Hilbert’s fourteenth problem over finite fields, and a

conjecture on the cone of curves

Burt Totaro

Hilbert’s fourteenth problem asks whether the ring of invariants of any repre-
sentation of a linear algebraic group is finitely generated over the base field. Nagata
gave the first counterexample, using a representation of (Ga )13 , where Ga denotes
the additive group [22]. In his example, the representation is defined over a field
of large transcendence degree over the prime field (of any characteristic). Mukai
simplified Nagata’s construction, showing that there are representations of (Ga )3
over the complex numbers whose ring of invariants is not finitely generated [21].
Mukai relates the problem of finite generation of rings of invariants to a natural
question in algebraic geometry: when is the total coordinate ring of a projective
variety finitely generated? The deepest known result is the Birkar-Cascini-Hacon-
Mc Kernan theorem that the total coordinate ring of a Fano variety in characteristic
zero is finitely generated [2, Cor. 1.3.1]. Nagata and Mukai give counterexamples to
finite generation for certain varieties (blow-ups of projective space) just outside the
realm of Fano varieties. The precise border between finite and infinite generation
remains to be understood.
Mukai’s construction yields representations whose coefficients are “general” com-
plex numbers. In this paper, we show that Mukai’s three best examples can all be
realized over finite fields and over the rational numbers, in fact with simple explicit
coefficients. We give examples over all fields, including the field of order 2. These
are the first published counterexamples to Hilbert’s fourteenth problem over finite
fields. (In retrospect, some examples over finite fields can be constructed using the
work of Manin [17], Tate [31], or Shioda [29] on elliptic surfaces.) For some classes
of examples, we characterize exactly when finite generation holds and when it does
not.
Theorem 0.1 Let k be any field. Then there are linear representations over k of
(Ga )3 on A18 , of (Ga )4 on A16 , and of (Ga )6 on A18 whose rings of invariants are
not finitely generated. The representations are defined explicitly.
Here is the geometric idea. The basic ingredient of Nagata’s examples (especially
as simplified by Steinberg [30]) is the existence of non-torsion line bundles of degree
zero on an elliptic curve. There are no such line bundles on curves over finite fields.
But the generic fiber of an elliptic fibration over a finite field can have infinite
Mordell-Weil group, and that turns out to be enough. Precisely, we encounter
known elliptic fibrations of blow-ups of P2 and P3 , and what seems to be a new
fibration of a blow-up of P5 by abelian surfaces (section 6). It would be interesting
to describe the geometry of these rational abelian fibrations of P5 in more detail.
For example, are these fibrations completely integrable systems with respect to some
Poisson structure?

1
 Our main results (Theorems 5.2 and 7.2) relate finite generation of the total
coordinate ring, in some situations, to finiteness of a certain Mordell-Weil group.
We conclude with a more general conjecture, saying that the cone of curves of any
variety with semi-ample anticanonical bundle is controlled by a group, which may
be infinite. This would follow from a generalization of the conjectures of Kawamata
and Morrison on Calabi-Yau fiber spaces to allow klt pairs (Conjecture 8.1). We
prove the conjecture for smooth projective surfaces with semi-ample anticanonical
bundle, the new case being that of rational elliptic surfaces (Theorem 8.2).
Thanks to Igor Dolgachev, Shigeru Mukai, and Greg Sankaran for their com-
ments.

1 Another example
The smallest known representation of an algebraic group for which finite generation
fails is Freudenburg’s 11-dimensional representation of a unipotent group (Ga )4 oGa
over the rational numbers [8], based on an example by Kuroda [14]. It would be
interesting to know whether there are such low-dimensional examples over finite
fields.
There are broader forms of Hilbert’s fourteenth problem, for example about
actions of algebraic groups on arbitrary affine varieties. Since even the most specific
form of the problem, about linear representations, has a negative answer, we focus
on that case.

2 Mukai’s method
In this section, we summarize Mukai’s geometric approach to producing counterex-
amples to Hilbert’s fourteenth problem.
We use the following result by Mukai [21]. (He works over the complex numbers,
but his proof is elementary and works over any field.)

Theorem 2.1 For any n ≥ r ≥ 3, let X be the blow-up of projective space Pr−1
at n distinct rational points p1 , . . . , pn , not contained in a hyperplane, over a field
k. Let G ∼= (Ga )n−r be the subgroup of (Ga )n which is the kernel of a linear map
A → A corresponding to the points pi . Let (Ga )n act on V = A2n by
n r

(t1 , . . . , tn )(x1 , . . . , xn , y1 , . . . , yn ) = (x1 , . . . , xn , y1 + t1 x1 , . . . , yn + tn xn ).

Then the ring of invariants O(V )G is isomorphic to the total coordinate ring of X,

T C(X) := ⊕a,b1 ,...,bn ∈Z H 0 (X, aH − b1 E1 − · · · − bn En ) ∼

= ⊕L∈PicX H 0 (X, L),

where H is the pullback of the hyperplane line bundle on Pr−1 and E1 , . . . , En are
the exceptional divisors in X.

Therefore, to give examples of rings of invariants which are not finitely generated,
it suffices to exhibit arrangements of points on projective space such that the total
coordinate ring of the blow-up X is not finitely generated.
Define a pseudo-isomorphism between smooth projective varieties to be a bi-
rational map which is an isomorphism outside subsets of codimension at least 2.

2
This notion can be useful for singular varieties, but our applications only involve
smooth varieties. (Example: a pseudo-isomorphism between smooth projective sur-
faces is an isomorphism.) A pseudo-isomorphism X 99K Y induces an isomor-
phism Pic(X) → Pic(Y ) by taking the proper transform of divisors. Define a
(−1)-divisor D on a projective variety X to be the proper transform under some
pseudo-isomorphism X 99K X 0 of the exceptional divisor for a morphism X 0 → Y
which is the blow-up of a smooth point. (For example, a (−1)-curve on a surface is
simply a curve isomorphic to P1 with self-intersection −1, by Castelnuovo.) Mukai
observes that every (−1)-divisor is indecomposable in the monoid of effective line
bundles on X (line bundles L with H 0 (X, L) 6= 0). Also, two different (−1)-divisors
represent different elements of Pic(X). Therefore, if a variety X contains infinitely
many (−1)-divisors, then the monoid of effective line bundles on X is not finitely
generated, and so the total coordinate ring of X is not finitely generated [21, Lemma
3].
Lemma 2.2 will give a way to ensure that a blow-up X of projective space
contains infinitely many (−1)-divisors. To state it, define a birational map Ψ :
Pr−1 99K Pr−1 by  
1 1
[x1 , . . . , xr ] 7→ ,..., .
x1 xr
It contracts the r coordinate hyperplanes to the r coordinate points. A birational
map between two projective spaces which is projectively equivalent to Ψ is called a
standard Cremona transformation. We note that Ψ lifts to a pseudo-isomorphism
from the blow-up of Pr−1 at the r coordinate points to itself. In modern terms, this
pseudo-isomorphism can be described as a composite of several flops.
We say that an arrangement of n points in Pr−1 , n ≥ r ≥ 3, is in linear general
position if no r of the points are contained in a hyperplane. (In particular, the
points are all distinct.) Given n points in linear general position, we can perform
the standard Cremona transformation on any r of the n points; this gives a different
arrangement of n points in a projective space. They need not be in linear general
position. We say that an arrangement of n points in Pr−1 is in Cremona general
position if they are in linear general position and this remains true after any finite
sequence of standard Cremona transformations on r-tuples of the points.
The key point of Mukai’s method is:

Lemma 2.2 Let p1 , . . . , pn , n ≥ r ≥ 3, be points in projective space Pr−1 over a

field k which are in Cremona general position. If
1 1 1
+ + ≤ 1,
2 r n−r
then the blow-up X of Pr−1 at p1 , . . . , pn contains infinitely many (−1)-divisors.
Therefore the total coordinate ring of X is infinitely generated, and the correspond-
ing representation of (Ga )n−r of dimension 2n over k has infinitely generated ring
of invariants.

For clarity, we recall the proof of Lemma 2.2. Clearly X contains n (−1)-divisors
over the n points p1 , . . . , pn . But since p1 , . . . , pn are in Cremona general position,
we can perform Cremona transformations on r-tuples of points in p1 , . . . , pn any
number of times. This shows that X is pseudo-isomorphic to a blow-up of Pr−1 in

3
many other ways, and so we find other (−1)-divisors on X. The inequality on n
ensures, by a purely combinatorial argument, that the resulting (−1)-divisors on X
have arbitrarily large degrees (when projected down to divisors on Pr−1 ), and so
there are infinitely many of them. To describe this argument in more detail, define
a symmetric bilinear form on Pic(X) by H 2 = r − 2, Ei2 = −1, H · Ej = 0, and
Ei · Ej = 0 for i 6= j. We have a canonical identification of Pic(X) with Zn+1 =
ZH ⊕ZE1 ⊕· · ·⊕ZEn for every blow-up X of an (r−1)-dimensional projective space
at an ordered set of n points. So the standard Cremona transformation at the points
p1 , . . . , pr determines a well-defined automorphism of Zn+1 , which we compute to be
the reflection sn orthogonal to H − E1 − · · · − Er . Switching points pi and pi+1 acts
on Zn+1 by the reflection si orthogonal to Ei − Ei+1 for 1 ≤ i ≤ n − 1. We compute
(it is also clear geometrically) that the Cremona action on Pic(X) = Zn+1 fixes the
anticanonical class of X, −KX = rH − (r − 2)E1 − · · · − (r − 2)En . Together the

d sn

d d d d d d d
s1 s2 sr−1 sr sr+1 sn−2 sn−1

Figure 1: T2,r,n−r Dynkin diagram

reflections s1 , . . . , sn−1 , sn generate the Weyl group with Dynkin diagram T2,r,n−r ,
and the inequality in Lemma 2.2 is just what is needed to ensure that this Weyl
group is infinite.
It is clear that very general n-tuples of points in projective space (that is, n-
tuples outside of countably many proper subvarieties of (Pr−1 )n ) are in Cremona
general position. Thus Lemma 2.2 gives counterexamples to Hilbert’s fourteenth
problem over the complex numbers (or any uncountable field). But this argument
leaves it unclear whether there are any arrangements in Cremona general position
over the rational numbers or over finite fields. In the next section we will construct
such arrangements. These will occur in the most interesting cases of Lemma 2.2,
(1)
where 12 + 1r + n−r 1
= 1, corresponding to the affine Weyl groups of type E8 (n = 9,
(1)
r = 3 or 6) or E7 (n = 8, r = 4).
As a historical note, Coble knew in 1929 that the blow-up of P2 at 9 very general
points contained infinitely many (−1)-curves [3, section 9], but the first proof to
modern standards was given by Nagata in 1960 [23, Lemma 2.5 and Theorem 4a].

3 The method, with elementary proofs

In this section we prove Theorem 0.1 over any sufficiently large finite field and over
any infinite field. This turns out to require only some simple geometry, following
the ideas of Coble and Dolgachev.
By Lemma 2.2, Theorem 0.1 will follow if we can produce explicit examples of
9-tuples of points in P2 in Cremona general position, or 8-tuples in P3 , or 9-tuples
in P5 . The following lemmas do that, over sufficiently large finite fields and over
infinite fields. The idea is to use arrangements of points which are special in one

4
way (for example, we take 9 points in P2 which are the intersection of two cubics),
but not too special.
In what follows, a cubic in P2 denotes an effective divisor of degree 3; it is
not assumed to be irreducible (likewise for conics, quadrics in higher-dimensional
projective spaces, and so on). We say that a cubic is irreducible if it is irreducible
and reduced (that is, the corresponding cubic form is irreducible). A pencil of cubics
in P2 means a linear system P1 of cubics; equivalently, it is a 2-dimensional linear
subspace of the vector space of cubic forms in 3 variables. We say that two cubics
in P2 over a field k intersect in 9 given points if those are all the intersection points
over the algebraic closure of k.

Lemma 3.1 Let p1 , . . . , p9 be 9 distinct rational points in P2 over a field k. Suppose

that there are cubics C1 and C2 whose intersection is the set {p1 , . . . , p9 }. Then no
three of p1 , . . . , p9 lie on a line if and only if all cubics in the pencil spanned by C1
and C2 are irreducible.

Proof. Suppose that a cubic C1 in the pencil is reducible. Then we can write
C1 = L + D for some line L and conic D. For any other cubic C2 in the pencil,

{p1 , . . . , p9 } = C1 ∩ C2 = (L ∩ C2 ) ∪ (D ∩ C2 ).

Since C2 intersects C1 transversely, three of these points are on the line L. Con-
versely, suppose that p1 , p2 , p3 lie on a line L. There is a unique cubic on L through
p1 , p2 , p3 . Since we have a pencil (a P1 ) of cubics in P2 through p1 , p2 , p3 , at least
one cubic C1 in the pencil must contain L. Thus C1 is reducible. QED

Corollary 3.2 Let p1 , . . . , p9 be 9 distinct rational points in P2 over a field k which

are the intersection of two cubics. Suppose that no three of p1 , . . . , p9 lie on a line.
Then p1 , . . . , p9 are in Cremona general position. Therefore the blow-up X of P2
at p1 , . . . , p9 has infinitely generated total coordinate ring, and the corresponding
18-dimensional representation of (Ga )6 over k has infinitely generated ring of in-
variants.

The first part of Corollary 3.2 was proved first by Manin [17, Lemma 3].
Proof. The Cremona action on Pic(X) ∼ = Z10 fixes the anticanonical class
−KX = 3H − E1 − · · · − E9 . Therefore, performing a standard Cremona transfor-
mation on any three of p1 , . . . , p9 transforms a cubic through p1 , . . . , p9 into a cubic
through the new points p01 , . . . , p09 . Since p1 , . . . , p9 are the intersection of two cubics
in P2 , this remains true after performing a standard Cremona transformation on
any three of p1 , . . . , p9 . Since no three of p1 , . . . , p9 lie on a line, all cubics in the
pencil of cubics through these points are irreducible by Lemma 3.1. The point is
that this remains true of the proper transforms of these cubics on the “new” P2 .
Therefore, no three of p01 , . . . , p09 lie on a line. Thus p01 , . . . , p09 satisfy the same prop-
erties we assumed for p1 , . . . , p9 . We can repeat this process any number of times.
So p1 , . . . , p9 are in Cremona general position. Lemma 2.2 gives the rest. QED
Corollary 3.2 gives counterexamples to Hilbert’s fourteenth problem over all
sufficiently large finite fields, and over all infinite fields; we give explicit examples
in section 4. The group involved is (Ga )6 . To get Mukai’s best example, with the
group (Ga )3 , to work over the same fields, we can use projective duality, as follows.

5
 Let q1 , . . . , q9 be 9 distinct points on P2 which are the intersection of two cubics.
These points can be represented by a linear map from A9 onto A3 . The kernel has
dimension 6, and so dualizing gives a linear map from A9 onto a 6-dimensional vector
space. This gives the dual arrangement of 9 points p1 , . . . , p9 in a 5-dimensional
projective space. Dolgachev gives an equivalent description of this arrangement,
under our assumption on q1 , . . . , q9 : p1 , . . . , p9 are the image of q1 , . . . , q9 under a
Veronese embedding P2 → P5 [5, Prop. 5.4].

Corollary 3.3 Let q1 , . . . , q9 be 9 distinct rational points in P2 over a field k which

are the intersection of two cubics. Suppose that no three of q1 , . . . , q9 lie on a line.
Let p1 , . . . , p9 be the dual arrangement of 9 points in P5 . Then p1 , . . . , p9 are in
Cremona general position. Therefore the blow-up X of P5 at p1 , . . . , p9 has infinitely
generated total coordinate ring, and the corresponding 18-dimensional representation
of (Ga )3 over k has infinitely generated ring of invariants.

Proof. Since q1 , . . . , q9 are in linear general position in P2 , we check by hand

that the dual arrangement p1 , . . . , p9 is in linear general position in P5 . Performing
the standard Cremona transformation on q1 , q2 , q3 in P2 and then dualizing corre-
sponds to performing the standard Cremona transformation on p9 , . . . , p4 in P5 , as
Dolgachev-Ortland computed [6, Theorem VI.4]. (This is elementary: the assump-
tion that q1 , . . . , q9 are the intersection of two cubics is not needed for this duality
statement.) Since q1 , . . . , q9 are in Cremona general position in P2 by Corollary 3.2,
it follows that p1 , . . . , p9 are in Cremona general position in P5 . Lemma 2.2 gives
the rest. QED
Thus we have counterexamples to Hilbert’s fourteenth problem using an 18-
dimensional representation of (Ga )3 over all sufficiently large finite fields and all
infinite fields. Finally, let us check that Mukai’s final class of examples, using the
group (Ga )4 , also works over all sufficiently large finite fields and all infinite fields.
A net of quadrics in P3 means a linear system P2 of quadrics, or equivalently a
3-dimensional linear subspace of the vector space of quadratic forms in 4 variables.

Lemma 3.4 Let p1 , . . . , p8 be 8 distinct points in P3 which are the intersection of

three quadrics, Q1 ∩ Q2 ∩ Q3 . Then no four of the points p1 , . . . , p8 lie on a plane
if and only if every quadric in the net spanned by Q1 , Q2 , Q3 is irreducible.

Proof. If some quadric Q in the net is reducible, then Q = S1 ∪ S2 for some

planes S1 and S2 in P3 . Then four of the points p1 , . . . , p8 lie on S1 (and the other
four lie on S2 ).
Conversely, suppose that p1 , p2 , p3 , p4 lie on a plane S in P3 . No three of the
points p1 , . . . , p8 lie on a line L in P3 ; otherwise every quadric in the net would
contain three points on L, hence would contain L, contradicting that Q1 ∩Q2 ∩Q3 =
{p1 , . . . , p8 }. So no three of the points p1 , p2 , p3 , p4 lie on a line. Therefore these
four points are the complete intersection of two conics in the plane S, p1 p2 ∪ p3 p4
and p1 p3 ∪ p2 p4 . It follows that there is only a pencil (a P1 ) of conics in S through
p1 , p2 , p3 , p4 . Since we have a net (a P2 ) of quadrics in P3 that contain p1 , p2 , p3 , p4 ,
at least one quadric Q in the net must contain the whole plane S. Thus Q is
reducible. QED

6
Corollary 3.5 Let p1 , . . . , p8 be 8 distinct points in P3 which are the intersection
of three quadrics, and suppose that no four of the points p1 , . . . , p8 lie on a plane.
Then p1 , . . . , p8 are in Cremona general position. Therefore the blow-up X of P3
at p1 , . . . , p8 has infinitely generated total coordinate ring, and the corresponding
16-dimensional representation of (Ga )4 over k has infinitely generated ring of in-
variants.

The first part of Corollary 3.5 was apparently known to Coble [3, section 44,
last paragraph].
Proof. The Cremona action on Pic(X) fixes half the anticanonical class,
− 21 KX = 2H − E1 − · · · − E8 . So performing a standard Cremona transforma-
tion on any four of p1 , . . . , p8 transforms a quadric through p1 , . . . , p8 into a quadric
through the new points p01 , . . . , p08 . Since p1 , . . . , p8 are the intersection of three
quadrics, p01 , . . . , p08 are also the intersection of three quadrics. By Lemma 3.4, since
no four of p1 , . . . , p8 lie on a plane, all the quadrics through p1 , . . . , p8 are irreducible.
Therefore their proper transforms, the quadrics through p01 , . . . , p08 , are also all ir-
reducible. So no four of p01 , . . . , p08 lie on a plane. Thus p01 , . . . , p08 satisfy the same
assumptions as p1 , . . . , p8 , and so we can repeat the process any number of times.
That is, p1 , . . . , p8 are in Cremona general position. QED

4 Examples over arbitrary fields

We now prove Theorem 0.1, showing that our three classes of representations with
infinitely generated rings of invariants all exist over arbitrary fields. In later sec-
tions, we will describe the geometry behind these representations in more detail.
One benefit will be to construct representations with infinitely generated ring of
invariants that are given by simpler formulas. Another benefit will be to prove par-
tial results about exactly when finite generation holds. For now, we will just prove
Theorem 0.1 as stated.
Theorem 2.1 associates a 2n-dimensional representation of (Ga )n−r over a field
k to any arrangement of n distinct k-points of Pr−1 not contained in a hyper-
plane. More generally, we can define such a representation associated to a smooth
0-dimensional subscheme of degree n defined over k, even if the individual points are
not defined over k, using the standard technique of “twisting” in Galois cohomology
[28, III.1.3]. Explicitly, given a Galois extension field K/k over which the points are
defined, let the Galois group Gal(K/k) act on K n by permuting the basis vectors
as Gal(K/k) permutes the n points, and by σ(ax) = σ(a)σ(x) for a ∈ K. We can
view K n with this Galois action as a “twisted form” H of the group (Ga )n over k,
but it is in fact isomorphic to (Ga )n over k because H(k) = (K n )Gal(K/k) is an n-
dimensional vector space over k. A linear map K n → K r associated to the n points
in Pr−1 is evidently Galois-equivariant, and so (taking Galois invariants) it gives a
linear map H(k) ∼ = k n → k r . The kernel is a vector space over k, which canonically
determines an algebraic group G over k, clearly isomorphic to (Ga )n−r . Likewise,
the 2n-dimensional representation of (Ga )n in Theorem 2.1 gives a 2n-dimensional
representation of its twisted form H and hence of the subgroup G ∼ = (Ga )n−r over
k.
Moreover, whether a ring of invariants O(V )G for a representation V of an
algebraic group G is finitely generated does not change under extension of the base

7
field, since the ring of invariants over an extension field K/k is just the ring of
invariants over k tensored with K. So we can apply the previous section’s results.
For example, given two plane cubics C1 and C2 over a field k whose intersection
is smooth of dimension 0, consider the associated 18-dimensional representation of
(Ga )6 over k. If no three points of C1 ∩C2 over the algebraic closure of k lie on a line,
then the corresponding representation of (Ga )6 over k has infinitely generated ring
of invariants, by Corollary 3.2. In practice, it is often easier to check the equivalent
hypothesis that all cubics in the pencil spanned by C1 and C2 are irreducible.
For example, consider the cuspidal cubics (y + z)3 + xz 2 and x3 + y 2 z over
a field Fp . For primes p ≤ 23 (not an optimal bound), we compute that the
intersection of these two cubics in P2 is smooth of dimension 0, and that all nonzero
linear combinations of these two cubics are irreducible. (This is particularly easy
to check for p = 2 or 3, where the given cuspidal cubics are the only singular
cubics in the pencil.) By Corollary 3.2, for p ≤ 23, the associated 18-dimensional
representation of (Ga )6 over Fp has infinitely generated ring of invariants. For fields
of characteristic greater than 23 (or characteristic zero), we can apply Corollary 3.5
directly. Namely, the 9 points in P2 which form the columns of the following matrix
are the base locus of a pencil of cubics, and are in linear general position over Q and
over Fp for all p > 23. (One can look for such examples by choosing eight points in
P2 and computing the ninth point on the pencil of cubics through the eight points.)
 
1 0 0 1 1 2 −3 −2 −7
0 1 0 1 −1 −1 4 −5 2 
0 0 1 1 2 1 1 1 −1

Thus we have constructed 18-dimensional representations of (Ga )6 with infinitely

generated ring of invariants over an arbitrary field.
Now consider again the intersection of the cuspidal cubics (y + z)3 + xz 2 and
x3 + y 2 z over a field Fp with p ≤ 23. The projective dual of these 9 points in
P2 gives a smooth subscheme of degree 9 in P5 over Fp . Consider the associated
representation of (Ga )3 over Fp . Since all cubics in this pencil of cubics in P2 are
irreducible when p ≤ 23, Corollary 3.3 shows that the ring of invariants for this
18-dimensional representation of (Ga )3 over Fp is infinitely generated. For fields of
characteristic greater than 23 or of characteristic zero, we can apply Corollary 3.3
directly, using the previous paragraph’s arrangement of 9 rational points. Thus we
have constructed 18-dimensional representations of (Ga )3 with infinitely generated
ring of invariants over an arbitrary field.
We now construct analogous representations of (Ga )4 . Over a field Fp , consider
the quadrics xy + y 2 + z 2 , xw + y 2 + w2 , xz − zw + z 2 + w2 . We compute that for
all primes p ≤ 7 (not an optimal bound), the intersection of these three quadrics is
smooth of dimension 0, and all nonzero linear combinations of these quadrics are ir-
reducible. Therefore, for p ≤ 7, the corresponding 16-dimensional representation of
(Ga )4 over Fp has infinitely generated ring of invariants. For fields of characteristic
greater than 7 (or characteristic zero), we can apply Corollary 3.5 directly. Namely,
the 8 points in P3 which form the columns of the following matrix are the base
locus of a net of quadrics, and are in linear general position over Q and over Fp for
all p > 7. (One can look for such examples by choosing seven points p1 , . . . , p7 in

8
P3 and computing the eighth point on the net of quadrics through p1 , . . . , p7 .)
 
1 0 0 0 1 −2 1 −6
0 1 0
 0 1 2 −4 −8 
0 0 1 0 1 1 −3 −7
0 0 0 1 1 −3 4 −4

Thus we have proved the existence of 16-dimensional representations of (Ga )4 with

infinitely generated ring of invariants over an arbitrary field. QED (Theorem 0.1)

5 Elliptic fibrations, and representations of (Ga )6 on A18

We now describe the geometry behind the representations of (Ga )6 in Theorem 0.1.
It turns out that finite generation depends upon the Mordell-Weil group of a certain
elliptic fibration of P2 . For a certain class of representations, we can say exactly
when the ring of invariants is finitely generated and when it is not. As a concrete
application, we give examples of non-finite generation where the coefficients of the
representation are very simple (in particular, simpler than the examples in section
4). As in section 4, we get examples of non-finite generation over all fields, even the
field of order 2.
The general problem here is to understand the border between finite generation
and non-finite generation (for rings of invariants, or for total coordinate rings). Sec-
tion 3 shows that if we assume a small amount of general position, we get examples
of non-finite generation. In this section we find that even for some more special ar-
rangements of points, which do not satisfy the hypotheses of section 3, we can still
prove non-finite generation. This case of 9 points in the plane has been intensely
studied, and many of the results of this section can be deduced from various earlier
works, as we will see. The main novelty is the precise characterization of finite
generation in Theorem 5.2.

Corollary 5.1 Let (Ga )n act on V = A2n by

(t1 , . . . , tn )(x1 , . . . , xn , y1 , . . . , yn ) = (x1 , . . . , xn , y1 + t1 x1 , . . . , yn + tn xn ).
Over any field k of characteristic not 2 or 3, consider the subgroup G = (Ga )6 of
(Ga )9 which is the kernel of the following linear map A9 → A3 :
 
−1 −1 −1 0 0 0 1 1 1
−1 0 1 −1 0 1 −1 0 1
1 1 1 1 1 1 1 1 1

Consider the restriction of the above 18-dimensional representation of (Ga )9 to the

subgroup (Ga )6 . The ring of invariants of this representation is not finitely generated
over k.
There are other (almost equally simple) 18-dimensional representations of (Ga )6
over F2 or F3 for which the ring of invariants is not finitely generated.

Corollary 5.1 is a consequence of the following theorem, which characterizes

exactly which intersections of two cubics in P2 yield infinitely generated rings of in-
variants (or infinitely generated total coordinate rings). We generalize this theorem
to all rational elliptic surfaces in Theorem 8.2.

9
Theorem 5.2 Let p1 , . . . , p9 be 9 distinct rational points in P2 over a field k which
are the intersection of two cubics. Let a be the number of collinear triples of points
p1 , . . . , p9 , let b be the number of partitions of p1 , . . . , p9 into three collinear triples,
and let
ρ = 8 − a + b.
Then ρ ≥ 0.
If ρ > 0, then the blow-up X of P2 at p1 , . . . , p9 has infinitely generated total
coordinate ring, and the corresponding 18-dimensional representation of (Ga )6 over
k has infinitely generated ring of invariants.
Conversely, if ρ = 0, then these rings are finitely generated.

Theorem 5.2 strengthens Corollary 3.2, which proves infinite generation assum-
ing that there are no collinear triples among p1 , . . . , p9 .
Proof. Let X be the blow-up of P2 at the points p1 , . . . , p9 . The pencil of
cubics through p1 , . . . , p9 makes the blow-up X an elliptic surface f : X → P1 (or
perhaps quasi-elliptic, in characteristics 2 and 3). We want to give criteria for X to
contain infinitely many (−1)-curves, without requiring that no three of p1 , . . . , p9
lie on a line (as in Corollary 3.2). There is in fact a complete classification of the
intersections of two plane cubics such that the blow-up X contains only finitely
many (−1)-curves (X is called an extremal rational elliptic surface), by Miranda-
Persson in characteristic zero and W. Lang in positive characteristic [18, 15, 16]; see
also Cossec-Dolgachev [4, section 5.6]. There is also more general work by Nikulin
classifying certain types of varieties whose cone of curves is finite polyhedral [24, 25].
We will not use any of these classification results, but just explain how to check
whether a given rational elliptic surface X has infinitely many (−1)-curves.
The (−1)-curves in any surface can be described as the smooth rational curves
C such that (−KX ) · C = 1. Since the fibers of the elliptic fibration f : X → P1 are
in the linear system | − KX |, it follows that the (−1)-curves in X are precisely the
sections of the elliptic fibration. Thus, to make X have infinitely many (−1)-curves,
it suffices to arrange that the Mordell-Weil group Pic0 (E) of the general fiber (an
elliptic curve over the field k(t)) has rank at least 1. We can define the Mordell-Weil
group as the group of sections of this elliptic fibration, with one section considered
as the zero section.
For each reducible fiber F of the elliptic fibration X → P1 , let rF + 1 be the
number of irreducible components of F . By Tate [31, 4.5]Pand Shioda [29, Cor.
1.5], the Mordell-Weil rank of E over k(t) is equal to 8 − rF . For the reader’s
convenience, here is a proof. The Picard group of the general fiber E over k(t) is
the quotient of the Picard group of the surface X by the classes of all irreducible
divisors in X which do not map onto P1 . So Pic(E) is the quotient of Pic(X) ∼ = Z10
1
by the class of −KX (the class of each irreducible fiber of f : X → P ) and by rF
classes for each fiber F withPrF + 1 irreducible components. Moreover, if we write
the reducible fibers as F = rj=1 F +1
mF j DF j , then the divisors DF j for all F and all
1 ≤ j ≤ rF , together with −KX , are linearly independent in Pic(X)Q (even modulo
numerical equivalence). This is a fact about any morphism P from a surface to a curve
[1, Cor. VIII.4]. Therefore Pic(E)P Q has rank 10 − 1 − rF and the Mordell-Weil
0
group Pic (E)Q has rank 8 − rF , as we want.
The only possible reducible cubics in the given pencil are the sum of a line and
a conic, or the sum of three lines. By the proof of Lemma 3.1, the lines occurring

10
in these reducible cubics are precisely
P the lines through collinear triples among the
points p1 , . . . , p9 . Therefore rF is equal to the number a of collinear triples,
minus the number b of partitions of p1 , . . . , p9 into three disjoint collinear triples.
(The second term comes in because P a cubic in the pencil which consists of three
lines contributes only 2, not 3, to rF .) So the Mordell-Weil rank of E over k(t)
is ρ = 8 − a + b, which is therefore nonnegative. If ρ is greater than 0, then the
blow-up X contains infinitely many (−1)-curves. It follows that the total coordinate
ring of X is infinitely generated, and that the corresponding ring of invariants for
an 18-dimensional representation of (Ga )6 over k is infinitely generated.
It remains to show that when ρ = 0, then the total coordinate ring of X is finitely
generated. Our main tools will be the cone and contraction theorems, proved for
log surfaces in any characteristic by Kollár and Kovács [12, 2.1.1, 2.3.3], and Hu
and Keel’s notion of a Mori dream space [10]. In particular, for any klt pair (X, ∆)
of dimension 2, we can contract any (KX + ∆)-negative extremal face of the closed
cone of curves N E(X), yielding a projective variety. Kollár and Kovács only state
this for contractions of extremal rays, but the result for higher-dimensional extremal
faces is a consequence. The point is that for any (KX + ∆)-negative extremal face
F , if the contraction f : X → Y of one ray of F is birational, then the pair (Y, f∗ ∆)
is klt and the image of F is a (KY + f∗ ∆)-negative extremal face [12, Lemma 2.3.5].
Then the cone theorem (including the contraction theorem) for higher-dimensional
extremal faces follows by induction on the Picard number of X.
The assumption ρ = 0 tells us that the subspace (KX )⊥ ∼ = Q9 of Pic(X) ⊗ Q is
spanned by the curves in X which are contained in fibers of the morphism X → P1 .
The sum of all the curves in a reducible fiber, with multiplicities, is numerically
equivalent to a general fiber −KX ; so (KX )⊥ is spanned just by the finitely many
irreducible components of reducible fibers of X → P1 . We can also describe these
curves as the (−2)-curves on X (curves isomorphic to P1 with self-intersection −2),
using the fact that −KX has degree 0 on any (−2)-curve.
Another consequence of the assumption ρ = 0 is that the Mordell-Weil rank of
the elliptic surface X → P1 is zero. Since the sections of X → P1 are exactly the
(−1)-curves, this means that X contains only finitely many (−1)-curves. By the
cone theorem, the extremal rays of the closed cone of curves in the KX -negative
half space (KX )<0 are all spanned by (−1)-curves, and so there are only finitely
many extremal rays in (KX )<0 . Since −KX is semi-ample (corresponding to the
contraction X → P1 ), we know that the closed cone of curves is contained in
(KX )≤0 .
Moreover, for any curve C on X which is not one of the finitely many (−2)-
curves DF j , we have C · DF j ≥ 0 for allP F and j. Because the curves DF j span
(KX )⊥ and are divided into subsets with j mF j DF j ∼ −KX , where mF j > 0, the
cone
{x ∈ N 1 (X) : x · DF j ≥ 0 for all F, j}
is a finite polyhedral subcone of (KX )≤0 whose intersection with (KX )⊥ is just
R≥0 · (−KX ). Since all curves except the (−2)-curves belong to this cone, it follows
that the closed cone of curves is finite polyhedral, spanned by the (−1)-curves
and (−2)-curves. (This consequence of ρ = 0 was proved already by Nikulin [24,
Example 1.4.1].)
By Hu and Keel, a projective variety has finitely generated total coordinate

11
ring if and only if it is a “Mori dream space” [10, Proposition 2.9]. For a smooth
projective surface, this means that the first Betti number is zero, the closed cone
of curves is finite polyhedral, and every codimension-1 face of the cone can be
contracted.
Thus, it remains to show that every codimension-1 face of the closed cone of
curves N E(X) can be contracted. A face in the KX -negative half space (equiva-
lently, a face spanned by (−1)-curves) can be contracted by the cone theorem. On
the other hand, because the (−2)-curves span the hyperplane (−KX )⊥ , the intersec-
tion N E(X) ∩ (KX )⊥ is a codimension-1 face (the span of all the (−2)-curves). We
can contract this face using the line bundle −KX , corresponding to the contraction
X → P1 .
It remains to contract a codimension-1 face A of N E(X) spanned by some
(−1)-curves and some (−2)-curves. Let C1 , . . . , Cr be the (−2)-curves in the face
A. These cannot includePall the irreducible components of any fiber F of X → P1 ,
because then −KX ∼ j mF j DF j would belong to the face A. But −KX is a
positive linear combination of all the (−2)-curves. Therefore −KX belongs to the
interior of the codimension-1 face (KX )⊥ ∩ N E(X), and so it cannot belong to any
other face.
Since the (−2)-curves C1 , . . . , Cr do not include all the irreducible components
of any fiber of X → P1 , the intersection pairing on ⊕ri=1 QC Pi is negative definite [1,
Cor. VIII.4]. So there are rational numbers ai such that ( ai Ci ) · Cj = −1 for all
P Kollár-Mori [13, Lemma 3.41], it follows that ai ≥ 0 for all i.
j = 1, . . . , r. By
Let ∆ =  ai Ci for some small positive number . The pair (X, ∆) is klt [12,
Definition 2.2.6] since X is smooth and  is small; so we can apply the contraction
theorem. Clearly KX + ∆ is negative on the (−1)-curves C in the face A, since
KX · C = −1 and  is small. Also, KX + ∆ has degree − on the (−2)-curves
C1 , . . . , Cr in A. Thus the face A is (KX + ∆)-negative. By the cone theorem, we
can contract the face A. This completes the proof that the total coordinate ring of
X is finitely generated. QED (Theorem 5.2)
We now prove Corollary 5.1. For any field k of characteristic not 2 or 3, consider
the 9 points in P2 (k) given in affine coordinates by (x, y) where x and y run through
the set {−1, 0, 1}. These points are the intersection of the two cubics x3 = xz 2 and
y 3 = yz 2 , each consisting of three lines through a point in P2 . We count (using
the assumption on the characteristic) that there are 8 collinear triples among these
points p1 , . . . , p9 , and 2 partitions of p1 , . . . , p9 into three collinear triples. (The
two partitions into three collinear triples correspond to two cubics x3 − xz 2 =
x(x − z)(x + z) and y 3 − yz 2 = y(y − z)(y + z) in our pencil which are unions of
three lines, and the other two collinear triples correspond to two other reducible
cubics in the pencil, (x3 − xz 2 ) − (y 3 − yz 2 ) = (x − y)(x2 + xy + y 2 − z 2 ) and
(x3 − xz 2 ) + (y 3 − yz 2 ) = (x + y)(x2 − xy + y 2 − z 2 ).) So ρ = 8 − 8 + 2 = 2 is greater
than 0. By Theorem 5.2, the blow-up X of P2 at this set of 9 points has infinitely
many (−1)-curves. Thus we get an 18-dimensional representation of (Ga )6 over
Fp for p ≥ 5 (or over Q) whose ring of invariants is not finitely generated. More
precisely, the Mordell-Weil group has rank 2 in this case.
We refer to the Appendix for simple examples of 18-dimensional representations
of (Ga )6 over F2 or F3 whose rings of invariants are not finitely generated. QED
(Corollary 5.1)

12
6 A new fibration of P5 by abelian surfaces, and repre-
sentations of (Ga )3 on A18
We now give a richer geometric explanation for the infinitely generated rings of
invariants for (Ga )3 constructed in Corollary 3.3: they are explained by a fibration
of a blow-up of P5 by abelian surfaces.

Theorem 6.1 Let q1 , . . . , q9 be nine distinct rational points in P2 over a field k

which are the intersection of two cubics. Suppose that no three of p1 , . . . , p9 lie
on a line. Let p1 , . . . , p9 be the dual arrangement of 9 points in P5 . Then the
blow-up X of P5 at p1 , . . . , p9 is pseudo-isomorphic to a smooth projective variety
W which is an abelian surface fibration over P3 with a section. The Mordell-Weil
group of W over P3 has rank 8. The translates by the Mordell-Weil group of the
(−1)-divisors E1 , . . . , E9 yield infinitely many (−1)-divisors on W , or equivalently
on X. As a result, the total coordinate ring of X is not finitely generated, and the
corresponding 18-dimensional representation of (Ga )3 over k has infinitely generated
ring of invariants.

Proof. We recall Dolgachev’s equivalent description of the dual arrangement,

under our assumption on q1 , . . . , q9 : p1 , . . . , p9 are the image of q1 , . . . , q9 under a
Veronese embedding P2 → P5 [5, Prop. 5.4].
There is a natural homomorphism from the group Z8 o Z/2 to the automor-
phism group of the blow-up Y of P2 at q1 , . . . , q9 [6, p. 124]. Explicitly, Y is
an elliptic surface, with the Z8 subgroup giving translations by differences of sec-
tions, and the Z/2 giving the map z 7→ −z with respect to some zero-section. The
construction of p1 , . . . , p9 by duality gives a corresponding action of Z8 o Z/2 by
pseudo-automorphisms of the blow-up X of P5 at p1 , . . . , p9 , by Dolgachev [5, 5.12].
Moreover, these pseudo-automorphisms are defined on the blown-up Veronese sur-
face Y inside X and give the action we mentioned of Z8 o Z/2 on Y . Our goal is
to give a more geometric interpretation of these pseudo-automorphisms of X.
The linear system of − 12 KX = 3H − 2E1 − · · · − 2E9 (that is, cubics on P5
singular at p1 , . . . , p9 ) is a P3 , with base locus a union of 45 curves: the 36 lines
through pairs of points pi and the 9 rational normal curves through any 8 of the
points p1 , . . . , p9 . Indeed, the base locus of the linear system | − 21 KX |, viewed on
P5 , is an intersection of 4 cubics. One checks that the base locus has dimension
1, which is easy in a particular example (and therefore holds for a general 9-tuple
p1 , . . . , p9 as above). Thus the base locus is a complete intersection of four cubics
in P5 and hence has degree 81. The 45 curves mentioned have total degree 81, and
they are all in the base locus, because − 21 KX has degree −1 on these curves. So
the base locus is precisely the union of these 45 curves, each with multiplicity 1.
We can describe the rational map X 99K P3 associated to − 12 KX . The base
locus of this linear system consists of 45 disjoint curves isomorphic to P1 in X, all
with normal bundle O(−1)⊕4 . Using the linear system | − 12 KX |, we can perform an
inverse flip on these 45 curves, giving another smooth projective 5-fold W in which
the 45 P1 ’s with normal bundle O(−1)⊕4 have been replaced by P3 ’s with normal
bundle O(−1)⊕2 . Explicitly, W can be described as Proj of the ring R(X, L) =
⊕a≥0 H 0 (X, aL) for any line bundle L which is a positive linear combination of
H + 5(− 12 KX ) and − 12 KX . On this pseudo-isomorphic variety W , the linear system

13
| − 12 KW | (= | − 12 KX |) is basepoint-free, giving a morphism W → P3 . The generic
fiber of W → P3 is a principally polarized abelian surface. The (−1)-divisors
E1 , . . . , E9 on W all induce the same polarization of the generic fiber, which is two
times a principal polarization. (Geometrically, each divisor Ei intersects the general
fiber in a curve of genus 5.) The 45 P3 ’s we have produced in W are sections of the
fibration, and their differences generate only a rank 8 subgroup of the Mordell-Weil
group, for p1 , . . . , p9 general as above.
To check that the smooth fibers of W → P3 are abelian surfaces, note that the
canonical bundle KW is trivial on all fibers, since − 21 KW is pulled back from P3 .
So the smooth fibers have trivial canonical bundle and hence (say, in characteristic
not 2 or 3) are either K3 surfaces or abelian surfaces. To see that they are abelian
surfaces, it suffices to compute that c2 (W )(− 12 KW )3 = 0, since K3 surfaces have
Euler characteristic c2 equal to 24 while abelian surfaces have c2 = 0. To do that
calculation, one can compute that c2 (X)(− 21 KX )3 = 45 and that performing an
inverse flip (replacing a P1 with normal bundle O(−1)⊕4 by a P3 with normal
bundle O(−1)⊕2 ) lowers c2 (X)(− 21 KX )3 by 1.
The following result helps to show what is going on.

Lemma 6.2 Let f : X → S be a proper morphism of smooth varieties over a field

such that KX has degree zero on all curves C with f (C) = point. Suppose that the
smooth locus of the generic fiber has a group structure (for example, the generic
fiber could be an abelian variety, or a cuspidal cubic curve in characteristic 2 or
3). Then for any sections S1 and S2 of f over the generic point of S, there is a
pseudo-automorphism of X over S that maps S1 to S2 .

Proof. We use a basic fact of minimal model theory: for any proper morphisms
fi : Xi → S, i = 1, 2, such that Xi is smooth and KXi has nonnegative degree on
all curves that map to a point in S, every birational map from X1 to X2 over S is a
pseudo-isomorphism. The proof works in any characteristic [13, Theorem 3.52], and
applies more generally to varieties with terminal singularities. Any section of the
morphism f : X → S over the generic point of S automatically lies in the smooth
locus of the generic fiber, which we assume has a group structure. Therefore, for any
sections S1 and S2 over the generic point, adding S2 − S1 using the group structure
is a pseudo-automorphism of X over S. QED
Since the Mordell-Weil group of the generic fiber of W → P3 has rank 8, it is in
particular infinite. By the fact of minimal model theory just stated, the Mordell-
Weil group acts on W , or equivalently on X, by pseudo-automorphisms. These
pseudo-automorphisms of X map the (−1)-divisor E1 to infinitely many other (−1)-
divisors. It follows that the total coordinate ring of X is not finitely generated,
and that the corresponding representation of (Ga )3 has infinitely generated ring of
invariants, as we already knew from Corollary 3.3. QED (Theorem 6.1).

7 An elliptic fibration of P3 , and representations of

(Ga )4 on A16
We now describe the geometry behind the representations of (Ga )4 in Theorem 0.1
in more detail. As in Theorem 5.2, the key concept is the Mordell-Weil group of

14
an elliptic fibration, in this case an elliptic fibration of a blow-up of P3 . For a
certain class of representations, we can show that the ring of invariants is infinitely
generated except in a very special situation. As a concrete application, we give
examples of non-finite generation where the coefficients of the representation are
very simple (in particular, simpler than the examples in section 4). As in section 4,
we get examples of non-finite generation over all fields, even the field of order 2.

Corollary 7.1 Over any field k of characteristic not 2, start with Nagata’s repre-
sentation of (Ga )8 on A16 as in Theorem 2.1, and restrict to the subgroup (Ga )4
spanned by the rows of the following 4 × 8 matrix.
 
0 0 0 0 1 1 1 1
0 0 1 1 0 0 1 1
 
0 1 0 1 0 1 0 1
1 1 1 1 1 1 1 1

Then the ring of invariants is not finitely generated over k. There are other (almost
equally simple) 16-dimensional representations of (Ga )4 over F2 for which the ring
of invariants is not finitely generated.

Corollary 7.1 is a consequence of the following theorem. Combined with A. Prendergast-

Smith’s results discussed below, Theorem 7.2 characterizes exactly which intersec-
tions of three quadrics in P3 yield infinitely generated rings of invariants (or in-
finitely generated total coordinate rings).

Theorem 7.2 Let p1 , . . . , p8 be 8 distinct rational points in P3 over a field k which

are the intersection of three quadrics. Let a be the number of coplanar quadruples
of points p1 , . . . , p8 , and define
a
ρ=7− .
2
Then ρ ≥ 0. If ρ is greater than 0, then the blow-up X of P3 at p1 , . . . , p8 has
infinitely generated total coordinate ring, and the corresponding 16-dimensional rep-
resentation of (Ga )4 over k has infinitely generated ring of invariants.

Theorem 7.2 strengthens Corollary 3.5, which proves infinite generation assum-
ing that there are no coplanar quadruples among p1 , . . . , p8 . A. Prendergast-Smith
has classified the intersections of three quadrics in P3 with ρ = 0 over an arbitrary
field, and strengthened Theorem 7.2 to show that finite generation holds if and only
if ρ = 0 [27].
Proof. It is classical that, after blowing up 8 points in P3 which are a
complete intersection of 3 quadrics, we have an elliptic fibration f : X → P2 ; a
reference is Dolgachev-Ortland [6, Theorem VI.9]. Let us prove that. We have
a net of quadrics through the given 8 points. So the vector space of sections of
− 12 KX = 2H − E1 − · · · − E8 has dimension 3, and it gives a rational map f from
X to P2 which is clearly defined outside E1 , . . . , E8 . Because our three quadrics
intersect in 8 distinct points, their intersection must be transverse at each of the 8
points, and so f is in fact a morphism X → P2 , with f ∗ H ∼ = − 12 KX . It follows that
the canonical bundle KX is trivial on all fibers of f , and so all smooth fibers are
curves with trivial canonical bundle, that is, elliptic curves. (Explicitly, all fibers

15
are complete intersections of two quadrics in P3 .) This construction also shows that
the exceptional divisors E1 , . . . , E8 are sections of f .
By Lemma 6.2, for any two sections S1 and S2 of the minimal elliptic fibration f
over the general point of P2 , adding S2 −S1 using the group structure on the general
fiber is a pseudo-automorphism of X which takes S1 to S2 . Since the exceptional
divisor E1 is a section of f and also a (−1)-divisor in Mukai’s sense, it follows that
all sections of f over the generic point of Y are (−1)-divisors in X.
Thus, suppose we can check that the general fiber E over k(P2 ) of our elliptic
fibration has infinite Mordell-Weil group. Then X has infinitely many (−1)-divisors,
and so the total coordinate ring of X is not finitely generated. Equivalently, the
ring of invariants for the 16-dimensional representation of (Ga )4 we are considering
is not finitely generated.
The Picard group of the general fiber E over the function field k(P2 ) is the
quotient of the Picard group of the 3-fold X by the classes of all irreducible divisors
in X which do not map onto P2 . Since f ∗ (H) = − 12 KX and Pic(P2 ) = ZH, the
pullback under f : X → P2 of every irreducible divisor in P2 is a multiple of
− 21 KX . So Pic(E) is the quotient of Pic(X) ∼ = Z9 by the class of − 21 KX together
2
with rF classes for each irreducible
PrF +1divisor in P whose inverse image in X has rF +1
irreducible components, say j=1 mF j DF j . Moreover, the divisors DF j for all F
and 1 ≤ j ≤ rF , together with − 12 KX , are linearly independent in Pic(X)Q (even
modulo numerical equivalence). This follows from the corresponding fact about
morphisms from a surface to a curve [1, Cor. VIII.4] by restricting the morphism
X → P2Pto the inverse image of a general line in P2 . Therefore Pic(E) P Q has rank
0
9 − 1 − rF , and the Mordell-Weil group Pic (E)Q has rank 7 − rF .
Let us analyze the subset of P2 over which f : X → P2 has reducible fibers.
Suppose that some fiber (the intersection of two quadrics Q1 and Q2 in our net)
contains a line L. Let Q3 be a quadric in our net which is not in the pencil spanned
by Q1 and Q2 . Since Q1 ∩ Q2 ∩ Q3 = {p1 , . . . , p8 } is smooth of dimension 0, Q3
must intersect L transversely in 2 points; so L must be the line through two of the
points p1 , . . . , p8 . Since Q3 does not contain L, the pencil spanned by Q1 and Q2 is
the
unique pencil in3 our net2 whose base locus contains L. Thus there are at most
8
2 = 28 fibers of P 99K P that contain a line.
So, apart from finitely many fibers, every fiber F of P3 99K P2 must be a union
of irreducible curves of degree at least 2, with some multiplicities, and with total
degree 4. The only reducible possibility is for F to be a union of two irreducible
conic curves (curves of degree 2), each with multiplicity 1. Since every conic in P3
is contained in a plane, F is contained (at least as a set) in a reducible quadric Q1 ,
the union of the two planes. Since the fiber F contains {p1 , . . . , p8 }, Q1 belongs to
the net of quadrics through {p1 , . . . , p8 }. It is clear that the intersection of Q1 with
every other quadric in our net is reducible. Equivalently, Q1 corresponds to a line
in P2 over which every fiber of X → P2 is reducible.
Thus we find something not at all clear a priori: the elliptic fibration X → P2 has
reducible fibers over the union of a finite set and finitely many lines. These lines are
in one-to-one correspondence with the reducible quadrics through p1 , . . . , p8 . Over
a general point P of each line, the fiber has exactly two irreducible components. Thus
the number rF is equal to the number of reducible quadrics through p1 , . . . , p8 .

16
 P
Since a reducible quadric is the union of two planes, the number rF is one
half the number of coplanar quadruples among p1 , . . . , p8 . (Recall from the proof
of Lemma 3.4 that p1 , . . . , p4 are coplanar if and only if p5 , . . . , p8 are coplanar.)
Thus the Mordell-Weil rank of the elliptic fibration X → P2 over the generic point
of P2 is ρ = 7 − a2 . Therefore ρ is nonnegative. If ρ is greater than 0, then the
Mordell-Weil group is infinite, X contains infinitely many (−1)-divisors, and the
total coordinate ring of X and the corresponding ring of invariants are infinitely
generated. QED (Theorem 7.2)
Proof of Corollary 7.1. Let k be any field of characteristic not 2. By
Theorem 2.1, the ring of invariants in Corollary 7.1 is the total coordinate ring of
the blow-up X of P3 at 8 points, namely the 8 points given in affine coordinates
w = 1 by (x, y, z) with x, y, z ∈ {0, 1}. These 8 points are the intersection of three
(very simple) quadric surfaces, x2 = xw, y 2 = yw, and z 2 = zw. To be precise,
it might appear that the above representation of (Ga )4 corresponds to the dual of
this arrangement of 8 points in P3 , but in fact the dual arrangement is projectively
isomorphic to the given one. That is a general property of complete intersections
of 3 quadrics in P3 , by Coble and Dolgachev-Ortland [6, Theorem III.3, Example
III.6].
Using that the field k has characteristic not 2, we count that there are exactly
12 coplanar quadruples among these 8 points, corresponding to 6 reducible quadrics
in our net: x(x − w), y(y − w), z(z − w), (x2 − xw) − (y 2 − yw) = (x − y)(x + y − w),
(x − z)(x + z − w), and (y − z)(y + z − w). Since ρ = 7 − 12 (12) = 1 is greater than
0, Theorem 7.2 shows that the Mordell-Weil group of the general fiber E of X over
k(P2 ) is infinite, and therefore the ring of invariants for the given 16-dimensional
representation of (Ga )4 , over Fp for p ≥ 3 or over Q, is not finitely generated.
The Appendix gives a simple example of a 16-dimensional representation of
(Ga )4 over F2 whose ring of invariants is not finitely generated. Corollary 7.1 is
proved. QED

8 A generalization of the Kawamata and Morrison con-

jectures on Calabi-Yau varieties
Theorems 5.2 and 7.2 relate finite generation of the total coordinate ring, in some
situations, to finiteness of a certain Mordell-Weil group. In this section, we make
a more general conjecture: the cone of curves of any variety with semi-ample an-
ticanonical bundle should be controlled by a certain group, which may be infinite.
(A line bundle is semi-ample if some positive multiple is basepoint-free.) This
would follow from a generalization of the conjectures of Kawamata and Morrison
on Calabi-Yau fiber spaces to allow klt pairs (Conjecture 8.1). We prove the conjec-
ture for smooth projective surfaces with semi-ample anticanonical bundle, the new
case being that of rational elliptic surfaces (Theorem 8.2).
For a projective morphism f : X → S of normal varieties with connected fibers,
define N 1 (X/S) as the real vector space spanned by Cartier divisors on X modulo
numerical equivalence on curves on X mapped to a point in S. Define a small
Q-factorial modification (SQM) of X over S to be a birational map f : X 99K X 0
over S, with X 0 projective over S and Q-factorial, which is an isomorphism in
codimension one. A Cartier divisor D on X is called f -nef (resp. f -movable, f -

17
effective) if D · C ≥ 0 for every curve C on X which is mapped to a point in S
(resp., if codim(supp(coker(f ∗ f∗ OX (D) → OX (D)))) ≥ 2, if f∗ OX (D) 6= 0). For an
R-divisor ∆ on a Q-factorial variety X, the pair (X, ∆) is klt if, for any resolution
π : X e → X with a simple normal crossing R-divisor ∆ e such that K e + ∆ e =
X
∗
π (KX + ∆), the coefficients of ∆ e are less than 1 [13, Definition 2.34].
The f -nef cone A(X/S) (resp. the closed f -movable cone M (X/S)) is the closed
convex cone in N 1 (X/S) generated by the numerical classes of f -nef divisors (resp.
f -movable divisors). The f -effective cone B e (X/S) is the convex cone, not nec-
essarily closed, generated by f -effective Cartier divisors. We call Ae (X/S) =
A(X/S) ∩ B e (X/S) and M e (X/S) = M (X/S) ∩ B e (X/S) the f -effective f -nef cone
and the f -effective f -movable cone, respectively. Finally, a finite rational polyhedral
cone in N 1 (X/S) means the closed convex cone spanned by a finite set of Cartier
divisors on X.

Conjecture 8.1 Let f : X → S be a projective morphism with connected fibers,

(X, ∆) a Q-factorial klt pair with ∆ effective. Suppose that KX + ∆ is numerically
trivial over S. Let Aut(X/S) and PsAut(X/S) denote the groups of automorphisms
or pseudo-automorphisms of X over the identity on S. Then:
(1) The number of Aut(X/S)-equivalence classes of faces of the cone Ae (X/S)
corresponding to birational contractions or fiber space structures is finite. Moreover,
there exists a finite rational polyhedral cone Π which is a fundamental domain for
the action of Aut(X/S) on Ae (X/S) in the sense that
(a) Ae (X/S) = ∪g∈Aut(X/S) g∗ Π,
(b) Int Π ∩ g∗ Int Π = ∅ unless g∗ = 1.
(2) The number of PsAut(X/S)-equivalence classes of chambers Ae (X 0 /S, α) on
the cone M e (X/S) corresponding to marked SQMs X 0 → S of X → S is finite.
Equivalently, the number of isomorphism classes over S of SQMs of X over S
(ignoring the birational identification with X) is finite. Moreover, there exists a
finite rational polyhedral cone Π0 which is a fundamental domain for the action of
PsAut(X/S) on M e (X/S).

Note that Conjecture 8.1 would not be true for Calabi-Yau pairs (pairs (X, ∆)
with KX + ∆ ≡ 0) that are log-canonical (or dlt) rather than klt. Let X be the
blow-up of P2 at 9 very general points. Let ∆ be the proper transform of the unique
smooth cubic curve through the 9 points; then KX + ∆ ≡ 0, and so (X, ∆) is a
log-canonical Calabi-Yau pair. The surface X contains infinitely many (−1)-curves
by Nagata [23], and so the nef cone is not finite polyhedral. But the automorphism
group Aut(X) is trivial and hence does not have a finite polyhedral fundamental
domain on the nef cone.
The conjecture also fails if we allow the R-divisor ∆ to have negative coefficients.
Let Y be a K3 surface whose cone of curves is not finite polyhedral, and let X be the
blow-up of Y at a very general point. Then (X, −E) is a klt Calabi-Yau pair, where
E is the exceptional curve. Here the cone of curves of X is not finite polyhedral,
but Aut(X) is trivial.
For X terminal and ∆ = 0, Conjecture 8.1 is exactly Kawamata’s conjecture
on Calabi-Yau fiber spaces, generalizing Morrison’s conjecture on Calabi-Yau vari-
eties [11, 19, 20]. (The group in part (2) can then be described as Bir(X/S), since
all birational automorphisms of X over S are pseudo-automorphisms when KX is

18
numerically trivial over S.) Assuming the base field has characteristic zero, Kawa-
mata’s conjecture is known for X of dimension at most 2, for X of dimension 3 with
S of positive dimension, and for a few classes of Calabi-Yau 3-folds [11, Remark
1.13]. Some evidence for Conjecture 8.1 in the case ∆ 6= 0 is provided by Coble
surfaces, smooth projective rational surfaces such that the linear system | − KX |
is empty but | − 2KX | is not empty. If there is a smooth divisor D equivalent to
−2KX , then (X, D/2) is a klt Calabi-Yau pair. Then Conjecture 8.1 is true for
(X, D/2). Indeed, the double cover of X ramified over D is a K3 surface [7, Lemma
6.2], and the conjecture for X follows from Oguiso-Sakurai [26, Corollary 1.9].
Conjecture 8.1 would have strong consequences for varieties with −KX semi-
ample, as considered in this paper. Let X be a Q-factorial klt variety with −KX
semi-ample. For any m > 1 such that −mKX is basepoint-free, let ∆ be 1/m times
a general divisor in the linear system | − mKX |; then (X, ∆) is klt [13, Lemma 5.17].
Moreover, KX + ∆ ≡ 0, and so (X, ∆) is a klt Calabi-Yau pair. Thus Conjecture
8.1 implies that conclusions (1) and (2) hold whenever −KX is semi-ample. If
−KX is ample, then the conjecture is known in characteristic zero: the nef cone is
finite rational polyhedral by the cone theorem [13], and the movable cone is finite
rational polyhedral, partitioned into the nef cones of the finitely many SQMs of X,
by Birkar-Cascini-Hacon-Mc Kernan [2].
We can now prove Conjecture 8.1 when X is a rational elliptic surface; this com-
pletes the proof for all smooth projective surfaces with −KX semi-ample. For ratio-
nal elliptic surfaces with no multiple fibers and Mordell-Weil rank 8 (the maximum
possible), Theorem 8.2 was already known, by Grassi and Morrison [9, Theorem
2.3].

Theorem 8.2 Let X be a rational elliptic (or quasi-elliptic) surface over an alge-
braically closed field. That is, X is a smooth projective surface with −KX semi-
ample such that T := Proj R(X, −KX ) has dimension 1. Then the number of
Aut(X/T )-equivalence classes of faces of the nef effective cone Ae (X) correspond-
ing to birational contractions or fiber space structures is finite. (In this case the
nef effective cone is closed, thus equal to the nef cone.) Moreover, there exists a
finite rational polyhedral cone Π which is a fundamental domain for the action of
Aut(X/T ) on Ae (X).
Also, the following are equivalent:
(1) The total coordinate ring of X is finitely generated.
(2) The nef cone of X is finite rational polyhedral.
(3) The Mordell-Weil group, Pic0 of the generic fiber of X over T , is finite.
(The Mordell-Weil group is automatically a finite-index subgroup of Aut(X/T ).)

Note that we only have to consider statement (1) in Conjecture 8.1; statement
(2) in the conjecture is vacuous for surfaces, because every movable divisor on a
surface is nef. (Or, related to that: minimal models of surfaces are unique.)
Proof. The basic point is to show that the group Aut(X/T ) has only finitely
many orbits on the set of (−1)-curves in X.
The generic fiber of X → T is a curve Xη of genus 1 over the function field of T .
The class in Pic(X) of a general fiber of X → T is −mKX for some positive integer
m. Here −KX is effective by Riemann-Roch, and so (if m > 1) there is a multiple
fiber of X → T , a curve in the class of −KX . (Here m = 1 if and only if X → T has

19
a section [4, Chapter 5, §6].) The irreducible components of the reducible fibers of
X → T , if any, are exactly the (−2)-curves on X (smooth curves C of genus 0 with
KX · C = 0). The Picard group Pic(Xη ) is the quotient of Pic(X) by −KX together
with all the (−2)-curves. The degree of a line bundle on X on a general fiber of
X → T is given by the intersection number with −mKX , and so the Mordell-Weil
group G := Pic0 (Xη ) is the subquotient of Pic(X) given by

G = (KX )⊥ /(−KX , (−2)-curves).

An element x of the group G acts by a translation on the curve Xη of genus

1 and hence by an automorphism ϕx on X, by Lemma 6.2. This gives an action
of G on Pic(X). We know how translation by an element x of Pic0 (Xη ) acts on
Pic(Xη ): by ϕx (y) = y + deg(y)x. Since −mKX ∈ Pic(X) is the class of a general
fiber of X → T , this means that ϕx acts on Pic(X) by ϕx (y) = y − (mKX · y)x
(mod − KX , (−2)-curves). Using that the action of G preserves the intersection
product, we compute the action of certain elements of G on Pic(X) by:
 
ϕx (y) = y − (mKX · y)x + y · (−x + (1/2)(x · x)mKX ) (−mKX )

for all x ∈ KX ⊥ with x · C = 0 for all (−2)-curves C , and all y ∈ Pic(X).

i i
Now let E1 and E2 be any two (−1)-curves on X such that E1 · Ci = E2 · Ci for
all (−2)-curves Ci and such that E1 ≡ E2 (mod mPic(X)). Let x = (E2 − E1 )/m ∈
Pic(X). Then x is in KX ⊥ , we have x · C = 0 for all (−2)-curves C , and
i i
 
ϕx (E1 ) = E1 − (mKX · E1 )x + E1 · (−mx + (1/2)(mx · mx)KX ) (−KX )
= E2 .

Every (−1)-curve E has E · (−mKX ) = m, which says that E is a multisection

of degree m of X → T . Therefore 0 ≤ E · Ci ≤ m for each (−2)-curve Ci . So the
(−1)-curves E are divided into finitely many classes according to the intersection
numbers of E with all (−2)-curves and the class of E in Pic(X)/m. By the previous
paragraph, the (−1)-curves on X fall into finitely many orbits under the action of
G. A fortiori, Aut(X/T ) has finitely many orbits on the set of (−1)-curves.
We now describe all the extremal rays of the cone of curves N E(X). By the
cone theorem [12, 2.1.1, 2.3.3], all the KX -negative extremal rays of N E(X) are
spanned by (−1)-curves. Since −KX is nef, it remains to describe N E(X) ∩ KX ⊥ . It

is a general fact, for any smooth projective surface X, that any extremal ray R>0 · x
of N E(X) with x2 < 0 is spanned by a curve C with C 2 < 0, and that such an
extremal ray is isolated. Both statements follow from the fact that any two distinct
curves C and D on X with C 2 < 0 and D2 < 0 have C · D ≥ 0 and hence are “far”
from each other.
For the rational elliptic surface X, the Hodge index theorem gives that the
intersection pairing on KX ⊥ is negative semidefinite, with x2 = 0 only on the line

spanned by −KX . Therefore, using the fact in the previous paragraph, all extremal
rays of N E(X) ∩ KX ⊥ are spanned by either −K
X or a (−2)-curve. There are
only finitely many (−2)-curves (the irreducible components of reducible fibers of
X → T ), and so the cone N E(X) ∩ KX ⊥ is finite rational polyhedral. We conclude

that every extremal ray of N E(X) is spanned by either a (−1)-curve or one of the
finitely many (−2)-curves (or −KX , if X contains no (−2)-curve).

20
 Since Aut(X/T ) has only finitely many orbits on (−1)-curves, it follows that
Aut(X/T ) has only finitely many orbits on the extremal rays of N E(X). Moreover,
if X contains infinitely many (−1)-curves, then the only possible limit ray of (−1)-
rays is R>0 (−KX ). Indeed, every (−1)-curve E has −KX · E = 1 and E 2 = −1,
while there are only finitely many (−1)-curves E with any given degree H · E (a
natural number), where H is a fixed ample divisor on X. Any limit ray R>0 x of
(−1)-rays has 0 < H · x < ∞, and so it must have (−KX ) · x = 0 and x2 = 0.
Since the intersection form is negative semidefinite on (KX )⊥ with kernel spanned
by −KX , it follows that x is a multiple of −KX , as claimed.
We can deduce that the nef cone A(X) is finite rational polyhedral near any
point x in A(X) not in the ray R≥0 (−KX ). First, such a point x has x2 ≥ 0 and
also (−KX ) · x ≥ 0, since x and −KX are nef. If (−KX ) · x were equal to zero,
these properties would imply that x is a multiple of −KX ; hence we must have
(−KX ) · x > 0. As a result, there is a neighborhood V of x and a neighborhood
W of −KX such that V · W > 0. Since the only possible limit ray of (−1)-rays is
R>0 (−KX ), almost all (all but finitely many) (−1)-curves are in the cone R>0 W .
So almost all (−1)-curves have positive intersection with the neighborhood V of x.
Since almost all extremal rays of N E(X) are spanned by (−1)-curves, we conclude
that the nef cone A(X) is finite rational polyhedral near x, as claimed.
In particular, for each (−1)-curve E, the face A(X) ∩ E ⊥ of the nef cone is finite
rational polyhedral, since it does not contain −KX . So the cone ΠE spanned by
−KX and A(X) ∩ E ⊥ is finite rational polyhedral.
Let x be any nef R-divisor on X. Let c be the maximum real number such that
y := x + cKX is nef. Then x and y have the same degree on all (−2)-curves, and
so there must be some (−1)-curve E with y ∈ E ⊥ . Therefore x is in the cone ΠE .
That is, the nef cone A(X) is the union of the finite rational polyhedral cones ΠE .
Moreover, for two distinct (−1)-curves E and F on X, the intersection Int ΠE ∩
Int ΠF is empty, since E and F are linearly independent in N 1 (X). Since Aut(X/T )
has only finitely many orbits on the set of cones ΠE , there is a finite rational
polyhedral cone Π which is a fundamental domain for the action of Aut(X/T ) on
Ae (X). (We can construct such a domain using the Dirichlet construction for the
hyperbolic metric on {x ∈ N 1 (X) : x2 > 0, H · x > 0}/R>0 , restricted to the
nef cone. That is, given a point x in the interior of the nef cone, we define a
fundamental domain as the set of points of the nef cone whose distance to x, in
the nef cone modulo scalars, is at most their distance to any other point in the
Aut(X/T )-orbit of x.)
Any rational point x in the nef cone A(X) is effective. This is clear from the
Riemann-Roch theorem χ(X, nL) = χ(X, O) + ((nL)2 − KX · nL)/2, since for a nef
divisor x on our surface X, either x2 > 0, x2 = 0 and (−KX ) · x > 0, or x is a
multiple of −KX . Since the cone ΠE is finite rational polyhedral for each (−1)-curve
E, ΠE is contained in the nef effective cone Ae (X). Since the whole nef cone A(X)
is the union of the cones ΠE , the nef effective cone Ae (X) is equal to its closure,
the nef cone A(X). This proves another statement of Theorem 8.2.
Next, we have to show that the number of Aut(X/T )-orbits of faces of the
nef effective cone Ae (X) corresponding to birational contractions or fiber space
structures is finite. Dually, it suffices to show that there are only finitely many
Aut(X/T )-orbits of faces in the cone of curves N E(X). Since there are only finitely
many orbits of (−1)-curves, and almost all extremal rays of N E(X) are spanned by

21
(−1)-curves, it suffices to show that each (−1)-curve E lies on only finitely many
faces of N E(X). This follows from the dual statement that the cone A(X) ∩ E ⊥ is
finite rational polyhedral, which we have proved.
Finally, let us prove the equivalence of statements (1), (2) and (3) in Theorem
8.2. That (1) implies (2) is easy; it is also part of Hu and Keel’s characterization of
varieties with finitely generated total coordinate ring [10, Proposition 2.9]. Also, (2)
easily implies (3), as follows. Suppose that the Mordell-Weil group Pic0 (Xη ) is infi-
nite. We know that X contains at least one (−1)-curve since X is a rational elliptic
surface. The Mordell-Weil group Pic0 (Xη ) acts on X, and only a finite subgroup
of it can map any given (−1)-curve into itself, since a (−1)-curve is a multisection
of X → T of degree a positive integer m. Thus, if Pic0 (Xη ) is infinite, then X has
infinitely many (−1)-curves, and so N E(X) is not finite rational polyhedral. So the
dual cone A(X) is not finite rational polyhedral. That is, (2) implies (3).
Conversely, suppose (3), that Pic0 (Xη ) is finite. This is a finite-index subgroup
of Aut(X/T ), and so Aut(X/T ) is finite. By the earlier parts of this theorem, it
follows that the nef effective cone Ae (X) is finite rational polyhedral. Statement
(2) is proved. It remains to prove (1), that the total coordinate ring of X is finitely
generated. By Hu-Keel’s theorem [10, Proposition 2.9], it suffices to show that
every codimension-1 face of N E(X) can be contracted. This follows from the cone
theorem (including the contraction theorem), by the same argument as in the proof
of Theorem 5.2. QED

9 Appendix: Infinitely generated rings of invariants

over F2 or F3
In this section we complete the proof of Corollaries 5.1 and 7.1 by giving simple ex-
amples of 18-dimensional representations of (Ga )6 over F2 or F3 , and 16-dimensional
representations of (Ga )4 over F2 , whose rings of invariants are not finitely generated.
Take the 9 points in P2 (F4 ) given in affine coordinates by (x, y) where x and
y run through the nonzero elements of a field F4 of order 4. These points are the
intersection of the two cubics x3 = z 3 and y 3 = z 3 , each consisting of three lines
through a point in P2 . There are 9 collinear triples among these points, and 3
partitions of them into disjoint collinear triples. (These partitions correspond to
three cubics in the pencil which are unions of three lines, x3 + z 3 , y 3 + z 3 , and
(x3 + z 3 ) + (y 3 + z 3 ) = x3 + y 3 .) Therefore ρ = 8 − 9 + 3 = 2 is greater than 0. By
Theorem 5.2, the blow-up X of P2 at these 9 points has infinitely many (−1)-curves.
(More precisely, this count shows that the Mordell-Weil rank is 2.) Thus we get an
18-dimensional representation of (Ga )6 over F4 such that the ring of invariants is
not finitely generated.
Since this pencil of cubics is defined over F2 , even though the 9 individual points
are not, this pencil gives an 18-dimensional representation of (Ga )6 over F2 such
that the ring of invariants is not finitely generated, as explained in section 4. We
will write out this representation explicitly, although the reader might be satisfied
to know that it can be done. Let ζ be a primitive cube root of unity in F4 . Our
representation of (Ga )6 on A18 := A91 ⊕ A92 over F4 can be defined by 6 commuting
linear maps 1 + Ai , 1 ≤ i ≤ 6, where Ai ∈ Hom(A92 , A91 ) are given by the following

22
diagonal matrices:

diag(1, ζ, ζ 2 , 0, 0, 0, 0, 0, 0), diag(1, 0, 0, ζ, ζ 2 , 0, 0, 0, 0),

diag(1, 0, 0, 0, 0, ζ, ζ 2 , 0, 0), diag(0, 1, 1, 1, 1, 1, 1, 0, 0),
diag(0, 0, 0, 0, 0, 1, 1, 1, 1), diag(1, 1, 1, 0, 0, 0, 0, ζ, ζ 2 ).

These 6 vectors are chosen as a basis for the kernel of the linear map A9 → A3
corresponding to our 9 points in P2 over F4 , in the order [1, 1, 1], [1, ζ, 1], [1, ζ 2 , 1],
[ζ, 1, 1], [ζ 2 , 1, 1], [ζ, ζ, 1], [ζ 2 , ζ 2 , 1], [ζ, ζ 2 , 1], [ζ 2 , ζ, 1]. We can define a form of
this representation over F2 by 6 commuting linear maps 1 + Ai , 1 ≤ i ≤ 6, where
Ai ∈ Hom(A92 , A91 ) are given by the following matrices:
   
0 1 0 1
diag(1, , 0, 0, 0, 0, 0, 0), diag(1, 0, 0, , 0, 0, 0, 0),
1 1 1 1
 
0 1
diag(1, 0, 0, 0, 0, , 0, 0), diag(0, 1, 1, 1, 1, 1, 1, 0, 0),
1 1
 
0 1
diag(0, 0, 0, 0, 0, 1, 1, 1, 1), diag(1, 1, 1, 0, 0, 0, 0, ).
1 1

This is an 18-dimensional representation of (Ga )6 over F2 with infinitely generated

ring of invariants.
A similar example works over F3 . Take the 9 points of P2 (F9 ) given in affine
coordinates by (x, y) where x and y run through the set of fourth roots of unity in F9
other than 1. These points are the intersection of the two cubics x3 + x2 z + xz 2 + z 3
and y 3 + y 2 z + yz 2 + z 3 . There are 7 lines through these points, and 2 partitions of
them into disjoint collinear triples. (The two partitions correspond to two cubics in
the pencil which are unions of three lines, x3 +x2 z +xz 2 +z 3 and y 3 +y 2 z +yz 2 +z 3 ,
and the other collinear triple corresponds to another reducible cubic in the pencil,
(x3 + x2 z + xz 2 + z 3 ) − (y 3 + y 2 z + yz 2 + z 3 ) = (x − y)(x2 + xy + y 2 + xz + yz + z 2 ).)
So ρ = 8 − 7 + 2 = 3. By Theorem 5.2, the blow-up X of P2 at these 9 points has
infinitely many (−1)-curves. This gives an 18-dimensional representation of (Ga )6
over F9 whose ring of invariants is not finitely generated.
Since this pencil of cubics is defined over F3 , we can define a form of the above
representation over F3 . Explicitly, let i be a square root of −1 in F9 . Then our
representation of (Ga )6 on A18 := A91 ⊕ A92 over F9 can be defined by 6 commuting
linear maps 1 + Aj , 1 ≤ j ≤ 6, where Aj ∈ Hom(A92 , A91 ) are given by the following
diagonal matrices:

diag(1, 1 + i, 1 − i, 0, 0, 0, 0, 0, 0), diag(1, 0, 0, 1 + i, 1 − i, 0, 0, 0, 0),

diag(1, 0, 0, 0, 0, 1 + i, 1 − i, 0, 0), diag(1, 1, 1, 1, 1, −1, −1, 0, 0),
diag(0, 0, 0, 0, 0, 1, 1, −1, −1), diag(0, −1, −1, 1, 1, −1, −1, 1 + i, 1 − i).

These 6 vectors are chosen as a basis for the kernel of the linear map A9 → A3
corresponding to our 9 points in P2 over F9 , in the order [−1, −1, 1], [−1, i, 1],
[−1, −i, 1], [i, −1, 1], [−i, −1, 1], [i, i, 1], [−i, −i, 1], [i, −i, 1], [−i, i, 1]. We can define
a form of this representation over F3 by 6 commuting linear maps 1 + Aj , 1 ≤ j ≤ 6,

23
where Aj ∈ Hom(A92 , A91 ) are given by the following matrices:
   
0 1 0 1
diag(1, , 0, 0, 0, 0, 0, 0), diag(1, 0, 0, , 0, 0, 0, 0),
1 −1 1 −1
 
0 1
diag(1, 0, 0, 0, 0, , 0, 0), diag(1, 1, 1, 1, 1, −1, −1, 0, 0),
1 −1
 
0 1
diag(0, 0, 0, 0, 0, 1, 1, −1, −1), diag(0, −1, −1, 1, 1, −1, −1, ).
1 −1

This is an 18-dimensional representation of (Ga )6 over F3 with infinitely generated

ring of invariants. Corollary 5.1 is proved.
We now finish the proof of Corollary 7.1 by exhibiting a simple example of a 16-
dimensional representation of (Ga )4 over F2 whose ring of invariants is not finitely
generated.
Let ζ be a primitive cube root of unity in F4 , and consider the 8 points in P3 (F4 )
defined in affine coordinates w = 1 by (x, y, z) where z is ζ or ζ 2 , y is 0 or z, and x is
0 or 1. These points are the complete intersection of the three quadrics x2 +xw = 0,
y 2 + yz = 0, and z 2 + zw + w2 = 0. Let X be the blow-up of P3 at these 8 points.
We have an elliptic fibration X → P2 . We count that there are exactly 10 coplanar
quadruples among the 8 points, corresponding to 5 reducible quadrics in our net:
x(x + w), y(y + z), z 2 + zw + w2 = (z + ζw)(z + ζ 2 w), y 2 + yz + z 2 + zw + w2 =
(y + w + ζz)(y + w + ζ 2 z), and x2 + xw + z 2 + zw + w2 = (x + z + ζw)(x + z + ζ 2 w).
Since ρ = 7 − 21 (10) = 2 is greater than 0, Theorem 7.2 shows that the Mordell-Weil
group of the general fiber E of X over k(P2 ) is infinite (in fact of rank 2), and
therefore the ring of invariants for the given 16-dimensional representation of (Ga )4
over F4 is not finitely generated.
Since this net of quadrics is defined over F2 , even though the 8 individual points
are not, we can define a form of the above representation over F2 . Our representation
of (Ga )4 on A16 := A81 ⊕ A82 over F4 can be defined by 4 commuting linear maps
1 + Ai , 1 ≤ i ≤ 4, where Ai ∈ Hom(A82 , A81 ) are given by the following diagonal
matrices:
diag(0, 0, 1, 1, 0, 0, 1, 1), diag(0, 0, 0, 0, ζ, ζ 2 , ζ, ζ 2 ),
diag(ζ, ζ 2 , ζ, ζ 2 , ζ, ζ 2 , ζ, ζ 2 ), diag(1, 1, 1, 1, 1, 1, 1, 1).

These 4 vectors are chosen as a basis for the kernel of the linear map A8 → A4
corresponding to our 8 points in P2 over F4 , in the order [0, 0, ζ, 1], [0, 0, ζ 2 , 1],
[1, 0, ζ, 1], [1, 0, ζ 2 , 1], [0, ζ, ζ, 1], [0, ζ 2 , ζ 2 , 1], [1, ζ, ζ, 1], [1, ζ 2 , ζ 2 , 1]. We can define a
form of this representation over F2 by 4 commuting linear maps 1 + Ai , 1 ≤ i ≤ 4,
where Ai ∈ Hom(A82 , A81 ) are given by the following matrices:
   
0 1 0 1
diag(0, 0, 1, 1, 0, 0, 1, 1), diag(0, 0, 0, 0, , ),
1 1 1 1
       
0 1 0 1 0 1 0 1
diag( , , , ), diag(1, 1, 1, 1, 1, 1, 1, 1).
1 1 1 1 1 1 1 1

This is a 16-dimensional representation of (Ga )4 over F2 with infinitely generated

ring of invariants. That completes the proof of Corollary 7.1.

24
References
[1] A. Beauville. Complex algebraic surfaces. Cambridge (1996).

[2] C. Birkar, P. Cascini, C. Hacon, and J. Mc Kernan. Existence of minimal models

for varieties of log general type. arXiv:math.AG/0610203

[3] A. Coble. Algebraic geometry and theta functions. AMS (1929).

[4] F. Cossec and I. Dolgachev. Enriques surfaces I. Birkhäuser (1989).

[5] I. Dolgachev. On certain families of elliptic curves in projective space. Ann.

Mat. Pura Appl. 183 (2004), 317–331.

[6] I. Dolgachev and D. Ortland. Point sets in projective spaces and theta functions.
Astérisque 165 (1988).

[7] I. Dolgachev and D.-Q. Zhang. Coble rational surfaces. Amer. J. Math. 123
(2001), 79–114.

[8] G. Freudenburg. A linear counterexample to Hilbert’s fourteenth problem in

dimension eleven. Proc. AMS 135 (2007), 51–57.

[9] A. Grassi and D. Morrison. Automorphisms and the Kähler cone of certain
Calabi-Yau manifolds. Duke Math. J. 71 (1993), 831–838.

[10] Y. Hu and S. Keel. Mori dream spaces and GIT. Michigan Math. J. 48 (2000),
331–348.

[11] Y. Kawamata. On the cone of divisors of Calabi-Yau fiber spaces. Int. J. Math.
8 (1997), 665–687.

[12] J. Kollár and S. Kovács. Birational geometry of log surfaces.

http://www.math.princeton.edu/~kollar

[13] J. Kollár and S. Mori. Birational geometry of algebraic varieties. Cambridge

(1998).

[14] S. Kuroda. A generalization of Roberts’ counterexample to the fourteenth prob-

lem of Hilbert. Tohoku Math. J. 56 (2004), 501–522.

[15] W. Lang. Extremal rational elliptic surfaces in characteristic p. I. Beauville

surfaces. Math. Z. 207 (1991), 429–437.

[16] W. Lang. Extremal rational elliptic surfaces in characteristic p. II. Surfaces

with three or fewer singular fibres. Ark. Mat. 32 (1994), 423–448.

[17] Y. Manin. The Tate height of points on an abelian variety: its variants and
applications. Izv. Akad. Sci. SSSR 28 (1964), 1363–1390; AMS Transl. 59
(1966), 82–119.

[18] R. Miranda and U. Persson. On extremal rational elliptic surfaces. Math. Z.

193 (1986), 537–558.

25
[19] D. Morrison. Compactifications of moduli spaces inspired by mirror symmetry.
Journées de géométrie algébrique d’Orsay (Orsay, 1992). Astérisque 218 (1993),
243–271.

[20] D. Morrison. Beyond the Kähler cone. Proceedings of the Hirzebruch 65 con-
ference on algebraic geometry (Ramat Gan, 1993), 361–376. Bar-Ilan Univ.
(1996).

[21] S. Mukai. Counterexample to Hilbert’s fourteenth problem for the 3-

dimensional additive group. RIMS preprint 1343 (2001).

[22] M. Nagata. On the fourteenth problem of Hilbert. Proc. ICM Edinburgh (1958),
459–462.

[23] M. Nagata. On rational surfaces. II. Mem. Coll. Sci. Univ. Kyoto Ser. A Math.
33 (1960/1961), 271–293.

[24] V. Nikulin. A remark on algebraic surfaces with polyhedral Mori cone. Nagoya
Math. J. 157 (2000), 73–92.

[25] V. Nikulin. On algebraic varieties with finite polyhedral Mori cone. The Fano
conference, 573–589. Univ. Torino (2004).

[26] K. Oguiso and J. Sakurai. Calabi-Yau threefolds of quotient type. Asian J.

Math. 5 (2001), 43–77.

[27] A. Prendergast-Smith. Finite generation of the total coordinate ring for certain
rational threefolds. Preprint (2008).

[28] J.-P. Serre. Cohomologie galoisienne. Lecture Notes in Mathematics 5, 5th

edition, Springer (1994).

[29] T. Shioda. On elliptic modular surfaces. J. Math. Soc. Japan 24 (1972), 20–59.

[30] R. Steinberg. Nagata’s example. Algebraic groups and Lie groups, Cambridge
(1997), 375–384. Also in Collected papers, AMS (1997).

[31] J. Tate. On the conjecture of Birch and Swinnerton-Dyer. Séminaire Bourbaki,

Vol. 9, 1964/65–1965/66, Soc. Math. de France (1995), Exposé 306, 415–440.

DPMMS, Wilberforce Road, Cambridge CB3 0WB, England

[email protected]

26