Abelian varieties over local and global fields

Alexei N. Skorobogatov

March 17, 2016

Contents
1 Geometry 1
1.1 Definition and basic properties . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Theorem of the cube . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Quotients and isogenies . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4 Dual abelian variety . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2 Abelian varieties over non-closed fields 14

2.1 Galois cohomology . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.2 Local fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Duality for abelian varieties over local fields . . . . . . . . . . . . . . 20
2.4 Global fields . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.5 Brauer–Manin obstruction . . . . . . . . . . . . . . . . . . . . . . . . 30

1 Geometry
1.1 Definition and basic properties
Let k be a field. A variety over k is a scheme over k which is geometrically integral
and of finite type.

Definition 1.1 An abelian variety over k is a proper variety A with a distinguished

element 0 ∈ A(k) (the origin of the group law) and morphisms m : A×A → A (group
law) and [−1] : A → A (the inverse) satisfying the axioms of a group.

So the set of points A(k) is a group. In this section k is assumed to be algebraically

closed.
Some results immediately follow from the definition.

1
Proposition 1.2 Any abelian variety is non-singular.

Proof. This follows from the existence of one non-singular point and the transitive
action of A on itself by translations. 

Proposition 1.3 Any abelian variety is commutative.

Proof. We need to prove that the action of A on itself by conjugations is trivial.

This action fixes 0 ∈ A, so it defines an action of A on the local ring O at 0. Let
m be the maximal ideal of O. (Then O/m = k.) The induced action on the finite-
dimensional k-vector space O/mn defines a homomorphism A → GL(O/mn ), which
is a morphism from the proper connected variety A to the affine variety GL(O/mn ).
Any such morphism is constant. Hence the induced action on O/mn is trivial. Since
the intersection of all powers of m is 0, we see that A acts trivially on O. But O
is a localisation of the ring of regular functions k[U ] on some affine neighbourhood
U ⊂ A of 0, in particular, k[U ] ⊂ O, hence the action on k[U ] and thus on U is
trivial. It follows that the action of A by conjugations is trivial on the irreducible
component of A that contains U , but this is just A. 

1.2 Theorem of the cube

Theorem 1.4 (of the cube) Let X, Y, Z be varieties over k, where X and Y are
proper. Then any line bundle on X × Y × Z that restricts to trivial line bundles on
x0 × Y × Z, X × y0 × Z and X × Y × z0 must be trivial.

See [3] for a proof.

The meaning of this theorem is as follows. Let F be a contravariant functor from
varieties to abelian groups. Then we have homomorphisms induced by projections
n
Y n
Y
F( Xi ) −→ F( Xi ).
i=1,i6=j i=1

Let Un ⊂ F( ni=1 Xi ) be the sum of the images of all such

Q
Qn homomorphisms induced
by the projections that
Q forget one of the factors of i=1 Xi . One would wish to
know when Un = F( ni=1 Xi ).
To answer this question fix a point on each variety Xi and call it xi . For a non-
empty subset I ⊂ {1, . . . , n} define VI as the intersection of kernels of the restriction
maps defined by assigning to the i-th coordinate the value xi :
Y
M Y

VI = Ker[F Xi −→ F ( X i ) × xj .
i∈I j∈I i∈I, i6=j

Write Vn for V{1,...,n} .

2
Lemma. We have F( ni=1 Xi ) = Un ⊕ Vn .
Q

Proof. Let us prove by induction on n that

n
Y M
F( Xi ) = F(k) ⊕ VI .
i=1 ∅(I⊆{1,...,n}

For n = 1 the restriction map F(X) → F(x) = F(k) has a section, so we have

F(X) = F(k) ⊕ Ker[F(X) −→ F(x)] = U1 ⊕ V1 .

Now assume the statement is proved for n − 1. The group Vn is the kernel of the
restriction map
n n j−1 n
Y M Y Y
F( Xi ) −→ F( Xi × xj × Xi ).
i=1 j=1 i=1 i=j+1

Using the inductive assumption one sees that this map factors through
n
Y M
F( Xi ) −→ F(k) ⊕ VI .
i=1 ∅(I({1,...,n}

The last map has a natural section whose image is contained in Un , so this proves
our statement and the lemma. 
The functor F has order n − 1 if Vn = 0. This definition makes sense since
although Vn depends on the choice of points xi , Un doesn’t. In this language the
theorem of the cube says that the functor Pic from proper varieties to abelian groups
is quadratic, that is, has order 2. In concrete terms this means that any line bundle
on X × Y × Z is a product of line bundles pulled back from X × Y , Y × Z and
X × Z.

Corollary 1.5 (i) Consider the following morphisms A × A × A → A: the sum of

all three coordinates s, the sum of two of the coordinates sij and the projections pi .
For any L ∈ P ic(A) we have

s∗ L ⊗ s∗12 L−1 ⊗ s∗23 L−1 ⊗ s∗13 L−1 ⊗ p∗1 L ⊗ p∗2 L ⊗ p∗3 L = 0 ∈ Pic(A × A × A).
2 2
(ii) [n]∗ L = L(n +n)/2 ⊗ [−1]∗ L(n −n)/2 .
(iii) (“Theorem of the square”) For x ∈ A let us denote by Tx the translation by
∗
x, that is, Tx (y) = x + y. Then for any x, y ∈ A we have Tx+y L ⊗ L = Tx∗ L ⊗ Ty∗ L.

Proof. Let us choose 0 as the base point in A. Then (i) immediately follows from
the theorem of the cube.

3
 (ii) Consider the morphism A → A × A × A given by x 7→ (x, x, −x). The pull-
back of the formula from (i) gives L3 ⊗ [−1]∗ L = [2]∗ L which is (ii) for n = 2. Now
consider the morphism A → A × A × A given by x 7→ (2x, x, −x). Then the pull-
back of the formula from (i) gives [2]∗ L2 ⊗ [−1]∗ L = [3]∗ L, which is (ii) for [3]∗ L.
Continue by induction.
(iii) Consider the morphism A → A × A × A given by z 7→ (x, y, z). Then the
pull-back of the formula from (i) gives (iii). 
To each line bundle L on A we attach a map from A to Pic(A) as follows:

ϕL (x) = Tx∗ L ⊗ L−1 ∈ Pic(A).

Then Corollary 1.5 says that ϕL is a homomorphism. Define

K(L) = Ker(ϕL ) ⊂ A.

Exercises Show that ϕL⊗M (x) = ϕL (x) + ϕM (x) and ϕTy∗ L (x) = ϕL (x).

1.3 Quotients and isogenies

We quote another result from algebraic geometry.

Theorem 1.6 (Seesaw principle) Let X be a proper variety and let L be a line
bundle over X × Y .
(i) The set of points y ∈ Y such that L restricts trivially to X × y is close in Y .
(ii) L = p∗2 M for a line bundle M on Y , where p2 : X × Y → Y is the second
projection, if and only if L restricts to a trivial line bundle over X ×y for any y ∈ Y .
(iii) L = 0 if and only if L restricts to a trivial line bundle over X × y for all
y ∈ Y and L restricts to a trivial line bundle over x×Y for at least one point x ∈ X.

It is clear that (iii) is an immediate consequence of (ii).

Corollary 1.7 K(L) is closed in A.

Proof. Recall that L is a line bundle on an abelian variety A. Consider the line
bundle m∗ L ⊗ p∗2 L−1 on A × A. It restricts trivialy to x × A if and only if x ∈ K(L).
So the statement follows from Theorem 1.6 (i). 

Corollary 1.8 Let B ⊂ A be the neutral connected component of K(L). Then the
line bundle m∗ L⊗p∗1 L−1 ⊗p∗2 L−1 on A×A restricts to a trivial line bundle on B ×B.

Proof. Use Theorem 1.6 (iii) for x = 0. 

4
Proposition 1.9 If L is ample, then K(L) is finite.

Proof. By Corollary 1.8 the line bundle m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1 is trivial on B × B.
Pulling it back via the morphism B → B × B given by x 7→ (x, −x) we see that
L ⊗ [−1]∗ L is trivial on B. However, both L and [−1]∗ L are ample, so this is a
contradiction unless B = 0. 
Remark When L is the line bundle associated with an effective divisor, then
|K(L)| < ∞ implies that L is ample, see [3, Ch. II, §6, Prop. 1]. One also
shows that if U ( A is an open affine subset, then the sum D of the irreducible
components of the closed set A \ U is an ample divisor (ibid.) In particular, ample
divisors on A exist. Thus every abelian variety is projective (and not just proper).

Proposition 1.10 For each n 6= 0 the multiplication by n morphism [n] : A → A is

surjective, so that the group of points A is divisible. The kernel A[n] of [n] is finite.

Proof. We can clearly assume that n > 1. Since A is irreducible, the surjec-
tivity of [n] is equivalent to the condition that dim([n]A) = dim(A). Since [n]
is a homomorphism, the fibres are cosets and so have the same dimension. Since
dim([n]A)+dim([n]−1 (0)) = dim(A) the two statements of the proposition are equiv-
2 2
alent. Take any ample line bundle L. Then L(n +n)/2 and [−1]∗ L(n −n)/2 are both
ample, [n]∗ L is ample by Corollary 1.5 (ii). But A[n] ⊂ K([n]∗ L), we conclude that
both these sets are finite by Proposition 1.9. 

Definition 1.11 A surjective homomorphism of abelian varieties f : A → B is

called an isogeny if its kernel is finite.

The degree deg(f ) of an isogeny f : A → B is defined as the degree of the finite

extension of function fields [k(A) : k(B)]. If E is the separable closure of k(B) in
k(A), then
degsep (f ) = [E : k(B)]
is the separable degree of f . The separable degree of f is the number of points in
each fibre of f . One says that f is separable if deg(f ) = degsep (f ). Next,

deginsep (f ) = [k(A) : E]

is the inseparable degree of f . The inseparable degree is always 1 if char(k) = 0. If

char(k) = p > 0, then deginsep (f ) = pm for some m ≥ 0.
The degree of f can be calculated in terms of the induced action of f on divisors:
if g = dim(A) = dim(B) and D is a divisor on B such that the self-intersection
index (D)gB > 0, then
(f ∗ D)gA = deg(f )(D)gB .

5
Proposition 1.12 The homomorphism [n] : A → A is an isogeny of degree n2g . If
n is prime to the characteristic of k, then [n] is separable. In this case the abelian
group A[n] is isomorphic to (Z/n)2g .

Proof. Take an ample divisor D on A. By passing to D + [−1]D we can assume

that D is symmetric, i.e. invariant under the antipodal involution [−1]. Then [n]∗ D
is linearly equivalent to n2 D. Hence deg([n]) = n2g . If n is prime to char(k), this
implies that p does not divide deg([n]), so [n] is separable. For r dividing n we have
A[r] ⊂ A[n]. Using this and the fact that deg([r]) = r2g for any such r we deduce
that the abelian group A[n] is isomorphic to (Z/n)2g . 
In particular, if ` is a prime not equal to char(k), then the projective limit

T` (A) = lim A[`n ]

←−
taken for n → ∞ with respect to the natural surjective maps A[`n ] → A[`m ], n ≥ m,
is isomorphic to Z2g
` , where Z` is the ring of `-adic integers. This abelian group is
called the `-adic Tate module of A.

Theorem 1.13 There is a bijection between finite subgroups of an abelian variety

A and separable isogenies f : A → B (considered up to an isomorphism of B).

Sketch of proof. The proof is based on a result from algebraic geometry which
states the existence and the uniqueness of the quotient variety by the action of a
finite group of automorphisms. More precisely, let X be a quasi-projective variety
and let G be a finite group of automorphisms of X. Then X has a covering by
G-invariant open affine subsets Spec(Ri ), where Ri is a k-algebra. There exists a
variety Y covered by open affine subsets Spec(RiG ) and a morphism π : X → Y
whose restriction to each Spec(Ri ) is the natural morphism Spec(Ri ) → Spec(RiG ).
The fibres of π are orbits of G. One also proves that the morphism π is finite (this is
essentially the statement that each Ri is a finitely generated RiG -module), separable
and surjective morphism. The finite extension k(X)/k(Y ) is a Galois extension with
the Galois group G.
If G acts freely on X, in the sense that the stabiliser of each point is trivial, then
π is étale (=flat and unramified).
Now if X is a commutative group and G ⊂ X is a finite subgroup, then the
set Y = X/G inherits the group structure from X. The composition law and the
operation of taking the inverse are morphisms. Indeed, the composition of Y comes
from the morphism X × X → X → Y which sends G × G to 0 and hence descends
to a morphism Y × Y → Y , and similarly for the inverse. The image of a proper
variety is again proper, so any quotient of an abelian variety by a finite subgroup G
is again an abelian variety. It is clear that the kernel of the resulting isogeny is G.

6
 If f : A → B is a separable isogeny, then, by the universal property of the
quotient, there is a morphism A/Ker(f ) → B which is separable and bijective on
points. It follows that it induces an isomorphism of function fields. Then it must
be an isomorphism. 

1.4 Dual abelian variety

We continue to explore the properties of the map ϕL : A → Pic(A) which sends
x ∈ A to Tx∗ L ⊗ L−1 ∈ Pic(A). Writing the group law of Pic(A) additively we have
ϕL⊗M (x) = ϕL (x) + ϕM (x).

Definition 1.14 Let Pic0 (A) be the subgroup of Pic(A) such that L ∈ Pic0 (A) if
and only if ϕL is the zero homomorphism.

Corollary 1.5 (iii) gives

Ty∗ (Tx∗ L ⊗ L−1 ) = Tx+y

∗
L ⊗ Ty∗ L−1 ∼
= Tx∗ L ⊗ L−1 ,

which says that ϕL (x) ∈ Pic0 (A). In other words, ϕL is actually a map A → Pic0 (A).
Thus we have an exact sequence of abelian groups

0 → Pic0 (A) → Pic(A) → Hom(A, Pic0 (A)), (1)

where the third map sends L to ϕL .

An equivalent definition is:

Proposition 1.15 L ∈ Pic0 (A) if and only if the line bundle m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1
on A × A is trivial.

Proof. Indeed, by the seesaw principle this happens if and only if the restrictions to
A × x are trivial for all x ∈ A (these are precisely Tx∗ L ⊗ L−1 ) and the restriction to
0 × A is trivial too (this is obviously true). 
Extensions of commutative algebraic groups. For commutative algebraic
groups A, B over k, the abelian group Ext1 (A, B) is defined as the set of equiv-
alence classes of extensions of A by B:

0 → B →? → A → 0,

where the arrows are morphisms of algebraic groups, i.e. homomorphisms of the
groups of k-points which are also morphisms of algebraic varieties. Two extensions
are equivalent when they are linked by the maps that are identities on A and B.
Pull-back and push-forward of exact sequences show that Ext1 (A, B) is covariant in
B and contravariant in A. Addition in Ext1 (A, B) is defined by pulling back the sum

7
of two extensions in Ext1 (A × A, B × B) via the diagonal A → A × A and pushing
it forward via the composition B × B → B. The neutral element in Ext1 (A, B) is
the direct product A × B. The key fact is that Ext1 (A, B) is an additive bi-functor
from the category of commutative algebraic groups over k to the category of abelian
groups, see [7, Ch. VII, §1]. Let us point out that this construction of Ext does
not use any cohomology theory. One directly constructs 6-term exact sequences
involving Hom and Ext associated to an extension. One also defines Extn (A, B) for
n ≥ 2 using n-fold extensions, using the same notion of equivalent extensions.
Picard group. Line bundles and torsors on a given variety are related in a canonical
way: removing the zero section from a line bundle on X produces an étale X-torsor
whose structure group is the multiplicative group Gm . More precisely, the Picard
group Pic(X) is the group of invertible coherent sheaves of OX -modules, so that
∗
Pic(X) = H1Zar (X, OX ). Let π : Xet → XZar be the continuous morphism of sites
induced by the identity on X. We have (R1 π∗ )(Gm ) = 0 (Grothendieck’s version of
Hilbert’s theorem 90, see [1]), and the Leray spectral sequence entails a canonical
isomorphism
Pic(X) = H1Zar (X, Gm,X )−→H
˜ 1et (X, Gm,X ).
Alternatively, to an invertible sheaf L one directly associates a torsor T for Gm,X
defined by T (U ) = IsomU (OU , f ∗ L), where f : U → X is étale. This gives an
equivalence of the category of invertible sheaves of OX -modules and the category of
étale X-torsors for Gm,X , see [Arcata], Prop. II.2.3.
Now we go back to the case when our variety is an abelian variety A. A particular
kind of A-torsors for Gm is given by commutative group extensions of A by Gm :

0 → Gm →? → A → 0.

In these terms we can state another equivalent definition:

Proposition 1.16 L ∈ Pic0 (A) if and only if the A-torsor L \ 0 for Gm has a
group structure of a commutative extension of A by Gm . Associating an A-torsor
for Gm to an extension gives an injective map Ext1 (A, Gm ) → Pic(A) whose image
is Pic0 (A).

Proof. Let us show that a torsor coming from an extension of A by Gm defines an

element of Pic0 (X). Let L be the line bundle on A such that this torsor is L \ 0. By
Proposition 1.15 we need to show that the line bundle m∗ L on A × A is isomorphic
to p∗1 L ⊗ p∗2 L. The pull-back m∗ defines a homomorphism

Ext1 (A, Gm ) −→ Ext1 (A × A, Gm ) = Ext1 (A, Gm ) ⊕ Ext1 (A, Gm ),

where the equality is due to the fact that Ext1 (A, Gm ) is an additive functor in
the first argument. Thus the line bundle m∗ L on A × A comes from an element of

8
Ext1 (A × A, Gm ) which is the composition of two extensions of A by Gm . Hence
m∗ L = p∗1 L1 ⊗ p∗2 L2 for some line bundles L1 and L2 on A. By restricting to A × 0
and 0 × A we see that L1 = L2 = L. Thus L ∈ Pic0 (A).
This gives a map Ext1 (A, Gm ) → Pic0 (A). Let us prove that this map is injective.
If the torsor defined by an extension E is trivial, it has a section σ : A → E, which
is a priori only a morphism of varieties and not necessarily a homomorphism. We
need to show that the extension E is trivial, i.e., there is an isomorphism of algebraic
groups E ∼ = A × Gm . By modifying σ by an element of Gm (k) we can assume that
σ(A) contain the origin of the group law of E. Since σ(A) is a proper subvariety of
E, the subgroup S ⊂ E generated by σ(A) is also proper. Hence E ∩ Gm must be
finite. This gives a finite morphism S → A; in particular, dim(S) = dim(A). But A
is irreducible, so σ(A) is an irreducible component of S. An irreducible component
of an algebraic group which contains the origin of the group law is a subgroup, hence
S = σ(A), so σ is in fact a homomorphism. This gives an isomorphism of algebraic
groups E ∼ = A × Gm .
To show that the map Ext1 (A, Gm ) → Pic0 (A) is surjective one uses the triviality
of m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1 on A × A to define a commutative group structure on L \ 0,
see [7, Ch. VII, §15]. 

Corollary 1.17 There is a canonical isomorphism of abelian groups

Pic0 (A) = Ext1 (A, Gm ), (2)

called the Barsotti–Weil formula.

Lemma 1.18 Let S be a variety and let M be a line bundle on A × S. For s ∈ S

write Ms for the restriction M |A×s . Then Ms1 ⊗ Ms−1
2
∈ Pic0 (A) for any s0 , s1 ∈ S.

Proof. We cover S by open sets, so that it is enough to prove the statement for
an open set that can be taken small enough. In particular, we can assume that
M |0×S = 0. By twisting M by a line bundle pulled back from A we can assume
that M |A×s0 = 0. To prove that Ms ∈ Pic0 (A) it is enough to show that the line
bundle m∗ Ms ⊗ p∗1 Ms−1 ⊗ p∗2 Ms−1 on A × A is trivial. This bundle is the restriction
to A × A = A × A × s of the obvious line bundle on A × A × S. But this is a trivial
line bundle by the theorem of the cube. 
The meaning of this lemma is that the continuous deformations of the trivial line
bundle are in Pic0 (A). Divisors D and D0 on a variety X are algebraically equivalent,
if there is a family of divisors parametrised by a (connected) variety of which D and
D0 are members. In this language Pic0 (A) is the group of divisor classes algebraically
equivalent to 0.

Proposition 1.19 For any ample line bundle L on A we have ϕL (A) = Pic0 (A).

9
 See [3, Ch. II, §8, Thm. 1] for a proof.
This gives an exact sequence of abelian groups
ϕ
L
0 −→ K(L) −→ A −→ Pic0 (A) −→ 0.

In characteristic zero we can use Theorem 1.13 to identify Pic0 (A) with the group
of k-points of an abelian variety, which is well defined up to isomorphism. Then ϕL
becomes an isogeny.

Definition 1.20 The dual abelian variety At is the abelian variety whose group of
k-points is Pic0 (A).

This result can be made more precise in several ways.

First of all, one shows that the line bundle m∗ L⊗p∗1 L−1 ⊗p∗2 L−1 on A×A descends
to A × At . In other words, it is the pull-back of a line bundle P on A × At :

m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1 = (id, ϕL )∗ P. (3)

To construct P one extends the action of K(L) by translations on A to the action

on A × A which is trivial on the first component. Translations by the points of
K(L) preserve L, hence this action of K(L) on A preserves m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1 ,
and thus extends to an action of K(L) on this line bundle. The restriction of
m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1 to 0 × A is the trivial bundle canonically isomorphic to the
fibre of L−1 at 0, ie., this is the direct product L−1 (0) × A. The above action of
K(L) on A × A extends to a well defined action of K(L) on m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1
normalised so that the action on the restriction of m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1 to 0 × A
is the action on L−1 (0) × A which is trivial on L−1 (0) and is the usual action by
translations on A. Passing to the quotient by this action we see that P |0×At = 0.
The restriction P |A×ϕL (x) is the restriction of m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1 to A × x, and
this is Tx∗ L⊗L−1 . This shows that for y ∈ At the restriction P |A×y is the line bundle
on A given by y ∈ At = Pic0 (A).

Definition 1.21 The Poincaré line bundle on A × At is a line bundle P such that
P |0×At = 0 and P |A×y is the line bundle on A given by y ∈ At = Pic0 (A). One also
requires P to satisfy the following universal property. For any variety S and any
line bundle M on A × S such that M |0×S = 0 and M |A×s ∈ Pic0 (A) for all s ∈ S
the natural map of sets f : S → At defined by M |A×s = P |A×f (s) is a morphism
of varieties and M = (1, f )∗ P is the pull-back of P with respect to the morphism
(1, f ) : A × S → A × At .

With this definition the pair (At , P ) is unique up to canonical isomorphism.

Secondly, one can adapt the definition of (At , P ) so that it works in arbitrary
characteristic. To a line bundle L on an abelian variety A one associates a group

10
subscheme K(L) ⊂ A defined as the maximal subscheme of A such that the re-
striction of m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1 to K(L) × A ⊂ A × A is trivial (see [3, Ch. III,
§10, §13]). In terms of the functor of points it is described as follows. Let S be a
scheme and write AS = A × S. An S-point of A is a morphism f : S → A. This
point is in K(L) if and only if Tf∗ L ∼= L ⊗ p∗2 M for some line bundle M on S. Here
Tf : AS → AS is the morphism Tf (x, s) = (x + f (s), s). The key point is that K(L)
need not be reduced.
Then one develops the procedure of forming the quotient by the action of a finite
group scheme and defines At = A/K(L) for an ample line bundle L. The Poincaré
bundle is defined as above and has similar properties.
Let us give an example of application of the universal property of (At , P ). By
applying Lemma 1.18 to the line bundle P on A × At we see that the line bundles
in Pic0 (A) are precisely the continuous deformations of the trivial line bundle on A.
Dual morphism. Let f : A → B be a morphism of abelian varieties. We have the
Poincaré bundles PA on A × At and PB on B × B t . Consider the pull-back (f, id)∗ PB
along (f, id) : A × B t → B × B t . This is a line bundle on A × B t whose restrictions
to A × x are in Pic0 (A) and whose restriction to 0 × At is trivial. We deduce the
existence of a morphism f t : B t → At such that (f, id)∗ PB = (id, f t )∗ PA :
(id,f t ) (f,id)
A × At ←− A × B t −→ B × B t .

The map f t : B t → At is called the dual morphism of f . The meaning of this

is that f t sends an element of B t = Pic0 (B) corresponding to a line bundle L (a
continuous deformation of the trivial bundle) on B to the element of At = Pic0 (A)
corresponding to the line bundle f ∗ L on A.
Duality of finite group schemes. Let G be a finite commutative group k-scheme.
The dual group scheme of G is definied as a group k-scheme GD that represents
the functor from the category of schemes to the category of abelian groups which
associates to a scheme S the group of homomorphisms of commutative group S-
schemes GS → Gm,S . There is a nice description of GD in terms of its algebra of
regular functions k[GD ]. The multiplication, the inverse and the neutral elements
are morphisms
G × G → G, G → G, Spec(k) → G.
They give rise to k-algebra homomorphisms

k[G] → k[G] ⊗k k[G], k[G] → k[G], k[G] → k.

One checks that the fact that G is a group implies that passing to the space of linear
functions k[G]∗ = Hom(k[G], k) we obtain a map

k[G]∗ ⊗k k[G]∗ → k[G]∗

11
which makes k[G]∗ an associative k-algebra with the unit coming from the dual map
k → k[G]∗ . If G is commutative, then this algebra is commutative, so we can define
GD = Spec(k[G]∗ ) so that k[GD ] = k[G]∗ = Hom(k[G], k).
Note that if G is reduced, then GD has the same cardinality as G. In this case the
k-vector space of regular functions on G is the same as the space of maps G → k
which is the dual to the k-vector space freely generated by the elements of G, that
is, the group algebra of G. Therefore, the k-algebra k[G]∗ is canonically isomorphic
to the group algebra of the group of k-points of G: the linear function given by the
value of a regular function on G at g ∈ G is identified with the canonical generator
g of the group ring of G. For example, if G = Z/pn is a reduced group scheme, then
the space of linear functions k[G] → k is the group algebra of G = Z/pn which is
n
k[t]/(tp − 1). (The action of a generator of Z/pn is encoded by the multiplication
n
by t.) Thus (Z/pn )D = µpn = Spec(k[t]/(tp − 1)), and hence (µpn )D = Z/pn . Note
that if p =char(k), then µpn is not reduced.
n
Exercise Assume that p =char(k) and define αpn = Spec(k[t]/(tp )). This is a non-
reduced subscheme of the additive group k-scheme Ga , whose reduced subscheme is
one point 0. One can check that αpn is a group subscheme of Ga , see [3, §11]. Show
that (αpn )D = αpn .

Proposition 1.22 Let f : A → B be an isogeny. Then the dual morphism f t

is an isogeny B t → At of the same degree deg(f ), called the dual isogeny of f .
Moreover, the finite groups schemes Ker(f ) and Ker(f t ) are dual to each other.

Proof. The exact sequence of group schemes

f
0 −→ K −→ A −→ B −→ 0

gives rise to the exact sequence of abelian groups

Hom(A, Gm ) −→ Hom(K, Gm ) −→ Ext1 (B, Gm ) −→ Ext1 (A, Gm ). (4)

Here the map Ext1 (B, Gm ) → Ext1 (A, Gm ) is given by the pull-back of extensions
along f : A → B so this is also the dual morphism f t : B t → At , because both
morphisms coincide on k-points: they send L to f ∗ L. We have Hom(A, Gm ) = 0
since A is proper and Gm is affine. The sequence (4) holds also after the base change
to any base scheme S. This shows that there is an exact sequence of commutative
group k-schemes
ft
0 −→ K D −→ B t −→ At , (5)
where we have used the Barsotti–Weil formula (Corollary 1.17) and the definition
of K D . We see that Ker(f t ) is finite, so f t is surjective because dim(At ) = dim(B t ).


12
Corollary 1.23 The dual isogeny of [n] : A → A is [n] : At → At . Thus the finite
group schemes A[n] and At [n] are dual to each other.

Proof. The map Ext1 (A, Gm ) → Ext1 (A, Gm ) induced by the multiplication by n
map on A is [n]. Thus we see from (4) that in our case the exact sequence (5) takes
the form
[n]
0 −→ A[n]D −→ At −→ At −→ 0.
Hence A[n]D is canonically isomorphic to At [n]. 
We obtain a perfect pairing A[n] × At [n] → µn , called the Weil pairing.

t t
Proposition 1.24 There is a canonical isomorphism A−→(A
˜ ).

Proof. The Poincaré bundles PA on A × At and PAt on At × (At )t = (At )t × At give

rise to a unique morphism c : A → (At )t such that PA = (c, id)∗ PAt . Choose a line
bundle L such that ϕL : A → At is an isogeny. One checks from the definition of
the dual morphism that ϕtL c = ϕL . It follows that c is an isogeny and in particular
is surjective. Now applying the snake lemma to the commutative diagram
ϕ
L
0 −→ K(L) −→ A −→ At −→ 0
↓ ↓ ||
ϕt
0 −→ K(L)D −→ (At )t −→
L
At −→ 0

we see that K(L) maps surjectively onto K(L)D . However, by Proposition 1.22
these finite group schemes have the same order, so the map K(L) → K(L)D is an
isomorphism. Thus c : A → (At )t is also an isomorphism. 
We always identify A with (At )t via the isomorphism c.
Warning. In view of the canonical identification of (At )t with A we have in fact
not one but two Weil pairings:

(, )A : A[n] × At [n] → µn , (, )At : At [n] × A[n] → µn .

It seems plausible that they should agree, however this is not quite true because in
fact they differ by sign:
(x, y)A = −(y, x)At ,
see, e.g. [4, §10.4].
The proof of Proposition 1.24 shows that if ϕL : A → At is an isogeny, then
ϕtL : A → At is equal to ϕL . Now we show that this holds for any line bundle L.

Proposition 1.25 Let L be any line bundle on A. Then the morphism ϕL : A → At

is self-dual.

13
Proof. By Proposition 1.24 the dual morphism ϕtL ∈ Hom(A, At ). Thus we have
morphisms
(ϕt ,id) (id,ϕL )
At × A ←−
L
A × A −→ A × At .
By (3) and the definition of the dual morphism we have canonical isomorphisms of
line bundles on A × A:

(ϕtL , id)∗ PAt = m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1 = (id, ϕL )∗ PA .

t t
The canonical isomorphism A−→(A ˜ ) of Proposition 1.24 identifies PA and PAt
when we swap the factors in A × A . Since m∗ L ⊗ p∗1 L−1 ⊗ p∗2 L−1 is invariant under
t

swapping the factors in A×A, we see that everything is invariant under the swapping
the factors, hence ϕtL = ϕL . 
Write Homself−dual (A, At ) for the subgroup of self-dual morphisms A → At . Then
(1) gives rise to an exact sequence of abelian groups

0 → At → Pic(A) → Homself−dual (A, At ) → 0. (6)

See [4, Thm. 13.7] for the proof of the surjectivity of the third arrow in this sequence.
The Néron–Severi group of a variety X over an algebraically closed field is defined as
the quotient of Pic(X) by the subgroup of classes of divisors agebraically equivalent
to zero. We see that the Néron–Severi group of an abelian variety A has a nice
interpretation
NS (A) = Homself−dual (A, At ).

Elliptic curves. An elliptic curve is an abelian variety of dimension 1. Let O ∈ E

be the origin of the group law on an elliptic curve E. The line bundle L = O(O) is
ample and the associated morphism ϕL sends x ∈ E to Tx∗ L ⊗ L−1 = O([−x] − [O]).
This map is injective, hence E = E t . Thus the Weil pairing is a non-degenerate
bilinear pairing
(, ) : E[n] × E[n] → µn .
From the Warning above we know that it is skew-symmetric: (x, y) = −(y, x) for
all x, y ∈ E[n]. Moreover, it is known that the Weil pairing is alternating, that is,
(x, x) = 0 for any x ∈ E[n]. (The second property is stronger when n is even!)

2 Abelian varieties over non-closed fields

2.1 Galois cohomology
Let G be a group. For a G-module M we have cohomology groups Hn (G, M ) which
are derived functors of the functor from the category of G-modules to abelian groups
given by M 7→ M G .

14
Exercise 1. If G is a cyclic group of order n with generator g, then H2n+1 (G, M ),
where n = 0, 1, 2, . . ., is the quotient of the kernel of the norm map N : M → M
Pn−1 i
by (g − id)M , where N = i=0 g . Similarly, H2n (G, M ) is the quotient of M G by
NM , where n = 1, 2, . . . (Check that . . . → Z[G] → Z[G] → . . ., where the arrows
are alternating N and g − 1, is a projective resolution of the G-module Z.)
We also have ExtnG (M, N ) which can be thought of either the right derived func-
tors of N 7→ HomG (M, N ) or the left derived functors of M 7→ HomG (M, N ).
Equivalently, ExtnG (M, N ) can be defined as the set of equivalence classes of n-fold
extensions of G-modules M by N . These groups can be computed using the spectral
sequence
p+q
Hp (G, ExtqZ (M, N )) ⇒ ExtG (M, N ).
From the definition in terms of extensions we get pairings

ExtpG (L, M ) × ExtqG (M, N ) → Extp+q

G (L, N )

defined by splicing extensions. This means composing the surjective (penultimate)

map in the first extension with the injective (second) map in the second extension.
From the right derived functor definition we see that ExtnG (Z, N ) = Hn (G, N ). In
particular, we obtain pairings

Hp (G, M ) × ExtqG (M, N ) → Hp+q (G, N ). (7)

Cohomology of profinite groups Let G be a profinite group, i.e. the inverse

limit lim Gi of a projective system of finite groups Gi , equipped with the topology
←−
for which the kernels of projections G → Gi form a basis of open sets. A G-module
M is a continuous discrete G-module if the action G × M → M is continuous when
M is given discrete topology. This means that the stabiliser of each element of M
is an open subgroup of G, or, equivalently, M = ∪U M U where U ⊂ G are open
subgroups. The cohomology groups Hn (G, M ), defined as the derived functors of
M → M G , can be calculated as a direct limit of cohomologies of finite groups:

Hn (G, M ) = lim Hn (G/U, M U ),

−→
where U ⊂ G ranges over all open normal subgroups of G. If U ⊂ U 0 , then
0
Hn (G/U 0 , M U ) → Hn (G/U, M U ) is an inflation homomorphism. Since G/U is fi-
nite, we see that for n ≥ 1 every element of Hn (G, M ) has finite order, that is,
Hn (G, M ) is a torsion group.
Exercise 2. Prove that if M is finite, then Hn (Ẑ, M ) = 0 for n ≥ 2. (To fix ideas let
n > 0 be even. We have Ẑ/mẐ = Z/m and Hn (Z/m, M mẐ ) = M Ẑ /Nm M mẐ , where
Nm = m−1 i
P
i=0 g and g is a topological generator of Ẑ. The inflation homomorphism

Hn (Z/m, M mẐ ) −→ Hn (Z/mr, M mrẐ )

15
is the map
M Ẑ /Nm M mẐ −→ M Ẑ /Nmr M mrẐ
given by Nr which induces the multiplication by r. So this map is zero if r is divisible
by |M |, so the direct limit is the zero group.)
Cohomology of commutative algebraic groups Let k be a perfect field with
an algebraic closure k̄ and the Galois group Γ = Gal(k̄/k). If G is a commutative
group scheme over a field k, then we write Hn (k, G) for the continuous cohomology
group Hn (Γ, G(k̄)), where the Galois group Γ has its natural profinite topology and
G(k̄) is given discrete topology. In particular, H0 (k, G) = G(k). The group H1 (k, G)
classifies k-torsors for G up to isomorphism.
Question: When does a self-dual isogeny ϕ : A → At of abelian varieties over k
come from a line bundle on A, i.e., when ϕ = ϕL for some L ∈ Pic(A)? (Variant: if
such a line bundle exists, can we choose it to be symmetric with respect to the action
of [−1]?) The first question can be restated as the question about the differential
∂(ϕ) attached to (6). Over some easy non-closed fields it has a positive answer. For
example, Lang’s theorem says that over any finite field k we have H1 (k, A) = 0.
The second question is a question about the differential attached to the subse-
quence of (6) obtained by taking invariants with respect to the induced action of
[−1] on A:
∗ t
0 → At [2] → Pic(A)[−1] → Homself−dual (A, A ) → 0. (8)
This sequence is exact because H1 (Z/2, At ) = 0. Let us denote by cϕ the image of
ϕ under the differential
t
Homself−dual (A, A )Γ −→ H1 (k, At [2]).

Poonen and Stoll showed that when k is a number field, the class cϕ plays an
important role in deciding whether the order of the n-torsion subgroup of X(A) is
a square or twice a square.
Let us define Ext-groups in the category of commutative group schemes over k.
The group Extnk (A, B) is the group of equivalence classes of n-fold extensions, as
was explained above in the case of an algebraically closed field. In that case we
obtained Extn (A, B) = Extnk̄ (A, B); this group also has the structure of a Γ-module.
The pairings
Hp (k, A) × Extqk (A, B) → Hp+q (k, B) (9)
are defined by forgetting the algebraic group structure, that is, by composing the
forgetful map Extqk (A, B) → ExtqΓ (A(k̄), B(k̄)) with the pairing between Ext-groups
in the category of discrete Γ-modules (7).

Proposition 2.1 Let k be a perfect field.

16
 (i) (J.S. Milne) There is a spectral sequence

Hp (k, Extq (A, B)) ⇒ Extp+q

k (A, B).

(ii) (F. Oort) If A is an abelian variety over k, then Extn (A, Gm ) = 0 for n ≥ 2.
(iii) (F. Oort) If N is a finite group scheme of order coprime to the characteristic
of k, then Extn (N, Gm ) = 0 for n ≥ 1.

Corollary 2.2 Let k be a perfect field and let n ≥ 0. For any abelian variety A we
have Extn+1
k (A, Gm ) = Hn (k, At ). For any finite group k-scheme N of order coprime
to the characteristic of k we have Extnk (N, Gm ) = Hn (k, N D ).

Proof This is immediate using Milne’s spectral sequence, Hom(A, Gm ) = 0 and the
Barsotti–Weil formula (Corollary 1.17). 
Thus we obtain pairings

Hn (k, A) × Ext2−n
k (A, Gm ) = Hn (k, A) × H1−n (k, At ) → H2 (k, Gm ) = Br(k), (10)

where n = 0 or 1, and

Hn (k, N ) × Ext2−n
k (N, Gm ) = Hn (k, N ) × H2−n (k, N D ) → H2 (k, Gm ) = Br(k), (11)

where n = 0, 1, 2. For example, for N = N D = Z/2 we are not getting anything

interesting when n = 0 or n = 2, but for n = 1 we get a pairing

k ∗ /k ∗2 × k ∗ /k ∗2 −→ Br(k)[2],

given by the symbol (a, b) ∈ Br(k) which is the class of the quaternion algebra
k ⊕ ik ⊕ jk ⊕ ijk, where i2 = a, j 2 = b and ij = −ji.

2.2 Local fields

As a warming-up for the duality over local fields let us first consider the case when k
is a finite field. Then Γ ∼
= Ẑ is the completion of the infinite cyclic group generated
by the Frobenius. If M is a Ẑ-module with a topological generator g, then

H0 (Ẑ, M ) = M Ẑ , H1 (Ẑ, M ) = M/(g − id) = MẐ

are the groups of invariants and co-invariants, respectively. Note that there is an
exact sequence
g−id
0 → M Ẑ → M −→ M → MẐ → 0,
hence if M is finite, then |M Ẑ | = |MẐ |.

17
 Assume that M is finite and define the dual module as M ∗ = Hom(M, Z/m),
where m = |M |. Then

H0 (Ẑ, M ∗ ) = HomẐ (M, Z/m) = Hom(MẐ , Z/m),

so there is a perfect duality of finite abelian groups

MẐ × (M ∗ )Ẑ −→ Z/m.

This can be thought of as a duality

Hr (Ẑ, M ) × H1−r (Ẑ, M ∗ ) −→ H1 (Ẑ, Z/m) = Z/m,

where r = 0 or 1. In Exercise 2 above we have seen that if M is finite, then

Hr (Ẑ, M ) = 0 for r ≥ 2.

Definition 2.3 We call a field k local if k is the field of fractions of a complete

discrete valuation ring with the finite residue field.

Let knr be the maximal unramified extension of k. Let Ok be the ring of integers
of k, let m be the maximal ideal of R and let κ = Ok /m be the residue field. The
Galois group Gal(knr /k) is canonically isomorphic to Gal(κ̄/κ) ∼
= Ẑ.
Recall the cohomological interpretation of the Brauer group of the field k:

Br(k) = H2 (k, Gm ) = H2 (Γ, k̄ ∗ ).

Lang’s theorem says that Br(knr ) = 0. This, Hilbert’s theorem 90 H1 (knr , k̄ ∗ ) = 0,

and the Hochschild–Serre spectral sequence

Hp (Ẑ, Hq (knr , k̄ ∗ )) ⇒ Hp+q (Γ, k̄ ∗ )

imply H2 (Γ, k̄ ∗ ) = H2 (Ẑ, knr

∗
). We obtain

∗ val
Br(k) = H2 (Ẑ, knr ) −→ H2 (Ẑ, Z)←−H
˜ 1 (Ẑ, Q/Z) = Hom(Ẑ, Q/Z) = Q/Z,

where all arrows are isomorphisms, see [6, Ch. XII]. This gives the local invariant
isomorphism
inv : Br(k)−→Q/Z.
˜

Theorem 2.4 (local Tate duality with finite coefficients) Let M be a finite
discrete Γ-module whose order is coprime to the characteristic of k. The paring

Hr (Γ, M ) × H2−r (Γ, M D ) −→ Br(k)−→Q/Z

˜

is a perfect duality of finite groups. We have Hr (Γ, M ) = 0 for r ≥ 3.

18
 The proof is based on local class field theory, more precisely, on the duality theo-
rem relative to a class formation. See [2, §I.2].
Example Take k = Qp , where p is a prime, and M = M D = Z/2. Then this pairing
is given by the Hilbert symbol (a, b)p ∈ {±1}, where a, b ∈ Q∗p /Q∗2p . The resulting
class in Br(Qp ) is the class of the quaternion algebra Qp ⊕ iQp ⊕ jQp ⊕ ijQp , where
i2 = a, j 2 = b and ij = −ji. Cyclic algebras can be used to prove this theorem in
the case when M = Z/m and M D = µm .
Remark A Γ-module M is called unramified if the inertia subgroup I = Gal(k̄/knr ) ⊂
Γ = Gal(k̄/k) acts trivialy on M , that is, M I = M . In this case we define the
unramified cohomology group Hnnr (Γ, M ) = Hn (Γ/I, M ). The most important of
these is H1nr (Γ, M ) which coincides with the image of the injective inflation map
H1 (Γ/I, M ) → H1 (Γ, M ).
Now assume that M is finite. Since Γ/I ∼ = Ẑ, the order of H1nr (Γ, M ) is equal to
the order of H0nr (Γ, M ) = M Γ . We have seen in Exercise 2 above that Hn (Ẑ, M ) = 0
for n ≥ 2. Let µ̄ be the group of roots of unity of order coprime to p. Then
µ̄I = µ̄, and thus M D = Hom(M, µ̄) is also unramified. We claim that H1nr (Γ, M )
and H1nr (Γ, M D ) are exact annihilators of each other for the local duality pairing

H1 (Γ, M ) × H1 (Γ, M D ) → H2 (Γ, k̄ ∗ ) = Q/Z.

Indeed, these subgroups annihilate each other because their pairing factors through
H2nr (Γ, µ̄) = H2 (Ẑ, µ̄) = 0, which is zero since µ̄ is unramified. The orders of these
subgroups are equal to |H0 (Γ, M )| and |H0 (Γ, M D )|, respectively. By the local dual-
ity with finite coefficients we have |H0 (Γ, M D )| = |H2 (Γ, M )|. Now the calculation
of the Euler–Poincaré characteristic in Proposition 2.5 below shows that

|H0 (Γ, M )| · |H0 (Γ, M D )| = |H1 (Γ, M )|,

so our claim is proved.

Proposition 2.5 For a finite Γ-module M of order m coprime to the characteristic

of k the Euler–Poincaré characteristic of M equals
|H 0 (Γ, M )| · |H 2 (Γ, M )| 1
χ(M ) = 1
= .
|H (Γ, M )| |Ok : mOk |

Proof. We only prove this under the additional assumption (m, p) = 1, where p is the
characteristic of the residue field κ = Ok /m. Let I ⊂ Γ be the inertia subgroup and
let Ip ⊂ I be the maximal pro-p-subgroup. Since (m, p) = 1 we have Hn (Ip , M ) = 0
for n > 0. Since Ip is normal in I, and I/Ip is isomorphic to the product of Z` for
all primes ` 6= p, the Hochschild–Serre spectral sequence

Hm (I/Ip , Hn (Ip , M )) ⇒ Hm+n (I, M )

19
gives similarly to Exercise 2 above that Hn (I, M ) = 0 for n ≥ 2.
Now Γ/I = Gal(knr /k) = Ẑ. We have H0 (Γ, M ) = H0 (Ẑ, M I ). The Hochschild–
Serre spectral sequence

Hm (Ẑ, Hn (I, M )) ⇒ Hm+n (Γ, M )

gives an exact sequence

0 → H1 (Ẑ, M I ) → H1 (Γ, M ) → H0 (Ẑ, H1 (I, M )) → H2 (Ẑ, M I ) = 0,

where the zero in the right hand side is due to Exercise 2. We also obtain H2 (Γ, M ) =
H1 (Ẑ, H1 (I, M )). But if N is a finite Ẑ-module, then the orders of finite groups
H0 (Ẑ, N ) and H1 (Ẑ, N ) are equal. This implies that if m is invertible in Ok , then
χ(M ) = 1. (See [2, Thm. I.2.8] for the proof in the general case.) 

2.3 Duality for abelian varieties over local fields

Recall that an abelian group G has cofinite type if it is a torsion group (every element
has finite order) and G[n] is finite for all n. Here is the main result of this section.

Theorem 2.6 (local Tate duality for abelian varieties) Let k be a local field
of characteristic zero and let A be an abelian variety over k. For n ≥ 2 we have
Hn (k, A) = 0. For n = 0, 1 the canonical pairing

Hn (k, A) × H1−n (k, At ) −→ Q/Z

gives rise to an isomorphism of compact groups

1
A(k)−→Hom(H
˜ (k, At ), Q/Z)

and an isomorphism of discrete groups of cofinite type

H1 (k, A)−→Hom(A
˜ t
(k), Q/Z).

As a preparation for the proof we prove a statement valid over an arbitrary ground
field. Let n be coprime to the characteristic of k. The exact sequence

1 → A[n] → A → A → 1 (12)

gives an exact sequence

0 → A(k)/n → H1 (k, A[n]) → H1 (k, A)[n] → 0. (13)

Recall that by the non-degeneracy of the Weil pairing we have At [n] = A[n]D , see
Corollary 1.23.

20
Lemma 2.7 The subgroups A(k)/n ⊂ H1 (k, A[n]) and At (k)/n ⊂ H1 (k, At [n]) =
H1 (k, A[n]D ) are orthogonal with respect to the pairing (11):

H1 (k, A[n]) × H1 (k, A[n]D ) → Br(k).

In particular, if E is an elliptic curve, then E(k)/n is a subspace of H1 (k, E[n]) such

that the restriction of the Weil pairing to it is trivial.

Proof. We shall construct several commutative diagrams that will be crucial in the
proof of the duality theorem for abelian varieties over a local field (Theorem 2.6
below).
Step 1 Let N and M be (discrete) G-modules, and let

0→C→B→A→0 (14)

be an exact sequence of G-modules. The differentials in the two long exact sequences
of Ext-groups (with respect to the first and to the second argument, respectively)
∂ ∂
ExtqG (C, M ) −→
1
Extq+1
G (A, M ), ExtpG (N, A) −→
2
Extp+1
G (N, C)

are (up to sign) obtained by splicing with the class of (14). Thus the pairings

Extp+1 q p+q+1
G (N, C) × ExtG (C, M ) −→ ExtG (N, M )

and
ExtpG (N, A) × Extq+1 p+q+1
G (A, M ) −→ ExtG (N, M )
are compatible in the obvious sense: ∂2 (x) ∪ y = x ∪ ∂1 (y).
Step 2 We apply Step 1 with (12) playing the role of (14), for p + q + 1 = 2,
M = Gm and N = Z. From the conclusion of Step 1 we deduce the commutativity
of the right hand square of the following diagram with exact rows, where r = 1 or
r = 0, possibly up to sign:

0 → Extrk (A, Gm )/n → Extrk (A[n], Gm ) → Extr+1

k (A, Gm )[n] → 0
↓ ↓ ↓ (15)
2−r ∗ 2−r ∗ 1−r ∗
0 → H (k, A)[n] → H (k, A[n]) → (H (k, A)/n) → 0

Here for an abelian group L we denote L∗ = Hom(L, Br(k)). The left hand square
commutes by functoriality.
Step 3 Now we take r = 1 and apply Corollary 2.2. We obtain a commutative
(up to sign) diagram with exact rows

0 → At (k)/n → H1 (k, At [n]) → H1 (k, At )[n] → 0

↓ ↓ ↓ (16)
0 → H1 (k, A)[n]∗ → H1 (k, A[n])∗ → (A(k)/n)∗ → 0

21
The horizontal maps in the top sequence come from [n] : At → At because this is
the dual morphism to [n] : A → A (we defined the dual morphism via Ext(·, Gm ),
see Proposition 1.22 and Corollary 1.23).
Now the statement of the lemma immediately follows from the commutativity of
(16) and the exactness of its rows. 
Proof of Theorem 2.6.
Step 1 The middle vertical map in diagram (16) is an isomorphism by The-
orem 2.4. Hence the map At (k)/n → H1 (k, A)[n]∗ is injective. The topolog-
ical groups A(k) and At (k) are extensions of a finite group by a subgroup iso-
morphic to (Ok )dim(A) , hence At (k) = lim At (k)/n is canonically isomorphic to its
←−
own profinite completion. Thus passing to the limit in n we obtain that the map
At (k) → H1 (k, A)∗ is injective.
We note that the representation of A(k) as an extension of a finite abelian group
by (Ok )dim(A) implies

|A(k)/n| = |A(k)[n]| · |Ok : nOk |dim(A) . (17)

(Use the snake lemma.)

Step 2 Using Corollary 2.2 the commutative diagram (15) for r = 0 can be
written as follows:
At [n](k) = At [n](k)
↓ ↓
0 → H2 (k, A)[n]∗ → H (k, A[n]) → (H (k, A)/n)∗ → 0
2 ∗ 1

The middle vertical map is again an isomorphism by Theorem 2.4. Since the right
hand vertical map is injective by Step 1, we see that H2 (k, A)[n] = 0. This holds for
any n ≥ 2, thus H2 (k, A) = 0. One deduces that Hm (k, A) = 0 for m ≥ 2. (Indeed,
Hm (k, A) is a torsion group, but Hm (k, A)[n] is a quotient of Hm (k, A[n]) which is
zero for m ≥ 3 by the last statement of Theorem 2.4.)
Step 3 Let us show that At (k) → H1 (k, A)∗ is surjective. By what was said in Step
1 this will follow if we show that the injective map At (k)/n → H1 (k, A)[n]∗ is also
surjective, for all n. We prove that these groups have the same order. This follows
from (17) and Proposition 2.5. On the one hand, we have

|At (k)/n| = |At (k)[n]| · |Ok : nOk |g = |H2 (k, A[n])| · |Ok : nOk |g ,

where the second equality follows from the local Tate duality with finite coefficients
for A[n] and At [n] = A[n]D . On the other hand,

1 |H1 (k, A[n])| |H1 (k, A[n])|

|H (k, A)[n]| = = g
= |H2 (k, A[n])| · |Ok : nOk |g ,
|A(k)/n| |A(k)[n]| · |Ok : nOk |

22
because |A[n]| = n2g and |Ok : n2g Ok | = |Ok : nOk |2g .
Step 4 The map in the last stement of the theorem is surjective by the diagram (16).
To show that it is an isomorphism we need to prove that the orders of H1 (k, At )[n]
and A(k)/n are equal. This is the same calculation as in Step 3 with At in place of
A. 
Complements. 1 We have proved that all vertical maps in diagram (16) are
isomorphisms. This implies that A(k)/n and At (k)/n are the exact annihilators for
the pairing
H1 (k, A[n]) × H1 (k, A[n]D ) → Z/n
discussed in Lemma 2.7.
2 If an abelian variety A over a local field k has good reduction and n is coprime
to the residual characteristic p of k, then the Galois module A[n] is unramified.
(Since (n, p) = 1 the multiplication by n is an étale morphism of the Néron model
A → A; its kernel is identified with A[n]I . The special fibre of the Néron model is
an abelian variety of the same dimension as A, hence its n-torsion subgroup has the
same cardinality as A[n], hence A[n]I = A[n].) Moreover, A(k)/n = H1nr (k, A[n]).
This can be seen as follows. Let P ∈ A(k). The set of k̄-points Q in A such that
nQ = P is a k-torsor for A[n] whose class is the image of P in H1 (k, A[n]). Since
A has good reduction, this torsor is unramified, i.e. it is the generic fibre of a
Spec(Ok )-torsor for A[n]. Thus its class in H1 (k, A[n]) comes from

H1 (Spec(Ok ), A[n]) = H1nr (k, A[n]) ⊂ H1 (k, A[n]).

2.4 Global fields

Let k be a global field, i.e. a finite extension of Q or Fp (t). Let k̄ be a separable
closure of k and let Γ = Gal(k̄/k). We write kv for a completion of k. This is a local
field or kv = R or C.
One van choose a valuation w of k̄ that extends the valuation v of k. Let Dw ⊂ Γ
be the decomposition group of w, i.e. the stabiliser of w in Γ. Let k̄w be the union of
completions of finite extensions of k at w. By Krasner’s lemma every finite extension
of kv has the form kv [x]/(f (x)) for an irreducible polynomial f (x) with coefficients
in k, so is a completion of a finite extension of k. Thus k̄w is a separable closure of
kv and the decomposition group Dw is identified with Γv = Gal(k̄w /kv ). (Choosing
w is equivalent to choosing an embedding of k̄ into a given separable closure of kv .)
This identification allows us to define the restriction maps

Hn (Γ, M ) −→ Hn (Γv , M )

simply by restricting to the subgroup Dw ∼= Γv of Γ. We shall abuse notation and

write k̄v for the algebraic closure k̄w .

23
 If G is an algebraic k-group, then H1 (k, G) classifies k-torsors for G up to isomor-
phism. In this case the restriction map H1 (k, G) → H1 (kv , G) sends the class of a
torsor X to the class of X ×k kv . In particular, the class of X is in the kernel of the
restriction map to kv if and only if X(kv ) 6= ∅.
Let A be an abelian variety over k. We define the Shafarevich–Tate group
X(A) = ∩v Ker[H1 (k, A) → H1 (kv , A)],
where v ranges over all places of k. Define the n-Selmer group as
Seln (A) = ∩v Ker[H1 (k, A[n]) → H1 (kv , A)],
then there is an obvious exact sequence
0 → A(k)/n → Seln (A) → X(A)[n] → 0.
It can be proved that Seln (A) is finite for any n, see [2, Remark I.6.7]. Thus X(A)[n]
is finite too. A conjecture of Shafarevich and Tate says that X(A) is finite for any
abelian variety A over any global field.
There is a bilinear Cassels–Tate pairing
h, i : X(A) × X(At ) → Q/Z
which is defined as follows. Any class of X(A) is represented by a k-torsor X for
A such that X has a kv -point Pv for each completion kv .
Let us write A for A×k k̄ and similarly for X. We defined Pic0 (A) as the subgroup
of Pic(A) consisting of the elements that are stable under the translations of all
points of A(k̄), see Definition 1.14. Later we proved that Pic0 (A) coincides with
the subgroup of Pic(A) formed by the classes of divisors algebraically equivalent
to 0 (this follows from Lemma 1.18 and the existence of the Poincaré line bundle).
This subgroup can be defined for any variety, in particular, for X. The torsor X is
constructed by twisting A by a cocycle Γ → A(k̄), where A(k̄) acts on A = A(k̄) by
translations. The induced action on Pic0 (A) = At (k̄) is trivial, therefore we obtain
an isomorphism of Γ-modules Pic0 (X) ∼ = At (k̄). Hence any class in X(At ) can be
realised as an (everywhere locally trivial) element ξ ∈ H1 (Γ, Pic0 (X)). There is an
exact sequence of Γ-modules
0 → k̄(X)∗ /k̄ ∗ → Div(X) → Pic(X) → 0.
Let Div0 (X) ⊂ Div(X) be such that the following sequence is exact:
0 → k̄(X)∗ /k̄ ∗ → Div0 (X) → Pic0 (X) → 0.
The differential ∂(ξ) ∈ H2 (Γ, k̄(X)∗ /k̄ ∗ ) comes from an element φ ∈ H2 (Γ, k̄(X)∗ )
because H3 (Γ, k̄ ∗ ) = 0. This φ is well defined up to an element in the image of
H2 (Γ, k̄ ∗ ) −→ H2 (Γ, k̄(X)∗ ).

24
Since ξ restricts to 0 in H1 (Γv , Pic0 (X v )), we see that ∂(ξ) goes to 0 ∈ H2 (Γv , k̄v (X)∗ /k̄v∗ ),
hence φ restricts to an element in the image of

H2 (Γv , k̄v∗ ) −→ H2 (Γv , k̄v (X)∗ ),

for each completion kv . Let cv ∈ H2 (Γv , k̄v∗ ) = Br(kv ) be such an element. We

can think of cv as the “value” of φ at Pv ∈ X(kv ). (By a small deformation of Pv
we can avoid the zerosP and poles of φ.) The collection (cv ) is well defined up to a
global element, hence v invv (cv ) ∈ Q/Z is well defined. This is the value of the
Cassels–Tate pairing of [X] ∈ X(A) and ξ ∈ X(At ). (See [2, Remark I.6.11]. A
definition that uses choices of cocycles instead of torsors can be found in [2, Prop.
I.6.9]. For more definitions and the proof that they are all equivalent, see [5].)
The main result about the Cassels–Tate pairing is this.

Theorem 2.8 The left and right kernels of h, i : X(A) × X(At ) → Q/Z are the
divisible subgroups of X(A) and X(At ), respectively.

Conjecturally, these divisible subgroups should be zero.

Recall the exact sequence of Γ-modules (8):
∗ t
0 → At [2] → Pic(A)[−1] → Homself−dual (A, A ) → 0.
t
We have NS (A) = Homself−dual (A, A ), so that NS (A)Γ = Homself−dual (A, At ). Let
us denote by cλ the image of a self-dual morphism λ : A → At (defined over k)
under the differential
t
Homself−dual (A, A )Γ −→ H1 (k, At [2]).

Any such λ that equals ϕL for some ample line bundle on A is called a polarisation
˜ t.
of A. A polarisation is principal if it is an isomorphism A−→A

Theorem 2.9 (Poonen–Stoll) Let A be an abelian variety over a global field k

and let λ ∈ NS (A)Γ . Then cλ ∈ Sel2 (At ).

Proof. This is a local fact: we need to show that ∂(λ) in H1 (k, At ) is zero when k is
a local field or k = R. Here ∂ is the differential in the long exact sequence of Galois
cohomology associated to the exact sequence

0 → At → Pic(A) → NS (A) → 0.

We shall use the fact that ∂(λ) goes to 0 under the map H1 (Γ, At ) → H1 (Γ, Pic(A)).
The proof of the theorem uses local duality for abelian varieties. The local duality
pairing
H1 (k, At ) × A(k) → Br(k),

25
or, equivalently, the pairing

H1 (k, At ) × Ext1k (At , Gm ) → Br(k),

is defined via the pairing

H1 (Γ, At ) × Ext1Γ (At , k̄ ∗ ) → Br(k),

which is the pairing

H1 (Γ, Pic0 (A)) × Ext1Γ (Pic0 (A), k̄ ∗ ) → Br(k).

By local duality (Theorem 2.6) our statement immediately follows from the obvious
fact that the last pairing is compatible with the pairing

H1 (Γ, Pic(A)) × Ext1Γ (Pic(A), k̄ ∗ ) → Br(k)

and, crucially, that any extension of At by Gm in the category of commutative alge-

braic k-groups gives (via the forgetful functor) an extension of Γ-modules Pic0 (A)
by k̄ ∗ which is a pull-back of some extension of Pic(A) by k̄ ∗ .
Let P ∈ A(k) be the point corresponding to our extension

0 → Gm →? → At → 0

in the category of commutative algebraic k-groups via the canonical identification

of A with (At )t . Let O0,P be the semilocal ring of {0, P } in A, where 0 is the origin
of the group law. This is the subring of k̄(A) consisting of the functions regular at
0 and P . Let Div0,P (A) be the group of divisors on A whose supports are disjoint
from {0, P }. We have an exact sequence of Γ-modules
∗
0 → O0,P /k̄ ∗ → Div0,P (A) → Pic(A) → 0.
∗
Its push-forward via the map O0,P /k̄ ∗ → k̄ ∗ that sends a function f to f (P )/f (0)
is an extension of Pic(A) by k̄ ∗ . We claim that the pull-back of this extension to
Pic0 (A) is an extension of Γ-modules At by k̄ ∗ obtained from our extension (?).
This can be checked by comparing the associated systems of rational factors as in
[7, VII.3 Thm. 6 and the remark after it]. We omit this verification. 

Proposition 2.10 (Poonen–Stoll) Let c̃λ be the image of cλ in X(A). For all
x ∈ X(A) we have hx, λ∗ x − c̃λ i = 0.

Proof. This follows straight from the definition of the Cassels–Tate pairing. Let X
be a k-torsor for A representing x. Choose P ∈ X(k̄). Then x is the class of the

26
cocycle g 7→ gP − P ∈ A(k̄) where g ∈ Γ. Next, suppose that λ = ϕD where D is a
divisor on A. Then λ∗ x is a k-torsor for At represented by the cocycle
∗ t 0
g 7→ [TgP −P D] − [D] ∈ A (k̄) = Pic (A).

However, λ also equals ϕg D because λ is the class of D in NS (A) and λ is Γ-invariant.

Thus λ∗ x can be also represented by the cocycle
∗ g g t 0
g 7→ [TgP −P ( D)] − [ D] ∈ A (k̄) = Pic (A).

The isomorphism of k̄-varieties A−→X

˜ that sends a to a + P induces a canoni-
∼
cal isomorphism of Γ-modules Pic (A) = Pic0 (X). Alternatively, we can take an
0

isomorphism a → a + gP . Using it we see that λ∗ x is represented by the cocycle

∗ ∗
g 7→ [T−P (g D)] − [T−gP (g D)] ∈ Pic0 (X).

(The sign is due to the fact that Tx∗ acts on divisors by −x, so a → a+gP corresponds
∗
to T−gP .) On the other hand, c̃λ is represented by the cocycle

g 7→ [g D] − [D] ∈ At (k̄) = Pic0 (A)

identified with the cocycle

∗ ∗
g 7→ [T−P (g D)] − [T−P D] ∈ Pic0 (X).

Thus λ∗ x − c̃λ is the class of the cocycle

∗ ∗
g 7→ [T−P D] − [g T−P D].
∗
It lifts to a cocycle g 7→ T−P ∗
D − g T−P D with coefficients in Div0 (X) because [D] −
[g D] ∈ Pic0 (A) and hence [T−P
∗
D]−[g T−P∗
D] ∈ Pic0 (X). It follows that ∂(λ∗ x−c̃λ ) =
0 in the notation used in the definition of the Cassels–Tate pairing. This implies
that φ ∈ H2 (Γ, k̄(X)∗ ) comes from an element of Br(k). The sum of local invariants
of a global element is zero, hence hx, λ∗ x − c̃λ i = 0. 
This proposition immediately implies that the bilinear form hx, λ∗ xi on X(A) is
antisymmetric. The proof of the following statement is elementary.

Corollary 2.11 (Poonen–Stoll) Let A be an abelian variety with a principal

polarisation λ. Define c ∈ X(A) as λ−1 ∗ (c̃λ ). Define hx, yiλ = hx, λ∗ yi. Then
assuming that X(A) is finite we have the following alternative.
If hc, ciλ = 0, then there is a finite abelian group T such that

X(A) ∼
= T × T.

If hc, ciλ = 1/2, then there is a finite abelian group T such that

X(A) ∼
= Z/2 × T × T.

27
Remark Let E be an elliptic curve. If O ∈ E is the origin of the group law, then
the line bundle L = O(O) is ample. At the end of Chapter 1 we have seen that the
associated isogeny ϕL is an isomorphism λ : E −→E ˜ t , so E is canonically principally
polarised. Moreover, the polarisation λ comes from a symmetric divisor O defined
over k.
For an elliptic curve Pic0 (E) is the kernel of the surjective map Pic(E) → Z given
by the degree. Hence NS (E) ∼ = Z and we have the exact sequence
∗
0 → E[2] → Pic(E)[−1] → Z → 0.
∗
The polarisation λ represents a generator of Z. It comes from [O] ∈ (Pic(E)[−1] )Γ ,
hence the associated long exact sequence of Galois cohomology groups shows that
cλ = 0. Thus in the case of elliptic curves we are always in the first case of the
alternative of Corollary 2.11. This proves

Corollary 2.12 Let E be an elliptic curve over a number field. If X(E) is finite,
then the Cassels–Tate pairing on X(E) is alternating and |X(E)| is a square.

This does not generalise to curves of higher genus!

Jacobians of curves Let C be a smooth, projective and geometrically integral
curve of genus g. The Jacobian of C is the abelian variety J = Pic0C/k over k whose
group of k̄-points is Pic0 (C). It is canonically principally polarised, so there is an
isomorphism λ : J −→J ˜ t.
The set of divisor classes on C ×k k̄ of degree n is denoted by Picn (C). This is the
group of k̄-points of an algebraic variety over k denoted by PicnC/k . The Jacobian
J acts on PicnC/k by translations, making PicnC/k a k-torsor for J. Note that PicnC/k
may or may not have k-points, but if there is a divisor on C of degree n defined over
k, then PicnC/k (k) 6= ∅.
Exercise Show that the converse is not true by taking C to be a conic without
k-points and n = 1.
If v is a place of k of large residual characteristic, then C(kv ) 6= ∅. (Indeed, C
has good reduction outside finitely many places. Then the reduction C̃ is a smooth
projective curve over a finite field Fq . When q is large enough compared to the genus
√
of C, there are Fq -points on C̃ by the Hasse–Weil formula |C̃(Fq ) − (q + 1)| ≤ 2g q.
A smooth Fq -point on C̃ lifts to a kv -point on C by Hensel’s lemma.) Thus X has
kv -points for almost all places v.
The canonical class KC is represented by the divisors of rational forms on C which
2g−2
are divisors over k. Hence KC is a k-point of PicC/k . The inverse image of KC under
the multiplication by 2 map

[2] : Picg−1 2g−2

C/k −→ PicC/k

28
is a finite k-scheme Θ whose k̄-points are divisor classes x ∈ Pic(C) such that
2x = KC . This is clearly a k-torsor for J[2], called the torsor of theta-characteristics.
We have the following crucial fact.

g−1
Lemma 2.13 cλ = [Θ], hence c̃λ = [PicC/k ] ∈ H1 (k, J).

If X = PicnC/k has points in each completion of k, then [X] ∈ X(J). By Theorem

2.9 and Lemma 2.13 we have

c̃λ = [Picg−1
C/k ] ∈ X(J),

so this always holds for n = g − 1. Poonen and Stoll prove that h[X], [X]iλ = N/2,
where N is the number of places v for which C has no divisor of degree n defined
over kv . Applying this to c̃λ = [Picg−1
C/k ] we obtain the following explicit form of
Corollary 2.11:

Corollary 2.14 If X(J) is finite, then |X(J)| is a square if and only if the number
of places v such that C does not have a divisor of degree g − 1 defined over kv is
even. Otherwise, |X(J)| is twice a square.

An explicit example is the following genus 2 curve over Q which has X(J) ∼
= Z/2:

y 2 = −3(x2 + 1)(x2 − 6x + 1)(x2 + 6x + 1).

The calculation uses the fact that J is isogenous to a product of CM elliptic curves
which makes it possible to apply a theorem of Rubin.
Remark Corollary 2.14 does not generalise to abelian varities without which are
not principally polarised! Indeed, for many small odd primes p (including p = 3)
William Stein [9] constructed an abelian variety A over Q such that X(A) is finite
of order pn2 for some n ∈ Z.
More precisely, take E p−1 to be the kernel of the homomorphism E p → E which
sends (M1 , . . . , Mp ) to M1 + . . . + Mp . Then Z/p acts on E p−1 by cyclic shifts.
The group H1 (k, Z/p) = Hom(Γ, Z/p) consists of characters of the Galois group
Γ → Z/p. The non-trivial characters bijectively correspond to the cyclic extensions
K/Q of degree [K : Q] = p. Thus we can twist E p−1 by such extensions. Let ` be
a prime such that ` ≡ 1 mod p and let K ⊂ Q(µ` ) be the unique subfield of degree
[K : Q] = p. Let A be the twist of E p−1 by K. Equivalently, A is the kernel of the
natural map RK/Q (EK ) → E, where RK/Q is the Weil restriction of scalars. The
resulting exact sequence

0 → A → RK/Q (EK ) → E → 0

29
gives us a map E(Q) → H1 (Q, A). William Stein, in the situation that he considers,
obtains from this map an injective map

E(Q)/p ,→ X(A)[p∞ ]

and proves that its cokernel is the kernel of a surjective map

X(EK )[p∞ ] → X(E)[p∞ ].

Thus Corollary 2.12 and the standard conjecture that the Shafarevich–Tate groups
are finite imply that |X(A)| cannot be a square when E has odd rank over Q and
the image of the Galois action on E[p] is the full group GL(Fp ).
For an explicit example one can take E to be the elliptic curve

y 2 + y = x3 − x

of conductor 37. Calculations (based on deep results of Rubin, Kolyvagin, Kato)

show that for every odd p ≤ 25000, p 6= 37, there exists a prime ` as above such
that X(A) has order pn2 for some n ≥ 1.

2.5 Brauer–Manin obstruction

The (cohomological) Brauer group of a scheme X is defined by Grothendieck as the
étale cohomology group H2 (X, Gm ). Let X be a proper geometrically integral variety
over a field k. Let k̄ be a separable closure of k, Γ = Gal(k̄/k) and X = X ×k k̄. By
Hilbert’s theorem 90 the spectral sequence

Hp (Γ, Hq (X, Gm )) ⇒ Hp+q (X, Gm )

gives rise to the exact sequence

0 → Pic(X) → Pic(X)Γ → Br(k) → Ker[Br(X) → Br(X)] → H1 (Γ, Pic(X)). (18)

Let k be a number field. Then H3 (Γ, k̄ ∗ ) = 0, and hence the last arrow in this exact
sequence is surjective.
Suppose that X has points everywhere locally, i.e., X(kv ) 6= ∅ for each place v.
Take Pv ∈ X(kv ) for each v. By functoriality an element A ∈ Br(X) gives rise to
A(Pv ) ∈ Br(kv ). We have the local invariant homomorphism invv : Br(kv ) → Q/Z,
see Section 2.2.

Lemma 2.15 Let A ∈ Br(X). For almost all places v we have invv (A(Pv )) = 0 for
any Pv ∈ X(kv ).

30
Sketch of proof. This uses the properness of X. There is a finite set of places S
of k and a proper morphism X → Spec(OS ) whose generic fibre is X such that A
extends to an element of Br(X ). By the valuative criterion of properness each point
Pv extends to an Ov -point of X . It follows that A(Pv ) ∈ Br(kv ) comes from an
element of Br(Ov ), but this group is zero, see [6, Ch. XII] or [1, IV.1]. 
Thus we can consider the (finite) sum
X
invv (A(Pv )) ∈ Q/Z. (19)
v

It defines the Brauer–Manin pairing:

Y
X(kv ) × Br(X) −→ Q/Z.
v

Definition 2.16 The Brauer–Manin set v X(kv )Br is the subset of v X(kv ) con-
Q Q

sisting of the elements orthogonal to Br(X) under the Brauer–Manin pairing.

The Albert–Brauer–Hasse–Noether theorem from global class field theory says that
the diagonal embedding of Br(k) into the direct sum of Br(kv ) fits into the exact
sequence
0 → Br(k) → ⊕v Br(kv ) → Q/Z → 0,
where
Q the third map the sum Qof local Br
invariants. Thus the diagonal image of X(k)
in v X(kv ) is contained in v X(kv ) . One says that the Brauer–Manin obstruc-
tion
Q is theBr only obstruction to the Hasse principle for a given class of varieties if
v X(kv ) is non-empty if and only if X(k) is.

Theorem 2.17 (Manin) Let A be an abelian variety over a number field k such
that X(A) is finite. Then the Brauer–Manin obstruction is the only obstruction to
the Hasse principle for k-torsors for A.

This implies that the well known failure of the Hasse principle for plane cubic
curves (e.g. Selmer’s counterexample 3x3 + 4y 3 + 5z 3 = 0) can be explained through
the Brauer–Manin obstruction.
Sketch of proof. The idea is to link (19) to the Cassels–Tate pairing. Let X be a
k-torsor for A with points everywhere locally. Then [X] ∈ X(A). For each place v
fix a point Pv ∈ X(kv ) in such a way that (Pv ) ∈ v X(kv )Br . The exact sequence
Q
(18) shows that
X(At ) ⊂ H1 (Γ, Pic0 (A)) = H1 (Γ, Pic0 (X))

31
naturally maps to a subgroup in Coker[Br(k) → Br(X)]. Using our explicit descrip-
tion of the Cassels–Tate pairing it is possible to check that if A ∈ Br(X) comes from
a class ξ ∈ X(At ), then (up to sign)
X
invv (A(Pv )) = h[X], ξi
v

(see [8, Ch. 6] for details). By the finiteness of X(A) the left kernel of the Cassels–
Tate pairing is zero. Thus if the Brauer–Manin set of X is not empty, then [X] = 0
so that X is a trivial torsor isomorphic to A. In particular, X(k) 6= ∅. 

References
[1] J. Milne. Étale cohomology. Princeton University Press, 1980. 8, 31

[2] J. Milne. Arithmetic Duality Theorems. Kea Books, 2006. 19, 20, 24, 25

[3] D. Mumford. Abelian varieties. Tata Institute of Fundamental Research Studies

in Mathematics 5. Oxford University Press, 1970. 2, 5, 10, 11, 12

[4] A. Polishchuk. Abelian varieties, theta functions and the Fourier transform.
Cambridge University Press, 2003. 13, 14

[5] B. Poonen and M. Stoll. The Cassels–Tate pairing on polarized abelian varieties.
Ann. Math. 150 (1999) 1109–1149. 25

[6] J.-P. Serre. Corps locaux. Hermann, 1968. 18, 31

[7] J.-P. Serre. Groupes algébriques et corps de classes. Hermann, 1959. 8, 9, 26

[8] A. Skorobogatov. Torsors and rational points. Cambridge University Press,

2001. 32

[9] W.A. Stein. Shafarevich–Tate groups of nonsquare order. Modular curves and
abelian varieties, 277–289, Progr. Math. 224 Birkhäuser, 2004. 29

32