, this “Tate-p-module”
is not really a good analogue of the Tate-"-modules. This is why we use a slightly different
notation for it. See further the discussion in § 2.

(10.3) It follows from (5.11) that T! X is (non-canonically) isomorphic to

% &
! ! ! !
lim {0} ←− (Z/"Z)2g ←− (Z/"2 Z)2g ←− (Z/"3 Z)2g ←− · · · = Z2g
! .

In other words, T! X is a free Z! -module of rank 2g. We also introduce

V! X := T! X ⊗Z! Q! ,

a Q! -vector space of dimension 2g.

But# T! is !not"$
just a Z! -module. We have a natural action of Gal(ks /k) on the projective
system X["n ] ks , and this gives rise to an integral "-adic representation

ρ! : Gal(ks /k) → GL(T! X) .

– 142 –
We refer to Appendix ?? for some basic notions on "-adic representations. If there is no risk of
confusion we use the same notation ρ! for the "-adic representation with values in GL(V! X).
Note that we can ! find
" back the group scheme X["n ] from T! X with its Galois action, since
T! X/"n T! X ∼
= X["n ] ks . Therefore, knowing the Tate-"-module with its action of Gal(ks /k) is
equivalent to knowing the full projective system of group schemes X["n ].

(10.4) The group Q! /Z! is the union of its subgroups "−n Z! /Z! . Phrased differently, Q! /Z! is
the inductive limit of the system {Z/"n Z}n!0 , where the transition maps are the homomorphisms
Z/"n Z $→ Z/"n+1 Z given by (1 mod "n ) )→ (" mod "n+1 ).
The definition of the Tate-"-module may be reformulated by saying that
! "
T! X = Hom Q! /Z! , X(ks ) ,

where we take homomorphisms of abstract groups. Indeed,

! " ! " ! "
Hom Q! /Z! , X(ks ) = lim n lim n
←− Hom Z/" Z, X(ks ) = ←− X[" ] ks ,
n n

where in the last term the transition maps are given by multiplication by ". Concretely, if
(0, x1 , x2 , . . .) with xn ∈ X["n ](ks ) is an element of T! X then the corresponding homomorphism
Q! /Z! → X(ks ) sends the class of "−n to xn . In this description the Gal(ks /k)-action on T! X
is induced by the Galois action on X(ks ).

(10.5) A homomorphism f : X → Y gives rise to a Z! -linear, Gal(ks /k)-equivariant

! map"
T! f : T! X → T! Y . It sends a point (0, x1 , x2 , . . .) of T! X to the point 0, f (x1 ), f (x2 ), . . .
of T! Y .
Suppose f is an isogeny with kernel N ⊂ X. Applying Hom(Q! /Z! , −) to the exact sequence
0 −→ N (ks ) −→ X(ks ) −→ Y (ks ) −→ 0 we obtain an exact sequence
T! f ! "
0 −→ T! X −−→ T! Y −→ Ext1 Q! /Z! , N (ks )
! " ! " (2)
−→ Ext1 Q! /Z! , X(ks ) −→ Ext1 Q! /Z! , Y (ks ) ,

where the Ext terms are computed in the category! Ab of abelian" groups.
Let us first try to understand the term Ext1 Q! /Z! , N (ks ) . We use that if A and B are
abelian groups, multiplication by an integer n on Ext1 (A, B) equals the map induced by [n]A
(multiplication by n on A), and also equals the map induced by [n]B .
Write N = N! × N ! with N ! a group scheme of order prime to " and N! a group scheme of
"-power order. If m is the order of N ! then multiplication by m kills N ! (ks ) but is a bijection
on Q! /Z! . Hence
! " ! " ! "
Ext1 Q! /Z! , N (ks ) = Ext1 Q! /Z! , N! (ks ) × Ext1 Q! /Z! , N ! (ks )
! "
= Ext1 Q! /Z! , N! (ks ) .

Next consider the long exact sequence

! " ! "
· · · −→ Hom Q! , N! (ks ) −→ Hom Z! , N! (ks )
! " ! " (3)
−→ Ext1 Q! /Z! , N! (ks ) −→ Ext1 Q! , N! (ks ) −→ · · · .

For a sufficiently big N! (ks ) is killed by "a , so multiplication by "a induces the zero map on
all terms in (3). On the other hand, multiplication by "a is a bijection on Q! and therefore

– 143 –
 i
! " ! "
induces
! an bijection
" on the terms Ext Q ! , N ! (k s ) . Hence the terms Hom Q ! , N ! (k s ) and
1
Ext Q! , N! (ks ) vanish, and the conclusion is that
! " ! "
Ext1 Q! /Z! , N (ks ) ∼= Hom Z! , N! (ks ) ∼= N! (ks ) . (4)
! " ! "
Write E 1 (f ): Ext1 Q! /Z! , X(ks ) → Ext1 Q! /Z! , Y (ks ) for the map induced by f . We
claim it is injective. If the ground field k is perfect, so that ks = k, then we know from
Cor. (5.10) that X(ks ) is a divisible group, !and is therefore
" an injective object in the category
1
of abelian groups. Hence in this case Ext Q! /Z! , X(ks ) = 0. In the general case, we first
choose an isogeny g: Y → X!such that g ◦ f" = [n]X for some positive integer n. Then E(g ◦ f ) is
multiplication by n on Ext1 Q! /Z! , X(ks ) . Now write n = "m · n! with " ! n! . Multiplication
by n! is a bijection on Q! /Z! ; so it suffices to show that E 1 ("m ) is injective. But if we take
f = "m then the sequence (2) becomes

!m δ ! "
0 −→ T! X −−→ T! X −→ Ext1 Q! /Z! , X["m ](ks )
! " E 1 (!m ) ! "
−→ Ext1 Q! /Z! , X(ks ) −−−−−→ Ext1 Q! /Z! , X(ks ) ,

and it follows from (4) that δ is surjective. This proves our claim.
Finally we remark that the maps in (2) are equivariant for the natural Galois actions on all
terms. To summarize, we have the following conclusion.

(10.6) Proposition. Let f : X → Y be an isogeny of abelian varieties over a field'k, with ker- (
nel N . If " is a prime number with " %= char(k) then we have an exact sequence of Z! Gal(ks /k) -
modules
T! f
0 −→ T! X −−→ T! Y −→ N! (ks ) −→ 0
where N! (ks ) is the "-Sylow subgroup of N (ks ).

(10.7) Corollary. If f : X → Y is an isogeny then for all " %= char(k) the induced map
V! f : V! X → V! Y is an isomorphism.

(10.8) The construction of the Tate module makes sense for arbitrary group varieties. Thus, if
G is a group variety over k and " %= char(k) then we can form
% ! " ! ! " ! ! " ! &
!
T! G := lim {0} ←− G["] ks ←− G["2 ] ks ←− G["3 ] ks ←− · · · .

In some cases the result is not very interesting. For instance, T! Ga = 0. But the Tate module of
the multiplicative group Gm is a fundamental object; so much so that it has a special notation:
we write
% ( )! ( )! ( )! ( )!
&
Z! (1) := T! Gm = lim {1} ←−−− µ! (ks ) ←−−− µ!2 (ks ) ←−−− µ!3 (ks ) ←−−− · · · .

(In this case we of course use multiplicative notation.) As a Z! -module, Z! (1) is free of rank 1.
The action of Gal(ks /k) is therefore given by a character
! "
χ! : Gal(ks /k) → Z! ∗ = GL Z! (1) ,

called the "-adic cyclotomic character.

– 144 –
 As discussed in Appendix ??, if T is any "-adic representation of Gal(ks /k) then we de-
fine T (n), called “T twisted by n”, to be
)
T ⊗Z! Z! (1)⊗n if n ! 0,
T ⊗Z! Z! (−1)⊗−n if n " 0,

where Z! (−1) := Z! (1)∨ and Z! (1)⊗0 is defined to be Z! with trivial Galois action. Concretely, if
ρ is the Galois action on T then T (n) is isomorphic to T as a Z! -module, but with σ ∈ Gal(ks /k)
acting via χ! (σ)n · ρ(σ).

(10.9) Proposition. We have a canonical isomorphism

! "
T! X t ∼
= (T! X)∨ 1 .

Proof. By Thm. (7.5) we have X t ["n ] ∼

= X["n ]D , and therefore
! " % ! " & % ! " &
X t ["n ] ks ∼
= Hom X["n ] ks , ks∗ = Hom X["n ] ks , µ!n (ks )

as groups with Galois action. Now take projective limits. #

§ 2. The p-divisible group.

If char(k) = p > 0 then the “Tate-p-module” Tp,ét X is in many respects not the right object
to consider. For instance, whereas T! X (for " %= char(k), as always) has rank 2g over Z! ,
independent of ", the rank of the module Tp,ét X equals the p-rank of X, and as we know this is
an integer with 0 " f (X) " g. In particular, Tp,ét X could be zero.
We have seen that the Tate-"-module captures the full system of group schemes X["n ]. That
this system can be encoded into a single Z! -module with Galois action is due to the fact that
X["n ] is étale for every n. So we should really consider the full system of group schemes X[pn ].
It turns out that it is most convenient to put these into an inductive system, and in this way
we arrive at the p-divisible group of an abelian variety.
Let us now first give the definition of a p-divisible group in a general setting.

(10.10) Definition. Let S be a base scheme. A p-divisible group over S, also called a Barsotti-
Tate group over S, is an inductive system
# $ i1 i
2 i
3
Gn ; in : Gn → Gn+1 n∈N
, in other words: G1 −→ G2 −→ G3 −→ ··· ,

where:
(i) each Gn is a commutative finite locally free S-group scheme, killed by pn , and flat when
viewed as a sheaf of Z/pn Z-modules;
(ii) each in : Gn → Gn+1 is a homomorphism of S-group schemes, inducing an isomorphism
∼
Gn −→ Gn+1 [pn ].
Homomorphisms of p-divisible groups are defined to be the homomorphisms of inductive
systems of group schemes.

– 145 –
 The flatness condition in (i) of the definition can be rephrased in more elementary terms,
as in the following lemma.

(10.11) Lemma. Let S be a scheme. Let p be a prime number. If H is an fppf sheaf of

Z/pn Z-modules on S then the following are equivalent:
(i) H is flat as a sheaf of Z/pn Z-modules;
(ii) Ker(pi ) = Im(pn−i ) for all i ∈ {0, 1, . . . , n}.
Proof. We closely follow Messing [1], Chap. I, § 1. For (i) ⇒ (ii), start with the exact sequence

pn−i pi
Z/pn Z −−−→ Z/pn Z −→ Z/pn Z .

If H is flat over Z/pn Z then − ⊗ H gives an exact sequence

pn−i pi
H −−−→ H −→ H (5)

and we see that (ii) holds.

For the proof of (ii) ⇒ (i) we use some results of Bourbaki [2]. These results are stated in
the context of modules over rings, but they carry over (with the same proofs) to the setting of
sheaves.
We use the flatness criterion, loc. cit., Chap. III, § 5, Thm. 1 together with ibid., Prop. 1.
This tells us that H is flat over Z/pn Z if and only if the following two conditions hold:
(a) H/pH is flat as a sheaf of Fp -modules;
Z/pn Z
(b) Tor1 (Z/pi Z, H) = 0 for all i ! 0.
But (a) is trivially true, as Fp is a field. To see that (b) holds, start with

pi
0 −→ Z/pn−i Z −→ Z/pn Z −→ Z/pi Z −→ 0 .

This gives a long exact sequence

Z/pn Z pi
0 −→ Tor1 (Z/pi Z, H) −→ H/pn−i H −→ H −→ H/pi H −→ 0 .

But assumption (ii), equivalent to the exactness of (5), says precisely that pi : H/pn−i H → H is
injective. #

(10.12) Let {Gn ; in } be a p-divisible group over S. If m and n are natural numbers then the
composition
m i im+1 im+n−1
im,n : Gm −−
→ Gm+1 −−−→ · · · −−−−−→ Gm+n
∼
gives an identification Gm −→ Gm+n [pm ]. Hence we may view Gm as a subgroup scheme
of Gm+n .
On the other hand, since Gm+n is killed by pm+n the map [pm ]: Gm+n → Gm+n factors
through Gm+n [pn ] = Gn . If there is no risk of confusion we simply write pm : Gm+n → Gn for
the induced homomorphism. By Lemma (10.11) this last map is an epimorphism. Hence the
sequence
im,n pm
0 −→ Gm −−−→ Gm+n −−→ Gn −→ 0 (6)
is exact.

– 146 –
(10.13) Given a p-divisible group as in the above definition, we may consider the Gn as fppf
sheaves on S and form the limit
G := lim
−→ n
G ,
n

in the category of fppf sheaves of abelian groups. We can recover Gn from G by Gn = G[pn ].
# $ # $
If Gn and Hn are two #p-divisible $ # groups
$ and we form G := lim−→ n
G and H := lim −→
Hn ,
then the homomorphisms from Gn to Hn are just the homomorphisms # $ from G to H as
fppf sheaves. In other words, by passing from the inductive system Gn to the limit G we can
identify the category of p-divisible groups over S with a full subcategory of the category of fppf
sheaves in abelian groups over S.
An fppf sheaf G is (or “comes from”) a p-divisible group if and only if it satisfies the
following conditions:
(i) G is p-divisible in the sense that [p]G : G → G is an epimorphism;
(ii) G is p-torsion, meaning that G = lim −→ n
G[pn ];
(iii) the subsheaves G[pn ] are representable by finite locally free S-group schemes.
To go back from a sheaf G satisfying these conditions to a p-divisible group as defined
in (10.10), take Gn := G[pn ], and let in : Gn → Gn+1 be the natural inclusion. It follows
from (i) that [pn ]G is an epimorphism for all n, and this implies that for all m and n we have
an exact sequence as in (6). By Lemma (10.11), we conclude that each Gn is flat as a sheaf of
Z/pn Z-modules; hence the system {Gn ; in }n is a p-divisible group. As a further simplification,
it can be shown that it suffices to require (iii) for G[p]; see Messing [1], Chap. I, § 1.
We can go one step further by remarking that, as a consequence of (ii), a p-divisible group G
has a natural structure of a sheaf in Zp -modules. More concretely, suppose we have a p-adic
number a = (a1 , a2 , . . .) with ai ∈ Z/pi Z. Then a acts on Gn as multiplication by an ; this gives
a well-defined Zp -module structure on the limit G because the diagrams

a ·
Gn −−n→ Gn
 

in +
i
+n
Gn+1 −−−−→ Gn+1
an+1 ·

are commutative. Homomorphisms of p-divisible groups are automatically Zp -linear. In partic-

ular, Hom(G, H) has a natural structure of a Zp -module.

(10.14) Remark. The name “p-divisible group” refers to condition (i) in (10.13). But we see
that the requirement for an fppf sheaf G to be a p-divisble group in the sense of Def. (10.10) is
stronger than only this condition. Thus, strictly speaking the terminology “p-divisible group” is
not correct. This is one of the reasons that some prefer the terminology “Barsotti-Tate group”,
after two of the pioneers in this area.

(10.15) If G = lim G is a p-divisible group over a connected base scheme S then, by definition,
−→ n
the group scheme G1 is locally free and killed by p. It follows that the rank of G1 equals ph for
some integer h. (Use Exercise (4.4).) This integer h = h(G) is called the height of G. It readily
follows from (6) and Lemma (4.46) that Gn , which is again locally free, has rank pnh .
Over an arbitrary basis S, we define the! height
" of a p-divisible group G as the locally
constant function |S| → Z!0 given by s )→ h G(s) .

– 147 –
(10.16) Definition. Let X be an abelian variety over a field k. Let p be a prime number.
Then we define the p-divisible group of X, notation X[p∞ ], to be the inductive system
# $
X[pn ] n!0

with respect to the natural inclusion homomorphisms X[pn ] $→ X[pn+1 ].

Note that X[p∞ ] has height 2g, where g = dim(X).

(10.17) A homomorphism f : X → Y of abelian varieties over k induces a homomorphism

f [p∞ ]: X[p∞ ] → Y [p∞ ] of p-divisible groups.
If we take f = [n]X for some integer n then the induced endomorphism of X[p∞ ] is mul-
tiplication by n, which for n %= 0 is surjective (as a homomorphism of fppf sheaves). Using
Prop. (5.12) it follows that if f is an isogeny then f [p∞ ] is an epimorphism of fppf sheaves.
Hence if f is an isogeny with kernel N we find an exact sequence of fppf sheaves

f [p∞ ]
0 −→ Np −→ X[p∞ ] −−−−→ Y [p∞ ] −→ 0 ,

where we write N = Np × N p with Np of p-power order and N p a group scheme of order prime
to p.

(10.18) Let us return to the general context of a p-divisible group G over a base scheme S.
Applying Cartier duality to (6) gives an exact sequence

0 −→ GD D D
n −→ Gm+n −→ Gm −→ 0 .

In particular,
# $ taking m = 1 this gives homomorphisms ιn : GD D
n → Gn+1 . The inductive system
D
Gn ; ιn is again a p-divisible group; it is called the Serre dual of G.
A homomorphism f : G → H induces a dual homomorphism f D : H D → GD ; in this way
G )→ GD gives a contravariant functor from the category of p-divisible groups over S to itself.
D ∼ D D ∼
The collection of isomorphisms (GD n ) −→ Gn give a canonical isomorphism (G ) −→ G.
It is immediate from the defnitions that the Serre-dual of G has the same height as G.

(10.19) Proposition. If X/k is an abelian variety then we have a canonical isomorphism

X t [p∞ ] ∼
= X[p∞ ]D .

Proof. Immediate from Thm. (7.5) and the definition of the Serre dual. #

(10.20) Like the construction of a Tate module, the definition of a p-divisible group also makes
sense for certain other commutative group varieties. Beyond abelian varieties, the main example
of interest is the p-divisible group Gm [p∞ ] associated to Gm . By definition, Gm [p∞ ] is the
inductive system of group schemes µpn with respect to the natural inclusions µpn $→ µpn+1 . If
we work over a field k and view Gm [p∞ ] as an fppf sheaf on Spec(k) then we have
! " # , n $
Gm [p∞ ] R = x ∈ R∗ , xp = 1 for some n ! 0 ,

for any k-algebra R. The height of Gm [p∞ ] is 1.

– 148 –
 The Serre-dual of Gm [p∞ ] is the p-divisible group Qp /Zp , i.e., the inductive limit of constant
∼
group schemes Z/pn Z with respect to the inclusion maps Z/pn Z −→ pZ/pn+1 Z ⊂ Z/pn+1 Z.

(10.21) As we have seen in Prop. (4.45), a finite commutative group scheme over a field k is, in
a canonical way, an extension of an étale group scheme by a local group scheme. An immediate
consequence of this is that any p-divisible group G = lim
−→ n
G over k is an extension

1 −→ Gloc −→ G −→ Gét −→ 1 (7)

of the “ind-étale” p-divisible group Gét = lim G

−→ n,ét
by the “ind-local” p-divisible group Gloc =
lim G
−→ n,loc
. To simplify terminology, the prefix “ind-” is often omitted; e.g., G is called an étale
∼
p-divisible group if G −→ Gét .
If k is perfect then the sequence (7) splits. See Exercise (10.1).
Combining the above with the Serre-duality functor G )→ GD of (10.18), we can further
decompose Gloc as an extension of a local-local p-divisble group by a local-étale one. Here we
extend the terminology introduced in (4.42) in an obvious way to p-divisible groups. Similarly,
Gét is an extension of an étale-local p-divisible group by an étale-étale one.

(10.22) If G is a p-divisible
! group over
" k, viewed as an fppf sheaf, then we define its Tate-p-
module by Tp G := Hom Qp /Zp , G(k) . Concretely, we take the limit of the projective system
π1,1 π1,2 π1,3
G1 (ks ) ←−−− G2 (ks ) ←−−− G3 (ks ) ←−−− · · · .

As usual, Tp G is a Zp -module that comes equipped with a continuous action of Gal(ks /k).
It is clear from the definitions that Tp G only sees the étale part of G, i.e., the canonical
map Tp G → Tp Gét is an isomorphism. It follows that Tp G is a free Zp -module of rank h(Gét ).
If p %= char(k) then clearly the Tate module of X[p∞ ] is the same as the Tate-p-module
of X as defined in (10.2). The Tate module of Gm [p∞ ] is Zp (1).

(10.23) Thus far we have not made any assumptions on the prime p in relation to the char-
acteristic of the ground field k. But if p %= char(k) then it follows from Prop. (4.47) that any
p-divisible group G over k is étale-étale. More precisely, Gks is non-canonically isomorphic to
(Qp /Zp )h(G) . In this case it is an easy exercise to show that the functor G )→ Tp G gives an
equivalence from the category of p-divisible groups over k to the category of free Zp -modules
of finite rank equipped with a continuous action of Gal(ks /k). This functor is compatible with
duality, in the sense that Tp (GD ) is canonically isomorphic to (Tp G)∨ (1).
In sum, for p %= char(k), a p-divisible group carries the same information as the correspond-
ing Tate module, and we typically work with the latter. (To stress that p %= char(k) we shall use
the letter " rather than p.) By contrast, if char(k) = p > 0 then a p-divisible group in general
contains finer information than the associated Tate module.

(10.24) To conclude this general section on p-divisible groups, let us discuss the relation with
formal groups. (??TO BE COMPLETED??)

§ 3. The algebraic fundamental group—generalities.

– 149 –
In Topology one defines the fundamental group π1 (X, x) of a space X with base point x ∈ X
as the group of homotopy classes (rel. {0, 1}) of paths γ: [0, 1] → X with γ(0) = γ(1) = x.
Now suppose we want to define the fundamental group of an algebraic variety over an arbitrary
field. Working with the Zariski topology does not give reasonable answers—for instance, any
two algebraic curves over the same field are homeomorphic as topological spaces! Further, the
above topological definition via paths does not have an obvious “algebraic” analogue that works
well. (In fact, an algebraic analogue of homotopy theory was developed only in the 1990’s; see
Morel and Voevodsky [1].)
Assuming that X is locally connected and locally simply connected, an alternative descrip-
tion of π1 (X, x) is that it is the automorphism group of the universal covering X̃ → X. See
for instance Massey [1] or Rotman [1]. In this description the fundamental group becomes the
group which classifies topological coverings of X. This is similar to Galois theory of fields, and it
was one of Grothendieck’s fundamental insights that it is possible to develop an abstract Galois
theory of which both are special instances. Using finite étale morphisms as coverings, this theory
also applies to algebraic schemes and gives rise to a notion of an algebraic fundamental group.
We shall now recall the definition of the algebraic fundamental group π1 (X, x̄), and some
basic properties. For further introduction we refer to SGA1. On a more advanced level, but
very readable, is Deligne [4], § 10. We shall write π1 for the algebraic fundamental group and
use the notation π1top for the fundamental group in the classical setting of topological spaces.

(10.25) Definition. Let X be a scheme. By an étale covering of X we mean a finite étale

morphism Y → X. (Do not confuse this with the notion of a covering for the étale topology.) We
write FEt/X ⊂ Sch/X for the full subcategory of such étale coverings. Note that the morphisms
in FEt/X are automatically again étale coverings. We say that an étale covering f : Y → X
dominates the étale covering g: Z → X if there exists a morphism h: Y → Z with f = g ◦ h.
Fix an algebraically closed field Ω and a geometric point x̄: Spec(Ω) → X. We define a
functor
Fx̄ : FEt/X → Sets
# , $
by Fx̄ (f : Y → X) = y ∈ Y (Ω) , f (y) = x . In other words, Fx̄ associates to an étale covering
of X the set of geometric points lying over x̄.

(10.26) Definition. (Grothendieck) Assume X to be locally noetherian and connected. Then

the algebraic fundamental group π1 (X, x̄) is defined to be the automorphism group of the func-
tor Fx̄ .

(10.27) Example. Suppose X = Spec(k) is the spectrum of a field. The geometric point x̄
corresponds to an embedding σ: k $→ Ω. An étale covering of X is a finite disjoint union
of schemes Spec(L), where k ⊂ L is a finite separable field extension. For such a scheme
Y = Spec(L) we have
# $
Fx̄ (Y ) = embeddings τ : L $→ Ω with τ|k = σ .

Write ks for the separable closure of k inside Ω. Clearly, every element of Gal(ks /k) gives
an automorphism of the functor Fx̄ . Conversely,
! "if α ∈ Aut(Fx̄ ) and ξ ∈ ks then the inclusion
k(ξ) ⊂! Ω gives an" Ω-valued point of Spec k(ξ) lying above x̄, in other words, an element
i ∈ Fx̄ Spec(k(ξ)) . Then α(i) is another embedding of k(ξ) into Ω that extends σ. Sending ξ

– 150 –
to its image under α(i) defines an element of Gal(ks /k). These two constructions are inverse to
each other, so we find a canonical isomorphism of pro-finite groups
! "
π1 Spec(k), x̄ ∼= Gal(ks /k) .
! "
Notice that the elements of! π1 Spec(k)," x̄ do not directly appear as automorphisms of
the field ks . Rather, if α ∈ π1 Spec(k), x̄ corresponds to β ∈ Gal(ks /k) then α describes
the effect that β has on all embeddings L $→ ks (= the geometric points lying over x̄ in the
covering Spec(L) → X). So, to phrase it in a more topological way, the point here is that an
automorphism of the “universal covering” of X is completely determined by its effect on the
points in the fibre over the base point x̄.

(10.28) Theorem. (Grothendieck) Assume X to be locally noetherian and connected. Then

π1 = π1 (X, x̄) is a pro-finite group, and Fx̄ induces an equivalence of categories
- .
eq finite
FEt/X −−→ ,
π1 -sets
where the right hand side denotes the category of finite sets with a continuous action of π1 (X, x̄).
For the proof of this theorem we refer to SGA1, in particular Exp. V. Note that in case
X = Spec(k) we have already seen this result in (3.25).
From now on, whenever we consider an algebraic fundamental group, it is assumed that the
scheme in question is locally noetherian and connected.
We shall briefly review some basic properties of the fundamental group. Proofs may be
found in SGA1. Note that some of the results discussed below are ingredients of the proof of
Thm. (10.28), rather than being consequences of it.

(10.29) Dependence on the choice of a base point. Suppose we have two geometric points
x̄1 : Spec(Ω1 ) → X and x̄2 : Spec(Ω2 ) → X; here the (algebraically closed) fields Ω1 and Ω2 may
be different, and may even have different characteristics. The theorem implies that there is an
equivalence of categories
- . - .
finite eq finite
−−→ . (8)
π1 (X, x̄1 )-sets π1 (X, x̄2 )-sets
Notice that this equivalence is not canonical, as it depends on the choice of a quasi-inverse of
the equivalence Fx̄2 . Now it is not difficult to show that the equivalence in (8) is induced by
∼
an isomorphism of topological groups π1 (X, x̄1 ) −→ π1 (X, x̄2 ). Hence up to isomorphism the
fundamental group of the (connected!) scheme X does not depend on the chosen base point.
As in topology, a more elegant way to express that the fundamental group does not depend
on the chosen base point is to work with the fundamental groupoid. See SGA1, Exp. V, sect. 5
or Deligne [4], § 10.

(10.30) Functoriality. Let f : Y → X be a morphism between connected, locally noetherian

schemes. Let ȳ be a geometric point of Y , and write x̄ = f (ȳ). Associating to an étale covering
X ! → X its pull-back Y ! := (X ! ×X Y ) → Y gives a functor f ∗ : FEtX → FEtY , and Fx̄ = Fȳ ◦ f ∗ .
In particular, every automorphism of the functor Fȳ induces an automorphism of Fx̄ , and this
gives a canonical homomorphism

f∗ : π1 (Y, ȳ) → π1 (X, x̄) .

– 151 –
If g: Z → Y is a second morphism then (f ◦ g)∗ = f∗ ◦ g∗ .
If f : Y → X is an étale covering (still with X and Y connected and locally noetherian),
one shows that f∗ gives an isomorphism
∼
π1 (Y, ȳ) −→ Stab(ȳ) ⊂ π1 (X, x̄) ,

where Stab(ȳ) is the stabilizer of the point ȳ ∈ f −1 (x̄) under the natural action of π1 (X, x̄)
on f −1 (x̄). Indeed, if g: Z → Y is an étale covering of Y , then f ◦ g is an étale covering of X
and g −1 (ȳ) ⊂ (f ◦ g)−1 (x̄). If σ ∈ Stab(ȳ) ⊂ π1 (X, x̄) then its natural action on (f ◦ g)−1 (x̄)
preserves the subset g −1 (ȳ); hence σ induces an automorphism of the functor Fȳ . This gives a
homomorphism Stab(ȳ) → π1 (Y, ȳ) inverse to f∗ .
Conversely, if H ⊂ π := π1 (X, x̄) is an open subgroup (equivalently, a subgroup of finite
index) then π/H is a finite set with a natural action of π by left multiplication, so by Thm. (10.28)
−1 ∼
there exists an étale covering fH : YH → X such that we have an isomorphism γ: fH (x̄) −→ π/H
−1
as π-sets. Since the π-action on π/H is transitive, YH is connected. If we let ȳ ∈ fH (x̄)
be the geometric point with γ(ȳ) = (1 mod H) then Stab(ȳ) = H as subgroups of π, and
the pair (Y, ȳ) is uniquely determined up to isomorphism over X. In this way we obtain a
bijective correspondence between pairs (Y, ȳ) up to X-isomorphism (with connected Y ) and
open subgroups of π1 (X, x̄). As a variant, we may forget the choice of a geometric point ȳ
above x̄; then we get a bijective correspondence between connected étale coverings Y → X up
to isomorphism over X and conjugacy classes of open subgroups of π1 (X, x̄).

(10.31) Geometric and arithmetic fundamental group. Let X be a geometrically connected

scheme of finite type over a field k. Let ks be a separable closure of k and write X̄ := X ×k ks .
Choose a geometric point x̄ of X̄, and write x̄! for its image in X. Then there is an exact
sequence
p∗ s∗
1 −→ π1 (X̄, x̄) −−→ π1 (X, x̄! ) −−→ Gal(ks /k) −→ 1 , (9)
where the homomorphisms are induced by the projection p: X̄ → X and the structural morphism
s: X → Spec(k), and where we use the isomorphism of (10.27). If x̄: Spec(Ω) → X factors
through a k-rational point x: Spec(k) → X then x∗ : Gal(ks /k) → π1 (X, x̄) is a section of s∗ .
The group π1 (X̄, x̄) is referred to as the geometric fundamental group of X. The “full” fun-
damental group π1 (X, x̄) is occasionally called the arithmetic fundamental group. If char(k) = 0
or if X is proper over k then π1 (X̄, x̄) does not change under extension of scalars to a bigger
separably closed field. More precisely, if L is a separably closed field containing ks such that x̄
lifts to a geometric point x̃ of X ×k L, then the natural map γ: π1alg (X ×k L, x̃) → π1alg (X ×k ks , x̄)
is an isomorphism. Note however that for char(k) > 0 and X not proper, γ need not be an
isomorphism; see SGA1, Exp. X, Sect. 1.
Writing Out(π) := Aut(π)/Inn(π) for the group of outer automorphisms of a group π, the
exact sequence (9) gives rise to a homomorphism
! "
Gal(ks /k) → Out π1 (X̄, x̄) .

In the special case that x̄ factors through x: Spec(k) → X, this naturally lifts (via the section x∗ )
to a homomorphism ! "
Gal(ks /k) → Aut π1 (X̄, x̄) .

(10.32) Comparison with the topological fundamental group. Let X be a variety over C.
Choose a base point x ∈ X(C). Let us write π1top (X, x) for the usual fundamental group

– 152 –
of X(C) with its analytic topology. If Y → X is an étale covering then the induced map on
points Y (C) → X(C) is a finite topological covering (taking the analytic topology on both
sides). Since π1top (X, x) naturally acts on the fibre of Y (C) over x, we obtain a homomorphism
π1top (X, x) → π1 (X, x). It can be shown that this map induces an isomorphism
' (∧ ∼
π1top (X, x) −→ π1alg (X, x) ,

where the left hand side denotes the pro-finite completion of π1top , that is, the projective limit of
all its finite quotients. The geometric content of this statement is that every finite topological
covering of X can be realised as an algebraic variety which is finite étale over X, and this
algebraic structure is unique up to isomorphism over X.
Note that π1top (X, x) may not be residually finite, i.e., it may happen that the natural
homomorphism π1top → [π1top ]∧ is not injective. (For examples, see Toledo [1].) Geometrically
this means that the natural map from the universal covering X̃ of X (in the context of topological
spaces) to the “algebraic universal covering” X̃ alg , obtained as the projective limit of all finite
étale coverings of X, is not injective.

(10.33) Galois coverings. As before, let X be a connected, locally noetherian scheme. Fix a
geometric base point x̄ ∈ X(Ω). If we claim that the theory of the fundamental group can be
viewed as an abstract Galois theory, one may expect that certain étale coverings Y → X play
the role of Galois extensions.
Consider an étale covering f : Y → X with Y connected. Choose a base point ȳ ∈ Y (Ω)
above x̄. For simplicity of notation, write π := π1 (X, x̄), and let H ⊂ π be the stabilizer of ȳ. As
∼
discussed in (10.30), we have an isomorphism f∗ : π1 (Y, ȳ) −→ H, and we get an identification
∼
π/H −→ Fx̄ (Y ) of finite sets with π-action. Write N := Nπ (H) ⊂ G for the normaliser of H.
Let G := Aut(Y /X) be the group of automorphisms of Y over X. Note that Y is affine
over X (as Y → X is finite), so any G-equivalence class in |Y | is contained in an affine subset,
and there exists a quotient of Y by G.
By Theorem (10.28), G maps isomorphically to the automorphism group of Fx̄ (Y ) as a π-
set. Using the above description we readily find that the latter group is isomorphic to (N/H)opp ,
the opposite group of N/H. Indeed, if a ∈ N/H then ϕa : π/H → π/H given by gH )→ gaH is
a well-defined automorphism of π-sets, any automorphism is of this form, and ϕb ◦ ϕa = ϕab .
We conclude that G is finite and that its natural action on Fx̄ (Y ) is faithful. As this holds
for any choice of the base point x̄, it follows that G acts freely on Y . Hence the morphism
Y → X factors as a composition of two étale coverings Y → (G\Y ) → X. From the given
description of G we then see that the following conditions are equivalent:
(i) the group G acts transitively on Fx̄ (Y );
(ii) the group G acts simply transitively on Fx̄ (Y );
(iii) the natural map f¯: G\Y → X is an isomorphism, i.e., X is the quotient of Y under G;
(iv) the subgroup H = f∗ π1 (Y, ȳ) ⊂ π1 (X, x̄) is normal.
If these conditions are satisfied we say that f : Y → X is a Galois covering with group G, and
we have an exact sequence of groups

1 −→ π1 (Y, ȳ) −→ π1 (X, x̄) −→ Aut(Y /X)opp −→ 1 .

(Caution: we here only consider étale coverings. The terminology “Galois covering” is also used
in the context of ramified coverings.)

– 153 –
 Using condition (ii), it readily follows from Theorem (10.28) that every étale covering Z →
X with connected Z is dominated by a Galois covering.
Suppose we have étale coverings g: Z → Y and f : Y → X, where all three schemes are
connected and locally noetherian. Suppose h := f ◦ g: Z → X is a Galois covering. Then g is
a Galois covering, too. Further, f is Galois if and only if Aut(Z/Y ) ⊂ Aut(Z/X) is a normal
subgroup, and if this holds then we have a short exact sequence

1 −→ Aut(Z/Y ) −→ Aut(Z/X) −→ Aut(Y /X) −→ 1 .

§ 4. The fundamental group of an abelian variety.

We now specialize to the case of an abelian variety. The key result of this paragraph is a theorem
of Lang and Serre which says that, for an abelian variety X, the finite étale coverings f : Y → X
with a rational point eY ∈ f −1 (eX ) are precisely the separable isogenies with target X.

(10.34) Proposition. Let X be a complete variety over a field k. Suppose given a point
e ∈ X(k) and a k-morphism m: X × X → X such that m(x, e) = x = m(e, x) for all x ∈ X.
Then X is an abelian variety with group law m and origin e.
Proof. Let g := dim(X),
! " write x · y for m(x, y). Consider the morphism τ : X × X → X × X
and
given by τ (x, y) = x · y, y . (If the proposition is true then τ is the universal right translation.)
We have τ −1 (e, e) = {(e, e)}, so the image of τ has dimension 2g. (We use a standard result
on the dimension of the fibres of a morphism; see HAG, Chap. II, Exercise 3.22.) As X × X is
complete and irreducible, it follows that τ is surjective.
We reduce the problem to the case that k is algebraically closed. Namely, suppose m induces
a group structure on X(k), with origin e. Then for every x ∈ X(k) the translation τx : y )→ x · y
is an automorphism of Xk as a variety, and by the argument of Prop. (1.5) it follows that X
is non-singular. It also follows that τ induces a bijection on k-valued points. Hence τ gives a
purely inseparable extension on function fields. On the other hand, by looking at the restrictions
of τ to {e} × X and X × {e} we see that the tangent map of τ at (e, e) is an !isomorphism. " The
conclusion is that τ is an isomorphism. Now define i: X → X by i(y) = p1 τ −1 (e, y) . Using
that X is geometrically reduced and that we know the group axioms to hold on X(k), it follows
that m, i and e define the structure of an abelian variety on X. Hence to complete the proof
of the proposition, we may assume that k = k and it suffices to prove that m gives a group
structure on X(k). ,
# $
Consider
! the closed
" subscheme Γ ⊂ X × X given by Γ := (x, y) , x · y = e . Then
Γ = τ −1 {e} × X , so the surjectivity of τ implies that the second projection p2 : Γ → X
is surjective. Let Γ1 ⊂ Γ be an irreducible component with p2 (Γ1#) = X. $ Notice that Γ1 is
−1 −1
complete, and that dim(Γ1 ) ! g. Further note that p1 (e) ∩ Γ = (e, e) = p2 (e) ∩ Γ; this
implies that (e, e) ∈ Γ1 . Again by comparing dimensions it follows that p1 : Γ1 → X is surjective,
too.
! " ! " ! "
Define f : Γ1 ×X×X → X by f (x, y), z, w = x· (y·z)·w . We have f Γ1 ×{e}×{e} = {e}.
Applying the rigidity lemma we find
! "
x · (y · z) · w = z · w for all (x, y) ∈ Γ1 and z, w ∈ X. (10)

– 154 –
As a particular case, taking w = e, we have

x · (y · z) = z for all (x, y) ∈ Γ1 and z ∈ X. (11)

Now fix y ∈ X(k). Choose any x ∈ X(k) with (x, y) ∈ Γ1 , and any z ∈ X(k) with (y, z) ∈ Γ1 .
(Such x and z exist, as we have shown the two projections pi : Γ1 → X to be surjective.) Then
(11) gives x = x · (y · z) = z · e = z. The conclusion is that y has a unique left and right inverse
in X(k). Finally, multiplying (10) from the left by y = x−1 , and using (11) gives
! ! ""
y · (z · w) = y · x · (y · z) · w = (y · z) · w ,

which shows that the group law on X(k) is associative. #

(10.35) Lemma. Let Z be a k-variety, let Y be an integral k-scheme of finite type, and let
f : Y → Z be a smooth proper morphism of k-schemes. If there exists a section s: Z → Y of f
then all fibres of f are irreducible.
Proof. As the fibres of f are non-singular, it suffices to show that they are connected. Write
Z ! := Spec(f∗ OY ), and consider the Stein factorization

f% g
f = g ◦ f ! : Y −−→ Z ! −→ Z .

By Zariski’s connectedness theorem (EGA III, Thm. 4.3.1) the morphism f ! has connected fibres.
The composition f ! ◦ s is a proper section of g, hence it induces an isomorphism of Z with a closed
subscheme of Z ! . As g is finite and Z ! is integral, it follows that g is an isomorphism. #

(10.36) Theorem. (Lang-Serre) Let X be an abelian variety over a field k. Let Y be a k-

variety and eY ∈ Y (k). If f : Y → X is an étale covering with f (eY ) = eX then Y has the
structure of an abelian variety such that f is a separable isogeny.
Proof. With Proposition (10.34) at our disposal, the main point of the proof is to construct
the group law mY : Y × Y → Y . Let ΓX ⊂ X × X × X be the graph of the multiplication
on X, and write Γ!Y ⊂ Y × Y × Y for the pull-back of ΓX via f × f × f . Let ΓY ⊂ Γ!Y be
the connected component containing the point (eY , eY , eY ), and if I ⊆ {1, 2, 3} write qI for
the restriction of the projection pI : Y 3 → Y I to ΓY . We want to show that the projection
q12 : ΓY → Y × Y is an isomorphism—if this is true then we can define the desired group law
−1
by taking mY := q3 ◦ q12 : Y × Y → Y . Note that q12 has a section s1 over {eY } × Y and a
section s2 over Y × {eY }, given on points by s1 (eY , y) = (eY , y, y) and s2 (y, eY ) = (y, eY , y).
This readily implies that the proposed group law mY satisfies the conditions of (10.34).
By construction we have a commutative diagram

ΓY −−−→ ΓX
 

q12 +
p12
+
f ×f
Y ×Y −−−→ X × X ,

in which both the upper arrow ΓY → ΓX and the morphism f × f are étale coverings, and the
right hand arrow p12 : ΓX → X × X is an isomorphism. Hence q12 : ΓY → Y × Y is an étale
covering, too.

– 155 –
 The projection q2 : ΓY → Y is a smooth proper morphism, being the composition of q12 and
p2 : Y × Y → Y . As s1 gives a section of q2 we conclude! from the"above lemma that all fibres
of q2 are irreducible. In particular, Z := q2−1 (eY ) = q12
−1
Y × {eY } is irreducible. Further, q12
restricts to an étale covering r: Z → Y = Y × {eY } of the same degree. But s2 gives a section
of r. Hence r is an isomorphism. It follows that the étale covering q12 has degree 1 and is
therefore an isomorphism. #

(10.37) Corollary. Let X be an abelian variety over a field k. Let Ω be an algebraically closed
field containing k, and regard 0 = eX as an Ω-valued point of X. Write ks for the separable
closure of k inside Ω. Then there are canonical isomorphisms
/0
! T! X if char(k) = 0,
π1 (Xks , 0) ∼
alg
= lim ∼
←− X[n](ks ) = 0
n Tp,ét X × !-=p T! X if char(k) = p > 0,

where the projective limit runs over all maps X[nm](ks ) → X[n](ks ) given by P )→ m · P , and
where " runs over the prime numbers. In particular, π1alg (Xks , 0) is abelian. Further there is a
canonical isomorphism
π1alg (X, 0) ∼ alg
= π1 (Xks , 0) " Gal(ks /k) ,
where Gal(ks /k) acts on π1alg (Xks , 0) through its natural action on the groups X[n](ks ).
Proof. For the proof of the first assertion we may assume that k = ks . Write π := π1alg (Xks , 0).
We have π = lim ←− (π/H) where H runs over the open subgroups of π. By (10.30), each H
corresponds to an étale covering fH : YH → X together with the choice of a point eH ∈ YH (Ω)
above 0, the pair (YH , eH ) being unique up to isomorphism over X. By the Lang-Serre theorem,
we have the structure of an abelian variety on YH with origin eH such that fH is a separable
isogeny. Further, it is clear that! a" separable isogeny f : Y → X is a Galois covering (in the sense
of (10.33)) with group Ker(f ) k . (Recall that we assume
! "opp ! " k = ks .) By what was explained
∼
in (10.33) we find that π/H = Ker(fH ) k = Ker(fH ) k , for any open subgroup H ⊂ π.
Let I be the set of isomorphism classes of separable isogenies f : Y → X, where we call
∼
f : Y → X and f ! : Y ! → X isomorphic if there is an isomorphism of abelian varieties α: Y −→ Y !
with f ! ◦ α = f . We partially order I by dominance; so f ! f ! if there is a homomorphism
of abelian varieties α: Y → Y ! with f ! ◦ α = f . If f ! f ! then we get a homomorphism
!
Ker(f ) →→ Ker(f
# ), independent
! "$ of the choice of α. In this way we have a projective system of
finite groups Ker(f ) k f ∈I , and the conclusion of the above discussion is that
∼ ! "
π −→ lim
←− Ker(f ) k (12)
f ∈I

as pro-finite groups.
f g
If n is a positive integer then [n] = [n]X factors as X −→ X/X[n]loc −→ X where f is
purely inseparable and g is separable. Of course, if char(k) = 0 or char(k) = p > 0 and p ! n
then f is the identity and g != "[n]. For the purpose of this discussion, write g = [n]sep . The
Galois group of [n]sep is X[n] k . Let I ! ⊂ I be the subset of all isogenies [n]sep for n ∈ Z!1 .
Then I ! is cofinal in I ; indeed, if f : Y → X is any separable isogeny, say of degree d, then by
Prop. (5.12) there is an isogeny g: X → Y with [d]X = f ◦ g, and then it follows from Cor. (5.8)
that [d]sep dominates f . Hence we may restrict the limit in (12) to the terms f ∈ I ! ; this gives
the desired isomorphism
∼ ! " ! "
π −→ lim lim
←− Ker(f ) k = ←− X[n] k ,
f ∈I % n

– 156 –
where n runs over the set Z!1 , partially ordered by divisibility.
The last assertion of the theorem (now again over an arbitrary ground field) follows by using
what was explained in (10.31), noting that 0 ∈ X(Ω) factors through a k-rational point. #

(10.38) As an application of this theorem, let us now discuss how the "-adic cohomology of an
abelian variety can be described in terms of its Tate-"-module.
First let X be any complete variety over a field k, say with dim(X) = g. Let ks be a
separable closure of k, and let " be a prime number different from char(k). The "-adic cohomology
H • (Xks , Z! ) = ⊕2g i
i=0 H (Xks , Z! ) is a graded-commutative Z! -algebra of finite type that comes
equipped with a continuous action of Gal(ks /k). If x̄ ∈ X(ks ) then the first "-adic cohomology
and the fundamental group of Xks are related by
! "
H 1 (Xks , Z! ) ∼
= Homcont π1 (Xks , x̄), Z! , (13)

where the right hand side is the !group of continuous" homomorphisms π1 (Xks , x̄) → Z! . The
homomorphism
! Gal(k
" s /k) → Out π 1 (X ks , x̄) of (10.31) induces a homomorphism
! Gal(ks /k)" →
ab
Aut π1 (X! ks , x̄) , and this
" gives a continuous Galois action on Homcont π1 (Xks , x̄), Z! =
Homcont π1 (Xks , x̄)ab , Z! . The isomorphism (13) is equivariant for the Galois actions on the
two sides.
Now we specialize this to the case where X is an abelian variety. As we shall prove later,
H (Xks , Z! ) is then the exterior algebra on H 1 (Xks , Z! ); see Cor. (13.32). Admitting this, we
•

find the following result.

(10.39) Corollary. Let X be an abelian variety over a field k, let k ⊂ ks be a separable

algebraic closure, and let " be a prime number with " %= char(k). Then we have

H 1 (Xks , Z! ) ∼
= (T! X)∨ := Hom(T! X, Z! )

as Z! -modules with continuous action of Gal(ks /k). Further we have an isomorphism of graded-
commutative Z! -algebras with continuous Gal(ks /k)-action
' (
H (Xks , Z! ) ∼
= ∧ (T! X)∨ .
• •

Exercises.
(10.1) Let G be a p-divisible group over a perfect field k. Show that for every n the square

Gn,red −−→ Gn+1,red

 
 
+ +
in
Gn −−→ Gn+1

is Cartesian. Conclude that the exact sequence (7) splits.

(10.2) Let K be a field, K ⊂ Ks a separable algebraic closure. Let K ⊂ L be a finite extension
inside Ks .

– 157 –
 (i) Let H be a finite étale group scheme over L, and consider G := ResL/K (H), the K-group
scheme obtained by Weil restriction of scalars from L to K. By definition of the Weil
restriction, G represents the functor Schopp
/K → Gr given by T )→ H(TL ). Show that G is
again a finite étale group scheme.
(ii) Assume (for simplicity) that H is commutative. Write ΓL := Gal(Ks /L) ⊂ ΓK :=
Gal(Ks /K). Show that G(Ks ) ∼ = IndΓΓK
L
H(Ks ) as representations of Gal(Ks /K).
(iii) Let X be an abelian variety over L, and write Y := ResL/K (X), which is an abelian variety
over K of dimension dim(X) · [L': K]. If " is (a prime number different from char(K), show
that T! (Y ) ∼
= IndΓΓKL
T! (X) as Z! Gal(Ks /K) -modules.

– 158 –
Chapter XI. Polarizations and Weil pairings.

In the study of higher dimensional varieties and their moduli, one often considers polarized
varieties. Here a polarization is usually defined as the class of an ample line bundle modulo a
suitable equivalence relation, such as algebraic or homological equivalence. If X is an abelian
variety then, as we have seen in (7.24), the class of an ample bundle L modulo algebraic equiv-
alence carries the same information as the associated homomorphism λ = ϕL : X → X t . And
it is in fact this homomorphism that we shall put in the foreground. One reason for this is
that λ usually has somewhat better arithmetic properties; for instance, it may be defined over a
smaller field than any line bundle representing it. The positivity of an ample bundle shall later
be translated into the positivity of the Rosati involution associated to λ; this is an important
result that shall be given in the next chapter.
The first Chern class of L only depends on L modulo algebraic equivalence, and we therefore
expect that it can be expressed directly in terms of the associated homomorphism λ = ϕL . This
is indeed the case. As we have seen before (cf. ??), the #-adic cohomology of X can be described
in more elementary terms via the Tate-#-module. The class c1 (L) then takes the form of an
alternating pairing E"λ : T" X × T" X → Z" (1), usually referred to as the Riemann form of L (or
of λ). It is obtained, by a limit procedure, from pairings eλn : X[n] × X[n] → µn , called the Weil
pairing.

§ 1. Polarizations.

(11.1) Proposition. Let X be an abelian variety. Let λ: X → X t be a homomorphism, and

consider the line bundle M := (id, λ)∗ PX on X. Then ϕM = λ + λt . In particular, if λ is
symmetric then ϕM = 2λ.
Proof. Immediate from Proposition (7.6) together with Exercise (7.5). !

(11.2) Proposition. Let X be an abelian variety over a field k. Let λ: X → X t be a homo-

morphism. Then the following properties are equivalent:
(a) λ is symmetric;
(b) there exists a field extension k ⊂ K and a line bundle L on XK such that λK = ϕL ;
(c) there exists a finite separable field extension k ⊂ K and a line bundle L on XK such that
λK = ϕL .
Proof. Assume (a) holds. Let M := (id, λ)∗ PX and N := M 2 . By the previous proposition
we know that ϕM = 2λ, so ϕN = 4λ. In particular, X[4] ⊂ K(N ) = Ker(ϕN ). We claim
that X[2]!⊂"X[4] is totally isotropic with respect to the commutator pairing eN . Indeed, if x,
x" ∈ X[2] T for some k-scheme T then possibly! after " passing to an fppf covering of T we can
write x = 2y and x" = 2y " for some y, y " ∈ X[4] T . Our claim now follows by noting that the
restriction of eN to X[4] × X[4] takes values in µ4 . By Corollary (8.11) we can find a line bundle

PolWp, 15 september, 2011 (812)

– 159 –
L on Xk such that N = ∼ [2]∗ L on X . But then 4λ = ϕ[2]∗ L = 4ϕL , using Corollary (7.25). As
k k
[4]X is an epimorphism, it follows that λk = ϕL . So (b) holds with K = k.
To see that the apparently stronger condition (c) holds, view λ as a k-valued point of
Hom (X, X t ). Let P (λ) ⊂ PicX/k be the inverse image of λ under the homomorphism
ϕ: PicX/k → Hom (X, X t ). As P (λ) is a closed subscheme of PicX/k , it is locally of finite
type. If T is a k-scheme then the T -valued points of P (λ) are the classes of line bundles M
on XT such that ϕM = λ. Note that P (λ) inherits a natural action of X t = Pic0X/k by trans-
lations. The exact sequence of (7.22) tells us that for every k-scheme T the set P (λ)(T ) is
either empty or it is a principal homogeneous space under X t (T ). Hence if L is a line bundle
∼
on Xk with ϕL = λk then x &→ [t∗x L] defines an isomorphism of k-schemes (X t )k −→ P (λ)k . In
particular, P (λ) is a geometrically integral k-scheme, so it has points with values in some finite
separable extension k ⊂ K.
Finally, it is clear that (c) implies both (a) and (b). !

(11.3) Corollary. Let X/k be an abelian variety. Then the homomorphism ψ: NSX/k →
Hom symm (X, X t ) of (7.26) is an isomorphism.
Proof. Both group schemes are étale and we already know that ψ is injective. Hence it suffices
to show that ψ is surjective on ks -valued points, and this follows from the preceding Proposi-
tion. !

(11.4) Proposition. Let X/k be an abelian variety. Let λ: X → X t be a symmetric homo-

morphism, and write M := (id, λ)∗ PX . Let k ⊂ K be a field extension and let L be a line
bundle on XK with λK = ϕL .
(i) We have: λ is an isogeny ⇔ L is non-degenerate ⇔ M is non-degenerate.
(ii) If λ is an isogeny then L is effective if and only if M is effective.
(iii) We have: L is ample ⇔ M is ample.
Proof. By Proposition (11.1) ϕMK = 2ϕL = ϕL2 , so MK and L2 are algebraically equivalent.
Now (i) is clear, and (ii) follows from Corollary (9.23) and part (ii) of Proposition (9.18). For
(iii), recall that a line bundle N on X is ample if and only if N is non-degenerate and effective;
this is just Proposition (2.22). !
Putting Propositions (2.22), (11.2) and (11.4) together we obtain the following corollary.

(11.5) Corollary. Let X/k be an abelian variety. Let λ: X → X t be a homomorphism. Then

the following properties are equivalent:
(a1) λ is a symmetric isogeny and the line bundle (id, λ)∗ P on X is ample;
(a2) λ is a symmetric isogeny and the line bundle (id, λ)∗ P on X is effective;
(b1) there exists a field extension k ⊂ K and an ample line bundle L on XK such that
λK = ϕL ;
(b2) there exists a finite separable field extension k ⊂ K and an ample line bundle L on XK
such that λK = ϕL .

(11.6) Definition. Let X be an abelian variety over a field k. A polarization of X is an isogeny

λ: X → X t that satisfies the equivalent conditions in (11.5).
By the Riemann-Roch Theorem (9.11) the degree of a polarization is always a square:
deg(λ) = d2 with d = χ(L) if λk = ϕL . If λ is an isomorphism (equivalent: λ has degree 1) then

– 160 –
we call it a principal polarization.
It is clear that the sum of two polarizations is again a polarization. But of course the
polarizations do not form a subgroup of Hom (X, X t ).
We also remark that if λ is a polarization, then for any line bundle L on XK with λK = ϕL
we have that L is ample. In fact, ampleness of a line bundle N on an abelian variety only
depends on the associated homomorphism ϕN , as is clear for instance from Proposition (11.4).

(11.7) Let X be an abelian variety over a field k. We have an exact sequence of fppf sheaves

0 −→ X t −→ PicX/k −→ Hom symm (X, X t ) −→ 0

which gives a long exact sequence in fppf cohomology

∂
0 −→ X t (k) −→ Pic(X) −→ Homsymm (X, X t ) −→ Hfppf
1
(k, X t ) −→ · · · .

For λ: X → X t a symmetric homomorphism, ∂(λ) is the obstruction for finding a line bundle L
on X (over k) with ϕL = λ. Now we know from Proposition (11.2) that ∂(2λ) = 0; hence ∂(λ)
lies in the image of ! "
1
Hfppf k, X t [2] → Hfppf
1
(k, X t ) .
(NOG VERDERE OPM OVER MAKEN, BV VGL MET GALOIS COHOM?)

(11.8) Proposition. Let f : X → Y be an isogeny. If µ: Y → Y t is a polarization of Y , then

f ∗ µ := f t ◦ µ ◦ f is a polarization of X of degree deg(f ∗ µ) = deg(f )2 · deg(µ).
Proof. It is clear that f ∗ µ is an isogeny of the given degree. By assumption there is a field
extension k ⊂ K and an ample line bundle M on YK such that µK = ϕM . Then f ∗ µK = ϕf ∗ M
and because f is finite f ∗ M is an ample line bundle on XK . !
See Exercise (11.1) for a generalization.

(11.9) Definition. Let X and Y be abelian varieties over k. A (divisorial) correspondence

∼
between X and Y is a line bundle L on X × Y together with rigidifications α: L|{0}×Y −→ OY
∼
and β: L|X×{0} −→ OX that coincide on the fibre over (0, 0).
Correspondences between X and Y form a group Corrk (X, Y ), with group structure ob-
tained by taking tensor products of line bundles. (Cf. the definition of PX/S,ε in Section (6.2).)
Note that the multiplicative groep Gm acts (transitively) on the choices of the rigidifications
(α, β). Moreover, if Y = X we can speak of symmetric correspondences.
The Poincaré bundle P = PX on X × X t comes equipped with a rigidification along
{0} × X t . There is a unique rigidification along X × {0} such that the two rigidifications agree
at the origin (0, 0). We thus obtain an element

[PX ] = (PX , αP , βP ) ∈ Corrk (X, X t ) .

The following proposition makes an alternative definition of the notion of polarization pos-
sible.
(11.10) Proposition. Let X/k be an abelian variety. Then we have a bijection
# $
t ∼ symmetric divisorial correspondences
{polarizations λ: X → X } −→
(L, α, β) on X × X such that ∆∗X L is ample

– 161 –
by associating to a polarization λ the divisorial correspondence (L, α, β) with L = (idX ×λ)∗ PX
and α and β the pull-backs under idX × λ of the rigidifications αP and βP .
Proof. This is essentially contained in Corollary (11.5). The inverse map is obtained by associat-
ing to (L, α, β) the unique homomorphism λ: X → X t such that (L, α) = (idX ×λ)∗ (PX , αP ) as
rigidified line bundles on X × X. The assumption that (L, α, β) is symmetric implies that λX is
symmetric, and because (idX , λ)∗ PX = ∆∗X (idX × λ)∗ PX = ∆∗X L is ample, λ is a polarization.
This establishes the correspondence. !
The alternative definition of a polarization suggested by Proposition (11.10) as “a sym-
metric self-correspondence such that restriction to the diagonal is ample” is evidently similar
in appearance to the definition of a positive definite symmetric bilinear form in linear algebra.
But, whereas in linear algebra one dominantly! views a bilinear
" form b as a map V × V → k
∗
rather than as a map V → V given by v &→ w &→ b(v, w) , in the theory of abelian varieties the
latter point of view dominates. Note further that the role of the evaluation map V × V ∗ → k
with (v, w) &→ w(v) is played in our context by the Poincaré bundle P.

§ 2. Pairings.

We now turn to the study of some bilinear forms attached to isogenies. In its most general form,
any isogeny f gives a pairing ef between Ker(f ) and Ker(f t ); this is an application of the duality
result Theorem (7.5). Of particular interest is the case f = [n]X . If we choose a polarization λ
we can map X[n] to X t [n], and we obtain a bilinear form eλn on X[n], called the Weil pairing.
The pairings that we consider satisfy a number of compatibilities, which, for instance, allow us
to take the limit of the pairings eλ"m , obtaining a bilinear form E λ with values in Z" (1) on the
Tate module T" X. In cohomological terms this pairing is the first Chern class of λ (or rather, of
any line bundle representing it). It is the #-adic analogue of what over C is called the Riemann
form associated to a polarization. (See also ???)

(11.11) Definition. Let f : X → Y be an isogeny of abelian varieties over a field k. Write

∼
β: Ker(f t ) −→ Ker(f )D for the isomorphism of Theorem (7.5).
(i) Define
ef : Ker(f ) × Ker(f t ) −→ Gm,k

to be the perfect bilinear pairing given (on points) by ef (x, y) = β(y)(x). Note that if Ker(f ) is
killed by n ∈ Z!1 then ef takes values in µn ⊂ Gm . In the particular case that f = nX : X → X
we obtain a pairing
en : X[n] × X t [n] → µn

which we call the Weil pairing.

(ii) Let λ: X → X t be a homomorphism. We write

eλn : X[n] × X[n] → µn

! "
for the bilinear pairing given by eλn (x1 , x2 ) = en x1 , λ(x2 ) . If λ = ϕL for some line bundle L
then we also write eL λ
n instead of en .

Recall that if A and B are finite commutative group schemes (written additively), a pairing
e: A×B → Gm is said to be bilinear if e(a+a" , b) = e(a, b)·e(a" , b) and e(a, b+b" ) = e(a, b)·e(a, b" )
for all points a and a" of A and b and b" of B. (Points with values in an arbitrary k-scheme.) The

– 162 –
 ∼
pairing e is said to be perfect if sending a to e(a, −): B → Gm gives an isomorphism A −→ B D .
∼
This is equivalent to the condition that b &→ e(−, b) gives an isomorphism B −→ AD . It is
clear from the construction that the pairings ef , in particular also the Weil pairings, are perfect
bilinear pairings. If n is relatively prime to the degree of λ then the pairing eλn is perfect, too.
There are various ways in which we can make the pairings defined above more explicit. We
shall give a couple of different points of view.

(11.12) Let us first try to unravel the definition of ef by going back to the proof of (7.5). This
leads to the following description. Let T be a k-scheme. Let L be a rigidified line bundle on YT
that represents a class η ∈ Ker(f t )(T ). Then f ∗ L ∼
= OXT . Hence the geometric line bundle L
corresponding to L can be described as a quotient of XT ×T A1T by an action of Ker(f )T . More
precisely, by what was explained in (7.3) there exists a character χ: Ker(f )T → Gm,T such that
the action of a point x of Ker(f ) on XT ×T A1T is given (on points) by
! "
(z, a) &→ z + x, χ(x) · a .
∼
The isomorphism Ker(f t ) −→ Ker(f )D of Theorem (7.5) sends η to χ. Hence the pairing ef is
given by ef (x, η) = χ(x).

(11.13) Next let us give a more geometric description of the Weil pairings en . Suppose D is
a divisor on X such that nD is linearly equivalent to zero. Write L = OX (D). As n∗ L ∼ = OX
∗
(cf. Exercise (7.2)), there exists a rational function g on X with divisor (g) = n D. But also
Ln ∼= OX , so there exists a rational function f with divisor (f ) = nD. Then n∗ f and g n both
have divisor n · n∗ D = n∗ (nD), so there is a constant c ∈ k ∗ with g n = c · (n∗ f ).
Let x ∈ X[n](k) be a k-rational n-torsion point. We find that
! " ! "n
g(ξ)n = c · f (nξ) = c · f n(ξ + x) = g(ξ + x)n = (t∗x g)(ξ)
! "
for all ξ ∈ X(k). So g/t∗x (g) is an n-th root of unity. We claim that in fact en x, [D] = g/t∗x (g).
∼
To see this, note that we have an isomorphism of line bundles n∗ L −→ OX given by
g &→ 1. As described in (11.12), there is a character χ: X[n] → Gm such that the natural
action of X[n] on n∗ L becomes the action of X[n] on OX given by the character χ. Note that
x ∈ X[n](k) acts on! the identity
" section 1 ∈ Γ(X, OX ) as multiplication by χ(x)−1 . Hence
∗
g/tx (g) = χ(x) = en x, [D] , as claimed.

(11.14) Example. We calculate the Weil pairing e3 on the elliptic curve E over F2 given by
2 3
the affine
! " equation y + y = x . This curve has 9 points over F4 which realise an isomorphism
∼
E[3] F4 = Z/3Z × Z/3Z. Let O = P∞ be the point at ∞, which we take as the identity
element on E. The bundle L = OE (P∞ ) is ample. The associated principal polarization
∼
λ: E −→ E t = Pic0E/F2 is given on points by R &→ OE (O − R). (Note that this is minus
the map given by R &→ OE (R − O); see Remark (2.11).)
Let us calculate eλ3 (Q, P ) for P = (0, 0) and Q = (1, α), where α is an element of F4 not
in F2 . First we note that the function y has divisor (y) = 3 · (P − O). Next we compute
a function g with divisor [3]∗ (O − P ). For this we compute the “triplication formula” on E
which expresses for a point R = (ξ, η) on E the coordinates of 3R in those of R. As we
have seen in Example (5.26), E is supersingular. The relative Frobenius π = FE/F2 : E → E
is an endomorphism of E. One can show that it satisfies π 2 = −2, for example by verifying

– 163 –
that for T ∈ E the point π 2 (T ) lies on the tangent line to E in T . As −1 on E is given by
(x, y) &→ (x, y + 1) we !find that 2R has coordinates (ξ 4 , η 4 + 1). Next " one calculates that the
coordinates of 3R are (ξ 9 + ξ 3 + 1)/(ξ + ξ 4 )2 , (ηξ 3 + 1)3 /(ξ + ξ 4 )3 . Hence the function

x4 + x
g=
yx3 + 1

has divisor (g) = [3]∗ (O − P ). (Use that 3 · (g) = [3]∗ (y) = 3 · [3]∗ (O − P ).)
Now we know that g/t∗Q g is constant and this constant can be computed by evaluating g
and t∗Q g at a suitable point T ; so

g/t∗Q g = g(T )/g(T + Q) .

For T we take a point rational over F64 . Let γ be a generator of F∗64 with γ 21 = α and such
that δ := γ 9 ∈ F∗8 satisfies δ 3 + δ = 1. Then the point T = (γ 3 , γ 18 ) is in E(F64 ). One easily
verifies that (γ 24 , γ 18 + 1) is again a point of E, and that it lies on the line ! through T" and Q;
24 18 λ
hence T + Q = (γ , γ ). By (11.13) we conclude that e3 (Q, P ) = e3 Q, (O − P ) equals
(γ 12 + γ 3 )/(γ 33 + γ 24 ) = 1/γ 21 = 1/α = α2 .
The value of eλ3 (P " , Q" ) for any pair (P " , Q" ) ∈ E[3] × E[3] can be computed from this using
the fact that e3 is bilinear and alternating; see Cor. (11.22) below.

(11.15) Let f : X → Y be an isogeny of abelian varieties over a field k. By definition, f t : Y t →

X t is the unique map such that (f × idY t )∗ PY ∼ = (idX × f t )∗ PX as line bundles on X × Y t
t
with rigidification along {0} × Y . Note that this isomorphism is unique, so without ambiguity
we can define Q := (f × idY t )∗ PY = (idX × f t )∗ PX . The diagram to keep in mind is

PX Q PY
(1)
id×f t f ×id
X × Xt ←−−−− X ×Yt −−−→ Y × Y t

On the line bundle Q we have an action of Ker(f ) × {0}, lifting the action on X × Y t by
∼
translations. This action is given by isomorphisms σx : QT −→ t∗(x,0) QT , for any k-scheme T
! " t
and x ∈ Ker(f ) T . Likewise, we !have " an action of {0} × Ker(f ), given by isomorphisms
∼ ∗ t
τq : QT −→ t(0,q) QT for q ∈ Ker(f ) T . Unless f is an isomorphism, these two group scheme
actions on Q do not commute, for if they did it would give us an action of Ker(f ) × Ker(f t )
and Q would descend to a line bundle L on (X × Y t )/Ker(f ) × Ker(f t ) = Y × X t . But then
we had (−1)g = χ(PX ) = deg(f ) · χ(L), which is possible only if deg(f ) = 1. We shall prove
that the extent to which the two actions fail to commute is measured by the pairing ef .
! "
Let Q " be the restriction of Q to X × Ker(f t ). We have Q " = (idX × f t )∗ (PX )|X×{0} ,
∼
so the natural rigidification of PX along X × {0} (see (7.7)) gives us a trivialisation Q " −→
OX×Ker(f t ) . The action of {0} × Ker(f t ) on Q restricts to the trivial action on Q " . It will be
useful to think of Q " as being the sheaf of sections of A1 over X ×Ker(f t ). Writing A1X×Ker(f t ) =
X × Ker(f t ) × A1 , the action of a point (0, q) ∈ {0} × Ker(f t ) on Q " corresponds to the action
on X × Ker(f t ) × A1 given by τq : (t, u, a) &→ (t, u + q, a).
Note that also the action of Ker(f ) × {0} restricts to an action on Q " . To describe this
action we apply what was explained in (11.12) in the “universal case”, i.e., with T = Ker(f t )
and η = idT . The corresponding line bundle L on YT = Y × Ker(f t ) is just the restriction
of PY to Y × Ker(f t ), so f ∗ L is precisely our bundle Q " . If we write a point of Ker(f )T =

– 164 –
Ker(f )×k Ker(f t ) as a pair (x, u) then the conclusion !of (11.12) is" that the character χ: Ker(f )×k
Ker(f t ) → Gm,k ×k Ker(f t ) is given by (x, u) &→ ef (x, u), u . Hence the action of a point
" t 1
! 0) ∈ Ker(f )×{0}"on Q corresponds to the action on X ×Ker(f )×A given by σx : (t, u, a) &→
(x,
t + x, u, ef (x, u) · a .
Now we can start drawing some conclusions. The first result is an interpretation of the
pairing ef as a measure for the extent to which the two group scheme actions on Q fail to
commute.

(11.16) Proposition. Let f : X → Y be an isogeny of abelian varieties over a field k, and

∗
consider the line bundle t f t )∗ PX on X × Y t . Let T be a k-
! " Q := (f × idt Y! ) "PY = (idX × ∼
scheme, x ∈ Ker(f ) T and q ∈ Ker(f ) T . Let σx : QT −→ t∗(x,0) QT be the isomorphism that
∼
gives the action of (x, 0) ∈ Ker(f ) × {0} on QT , and let τq : QT −→ t∗(0,q) QT be the isomorphism
that gives the action of (0, q) ∈ {0} × Ker(f t ). Then we have a commutative diagram
σ t∗
(x,0) τq
QT x
−−→ t∗(x,0) QT −−−−−→ t∗(x,q) QT
% 
% multiplication by e (x, q)
% ' f

τq t∗
(0,q) σx
QT −−→ t∗(0,q) QT −−−−−→ t∗(x,q) QT

Proof. A priori it is clear that there exists a constant c ∈ Gm (T ) such that (t∗(0,q) σx ) ◦ τq =
c · (t∗(x,0) τq ) ◦ σx , so all we need to show is that c = ef (x, q). For this we may restrict everything
to X × Ker(f t ). As in the above discussion, we think of Q " as the sheaf of sections ! of A1
over X × Ker(f t ). We have seen that (t∗(x,0) τq ) ◦ σx is given on points by (t, u, a) &→ t + x, u +
" ! "
q, ef (x, u) · a , whereas (t∗(0,q) σx ) ◦ τq is given by (t, u, a) &→ t + x, u + q, ef (x, u + q) · a . Because
ef is bilinear, the result follows. !
Next we prove a compatibility result among the two main duality theorems that we have
proved in Chapter 7.

(11.17) Proposition. Let f : X → Y be an isogeny of abelian varieties. Let κX : X → X tt be

the canonical isomorphism.
! " t
! "
" k-scheme−1T and points x ∈ Ker(f ) T and η ∈ Ker(f ) T we have the relation
! (i) For any
ef t η, κX (x) = ef (x, η) .
∼ ∼
(ii) Let β1 : Ker(f t ) −→ Ker(f )D and β2 : Ker(f tt ) −→ Ker(f t )D be the canonical isomor-
∼
phisms as in Theorem (7.5), and let γ: Ker(f )DD −→ Ker(f ) be the isomophism of Theo-
∼
rem (3.22). Then the isomorphism Ker(f ) −→ Ker(f tt ) induced by κX equals −β2−1 ◦ β1D ◦ γ −1 .
Proof. (i) Consider the commutative diagram
id×f t f ×id
X × Xt ←−−−− X ×Yt −−−→ Y ×Yt
  

κX ×id'

κX ×id'
κ ×id
' Y (2)
t tt
id×f f ×id
X tt × X t ←−−−− X tt × Y t −−−−→ Y tt × Y t .

If we read the lower row from right to left (term by term!), we get the row

id×f tt f t ×id
Y t × Y tt ←−−−− Y t × X tt −−−−→ X t × X tt

– 165 –
which is precisely (1) for the morphism f t : Y t → X t . Now the result follows from the previous
proposition, with the −1 in the exponent coming from the fact that we are reading the lower
row in (2) from right to left, thereby switching factors. ! " ! "
D ◦ −1
! (ii) This" follows
! from"!(i)" using the relations ef (x, η) = β1 (η) x = (β1 γ )(x) η and
ef t η, κX (x) = β2 κX (x) η . !

(11.18) Example. Let X be an abelian variety over k. Let P = PX be its Poincaré bundle.
Let n be a positive integer, and let en : X[n] × X t [n] → µn be the Weil pairing.
The geometric line bundle on X × X t [n] that corresponds to P|X×X t [n] is the quotient of
A1X×X t [n] = X×X t [n]×A1 under the action of X[n]×{0}, with x ∈ X[n] acting on X×X t [n]×A1
! "
by σx : (t, u, a) &→ t + x, u, en (x, u) · a .
To make this completely explicit, suppose k = k and char(k) ! n, so that X[n] and X t [n]
are constant group schemes, each consisting of n2g distinct points. Then for ξ ∈ X t [n](k), the
restriction of the Poincaré bundle to X × {ξ} is given by
! " ( ) *
P|X×{ξ} U = f ∈ OX (n−1 U ) ) f (v + x) = en (x, ξ) · f (v) for all v ∈ n−1 U and x ∈ X[n] .

For the restriction of PX to X[n] × X t we have an analogous description; namely, the

corresponding geometric line bundle is the quotient of A1X[n]×X t = X[n] × X t × A1 under the
t t t 1
!
action of {0} × " X [n], with ξ ∈ X [n] acting on X[n] × X × A by τ ξ : (t, u, a) →
& t, u +
ξ, en (t, ξ)−1 · a . Note, however, that whereas our description of P|X×X t [n] is essentially a
reformulation of the definition of the Weil pairing, to arrive at our description of P|X[n]×X t we
use (i) of Proposition (11.17).

(11.19) Let L be a non-degenerate line bundle on an abelian variety X. As the associated

isogeny ϕL : X → X t is symmetric, we have K(L) = Ker(ϕL ) = Ker(ϕtL ), and we obtain a
pairing
eϕL : K(L) × K(L) → Gm .
On the other hand we have the theta group 1 −→ Gm −→ G (L) −→ K(L) −→ 0, and this, too,
gives a pairing
eL : K(L) × K(L) → Gm .

(11.20) Proposition. We have eϕL = eL .

Proof. We apply what was explained in (11.15) to the isogeny ϕL : X → X t . We identify
∼
X × X tt with X × X via the isomorphism id × κX : X × X −→ X × X tt . The line bundle
Q := (ϕL × κX )∗ PX t = (id × ϕL )∗ PX is none other than the Mumford bundle Λ(L) associated
to L. Let Q " := Q|X×K(L) = Λ(L)|X×K(L) which, as we already knew from Lemma (2.17), is
trivial. ! "
Let T be a k-scheme, and consider T -valued points x, y ∈ K(L) T . Possibly after replac-
∼ ∼
ing T by a covering we can choose isomorphisms ϕ: LT −→ t∗x LT and ψ: LT −→ t∗y LT . Then
(x, ϕ) and (y, ψ) are T -valued points of G (L), and by definition of the pairing eL we have the
relation
(t∗y ϕ) ◦ ψ = eL (x, y) · (t∗x ψ) ◦ ϕ . (3)
We can also view ψ as the trivialisation
∼
ψ: OXT ×{y} −→ Λ(LT )XT ×{y} = t∗y LT ⊗ L−1
T

– 166 –
that sends 1 ∈ Γ(XT , OXT ×{y} ) to the global section ψ of t∗y LT ⊗ L−1 ∗
T . If σx : QT → t(x,0) QT
is the isomorphism that gives the action of (x, 0) ∈ K(L) × {0} on Q then it follows from what
we have seen in (11.15) that we have a commutative diagram
(σx )|X ×{y}
Λ(L)XT ×{y} −−−−−T−−−→ t∗(x,0) Λ(L)|XT ×{y}
+ + ! "

ψ

eϕL (x,y)· t∗(x,0) ψ
can
OXT ×{y} −−−−−−−−→ t∗(x,0) OXT ×{y} .

We have t∗(x,0) Λ(LT ) = m∗ (t∗x LT ⊗ L−1 ∗ ∗ −1 −1

T ) ⊗ p1 (tx LT ⊗ LT ) ⊗ Λ(LT ). Taking this as an
∗ ∗ −1
identification, σx is given on sections by s &→ m ϕ ⊗ p2 ϕ ⊗ s. (Note that this does not depend
on the choice of ϕ.) Now restrict to XT × {y} and use the natural identification

t∗(x,0) Λ(LT )|XT ×{y} = t∗x+y LT ⊗ t∗x L−1 ∗ ∗

T = Hom(tx LT , tx+y LT ) .

we find that σx ◦ ψ maps 1 ∈ Γ(XT , OXT ×{y} ) to the homomorphism t∗y ϕ ◦ ψ ◦ ϕ−1 : t∗x LT →
t∗x+y LT . On the other hand, the composition (t∗(x,0) ψ) ◦ can sends 1 to t∗x ψ. Hence we have

t∗y ϕ ◦ ψ ◦ ϕ−1 = eϕL (x, y) · t∗x ψ

and comparison with (3) now gives the result. !

(11.21) Proposition. (i) Let f : X → Y be a homomorphism of abelian varieties over k. Then

for any integer n " 1 the diagram
1×f t
X[n] × Y t [n] −−−→ X[n] × X t [n]
 

f ×1'
e
'n
e
Y [n] × Y t [n] n
−−−→ µn
! " t
! "
is !commutative.
" ! In other
" words: if T is a k-scheme, x ∈ X[n] T and η ∈ Y [n] T then
t
en f (x), η = en x, f (η) .
(ii) Let f : X → Y and g: Y → Z be isogenies, and write h := g ◦ f : X → Z. Then we have
“commutative diagrams”
ef eg
Ker(f ) × Ker(f t ) −−→ Gm Ker(g) × Ker(g t ) −−→ Gm
 + % +  %
  t %   %
i' g % and f 'i %
e e
Ker(h) × Ker(ht ) h
−−→ Gm Ker(h) × Ker(ht ) h
−−→ Gm

where the maps labelled “i” are the natural inclusion homomorphisms. By ! our
" assertion that !the"
t
first diagram
! is commutative
" ! " mean that if T is a k-scheme, x ∈ Ker(f ) T and η ∈ Ker(h ) T
we
then ef x, g t (η) = eh i(x), η ; similarly for the second diagram.
Proof. (i) Let χ: Y [n]T → Gm,T be the character corresponding to η, as in (11.12). Then the
character corresponding to ht (η) is χ ◦ h: X[n]T → Gm,T . By (11.12) we find
! " ! " ! "
en h(x), η = χ h(x) = χ ◦ h(x) = en x, ht (η) .

(ii) Let χ: Ker(h)T → Gm,T be the character corresponding to η. Then the character
Ker(f )T → Gm,T corresponding to g t (η) is simply χ ◦ i. Hence by what was explained in (11.12),

– 167 –
 ! " ! " ! "
eh i(x), η = χ i(x) = χ ◦ i(x) = ef x, g t (η) . This gives the first commutative diagram. For
the second, apply the first diagram to the composition f t ◦ g t : Z t → Y t → X t ; then apply (i) of
Proposition (11.17). !

(11.22) Corollary. Let λ: X → X t be a polarization, and let n be ! a" positive integer. Then
λ
the pairing en : X[n] × X[n] → µn is alternating: for any x ∈ X[n] T with T a k-scheme we
have eλn (x, x) = 1.
Proof. Without loss of generality we may assume that k = k and write λ = ϕL for some
ample L. Consider the composition nλ = λ ◦ [n]X . Applying (ii) of Proposition (11.21) we find
a commutative diagram e
X[n] × X t [n] −−n→ Gm
 + %
  %
i' λ %
nλ e
Ker(nλ) × Ker(nλ) −− → Gm
! " ! "
This gives eλn (x, x) = en x, λ ◦ i(x) = enλ i(x), i(x) = 1, where in the last step we use Propo-
sition (11.20) together with the remark that nλ = ϕLn . !
In particular, we find that the pairing eλn is skew-symmetric: eλn (x, y) = eλn (y, x)−1 . Note,
however, that skew-symmetry is weaker in general than the property of being alternating.

(11.23) Let X be an abelian variety over a field k. Fix a separable closure k ⊂ ks . As usual,
# denotes a prime number different from char(k). Let x = (0, x1 , x2 , . . .) be an element of T" X
and ξ = (0, ξ1 , ξ2 , . . .) and element of T" X t . Applying (ii) of Proposition (11.21) we find that

e"m (xm , ξm ) = e"m+1 (# · xm+1 , ξm+1 ) = e"m+1 (xm+1 , ξm+1 )" .

This means precisely that

! "
E(x, ξ) = 1, e" (x1 , ξ1 ), e"2 (x2 , ξ2 ), . . .

is a well-defined element of Z" (1) = T" Gm . The map (x, ξ) &→ E(x, ξ) defines a perfect bilinear
pairing
E: T" X × T" X t → Z" (1) .
∼
If β: T" X t −→ (T" X)∨ (1) is the canonical isomorphism as in Proposition (10.9) then the pair-
ing E is nothing else but the composition
id×β ev
T" X × T" X t −−−→ T" X × (T" X)∨ (1) −−→ Z" (1)

where the map “ev” is the canonical pairing, or “evaluation pairing”. Note that the pairing E
is equivariant with respect to the natural action of Gal(ks /k) on all the terms involved.
If λ: X → X t is a polarization, we obtain a pairing
! "
E λ : T" X × T" X → Z" (1) by E λ (x, x" ) := E x, T" λ(x" ) .

If λ = ϕL we also write E L for E λ . It readily follows from Corollary (11.22) that the pairing E λ
is alternating. ! "
Putting everything together, E λ is a Gal(ks /k)-invariant element in ∧2 (T" X)∨ (1). The
λ
cohomological interpretation
! 2is that ∨E" is the first
! Chern class
" of λ, or rather of any line bundle
representing λ. Note that ∧ (T" X) (1) = H 2 Xks , Z" (1) , see Corollary (10.39).

– 168 –
§ 3. Existence of polarizations, and Zarhin’s trick.

(11.24) Suppose we have an abelian variety X of dimension g over a field k. If g = 1 then

X is an elliptic curve, and the origin O (as a divisor on X) gives a principal polarization (via
Q &→ O − Q). If g " 2 then in general X does not carry a principal polarization, not even if we
allow an extension of the base field. Let us explain why this is so.
Fix g " 2. We shall use the fact that there exists an algebraically closed field k and an
abelian variety Y of dimension g over k such that End(Y ) = Z. A proof of this shall be given
later; see ??. Note that this does not work for arbitrary k; for instance, every abelian variety
over Fp has Z " End(Y ), as we shall see in ??.
If Y carries no principal polarization then we have the desired example. Hence we may
assume there is a principal polarization λ: Y → Y t . As k = k there is a line bundle L with
λ = ϕL . Because λ is principal and End(X) = Z the only polarizations of Y are those of the
form ϕLn = n · λ, of degree n2g .
On the other hand, if # is any prime number different from char(k) then Y [#] ∼ = (Z/#Z)2g
as group schemes. Hence Y has a subgroup scheme H of order #. Let q: Y → X := Y /H be
the quotient. If µ: X → X t is a polarization then q ∗ µ is a polarization of Y , with deg(q ∗ µ) =
#2 · deg(µ). But as just explained, any polarization of Y has degree equal to n2g for some n ∈ N.
Hence µ cannot be principal.
With a similar construction we shall see later that an abelian variety of dimension g " 2
over a field of characteristic p in general does not even carry a separable polarization; see ??.
To arrive at some positive results, we shall now first give a very useful criterion for when a
polarization λ: X → X t descends over an isogeny f : X → Y . If L is a line bundle on X then
by Theorem (8.10) there exists a line bundle M on Y with L ∼ = f ∗ M if and only if the following
conditions are satisfied:
(a) Ker(f ) is contained in K(L) and is totally isotropic with respect to the pairing eG (L) = eϕL ;
(b) the inclusion map Ker(f ) 1→ K(L) can be lifted to a homomorphism Ker(f ) 1→ G (L).
(The second condition in (a) is in fact implied by (b).) As we shall prove now, in order for a
polarization to descend, it suffices that the analogue of condition (a) holds.

(11.25) Proposition. Let λ: X → X t be a symmetric isogeny, and let f : X → Y be an isogeny.

(i) There exists a symmetric isogeny µ: Y → Y t such that λ = f ∗ µ := f t ◦ µ ◦ f if and only
if Ker(f ) is contained in Ker(λ) and is totally isotropic with respect to the pairing eλ : Ker(λ) ×
Ker(λ) → Gm . If such an isogeny µ exists then it is unique.
(ii) Assume that an isogeny µ as in (i) exists. Then µ is a polarization if and only if λ is a
polarization.
Note that the “only if” in (ii) was already proven in Proposition (11.8). For this implication
the assumption that f is an isogeny can be weakened; see Exercise (11.1).
Proof. (i) If λ = f t ◦ µ ◦ f then Ker(f ) ⊂ Ker(λ) and it follows from (ii) of Proposition (11.21),
applied with g = (f t ◦ µ) and h = λ, that Ker(f ) is totally isotropic for the pairing eλ .
For the converse, assume Ker(f ) is contained in Ker(λ) and is totally isotropic with respect
to eλ . Consider the line bundle M := (1 × λ)∗ PX on X × X. Recall from Example (8.26) that
the theta group G (M ) is naturally isomorphic to the Heisenberg group associated to the group
scheme Ker(λ). We have natural actions of Ker(λ) × {0} and {0} × Ker(λ) on M ; for the first
action note that M can also be written as (λ × 1)∗ PX t . The assumption that Ker(f ) ⊂ Ker(λ)

– 169 –
is totally isotropic for eλ means precisely that the actions of Ker(f ) × {0} and of {0} × Ker(f )
commute, and therefore define an action of Ker(f )×Ker(f ) on M . This gives us a line bundle N
on Y × Y such that M ∼ = (f × f )∗ N . If µ: Y → Y t is the (unique) homomorphism such that
N = (1×µ)∗ PY then we get the desired relation λ = f t ◦ µ ◦ f . The uniqueness of µ is immediate
from Lemma (5.4). But we also have λ = λt = (f t ◦ µ ◦ f )t = f t ◦ µt ◦ f . Hence µ = µt .
(ii) By Proposition (11.2) there exists a field extension k ⊂ K and a line bundle L on YK
with µK = ϕL , and then λK = ϕf ∗ L . Because f is finite, L is effective if and only if f ∗ L is
effective. !

(11.26) Corollary. Let X be an abelian variety over an algebraically closed field. Then X is
isogenous to an abelian variety that admits a principal polarization.
Proof. Start with any polarization λ: X → X t . By Lemma (8.22) there exists a Lagrangian
subgroup H ⊂ Ker(λ). (There clearly exists a subgroup H ⊂ Ker(λ) satisfying condition (i) of
that Lemma.) By the previous Proposition, λ descends to a principal polarization on X/H. !
The conclusion of the Corollary no longer holds in general if we drop the assumption that the
ground field is algebraically closed. For examples, see e.g. Howe [1], [2] and Silverberg-Zarhin [1].

(11.27) Before we turn to Zarhin’s trick, we recall from Exercise (7.8) some notation.
Suppose X is an abelian variety and α = (aij ) is an r × s matrix with integral coefficients.
Then we denote by [α]X : X s → X r the homomorphism given by

s
,
! ! "
[α]X x1 , . . . , xs ) = a11 x1 + a12 x2 + · · · + a1s xs , . . . , aij xj , . . . , ar1 x1 + ar2 x2 + · · · + ars xs .
j=1

For r = s = 1 this just gives our usual notation [n]X for the “multiplication by n” maps. - As .
1
another example, the 1 × 2 matrix (1 1) gives the group law on X while the 2 × 1 matrix
1
gives the diagonal.
If β is a q × r matrix with integral coefficients then [β · α]X = [β]X ◦ [α]X : X s → X q . It
follows that if α is an invertible r × r matrix then [α]X is an automorphism of X r . Further, if
f : X → Y is a homomorphism of abelian varieties then for any integral r × s matrix α,

[α]Y ◦ (f, . . . , f ) = (f, . . . , f ) ◦ [α]X : X s → Y r .

/ 01 2 / 01 2
s r

(11.28) Proposition. Let X be an abelian variety of dimension g.

(i) If α ∈ Mr (Z) then [α]X : X r → X r has degree det(α)2g .
! "t 3 4
(ii) Let β be an r × s matrix with integral coefficients. Then [β]X = t β X t , where t β is
the transposed matrix.
Proof. (i) If !det(α)
" = 0 then it is readily seen that [α]X has infinite kernel, so by convention
we have deg [αX ] = 0. Now assume det(α) -= 0, and let {e1 , . . . , er } be the standard ordered
basis of Zr . By the theory of elementary divisors, there is an ordered basis {f1 , . . . , fr } for Zr
and a sequence of nonzero integers (n1 , . . . , nr ) such that α(ei ) = ni · fi . Let β ∈ GLr (Z) be the
matrix with β(ei ) = fi , and let γ = diag(n1 , . . . , nr ) be the diagonal matrix with coefficients ni .
Then [β]X is an automorphism of X r and it is clear that [γ]X : X r → X r , which is given by

– 170 –
(x1 , . . . , xr ) &→ (n1 x1 , . . . , nr xr ), has degree (n1 · · · nr )2g = det(α)2g . As [α]X = [γ]X ◦ [β]X the
claim follows.
(ii) Write β = (bij ). Any line bundle L on X r with class in Pic0 can be written as L =
p∗1 L1 ⊗· · ·⊗p∗r Lr , where the pi : X r → X are the projection maps and the Li are line bundles on X
with class in Pic0 . Because (X s )t ∼ = (X t )s (cf. Exercise (6.2)) it suffices to know the restriction
of [β]∗X L to each of the coordinate axes {0} × · · · × {0} × X × {0} × · · · × {0}. But the restriction
of [β]X to the j-th coordinate axis is the map X → X r given by x &→ (b1j x, b2j x, . . . , brj x) and
the pull-back of L under this map is
⊗b1j
b∗1j L1 ⊗ · · · ⊗ b∗rj Lr = L1 ⊗ · · · ⊗ L⊗b
r
rj
.

This means precisely that [β]tX : (X r )t = (X t )r → (X s )t = (X t )s is the map given by the matrix
 
b11 ··· bi1 ··· br1
 .. .. .. 
 . . . 
  t
 b1j ··· bij ··· brj 
 = β,
 .. .. .. 
 . . . 
b1s ··· bis ··· brs

as claimed. !

(11.29) Theorem. (Zarhin’s trick) Let X be an abelian variety over a field k. Then X 4 ×(X t )4
carries a principal polarization.
Proof. Suppose we have an abelian variety Y , a polarization µ: Y → Y t , and! an endomorphism "
α: Y → Y . Consider the isogeny f(: Y × Y →) Y × Y t given * by (y 1 , y 2 ) →
& y 1 − α(y 2 ), µ(y 2 ) .
)
The kernel is given by Ker(f ) = (α(y), y) y ∈ Ker(µ) . In particular, deg(f ) = deg(µ).
Proposition (11.25) tells us under what conditions the polarization µ × µ: (Y × Y ) → (Y t × Y t )
descends to a polarization on Y × Y t via the isogeny f . Namely: there exists a polarization ν
t ∗
on Y ×! Y with" f ν = (µ × µ) if and only if
(a) α !Ker(µ) ⊆ Ker(µ),
" and
(b) eµ α(y1 ), α(y2 ) · eµ (y1 , y2 ) = 1 for all (scheme valued) points y1 , y2 of Ker(µ).
Note that if such a descended polarization ν exists then it is principal.
Condition (a) means that there exists an endomorphism β: Y t → Y t such that β ◦ µ = µ ◦ α.
By (ii) of Proposition (11.21),
! " ! " ! " ! "
eµ α(y1 ), α(y2 ) = eµ ◦ α y1 , α(y2 ) = eβ ◦ µ y1 , α(y2 ) = eµ y1 , β t α(y2 ) ,
! "
so (b) is equivalent to the condition that eµ y1 , (1 + β t α)(y2 ) = 1 for all y1 , y2 in Ker(µ). As
eµ is a pefect pairing on Ker(µ), this is equivalent to the condition that (1 + β t α) ∈ End(Y )
kills Ker(µ).
We now apply this with Y = X 4 . Choose any polarization λ on X, and take µ = λ4 (so
µ = λ × λ × λ × λ). For α we take the endomorphism [α]X given by a 4 × 4 matrix α with
integral coefficients. As λ4 ◦ [α]X = [α]X t ◦ λ4 , condition (a) is automatically satisfied, and we
have β = [α]X t in the above. Using (ii) of Proposition (11.28) we find that the only condition
that remains is that [id4 + t αα]X kills Ker(µ) = Ker(λ)4 , where id4 is the 4 × 4 identity matrix.
Choose an integer m such that Ker(λ) ⊂ X[m]. We are done if we can find an integral 4 × 4
matrix α such that id4 + t αα ≡ 0 mod m. To see that such a matrix can be found we use the

– 171 –
fact that every integer can be written as a sum of four squares. In particular there exist integers
a, b, c, d with a2 + b2 + c2 + d2 = m − 1. Now take
 
a −b −c −d
b a −d c 
α=
c d
, (4)
a −b 
d −c b a

for which we have id4 + t αα = m · id4 . !

(11.30) Remarks. (i) The choice of the matrix α can be explained as follows. Consider the
Hamiltonian quaternion algebra H = R · 1 + R · i + R · j + R · k, which is a central simple algebra
over R. For x = a·1+b·i+c·j +d·k we define its complex conjugate by x̄ = a·1−b·i−c·j −d·k.
The reduced trace and norm of H over R are given by

TrdH/R (x) = x + x̄ = 2a and NrdH/R (x) = xx̄ = a2 + b2 + c2 + d2 .

Further, taking {1, i, j, k} as a basis of H, left multiplication by x is given precisely by the

matrix (4). The map h: H → M4 (R) sending x to this matrix! is an" injective homomorphism of
R-algebras, and we have h(x̄) = t h(x) and NrdH/R (x) = det h(x) . Further it is clear that h
maps the subring Z · 1 + Z · i + Z · j + Z · k into M4 (Z). In sum, we can think of α as being
the (left) multiplication by a · 1 + b · i + c · j + d · k, where a, b, c, d are chosen such that
a2 + b2 + c2 + d2 = m − 1.
(ii) In general there is no positive n such that for any abelian variety X the nth power X n
admits a principal polarization. To see this we go back to the example in (11.24). We start
with an abelian variety Y of dimension g " 2 over a field k = k such that End(Y ) = Z and
such that Y does admit a principal polarization; see ?? for the existence. Any homomorphism
Y n → (Y t )n is of the form λn ◦ [α]Y = [α]Y t ◦ λn for some α ∈ Mn (Z), and it easily follows
from (ii) of Proposition (11.28) that this homomorphism is symmetric if and only if α = t α.
Now choose a prime number # different from char(k), and choose a subgroup H ⊂ Y of order #,
generated by a point of order #. Let π: Y → X := Y /H be the quotient.
Let µ be any polarization on X n . By what was just ! explained
" we have (π n )∗ µ = λn ◦ [α]Y
for some α ∈ Mn (Z). Moreover, H × · · · × H ⊂ Ker [α]Y , which ! readily
" implies that α is
divisible by #, say α = # · β. Further we have deg(µ) · #2n = deg [α]Y = #2ng · det(β)2g , so
deg(µ) = #2n(g−1) · det(β)2g . In particular, X n does not carry a principal polarization.

Exercises.
(11.1) Let f : X → Y be a homomorphism of abelian varieties with finite kernel. If µ: Y → Y t
is a polarization, show that f ∗ µ := f t ◦ µ ◦ f is a polarization of X.
(11.2) Let X be an abelian variety over a field k. Suppose there exists a polarization λ: X → X t
with deg(λ) = m odd.
(i) Show that there exist integers a and b with 1 + a2 + b2 ≡ 0 mod m. [Hint: Use the Chinese
remainder theorem. First find a solution modulo p for any prime p dividing m. Then use
the fact that the curve C ⊂ A2 given by 1 + x2 + y 2 = 0 is smooth over Zp (p -= 2 !) to see
that the solutions can be lifted to solutions modulo arbitrarily high powers of p.]

– 172 –
(ii) Adapting the proof of Zarhin’s trick, show that X 2 × (X t )2 admits a principal polarization.
(11.3) Let L be a line bundle on an abelian variety X over a perfect field k. Write Y := K(L)0red ,
which is an abelian subvariety of X, and let q: X → Z := X/Y be the quotient.
(i) Show that ϕL : X → X t factors as ϕL = q t ◦ ψ ◦ q for some homomorphism ψ: Z → Z t .
(ii) Show that there is a finite separable field extension k ⊂ K and a line bundle M on ZK such
that ψK = ϕM .
(iii) With K and M as in (ii), conclude that the class of L ⊗ q ∗ M −1 lies in Pic0X/k (K).

– 173 –