Finiteness theorems for K3 surfaces over arbitrary fields

Over an algebraically closed field, various finiteness results are known regarding the automorphism group of a K3 surface and the action of the automorphisms on the Picard lattice. We formulate and prove versions of these results over arbitrary base fields, and give examples illustrating how behaviour can differ from the algebraically closed case.


Introduction
The geometry of K3 surfaces over the complex numbers has a long history, with many results known about the cohomology, the Picard group, and the automorphism group of an algebraic K3 surface, and how these objects interact. Such results over the complex numbers carry over to other algebraically closed fields of characteristic zero, and similar results are also known over algebraically closed fields of other characteristics. For a comprehensive treatment of the geometry of K3 surfaces, we refer the reader to the lecture notes of Huybrechts [11].
K3 surfaces are also interesting from an arithmetic point of view, with much recent work on understanding the rational points, curves, Brauer groups and other invariants of K3 surfaces over number fields. In this article, we investigate the extent to which some standard finiteness results for K3 surfaces over algebraically closed fields remain true over more general base fields. In particular, we show how to define the correct analogue of the Weyl group, and give an explicit description of it. This allows us to formulate and prove finiteness theorems over arbitrary fields, modelled on those already known over algebraically closed fields. The tools we use include representability of the Picard and automorphism schemes, classification of transitive group actions on Coxeter-Dynkin diagrams, and an explicit description of the walls of the ample cone. We follow these theoretical results with several detailed examples, showing how the relationship between the Picard group and the automorphism group can be different from the geometric case. We end by proving that a general surface over Q of the form x 4 − y 4 = c(x 4 − w 4 ) has finite automorphism group.
The specific finiteness results we address go back to Sterk [18]. To state them, we need some definitions; we follow the notation of [11]. In this article, by a K3 surface we will always mean an algebraic K3 surface, which is therefore projective.
Let k be a field, and let X be a K3 surface over k. Denote the group of isometries of Pic X by O(Pic X). Reflection in any (−2)-class in Pic X defines an isometry of Pic X, and we define the Weyl group W (Pic X) ⊂ O(Pic X) to be the subgroup generated by these reflections.
We recall the definitions of the positive, ample and nef cones associated to the K3 surface X; see also [11,Chapter 8]. Let (Pic X) R denote the real vector space (Pic X) ⊗ Z R. By the Hodge index theorem, the intersection product on (Pic X) R has signature (1, ρ − 1); so the set {α ∈ (Pic X) R | α 2 > 0} consists of two connected components. The positive cone C X ⊂ (Pic X) R is the connected component containing all the ample classes.
The ample cone Amp(X) is the cone in (Pic X) R generated by all classes of ample line bundles. The nef cone Nef(X) is defined as Nef(X) = {α ∈ (Pic X) R | α · C ≥ 0 for all curves C ⊂ X}.
Finally, we define Nef e (X) to be the real convex hull of Nef(X) ∩ Pic X. An application of the criterion of Nakai-Moishezon-Kleiman shows that Amp(X) is the interior of Nef(X) and Nef(X) is the closure of Amp(X): see [11,Corollary 8.1.4].
The following finiteness theorems are due to Sterk [18] for k = C and to Lieblich and Maulik [12] when k has positive characteristic not equal to 2. As in Huybrechts [11,Chapter 8], a fundamental domain for the action of a discrete group G acting continuously on a topological manifold M is defined as the closure U of an open subset U ⊂ M such that M = ∪ g∈G gU and such that for g = h ∈ G the intersection gU ∩ hU does not contain interior points of gU or hU . Theorem 1.1. Let k be an algebraically closed field of characteristic not equal to 2, and let X be a K3 surface over k.
(1) [11,Corollary 8.2.11] The cone Nef(X) ∩ C X is a fundamental domain for the action of W (Pic X) ⊂ O(Pic X) on the positive cone C X . The action of Aut X on Nef e X admits a rational polyhedral fundamental domain. (4) [11,Corollary 8.4.6] The set of orbits under Aut X of (−2)-curves on X is finite. More generally, for any d there are only finitely many orbits under Aut X of classes of irreducible curves of self-intersection 2d.
In Section 3 we will prove analogues of the various statements of Theorem 1.1 when k is replaced by an arbitrary base field of characteristic different from 2.
A consequence of Theorem 1.1 (2) is that the finiteness of Aut X depends only on the lattice Pic X. Those possible Picard lattices for which O(Pic X)/W (Pic X) is finite have been classified [14]. Over an arbitrary base field we will see that, instead of using the Weyl group W (Pic X), we must use the Galois-invariant part of the geometric Weyl group. This means that the finiteness of Aut X is no longer determined purely by the Picard lattice Pic X; rather, it depends on the geometric Picard lattice together with the Galois action. In Section 4, we give several examples that illustrate this difference to the classical case.

Lemmas on lattices
In this section we will study lattices with the action of a finite group. Given a lattice Λ with the action of a finite group H, we consider the group of automorphisms that commute with H, and the group of automorphisms that preserve the sublattice fixed by H. be the subgroup of Λ consisting of the elements fixed by every element of H (note that according to our conventions Λ H may not be a lattice, because the quadratic form on Λ may be degenerate when restricted to L H ).
Our goal is to prove the following proposition. Our interest in this situation arises from the geometry of K3 surfaces. Let X be a projective K3 surface defined over a field F , and let K/F be a Galois extension. Let Λ = Pic X K with the intersection pairing, and let H be the image of Gal(K/F ) in O(Λ). If X has a rational point over F , we have Pic X F = Λ H . (See Section 3 for more details. This statement holds slightly more generally: for example, if F is a number field and X has points everywhere locally over F .) The Hodge index theorem states that Λ has signature (1, n) and Λ H has signature (1, m): therefore M ⊥ is definite.
Before giving the proof we first collect a few helpful statements.
Proof. Since the pairing on Λ is nondegenerate, the same is true for the pairing on Λ ⊗ Q. Therefore Finally we prove (4). Choose φ in the kernel and h ∈ H, and let x ∈ Λ. We will view φ and h as automorphisms of Λ⊗Q. In Λ⊗Q we may write x = x 1 +x 2 , where In part (2) of the lemma just proved we showed that the first map has image of finite index. The second map is surjective, so the composition has image of finite index as well.
We now suppose that M ⊥ is definite to prove the second statement. Then is a composition of an injective map with a map with finite kernel and so its kernel is finite.

Finiteness results for K3 surfaces
In this section we formulate and prove analogues of the statements of Theorem 1.1 when k is an arbitrary field. We first look at the case of k separably closed, which is straightforward.
Lemma 3.1. Let k be a separably closed field, and letk be an algebraic closure of k. Let X be a K3 surface over k, and letX be the base change of X tok. Then the natural maps Pic X → PicX and Aut X → AutX are isomorphisms.
Proof. As X is projective, the Picard scheme Pic X/k exists, is separated and locally of finite type over k, and represents the sheaf Pic (X/k)(ét) , which is defined to be the sheafification on the bigétale site over k of the presheaf T → Pic(X × k T )/ Pic T (see [8,Theorem 9.4.8]). In particular, because k andk are both separably closed, we have Pic X/k (k) = Pic X and Pic X/k (k) = PicX. From H 1 (X, O X ) = 0 it follows from [8,Theorem 9.5.11] that Pic X/k isétale over k. So everyk-point of Pic X/k is defined over k, and Pic X → PicX is an isomorphism.
The functor taking a k-scheme S to the group Aut(X × k S) is represented by a scheme Aut X/k : see [8,Theorem 5.23]. A standard argument in deformation theory shows that the tangent space at the identity element is isomorphic to H 0 (X, T X ), where T X denotes the tangent sheaf on X. Indeed, an element of the tangent space is given by a morphism S = Spec k[ǫ]/(ǫ 2 ) → Aut X/k extending the morphism sending Spec k to the identity automorphism. Such a morphism corresponds to an automorphism of X × k S restricting to the identity on the central fibre. By [8,Theorem 8.5.9], the set of these morphisms forms an affine space under H 0 (X, T X ). In our case, the group H 0 (X, T X ) is zero [11,Theorem 9.5.1], so the scheme Aut X/k isétale over k, and Aut X → AutX is an isomorphism. Proof. LetC be a (−2)-curve onX. Then Lemma 3.1 shows that there is a line bundle L on X whose base change toX is isomorphic to OX (C). The Riemann-Roch theorem and flat base change give h 0 (X, L) = h 0 (X, OX (C)) = 1. So all nonzero sections of L cut out the same divisor C ⊂ X and the base change of C tō X must coincide withC. In other words,C is defined over k.
We now pass to the case of a general field. Let k be a field; fix an algebraic closurek of k, and let k s be the separable closure of k ink. Let X be a K3 surface over k, and let X s andX denote the base changes of X to k s andk, respectively. Write Γ k = Gal(k s /k).
The group Γ k acts on Pic X s preserving intersection numbers, giving a representation Γ k → O(Pic X s ). Let Γ k act on O(Pic X s ) by conjugation, that is, such that ( σ f )(x) = σ(f (σ −1 x)) for all x ∈ Pic X s . For a (−2)-class α ∈ Pic X s , denote the reflection in α by r α ; then we have ( σ r α ) = r σα . So the action of Γ k on O(Pic X s ) restricts to an action on W (Pic X s ).
Recall that Pic X is contained in, but not necessarily equal to, the fixed subgroup (Pic X s ) Γ k . The Hochschild-Serre spectral sequence gives rise to an exact sequence If X has a k-point, then evaluation at that point gives a left inverse to Br k → Br X, showing that Pic X → (Pic X s ) Γ k is an isomorphism. In general this does not have to be true. However, because Br k is torsion and Pic X s is finitely generated, the above sequence shows that Pic X is of finite index in (Pic X s ) Γ k .
It is easy to see that the action of R X on Pic X s preserves (Pic X s ) Γ k , but it is not immediately obvious that this action preserves Pic X. To show that this is the case, we use an explicit description of R X provided by a theorem of Hée and Lusztig, for which Geck and Iancu gave a simple proof. Before stating their theorem, we establish some notation and conventions for Coxeter systems.
Definition 3.4. Let W be a group generated by a set T ⊂ W of elements of order 2. For t i , t j ∈ T , let n i,j = n j,i be the order of t i t j if t i t j has finite order and 0 otherwise. Suppose that the relations t 2 i = 1, (t i t j ) ni,j = 1 for i, j with n i,j = 0 are a presentation of W . Then (W, T ) is a Coxeter system.
Let G be a graph with vertices T and such that t i , t j are adjacent in G if and only if t i does not commute with t j ; in this case, label the edge joining t i to t j with n i,j − 2 for n i,j > 0 and 0 otherwise. We refer to G as the Coxeter-Dynkin diagram of (W, T ). The Coxeter system (W, T ) is said to be irreducible if its Coxeter-Dynkin diagram is connected.
Let the length ℓ(w) of an element w ∈ W be the length of a shortest word in the t i that represents it. If W is finite, there is w 0 ∈ W such that ℓ(w 0 ) > ℓ(w) for all w = w 0 ∈ W (see [2, Proposition 2.3.1]). We refer to w 0 as the longest element of W .
Let σ be a permutation of T . Then there is at most one way to extend σ to a homomorphism W → W , because T generates W . If there is such an extension, it is an automorphism, because σ −1 extends to its inverse, and we speak of it as the automorphism induced by σ.
If (W, T ) is a Coxeter system, and I is a subset of T , let W I denote the subgroup of W generated by the elements of I. Then (W I , I) is a Coxeter system: see 9], Theorem 1). Let (W, T ) be a Coxeter system. Let G be a group of permutations of T that induce automorphisms of W . Let F be the set of orbits I ⊂ T for which W I is finite, and for I ∈ F let w I,0 be the longest element of (W I , I). Then (W G , {w I,0 : I ∈ F }) is a Coxeter system.
We will apply this theorem with W = W (Pic X s ) and T being the set of reflections in (−2)-curves on X s . Proposition 3.6. Let F be the set of Galois orbits I of (−2)-curves on X s of the following two types: (i) I consists of disjoint (−2)-curves; (ii) I consists of disjoint pairs of (−2)-curves, each pair having intersection number 1.
Then the following statements hold.
(1) For each I ∈ F , let W I be the subgroup of W (Pic X s ) generated by reflections in the classes of curves in I, and let r I be the longest element of the Coxeter system (W I , I). Then (R X , {r I : I ∈ F }) is a Coxeter system. (2) For each I ∈ F , let C I ∈ (Pic X s ) Γ k be the sum of the classes in I. Then r I acts on (Pic X s ) Γ k as reflection in the class C I . (3) The action of R X on Pic X s preserves Pic X.
Proof. Let I be a Galois orbit of (−2)-curves, and suppose that the subgroup W I is finite. We will show that I is of one of the two types described. Firstly, no two (−2)-curves in I have intersection number greater than 1, for then the corresponding reflections would generate an infinite dihedral subgroup of W I . Since W I is finite, its Coxeter-Dynkin diagram is a finite union of trees [2, Exercise 1.4]. In particular, it contains a vertex of degree ≤ 1. However, the Galois group Γ k acts transitively on the diagram, so we conclude that either every vertex has degree 0, or every vertex has degree 1. These two possibilities correspond to the two types of orbits described. Now (1) follows from Theorem 3.5.
We prove (2) separately for the two types of orbits. In the first case we have I = {E 1 , . . . , E r }. The reflections in the E i all commute, so W I is isomorphic to the Coxeter group A r 1 . The longest element is r I = s E1 • · · · • s Er . For D ∈ (Pic X s ) Γ k , the intersection numbers D · E i are all equal, and one calculates r I (D) = D + (D · E 1 )(E 1 + · · · + E r ), that is, r I coincides on (Pic X s ) Γ k with reflection in the class C I = E 1 + · · · E r , of self-intersection −2r.
In the second case, write I = {E 1 , E ′ 1 , . . . , E r , E ′ r }, where E i · E ′ i = 1 and the other intersection numbers are all zero. The two reflections s Ei and s E ′ i generate a subgroup isomorphic to the Coxeter group A 2 , in which the longest element is 2 and the longest element in W I is the product of the longest elements in the factors A 2 , that is, it equals Finally, each class C I is, by construction, the class of a Galois-fixed divisor on X s , so lies in Pic X. So in both cases the formula for r I given above shows that reflection in C I preserves Pic X, and therefore the action of R X preserves Pic X, proving (3).
We now turn to the ample and nef cones. As ampleness is a geometric property, and the nef cone is the closure of the ample cone over any base field, it follows that Amp(X) = Amp(X) ∩ (Pic X) R and Nef(X) = Nef(X) ∩ (Pic X) R , the intersections taking place inside (PicX) R . The following result is well known when k is algebraically closed (see [11,Corollary 8.2.11]), and descends easily to arbitrary k. Proposition 3.7. Let X be a K3 surface over k. The cone Nef(X) ∩ C X is a fundamental domain for the action of R X on the positive cone C X , and this action is faithful.
Proof. We will prove two things: first, that every class in C X is R X -equivalent to an element of Nef(X) ∩ C X ; and second, that the translates of Nef(X) ∩ C X by two distinct elements of R X meet only along their boundaries. The second of these shows in particular that the action is faithful. (When we refer to the boundary of Nef(X) or one of its translates, we mean the boundary within (Pic X) R . The boundary of Nef(X) ∩ C X in C X is just the boundary of Nef(X) intersected with C X , so that distinction is not so important.) To prove the first statement, let D ∈ C X . Suppose first that D has trivial stabilizer in W (PicX). Then, by [11,Corollary 8.2.11], there exists a unique g ∈ W (PicX) = W (Pic X s ) such that gD lies in the interior of Nef(X) ∩ CX . We claim that g lies in R X . For any σ ∈ Γ k , we have since the Galois action preserves the properties of being nef and positive. By uniqueness of g, we conclude that g = σ g, that is, g lies in R X . It then follows that gD lies in ((Pic X s ) R ) Γ k = (Pic X) R and therefore in Nef(X) ∩ C X . Now suppose that D ∈ C X has non-trivial stabilizer. Then D lies on at least one of the walls defined by the action of W (PicX) on (PicX) R ; see [11,Section 8.2]. By [11,Proposition 8.2.4], the chamber structure of this group action is locally polyhedral within CX , so a small enough neighbourhood of D meets only finitely many chambers. Also note that Pic X is not contained in any of the walls, because X admits an ample divisor. This allows us to construct a sequence (D i ) ∞ i=1 of elements of C X , tending to D and all lying in the interior of the same chamber of CX . As in the previous paragraph, there is a unique g ∈ R X satisfying gD i ∈ Nef(X) ∩ C X for all i. By continuity, gD also lies in Nef(X) ∩ C X .
We now prove the second statement, that Nef(X) ∩ C X intersects the translate by any non-trivial element of R X only in its boundary. Suppose that x ∈ C X lies in the intersection Nef(X) ∩ g Nef(X), for some non-trivial g ∈ R X . By [11,Corollary 8.2.11], we see that x lies in the boundary of Nef(X). The following lemma shows that x lies in the boundary of Nef(X).
Lemma 3.8. Let V be a real vector space, let C ⊂ V be a closed convex cone, and let W ⊂ V be a subspace having non-empty intersection with the interior of C.
Proof. The dual cone C * ⊂ V * is defined by By the supporting hyperplane theorem, C * has the property that a point x ∈ C lies in ∂ V (C) if and only if there exists a non-zero φ ∈ C * satisfying φ(x) = 0. (The hyperplane φ = 0 is called a supporting hyperplane of C at x.) Let f : V * → W * be the natural restriction map; then the dual (C ∩ W ) * is equal to f (C * ) (see [15,Corollary 16.3.2]).
Let x be a point of ∂ V (C) ∩ W . Then there is a supporting hyperplane to C at x, that is, there exists a non-zero φ ∈ C * satisfying φ(x) = 0. The condition that W meet the interior of C implies that φ does not vanish identically on Remark 3.9. We can also give an explicit description of the walls of Nef(X) ∩ C X . According to [11, Corollary 8.1.6], a class in C X lies in Nef(X) if and only if it has non-negative intersection number with every (−2)-curve on X s , or, equivalently, with every Galois orbit of (−2)-curves. The question is to determine which Galois orbits are superfluous, and which actually define walls of Nef(X) ∩ C X .
Let I be a Galois orbit of (−2)-curves on X s , and suppose that the subgroup W I ⊂ W (Pic X s ) generated by reflections in the elements of I is finite, that is, I is as described in Proposition 3.6. The longest element of W I acts on Pic X by reflection in the class C I = E∈I E, which has negative self-intersection. The hyperplane orthogonal to C I is a wall of Nef(X) ∩ C X : by the same argument as in [18], given a class D ∈ Amp(X), the class is orthogonal to C I but has positive intersection number with all (−2)-curves outside I.
On the other hand, let I be a Galois orbit of (−2)-curves such that the subgroup W I is infinite. Then, for a (−2)-curve C ∈ I, the hyperplane orthogonal to C in Pic(X s ) R does not meet Nef(X) ∩ C X , as the following argument shows. Let x be a class in C X orthogonal to C; then x is also orthogonal to all the other curves in I, and so is fixed by W I . Therefore x is also orthogonal to the infinitely many images of C under the action of W I , contradicting the fact that the chamber structure induced by W (Pic X s ) on C s X is locally polyhedral. Next we would like to prove an analogue of Theorem 1.1 (2). However, while the authors do not have an explicit counterexample, there seems to be no reason for the image of R X in O(Pic X) to be normal. Instead, we will see that there is a natural homomorphism from a semidirect product Aut X ⋉ R X to O(Pic X) having finite kernel and image of finite index.
Note that the natural action of Aut X ⊂ AutX on O(PicX) by conjugation fixes W (PicX) and commutes with the Galois action, so fixes R X . This gives an action of Aut X on R X , and a homomorphism from the associated semidirect product Aut X ⋉ R X to O(PicX). Since Aut X and R X both fix Pic X, we also obtain a homomorphism Aut X ⋉ R X → O(Pic X). Proposition 3.10. Let X be a K3 surface over a field k of characteristic different from 2. Then the natural map has finite kernel and image of finite index.
Remark 3.11. In the literature, this statement appears in various different but equivalent forms.
• Because the action of Aut X fixes the ample cone, the image of Aut X in O(Pic X) meets R X only in the identity element. Therefore the finiteness of the kernel in Proposition 3.10 is equivalent to the finiteness of the kernel of Aut X → O(Pic X). • Let Aut s X ⊂ Aut X be the subgroup of symplectic automorphisms. Over an algebraically closed field k of characteristic zero, the induced map from Aut s X to O(Pic X) is injective, so that Aut s X ⋉ W (Pic X) can be viewed as a subgroup of O(Pic X). Instead of our Proposition 3.10, Huybrechts [11,Theorem 15.2.6] makes the statement that Aut s X ⋉ W (Pic X) has finite index in O(Pic X). This is equivalent to our formulation, because Aut s X is of finite index in Aut X. However, when k is not algebraically closed there is no reason for Aut s X → O(Pic X) to be injective, so there is no advantage to stating the result in terms of Aut s X. • Lieblich and Maulik [12, Theorem 6.1] define Γ ⊂ O(Pic X) to be the subgroup of elements preserving the ample cone, and then show that Aut X → Γ has finite kernel and cokernel. Over an algebraically closed base field k, the subgroups Γ and W (Pic X) generate O(Pic X): any element of O(Pic X) permutes the (−2)-classes in Pic X and therefore takes the ample cone to one of its translates under W (Pic X); so composing with an element of W (Pic X) gives an element of Γ. Therefore that condition is also equivalent to the condition of Proposition 3.10.
Before proving Proposition 3.10, we first state two lemmas which are well known in the case of abelian groups. Proof. The kernel of f G is contained in the kernel of f , so is finite. For the statement about the image, consider the short exact sequence The associated exact sequence of cohomology gives Proof. We leave this as an exercise for the reader.
Proof of Proposition 3.10. By [11,Theorem 15.2.6] in characteristic zero, or [12] in characteristic p > 2, the natural map AutX → O(PicX)/W (PicX) has finite kernel and image of finite index. Lemma 3.1 shows that the same is true for The action of Γ k on all of these groups factors through a finite quotient, so Lemma 3.12 shows that the induced homomorphism also has finite kernel and image of finite index. Because the automorphism group functor is representable, we have (Aut X s ) Γ k = Aut X.
There is an exact sequence Therefore, by Lemma 3.13, the map has finite kernel and image of finite index. Since the image of Aut X in O(Pic X s ) Γ k meets R X only in the identity element, this shows that the natural map also has finite kernel and image of finite index. We apply Proposition 2.2 with Λ = Pic X s and H being the Parts (1) and (2) has finite kernel and image of finite index; by Lemma 3.13 so does the map Composing with (2) and applying Lemma 3.13 shows that has finite kernel and image of finite index. Since the actions of both Aut X and R X on (Pic X s ) Γ k preserve Pic X, this last map factors as As the second map in this composition is clearly injective, Lemma 3.13 shows that the first one has finite kernel and image of finite index. Finally, Pic X is of finite index in (Pic X s ) Γ k , so its orthogonal complement in the latter is trivial. Lemma 2.3 (2) shows that the natural map is injective, and its image has finite index. Composing with the first map of (3) and applying Lemma 3.13 again, we deduce that Aut X ⋉ R X → O(Pic X) has finite kernel and image of finite index.
Remark 3.14. The finiteness of the kernel can also be proved directly, by the same proof as over an algebraically closed field.
Having proved Proposition 3.10, we can deduce the remaining results exactly as in the classical case. Define the cone Nef e (X) to be the real convex hull of Nef(X) ∩ Pic X. Proof. This is as in the case of k = C; we briefly recall the argument of Sterk [18], making the necessary adjustments.
Let Λ be a lattice of signature (1, ρ − 1) and let Γ ⊂ O(Λ R ) be an arithmetic subgroup (for example, a subgroup of finite index in O(Λ)). Let C ⊂ Λ R be one of the two components of {x ∈ Λ R | (x · x) < 0}; this is a self-adjoint homogeneous cone [1,Remark 1.11]. Let C + be the convex hull of C ∩ Λ Q . Pick any y ∈ C ∩ Λ; the argument of [18, p. 511] shows that the set is rational polyhedral (for this, the reference to Proposition 11 of the first edition of [1] has become Proposition 5.22 in the second edition) and is a fundamental domain for the action of Γ on C + . (Sterk does not explicitly prove that two translates of Π intersect only in their boundaries, but this is easy to show from the description above.) In our case, applying this with Λ = Pic X and Γ being the image of Aut X ⋉ R X → O(Pic X) gives a rational polyhedral fundamental domain Π for the action of Aut X ⋉ R X on (C X ) + .
If we choose y to be an ample class in Pic X, then the resulting Π is contained in Nef(X), as we now show. Let I be a Galois orbit of (−2)-curves on X s such that the corresponding group W I ⊂ W (Pic X s ) is finite. Proposition 3.6 states that the longest element w of W I acts on Pic X as reflection in the class C I = E∈I E, and that these elements generate R X . Taking γ = w in the definition of Π shows that Π is contained in the half-space {x | x.C I ≥ 0}. As this holds for all such I, Remark 3.9 shows that Π is contained in Nef(X).
We conclude as in [18]. If x is a class in Nef e (X) then, since Π is a fundamental domain for the action of Aut X ⋉ R X on (C X ) + , we can find φ ∈ Aut X and r ∈ R X such that rφ(x) lies in Π. But now φ(x) and rφ(x) both lie in Nef e (X), so they are equal and lie in Π. This shows that Π is a fundamental domain for the action of Aut X on Nef e (X). (1) There are only finitely many Aut X-orbits of k-rational (−2)-curves on X.
(2) There are only finitely many Aut X-orbits of primitive Picard classes of irreducible curves on X of arithmetic genus 1.
(3) For g ≥ 2, there are only finitely many Aut X-orbits of Picard classes of irreducible curves on X of arithmetic genus g.
Proof. Let Π be a rational polyhedral fundamental domain for the action of Aut X on Nef e (X), as in Corollary 3.15. Every k-rational (−2)-curve on X defines a wall of Nef(X), by Remark 3.9. Since Π meets only finitely many walls of Nef(X), this proves (1). Gordan's Lemma states that the integral points of the dual cone of a rational polyhedral convex cone form a finitely generated monoid. Applying this to the dual cone of Π, let D 1 , . . . , D r be a minimal set of generators for Π ∩ Pic X. Since these all lie in Nef(X), we have D i · D i ≥ 0 for all i, and D i · D j > 0 for i = j. As observed in [18], this implies that, for any n > 0, there are only finitely many classes in Π ∩ Pic X of self-intersection n; and there are only finitely many primitive classes in Π ∩ Pic X of self-intersection zero. The class of an irreducible curve of arithmetic genus g ≥ 1 has self-intersection 2g − 2 and therefore lies in Nef(X), so this proves (2) and (3).
Remark 3.17. It is not true that every irreducible curve on X of arithmetic genus 0 is a k-rational (−2)-curve. However, there are not many possibilities, as we now show. Let C be such a curve, and let C 1 , . . . , C r be the geometric components of C; the Galois group Γ k acts transitively on them. In order to achieve C 2 = −2, we must have C 2 i < 0 for all i, so each C i is a (−2)-curve. Consider the intersection matrix (C i · C j ): the sum of the entries in the matrix is C 2 = −2, and the Galois action shows that every row sum is the same. Therefore the number r of rows divides 2, and there are only two options: r = 1, so that C is a rational (−2)-curve; or r = 2, and C = C 1 ∪ C 2 is the union of two conjugate (−2)-curves meeting transversely in a single point. By Corollary 3.15, both types of curves define walls of Nef(X), so in fact both types fall into finitely many orbits under the action of Aut(X).
Remark 3.18. In the case g = 1, the condition that the class be primitive cannot be omitted, for the following reason. Take for example k = Q, and suppose that X admits an elliptic fibration π : X → P 1 . A general fibre of such a fibration is a smooth, geometrically irreducible curve E of genus 1, whose class in Pic X has self-intersection 0 and is primitive. (In fact, the class is even primitive in PicX.) If s ∈ P 1 is a point of degree m > 1, then in general the fibre π −1 (s) will be an irreducible curve on X of arithmetic genus 1, linearly equivalent to mE. As m varies, this construction gives infinitely many such classes that are clearly not Aut X-equivalent.

Examples
In this section we give three examples that illustrate some of the theory developed up to this point. First, we will give a K3 surface T with finite automorphism group, even though all K3 surfaces X overQ with Pic X ∼ = Pic T have infinite automorphism group. Second, we will construct a surface U with finite automorphism group, even though, for all extensions k/Q, all K3 surfaces X overQ with Pic X ∼ = Pic U k have infinite automorphism group. For a third example, we will prove that the quartic surface in P 3 defined by x 4 − y 4 = c(z 4 − w 4 ) over a field k of characteristic 0 has finite automorphism group when c ∈ k is such that the Galois group of the field of definition of the Picard group has degree 16, the largest possible.
4.1. First example. We construct a surface T such that AutT is finite, and thus so is Aut T , while any K3 surface overQ having the same Picard lattice as T has infinite automorphism group. This is in contrast to the situation over an algebraically closed field, where finiteness of the automorphism group depends only on the isomorphism type of the Picard lattice.
Let M and N be the block diagonal matrices where I 4 is the 4 × 4 identity matrix. Let L N be a lattice with basis (e 1 , . . . , e 6 ) and Gram matrix N with respect to that basis. Let L M ⊂ L N be the sublattice generated by e 1 , e 2 and e 3 + e 4 + e 5 + e 6 . The Gram matrix for L M with respect to this basis is M . The surface T that we will construct will have compatible isomorphisms PicT ∼ = L N and Pic T ∼ = L M . Proof. There is an obvious embedding of the hyperbolic lattice U with Gram matrix ( 0 1 1 0 ) into L M ∼ = Pic X. By [11,Remark 11.1.4], this implies that there is an elliptic fibration π : X → P 1 with a section. Let E be the class of a fibre of π and S the class of a section; then E and S generate a sublattice of Pic X isomorphic to U (though not necessarily the obvious one). If π were to have a reducible fibre, then any component of that fibre other than the one meeting S would lie in E, S ⊥ and have self-intersection −2. However, E, S is a unimodular lattice of rank 2, so E, S ⊥ has the same determinant as Pic X and rank 2 less. Thus it is generated by a single vector of norm −8 and so there are no reducible fibres. It follows by the Shioda-Tate formula [11,Corollary 11.3.4] that the Mordell-Weil group of the fibration has rank 1. Translation by a non-torsion section gives an automorphism of X of infinite order. In contrast with Proposition 4.1, we will see that there exist K3 surfaces over Q with Picard lattice isomorphic to L M for which the automorphism group is finite. Proposition 4.3. Let X be a K3 surface overQ with an elliptic fibration π : X → P 1 that has a section. Suppose that π has four fibres each of type either I 2 or III. Then the following statements are equivalent.
(1) The rank of Pic X is at most 6.
(2) The Picard lattice Pic X is isomorphic to L N .
(3) The Mordell-Weil group associated to π is trivial, and there are only four reducible fibres. If these equivalent conditions hold, then Aut X is finite.
Proof. The equivalence of (2) and (3) is proved as follows. By [11, Proposition 11.3.2], there is an exact sequence where A ⊂ Pic X is the subgroup generated by vertical divisors and a chosen section, and G is the Mordell-Weil group of the elliptic fibration. Let E be a fibre of π, let S be a chosen section, and let W 1 , . . . , W 4 be the components of the four given fibres that do not meet S. ] lie in A and have intersection matrix equal to N , so they generate a sublattice of A isomorphic to L N . If G is trivial and there are no other reducible fibres, then these six classes generate Pic X and so we have Pic X ∼ = L N . Conversely, if Pic X is isomorphic to L N then these six classes must generate Pic X, so G is trivial and there are no further reducible fibres.
The implication (2)⇒(1) is trivial; we now prove (1)⇒(2). As above, we have an embedding L N ֒→ Pic X, so Pic X must have rank exactly 6. Since Pic X has rank 6 this embedding has finite index, and we must prove it to be an isomorphism. The determinant of N is −16; the square of the index [Pic X : L N ] must divide this, so the index is 1, 2, or 4. If it is not 1, there is some element of L N that can be divided by 2 in Pic X but not in L N . We take this element to be of the form where all the coefficients are 0 or 1 and not all are 0. Then a = b = 0, for otherwise the intersection number with [S] or [E] would be odd; and all the c i must be equal, because the self-intersection of any divisor on a K3 surface is even and hence that of any divisor that can be divided by 2 is a multiple of 8. Thus all c i are equal to 1. However, Lemma 3 of [13] shows that  Proposition 4.3 also shows that AutT is finite, and a fortiori that Aut T is finite.
We now construct a K3 surface T over Q satisfying the conditions of Proposition 4.4, as the Jacobian of a genus-1 fibration on a quartic R in P 3 . See [11,Section 11.4] for the properties of the Jacobian of a genus-1 fibration on a K3 surface.
Let R be a smooth quartic surface in P 3 over Q containing a line L. Projection away from L induces a morphism π L : R → P 1 whose fibres are the residual intersections with R of planes containing L. The generic fibre is a smooth curve of genus 1, and the induced morphism L → P 1 has degree 3.
Proposition 4.5. Let R be a smooth quartic surface in P 3 Q containing a rational line L and a Galois orbit L of four lines that meet L. Suppose in addition that each of the four planes containing L and a line in L meets R in one further component, which is a smooth conic. Then π L has four conjugate fibres of type I 2 or III.
Let T → P 1 be the relative Jacobian of π L , and suppose in addition that Pic(R) has rank at most 6. Then there are compatible isomorphisms PicT ∼ = L N and Pic T ∼ = L M , and Aut T is finite.
Proof. Let H be one of the four conjugate planes containing L and a line L ′ ∈ L, and let C denote the residual smooth conic. The union L ′ ∪ C is a fibre of π L . We have L ′ · C = 2, because L ′ , C are a line and a conic in the plane H. Either L ′ is tangent to C, in which case we have a fibre of type III, or they intersect in two distinct points, and the fibre is of type I 2 . The same description holds for the other three planes that are Galois conjugates of H.
The relative Jacobian T is a K3 surface [11,Proposition 11.4.5] that has the same geometric Picard number as R [11, Corollary 11.4.7 and the discussion following it], and T → P 1 has the same geometric fibres as π L [11, Chapter 11, equation (4.1)]. Now apply Proposition 4.4.
We expect that a very general surface R constructed according to this proposition will have Picard group generated by the classes of L, the lines in L and a fibre of π L . This does not imply that such a surface exists over Q, but as we will see it is not difficult to find an example. Example 4.6. We claim that the surface R ⊂ P 3 Q given by the equation  z = (α 3 − α 2 + α + 1)w, and one verifies that R ∩ H consists of L, L ′ and a smooth conic.
Let R 3 ⊂ P 3 F3 be the surface defined by the equation (4), which is smooth. LetR 3 be the base change of R 3 to an algebraic closure of F 3 , and F :R 3 →R 3 the geometric Frobenius morphism, defined by (x, y, z, w) → (x 3 , y 3 , z 3 , w 3 ). Choose a prime ℓ = 3 and let F * be the endomorphism of H 2 (R 3 , Q ℓ ) induced by F . By [19,Proposition 6.2] the Picard rank ofR is bounded by that ofR 3 , which in turn is at most the number of eigenvalues λ of F * for which λ/3 is a root of unity by [19,Corollary 6.4]. As in [19], we find the characteristic polynomial of F * by counting points on R 3 . As we already know the Galois action on the subspace of dimension 6 of H 2 (R 3 , Q ℓ ) coming from L N ⊂ PicR 3 , it suffices to count points over the fields F 3 n for n = 1, . . . , 8. We obtain point counts as in Table 1. We find that the characteristic polynomial of F * is 3 15 The polynomial p is irreducible, primitive and not monic, so its roots are not roots of unity. Thus we obtain an upper bound of 6 for the Picard rank ofR.
It must have nonnegative intersection with the known curves, which implies the inequalities Let m = min({c i }): Then ab ≥ b 2 ≥ 4m 2 . We thus find which contradicts the fact that C 2 = −2.
We now explain why we do not construct the Jacobian of R directly as a smooth surface in a projective space.
So b ≥ 3 and b + 2c i ≥ 5. We thus find Equality is attained with a = 3, b = 3, c i = 1, and this gives an ample divisor class by [11,Corollary 8.1.7].
Thus an ample divisor class cannot give an embedding into P n for n < 9.

4.2.
Second example. Now we give an example that is perhaps more surprising: a K3 surface S/Q for which Aut S is finite even though, for all field extensions L/Q, a K3 surface overQ whose Picard lattice is isomorphic to Pic S L would have infinite automorphism group. We will choose S to be a K3 surface whose Picard lattice over Q is isomorphic to L M , while overQ, and indeed over a certain quadratic extension K S , the Picard lattice is isomorphic to L N . The Galois group will act through the quotient Z/2Z by exchanging the second and third generators.
Proposition 4.11. Let S be a K3 surface in P 4 Q given as the intersection of a quadric and a cubic, containing two disjoint Galois-conjugate conics C 1 , C 2 defined over a quadratic field K S , and having Picard number 3 overQ. Then Pic S ∼ = L M and Pic SQ ∼ = L N .
Proof. Let H be a hyperplane section. Then the divisors H + C 1 + C 2 , C 1 , C 2 have the intersection matrix N , whose determinant is −40. Thus we have an embedding L N ֒→ Pic SQ whose image has index 1 or 2.   (1) Let Y be a K3 surface over an algebraically closed field having Picard lattice isomorphic to L M . Then Aut Y is infinite. In addition, if α ∈ O(Pic Y ) is of infinite order and A is a normal subgroup of O(Pic Y ) containing α then O(Pic Y )/A is finite.
(2) Let Z be a K3 surface over an algebraically closed field having Picard lattice isomorphic to L N . Then Aut Z is infinite.
Proof. We first prove (1). Working mod 5 we see that Pic Y has no vectors of norm   10) . (This generator is 19 + 6 √ 10 and it takes (x, y) to (19x + 12y, 30x + 19y).) Thus O(Pic Y ) has a subgroup of finite index isomorphic to Z, so the quotient by any infinite normal subgroup is finite.
We now turn to (2). Let D 1 , D 2 , D 3 be divisors on Z whose intersection matrix is N , and let D = D 1 − 2D 2 − D 3 . Since D 2 = 0, for some α ∈ O(L N ) there is a genus-1 fibration π with fibres of class α(D). There is no section of π (indeed, no two curves on Z have odd intersection), but there is a 2-section. The Jacobian J of π is a K3 surface of Picard number 3 [11,Corollary 11.4.7] on which the determinant of the intersection pairing is (det N )/2 2 = −10 [11,Equation (11.4.5)]. As in Proposition 4.1, this shows that J has no reducible fibres: the non-identity component of a reducible fibre would be orthogonal to the classes of both a fibre and the zero-section, and have self-intersection −2, which is incompatible with the required determinant. The Shioda-Tate formula then shows that the Mordell-Weil group of J has rank 1. This Mordell-Weil group acts faithfully on Z, so it follows as in Proposition 4.1 that Aut Z is infinite. Proposition 4.13 states that any K3 surface over an algebraically closed field having Picard lattice isomorphic to either L M or L N has infinite automorphism group. In contrast, the following proposition shows that a K3 surface over Q having Picard lattice isomorphic to L M or L N over any algebraic extension of Q may have finite automorphism group. Indeed, the last part of this section will be devoted to finding an example of such a surface.
Proposition 4.14. Let S be a K3 surface as in Proposition 4.11. Then Aut S is finite.
, so that D 1 , D 2 have the intersection matrix M . Since C 1 , C 2 are disjoint conjugate rational curves, the product of the reflections in their classes is in our reflection group R X (Definition 3.3). In the given basis, it acts on Pic X with matrix For 1 ≤ a ≤ 6, we now give lower bounds on b 1 , b 2 for a[H] + b 1 [C 1 ] + b 2 [C 2 ] to be effective. The point is simply that an irreducible effective divisor with a > 0 must have nonnegative intersection with C 1 , C 2 and self-intersection at least −2.
We define a partial order on Z 2 for which (x 1 , x 2 ) ≤ (y 1 , y 2 ) if x 1 ≤ y 1 and x 2 ≤ y 2 , while (x 1 , x 2 ) and (y 1 , y 2 ) are incomparable if x 1 < y 1 and y 2 < x 2 . For example, the minimal elements (b 1 , b 2 ) of Z 2 for which 10 − 2b 2 1 − 2b 2 2 ≥ −2 are (−2, −1) and (−1, −2), so for −2). Repeating this calculation for a = 2, 3, 4, 5, 6, we obtain the list of minimal coefficients of b 1 , b 2 for effective divisors ( Table 2).  Let σ be an automorphism ofQ that exchanges C 1 and C 2 . Then the class of R σ is 6D 1 − 10[C 1 ] − 9[C 2 ], and one checks immediately that (R, R σ ) = 0. The product of the reflections in R and R σ has matrix R 2 = 721 1140 −456 −721 with respect to the given basis of Pic X. Now, R 1 R 2 is an element of R X of infinite order; as in Proposition 4.13, this implies that O(Pic X)/R X is finite and hence that Aut X is finite.
Remark 4.15. Given equations for S, the curve R used in the proof can be made explicit. Let I(V ) be the ideal in the coordinate ring of P 4 of polynomials that vanish on V . Then the dimension of the degree-6 part of ((I(S) + I(C 1 ) 3 ) ∩ (I(S) + I(C 2 ) 4 ))/I(S) is 1; let f 6 be a generator, so that the divisor cut out on S by f 6 is of the form C + 3C 1 + 4C 2 . Using Magma one checks that C is indeed an irreducible curve of arithmetic genus 0. The remainder of this section will be devoted to constructing a K3 surface S satisfying the conditions of Proposition 4.11. The verification will be more complicated than that of Example 4.6, because the Picard number overQ in this example is odd, so that the reduction map cannot induce an isomorphism of Pic SQ to Pic(S mod p)F p . We use the following proposition.
Proposition 4.17. Let S be a smooth intersection of a quadric and a cubic in P 4 (Q) containing two disjoint conics C 1 , C 2 that are defined and conjugate over a quadratic field K S . Let p = 2, q be primes such that: (1) S has good reduction at p and q; (2) the reduction of S mod p contains a line L disjoint from the reductions of C 1 , C 2 , and its Picard group overF p has rank 4; (3) the reduction of S mod q base changed toF q contains no lines disjoint from the reductions of C 1 and C 2 . Then SQ has Picard number 3 and hence S satisfies the conditions of Proposition 4.11.
Proof. Let s p be the specialization homomorphism Pic SQ → Pic(S mod p)F p . Then by [7,Theorem 1.4] the cokernel of s p is torsion-free. If s p is surjective, the class of L is in the image. The class L 0 with s p (L 0 ) = [L] has positive intersection with the hyperplane class and self-intersection −2, so it is the class of a line on SQ. Its intersection with the classes of the C i is 0, so it is disjoint from them. But L 0 would reduce mod q to a line disjoint from the reductions of the C i , contradicting hypothesis (3). Since s p is injective [19,Proposition 6.2] it follows that rk Pic SQ < rk Pic(S mod p)F p = 4. On the other hand, the hyperplane class in P 4 and the classes of C 1 , C 2 are independent, so rk Pic SQ ≥ 3.
In our construction, we will use p = 3, q = 5. We will take the conics C 1 , C 2 to be defined over Q ( √ 19), the smallest quadratic field in which both p and q split. Let ρ = √ 19, and denote the coordinates on P 4 by x 0 , . . . , x 4 . We choose the conic C 1 defined by x 4 + 2x 2 4 = 0 and let C 2 be its Galois conjugate. Next, let L be the line in P 4 (F 3 ) through (1 : 1 : 2 : 0 : 1), (2 : 1 : 1 : 1 : 0) and let C 3,1 , C 3,2 be the reductions of C 1 and C 2 , respectively, at one of the two primes above 3.
Proof. The verification that S 3 contains the given curves is routine. To see that S 3 has Picard number 4, we proceed as in Example 4.6. In order to count the number of points on S 3 over F 3 n for 1 ≤ n ≤ 9, we write S 3 as an elliptic surface using the fibration of Picard class H − C 3,1 , for which L is a section, and we sum the number of points of the fibres. We obtain point counts as in Table 3.
From the Lefschetz fixed point formula we find that the trace of the nth power of Frobenius acting on H 2 The trace on the Tate twist H 2 et (S 3 , Q ℓ (1)) is obtained by dividing by 3 n , while on the subspace V generated by H, C 3,1 , C 3,2 , and L, the trace is equal to 4; hence, on the 18-dimensional quotient Q = H 2 et (S 3 , Q ℓ (1))/V , the trace equals #S 3 (F 3 n )/3 n − 3 n − 3 −n − 4. These traces are sums of powers of eigenvalues, and we use the Newton identities to compute the elementary symmetric polynomials in these eigenvalues, which are the coefficients of the characteristic polynomial f of Frobenius acting on Q. This yields the first half of the coefficients of f , including the middle coefficient, which turns out to be nonzero. This implies that the sign in the functional equation t 18 f (1/t) = ±f (t) is +1, so this functional equation determines f , which we calculate to be f = 1 3 3t 18 − 4t 17 + 5t 16 − 4t 15 + 4t 14 − 4t 13 + 5t 12 − 5t 11 + 5t 10 − 6t 9 + 5t 8 − 5t 7 + 5t 6 − 4t 5 + 4t 4 − 4t 3 + 5t 2 − 4t + 3 .
As the characteristic polynomial of Frobenius acting on V is (t − 1) 4 , we find that the characteristic polynomial of Frobenius acting on H 2 et (S 3 , Q ℓ (1)) is (t−1) 4 f . The polynomial 3f is irreducible, primitive and not monic, so its roots are not roots of unity. Thus we obtain an upper bound of 4 for the Picard rank ofS 3 . The rank is equal to 4 as the Picard group contains the linearly independent classes of H, L, C 3,1 , and C 3,2 . Now let C 5,1 , C 5,2 be the reductions of C 1 and C 2 , respectively, at one of the two primes above 5.
Proof. Again, it is easy to check that the C 5,i are on S 5 . It is also easy to verify directly that S 5 contains no lines defined over F 5 by checking that no line through two distinct F 5 -points of S 5 is contained in S 5 ; however, this is not sufficient. Suppose that L is a line contained in S 5 and disjoint from C 5,1 . We embed S 5 in P 5 (F 5 ) by the divisor class H + [C 5,1 ], where H is the hyperplane class in P 4 (F 5 ). In this embedding, L is still a line, because L · (H + [C 5,1 ]) = 1.
The image is a surface T 5 defined by three quadrics Q 1 , Q 2 , Q 3 . For each Q i , let Y i be the subscheme of P 5 × P 5 consisting of pairs of points P, P ′ for which P = P ′ or the line joining P to P ′ is on the variety defined by Q i = 0. The image of Y i in P 14 under the Plücker embedding is isomorphic to the Fano scheme of lines on Q i . It is easy to check in Magma that the intersection of the images of the Y i in P 14 is empty, which implies that there are no lines on T 5 even overF 5 .
Remark 4.20. In fact S 5 contains no lines at all. A similar calculation shows that there are no lines on S 5 disjoint from C 5,2 . To prove that there are no lines that meet both, we write down the subschemes Z 2 , Z 3 of P 4 × P 4 parametrizing lines on the quadric and cubic defining S 5 and check in Magma that Z 2 ∩ Z 3 ∩ (C 5,1 × C 5,2 ) is empty. This is done in our Magma scripts [5].
In principle one can prove directly that S 5 has no lines by checking that the images of Z 2 and Z 3 under the Plücker embedding in P 9 are disjoint, but this calculation is very slow.
It remains to show that there exists a surface S ⊂ P 4 Q containing the conics C 1 , C 2 and lifting the surfaces S 3 , S 5 described above. Denote by V 2 (C 1 ∪C 2 , Q) the vector space of homogeneous forms of degree 2, with coefficients in Q, vanishing on C 1 ∪C 2 . One easily computes that V 2 (C 1 ∪C 2 , Q) has dimension 5; let q 1 , . . . , q 5 be a Z-basis for the free abelian group V 2 (C 1 ∪ C 2 , Z) of forms having integer coefficients. One also easily computes that the analogous spaces V 2 (C 3,1 ∪ C 3,2 , F 3 ) and V 2 (C 5,1 ∪ C 5,2 , F 5 ) both have dimension 5 over F 3 and F 5 respectively, which implies that the images of q 1 , . . . , q 5 span these spaces. Now the Chinese remainder theorem guarantees the existence of a quadratic form F ∈ V 2 (C 1 ∪ C 2 , Z) satisfying both F ≡ F 3 (mod 3) and F ≡ F 5 (mod 5). Similarly, the analogous spaces of cubic forms vanishing on C 1 ∪ C 2 are all of dimension 21, which implies the existence of a cubic form G satisfying G ≡ G 3 (mod 3) and G ≡ G 5 (mod 5). Remark 4.21 below show that one can avoid explicit computations to calculate these dimensions. The surface S given by F = G = 0 reduces modulo 3 and 5 to S 3 and S 5 , respectively. Remark 4.21. Let k be any field. Let C 1 , C 2 ⊂ P 4 k be two conics; suppose that the two planes spanned by C 1 , C 2 respectively meet only in one point P , and suppose that P does not lie on either C 1 or C 2 . Then the space of forms of degree d vanishing on C 1 and C 2 has dimension

4.3.
Finite automorphism group for a family of diagonal surfaces. In this example, for which we summarize the calculations in a Magma script [5], we consider a quartic surface X over a field k of characteristic 0 defined by for some nonzero c ∈ k. Conjecture 2.5 of [20] states that if k is the field Q of rational numbers, then for every nonzero c ∈ Q, the set of rational points on the associated surface X is dense. The 48 lines on a surface in this family are defined over the field k(ζ 8 , 4 √ c). We assume that k, c are such that this is a Galois extension of k of degree 16. (This Galois action is case A45 of [4, Appendix A].) We will prove under this assumption that Aut X is finite. The argument will proceed by the following steps.
(1) Determination of Pic X, and comparison to the Picard lattice of a K3 surface T with finite automorphism group overk.
(2) Determination of O(Pic T ) and O(Pic X) up to finite index.
(3) Construction of elements of R X and verification that O(Pic X)/R X is finite.
There is a sublattice of Λ of index 4 that is isomorphic to Pic X.
Proof. The Picard group ofX is generated by the classes of the 48 lines. (See [16, Section 6.1] for a history of proofs of this fact.) One easily verifies that the subspace fixed by the action of Galois is of rank 6 and that the intersection form has discriminant −256. Since the surface has rational points such as (1, 1, 0, 0), the fixed subspace of the Picard group ofX is in fact the Picard group of X. Let M be the lattice obtained from Pic X by adjoining v/2 for each v ∈ Pic X with 8 | (v, v) and 2 | (v, w) for all w ∈ Pic X: then M contains Pic X with index 4 and hence has discriminant −16. It is verified in [5] that Λ and M are isomorphic.
We now proceed to step 2, determining O(Λ). Let T be a K3 surface overk with Picard lattice Λ, as in Proposition 4.5. We choose a basis e 1 , . . . , e 6 for Λ ∼ = Pic T to consist of the classes E, E + S, C 1 , C 2 , C 3 , C 4 , where E is the class of an elliptic fibration with zero section S and the C i are the components of the reducible fibres that do not meet S. In this basis, the Gram matrix of Pic T is as above. Let G ⊂ O(Λ) be the image of the symmetric group S 4 by the homomorphism taking a permutation π to the automorphism fixing E and E + S and taking C i to C π(i) . Let A Λ be the discriminant group of Λ. Proof. The e i /2 for 3 ≤ i ≤ 6 belong to Λ * and they generate a group that contains Λ with index 16. Since Λ has discriminant 16, they generate Λ * and their classes generate A L . The quadratic form takes 6 i=3 a i e i /2 to 6 i=3 a 2 i e 2 i /4 mod 2 = − 6 i=3 a 2 i /2. It follows that the only elements of A Λ taken to −1/2 by the quadratic form are the basis vectors, so every automorphism of A Λ must permute these. Conversely it is clear that every permutation of the basis vectors extends to an automorphism of A Λ and that this automorphism is the image of a unique element of G. Proof. By Lemma 4.8, every automorphism of T must preserve the configuration of the 9 rational curves. The only automorphisms are those that permute the reducible fibres, and these act on Pic T by elements of G. This proves the first statement.
It is easy to verify that an element of G acts as ±1 on the discriminant group of Pic T if and only if it is the identity. Now, a symplectic automorphism of a K3 surface T in characteristic 0 acts as the identity on the transcendental lattice H 2 (T, Z)/ Pic T [11, Remark 15.1.2] and hence on its discriminant group, so in addition it acts trivially on the discriminant group of Pic T by [11,Lemma 14.2.5]. So it acts trivially on Pic T and H 2 (T, Z)/ Pic T and therefore on H 2 (T, Z). It is therefore the identity by [11,Corollary 15.1.6].
Since we are in characteristic 0, there is an exact sequence [11, (15.1.3)] for some m, where Aut s T is the group of symplectic automorphisms that we have just shown to be trivial. It follows that Aut T is finite cyclic.
Remark 4.25. There is always an automorphism of order 2 of T given by negation.
In general Aut T is of order 2, but it might be possible for Aut T to be of order m = 4, 6, or 8 if there is an automorphism that permutes the reducible fibres of the fibration nontrivially. In this case the field of definition of Pic T would contain µ m , which does not happen generically.
By [6, Proposition 5.10], we know that O(Λ) is generated by the automorphisms of the ample cone and the reflections in the classes of rational curves, together with −1; the automorphisms of the ample cone are just the G of Lemma 4.23. Since G is generated by 2 elements and has 3 orbits on the set of rational curves, we have a set of 6 generators for O(Λ).
To pass to O(Pic X), we use a simple lemma.  Using Magma [3] to compute the permutation representation of O(Λ) on sublattices with quotient (Z/2Z) 2 , we find a set of 30 generators for O(Λ) ∩ O(Pic X). However, some of these are products of others, so we easily reduce to a set of 9 generators. We now perform step 3, the construction of elements of R X . Proof. We list the conics by observing that their Picard classes are of the form H/2 + M , where M ∈ 1 2 PicX is such that H · M = 0 and M 2 = −3. Since H ⊥ ⊂ 1 2 PicX is a definite lattice, the set of M with these properties is finite. Such a class is represented by a conic if and only if it is integral and not the sum of two classes of lines, both of which are easily checked.
We finish the proof by referring to Proposition 3.6 for the configurations of curves that give finite Coxeter groups.  Proof. We use Proposition 3.10. To do so, we search for relations among the generators of O(Λ) ∩ O(Pic X) and the 22 elements of R X that we have found, simply by checking which products of up to 5 generators of O(Λ) ∩ O(Pic X) and up to 4 of these together with the generators of R X give the identity matrix. Magma quickly verifies that these relations are sufficient to show that O(Λ) ∩ O(Pic X), R X /R X is finite, from which we conclude by Proposition 3.10 that Aut X is finite. Proof. This follows by combining the theorem with Corollary 3.16.
We conclude that in order to prove that the rational points of X are Zariskidense, we do not have an infinite automorphism group Aut X to our disposal. In particular, we will not be able to find an elliptic fibration on X with a section of infinite order, as translation by such a section would be an automorphism of infinite order.