Preprint
Article

This version is not peer-reviewed.

Hodge Classes on K3 Powers and Sixfolds

Submitted:

13 November 2026

Posted:

17 November 2026

You are already at the latest version

Abstract
The Hodge conjecture holds on all powers of every K3 surface of Picard number at least seventeen and on a dense union of curves in every family of K3 surfaces of transcendental rank at least eight with real quadratic multiplication. Weil classes are algebraic on every abelian sixfold of Weil type with complex multiplication and on a dense union of special subvarieties of every family of non-split Weil sixfolds.
Keywords: 
;  ;  

MSC:  14C30 (primary); 14J28; 14K20; 11E12

1. Introduction

The Hodge conjecture asks that every rational cohomology class of type ( p , p ) on a smooth complex projective variety be a rational combination of classes of algebraic subvarieties. Hodge raised it in 1950 [54]; Lefschetz had proved it for divisors [70]. It is known in a scattered family of cases and open in general. This paper proves it on new classes of varieties built from K3 surfaces and abelian sixfolds, inside a reduction of the conjecture to two statements (F2) and (F3′) about abelian varieties, whose known loci have the same shape in every family: dense, and either everything or meagre. On the side of (F3′) we prove the Hodge conjecture for all powers and Hilbert schemes of points of every K3 surface of Picard number at least seventeen, determine exactly which K3 surfaces of Picard number sixteen Floccari’s theorem reaches, and make real multiplication algebraic by Q ( 6 ) , by the cyclic cubic fields Q ( ζ 7 ) + and Q ( ζ 9 ) + on the surfaces of Mumford type, by every real quadratic field in transcendental rank six, and on a dense union of curves in every family of transcendental rank at least eight with real multiplication of degree two; the conjecture holds on every power of the Fricke K3 surfaces of levels 5 and 7. On the side of (F2) we prove the Weil classes of every abelian sixfold of Weil type with complex multiplication algebraic, and the conjecture for every abelian sixfold with complex multiplication by a CM field of degree twelve containing an imaginary quadratic field; in every family of non-split Weil sixfolds the Weil classes are algebraic on a dense union of six-dimensional loci with quaternionic multiplication, while Markman’s construction produces only split families; for a Mumford fourfold we attach an abelian sixfold whose real multiplication is algebraic, show that a single link class decides the conjecture for all products of their powers, that one algebraic class through an elliptic curve with complex multiplication suffices, and that divisor classes, Weil classes of every field and homomorphisms cannot produce it; at the point E 0 6 we exclude every twisted complex of line bundles and distinct threefolds with at most seven line bundles, and all but two arrangements of eight; and at E 0 6 and E 0 8 no convolution of the line bundles that realise a Weil class meets the polarised criterion, with or without a group action. For real multiplication of degree two on a product of two K3 surfaces the Hodge locus of the relevant class is a Hecke graph, algebraic at a dense set of its points. Figure 1 draws the plan, and Table 1 lists the main results.

1.1. The Two Statements

Write HC for the Hodge conjecture for all smooth complex projective varieties and AB for its restriction to abelian varieties. For a smooth projective variety X and an integer p let Hdg p ( X ) be the Hodge classes of degree 2 p , and let Ab p ( X ) ⊆ Hdg p ( X ) be the span of the classes Γ * γ , where B runs over the complex abelian varieties (the point included), γ over the Hodge classes on B, and Γ over the algebraic correspondences from B to X. The two statements are
(F2)
every Hodge class on every complex abelian variety is algebraic, in the form in which it first appears in the closure computation of Appendix A.8: the Hodge classes of the smallest known instance, the exceptional classes on the square of a Mumford fourfold, are algebraic together with everything that can be built from them;
(F3′)
Hdg p ( X ) = Ab p ( X ) for every X and every p: every Hodge class is algebraic modulo the images, under algebraic correspondences, of Hodge classes of abelian varieties.
Theorem 1
(Reduction). HC holds if and only if (F2) and (F3′) both hold; (F2) is equivalent to AB , and (F3′) holds on every abelian variety, on every variety dominated by a product of curves or by an abelian variety, and on every variety in degrees 2 p with p ∈ { 0 , 1 , dim X − 1 , dim X } .
The content of the theorem is in the second sentence: (F2) is not a special case of the conjecture for abelian varieties but the whole of it, carried by its hardest classes, and (F3′) is exactly what the conjecture asks of the part of the cohomology of X that no abelian variety reaches. Section 3 proves it and places the two statements in the closure graph of the known reductions, with the Lefschetz standard conjecture, André’s motivated classes and the variational Hodge conjecture as the other nodes (Figure 2).

1.2. The Shape of the Known Locus

For a smooth projective family π : X → S and a flat Hodge class ξ let Σ ( ξ ) ⊆ S be the set of points at which ξ is algebraic.
Theorem 2
(Dichotomy and propagation).
(i)
Σ ( ξ ) is a countable union of Zariski closed subsets of S, and either Σ ( ξ ) = S or Σ ( ξ ) is meagre.
(ii)
(F2) holds if and only if, for every polarised abelian scheme and every flat Hodge class ξ, algebraicity of ξ at one point of complex multiplication forces Σ ( ξ ) = S .
(iii)
For the Weil classes of every CM field, every n ≥ 2 and every discriminant, Σ ( ξ ) contains an explicit point of complex multiplication and is dense.
(iv)
(F3′)holds for X in degree 2 p if and only if a Hodge complement of the abelian part of H 2 p ( X , Q ) carries no nonzero Hodge class.
So both statements have the same form. There is a dense set of points, the points of complex multiplication, at which the class is known to be algebraic by explicit cycles; the locus of algebraicity is a countable union of closed strata; and the conjecture is the statement that this union is not meagre. Figure 3 draws it. Section 4 proves the theorem, with the results on Weil type of Appendix A.2.

1.3. Constructions at the First Class

The rest of the paper is constructive. At each statement we go to the first class that is not known and build the object that would make it algebraic.
At (F3′) the first class lives on powers of a K3 surface S with real multiplication: the transcendental lattice T ( S ) Q is a vector space over a totally real field E = End Hdg T ( S ) Q , and the elements of E are Hodge classes on S × S . Section 5 proves:
Theorem 3
(Chains of symplectic quotients). Let S be a projective K3 surface and e ∈ E with e 2 = n ∈ Q > 0 . If there is a chain of projective K3 surfaces S = S 0 , S 1 , ⋯ , S r joined by dominant rational maps or rational Hodge isometries whose degrees multiply to n modulo squares, then e is algebraic on S × S . In particular, for every projective K3 surface S with Q ( 6 ) ⊆ E whose transcendental space embeds in the invariant lattice of a symplectic automorphism of order three, the real multiplication by 6 is algebraic, the Hodge conjecture holds for S × S and S [ 2 ] , and such surfaces form two-dimensional families whose very general member has E = Q ( 6 ) exactly.
Varesco [115] proved the case of a single quotient by a symplectic automorphism of prime order, which gives 2 and 3 ; van Geemen and Schütt [121] reach 2 , 3 , 5 , 7 through rational self-maps and the dihedral families. The chain adds 6 , the first composite value, with the lattice condition made explicit.
A second construction starts from the universal elliptic curve over the modular curves X 0 ( 5 ) and X 0 ( 7 ) : the Fricke involution lifts to an isogeny of degree 5 or 7 between the elliptic K3 surface over the base and the one over its image, and on the quadratic twists of the two surfaces by the same pair of parameters it becomes a rational map of that degree between two K3 surfaces of the same family. Lemma 9 determines the sign of its square: on the twists at a pair exchanged by the involution the map acts on the transcendental lattice with square − N , a complex multiplication by Q ( − 5 ) or Q ( − 7 ) , while at the fixed points of the involution, the surfaces of [121], its square is N. Theorem 13 shows that every member of the two families has that complex multiplication, algebraic because it is the pullback of the isogeny, and that the Hodge conjecture and (F3′) hold for every power and every Hilbert scheme of points of every member. Proposition 9 shows why the method stops: every correspondence component of geometric genus at most one acts on the transcendental lattice as a similarity, and these span only the multiquadratic subfield of E.
The third construction reaches every real quadratic field at once. Floccari proved that a K3 surface whose rational transcendental space embeds isometrically in V m = U Q 3 ⊕ 〈 − m 〉 for some integer m > 0 has an algebraic Kuga–Satake correspondence and satisfies the Hodge conjecture on all its powers [42,43]. Section 5.8 determines which spaces embed (Lemma 11): all of rank at most five, none of rank at least seven, and in rank six exactly those of Witt index two. For real multiplication of degree two the Witt index is a sum of local invariants of a quaternion algebra over the field of multiplication (Theorem 17), and we obtain:
Theorem 4
(K3 surfaces through Floccari’s theorem).
(i)
Every projective K3 surface of Picard number at least 17 has an algebraic Kuga–Satake correspondence, and the Hodge conjecture and (F3′) hold on every power and every Hilbert scheme of points of it.
(ii)
A projective K3 surface S of Picard number 16 satisfies the hypothesis of Floccari’s theorem if and only if T ( S ) Q has Witt index two, if and only if its even Clifford algebra is M 4 ( Q ( − d ) ) , d = d ( T ( S ) Q ) ; then the conclusions of (i) hold for S.
(iii)
For every squarefree integer d > 1 there is a one-dimensional family of projective K3 surfaces, of Picard number 16 at the very general member, whose endomorphism field is exactly Q ( d ) there, and on every power and every Hilbert scheme of points of every member the Hodge conjecture holds. For every real quadratic field there are also such families of Witt index one, which Floccari’s theorem does not reach.
(iv)
In every family of K3 surfaces of transcendental rank 2 m ≥ 8 with real multiplication by a real quadratic field E, the members with algebraic real multiplication by E on which the Hodge conjecture holds for all powers and Hilbert schemes of points contain a dense union of curves. For E = Q ( 2 ) and 2 m ≤ 10 the real multiplication is algebraic, and the Hodge conjecture holds on S × S and S [ 2 ] , at every member (Theorem 12).
For the cubic field of a Mumford fourfold the multiquadratic subfield is Q , so neither quotient chains nor Fricke maps see it. For the two cyclic cubic fields Q ( ζ 7 ) + and Q ( ζ 9 ) + the quaternion algebra over F ramified at exactly two real places is the even Clifford algebra of a two-dimensional dihedral family of K3 surfaces of Picard number ten, and Theorem 14 makes the real multiplication by F algebraic there, through Buskin’s theorem; Proposition 11 says exactly what the mechanism cannot reach.
At (F2) the first class is a Weil class of a very general abelian 2 n -fold of Weil type with n ≥ 3 , or an exceptional class of a Mumford fourfold. For the Weil classes the tool is the polarised semiregularity criterion (Theorem 20, proved in Appendix A.3): a perfect complex E whose Chern character is a nonzero multiple of a Weil class plus a polynomial in the polarisation, with dim Ext 2 ( E , E ) equal to the lower bound r ( γ ) = ( 4 + ρ ) n 2 − 2 n , makes the Weil classes algebraic on every member with full Hodge group. Section 6 proves:
Theorem 5
(The product object). Let A, A ′ be very general polarised abelian varieties of Weil type for the same imaginary quadratic field, of dimensions 2 m , 2 m ′ , with algebraic Weil classes, and let n = m + m ′ ≥ 3 . There is a perfect complex E on A × A ′ , a direct sum with shifts of exterior products of secant sheaves and powers of the polarisations, whose Chern character has the polarised shape with Hankel rank ρ = 3 , so that r ( γ ) = 7 n 2 − 2 n , and whose first-order Hodge locus is the whole tangent space of the polarised Weil family. Every such direct sum has dim Ext 2 ( E , E ) > r ( γ ) , and its semiregularity map is not injective. The criterion can therefore be met at a product point only by a convolution of these pieces with nonzero differentials.
At the point E 0 6 of complex multiplication the Ext groups are computable (Lemma 15), and Section 6 excludes, for every polarisation, every convolution of distinct threefolds and at most five line bundles (Theorem 23); a sharper Maurer–Cartan rule (Lemma 23) and an enumeration extend this to seven line bundles and to all but two arrangements of eight (Remark 9), and Section 6.8 reads the support of an object meeting the criterion off its localised obstruction.
For sixfolds the polarisation is a free parameter: Markman proved the Weil classes algebraic on every abelian sixfold of split Weil type [78], and the discriminant belongs to the polarisation, not to the abelian variety with its action of K. Section 6.9 uses this freedom.
Theorem 6
(Weil sixfolds). Let K be an imaginary quadratic field.
(i)
If the Weil classes are algebraic on every polarised abelian ( 2 n + 2 ) -fold of split Weil type for K, they are algebraic on every abelian 2 n -fold of Weil type for K.
(ii)
The Weil classes of an abelian sixfold of Weil type for K are algebraic if it has complex multiplication, or a K-stable sub-Hodge structure of odd K-dimension in H 1 ; in every period domain of Weil sixfolds for K the latter form a dense union of six-dimensional subdomains, and in every non-split one so do the sixfolds with quaternionic multiplication, on which the Weil classes are products of divisor classes.
(iii)
The Hodge conjecture holds for every abelian sixfold with complex multiplication by a CM field of degree twelve that contains an imaginary quadratic field.
For the exceptional classes of a Mumford fourfold X we replace the K3 surface that carries their transcendental part by an abelian sixfold Z K over the same Shimura curve, so that the whole problem stays inside abelian varieties.
Theorem 7
(The sixfold link). Let X be a Mumford fourfold with cubic field F whose Hodge group is the group of Mumford’s construction, as it is outside a countable subset of its Shimura curve C, let K = F ( − δ ) be a CM extension inside the quaternion algebra, and let Z K be the abelian sixfold of Shimura type attached to K over C. The Hodge classes in H 1 ( X ) ⊗ H 1 ( X ) ⊗ H 2 ( Z K ) that identify the two copies of the nine-dimensional representation T form a line over F, spanned by a class ξ.
(i)
The real multiplication by F on the copy T Z of T in H 2 ( Z K ) is algebraic.
(ii)
If ξ is algebraic, the exceptional classes of X × X are algebraic, and so is every Hodge class on every power of X.
(iii)
At a very general point of C, ξ is algebraic if and only if the Hodge conjecture holds for every X a × Z K b ; otherwise the algebraic classes on X a × Z K b are exactly the sums of products of algebraic classes on X a and Hodge classes on Z K b .
(iv)
If K = F · k with k imaginary quadratic, then ξ t is algebraic at every point t of a countable dense subset of C, and so is every Hodge class on every product of powers of X t and Z K , t .
Parts (i)–(iii) are Theorems 37 and 38; the proof of (iii) runs through the motivic Galois groups of X and Z K . Part (iv) is Theorem 41: at those points X t is isogenous to E × A and Z K , t to E 3 × A , with E a CM elliptic curve and A a CM threefold, the Hodge ring of every E a × A b is generated by divisors and pull-backs of the Weil plane of the fourfold A × E , and that fourfold carries a polarisation of discriminant one, so that Markman’s theorem [76, Thm. 1.5] applies. By the dichotomy the algebraic locus of ξ is then either countable and dense or all of C; which one holds is open.

1.4. What Is Proved and What Is Not

Table 8 in Section 7 lists every statement of the paper with its status. The reduction, the dichotomy, the theorems on K3 surfaces, on real multiplication and on the Fricke families, the theorems on Weil sixfolds, the sixfold link and its dichotomy, the link at the points of complex multiplication, the exclusions at E 0 6 and E 0 8 , the Hecke graph and the no-go theorem for the link are proved; the semiregularity count on a product of two K3 surfaces and the conjectures of Bouchet, Hanselman, Pieper and Schiavone are proved under named inputs. The Hodge conjecture, (F2) and (F3′) remain open. The value of the constructions lies in the precision of the question each one leaves.
Appendix A proves the results on Weil classes, on the polarised criterion, on convolutions of line bundles and of pieces of any kind, on the closure graph and on Mumford fourfolds that the body uses, so that the paper depends only on published work. Appendix D describes the machine checks: a Lean 4 file that verifies the closure graph and the finite combinatorics of Section 6, and programs in Python and Julia that recompute every finite quantity used in the proofs. No proved statement of the paper rests on a computation; the statements that do are marked as such.

2. Conventions

Varieties are smooth, projective and connected over C unless said otherwise. Cohomology is singular cohomology with rational coefficients, H k ( X ) = H k ( X , Q ) , with its Hodge structure. A Hodge class of degree 2 p is an element of H 2 p ( X ) ∩ H p , p ; their space is Hdg p ( X ) . A class is algebraic if it is a rational combination of fundamental classes of subvarieties of codimension p; by Fulton [44], intersection products, pull-backs and push-forwards along morphisms preserve algebraic classes. An algebraic correspondence from Y to X is an algebraic class Γ on Y × X , acting by Γ * γ = pr X * ( [ Γ ] ∪ pr Y * γ ) . The Hodge group Hg ( V ) of a rational Hodge structure V is the smallest algebraic subgroup of GL ( V ) defined over Q whose real points contain the circle h ( U ( 1 ) ) ; for a polarisable V it is reductive, the Hodge classes in every tensor construction on V are its invariants, and the Hodge group of a variety is that of its total cohomology [35]. The Hodge conjecture for X is the statement that every Hodge class on X is algebraic; HC is the conjecture for all varieties and AB for all abelian varieties.
For an abelian variety A of dimension g with H 1 ( A ) = V , the cohomology ring is ⋀ • V . An abelian 2 n -fold A is of Weil type for the imaginary quadratic field K if K acts on A up to isogeny with V a K-vector space of dimension 2 n on which the two complex embeddings have equal multiplicity n in H 1 , 0 . The Weil classes are the elements of ⋀ K 2 n V ⊆ ⋀ 2 n V , a two-dimensional rational space of Hodge classes of degree 2 n  [125]; K × acts on it by τ ↦ τ 2 n , and η k , for η the class of a K-hermitian polarisation, carries τ ↦ N K / Q ( τ ) k . For a general CM field we use the setting of Appendix A (Setting 8): a family of Weil type for a CM field F of degree 2 m with n ≥ 2 and discriminant δ has a connected polarised period domain D F of dimension m n 2 (Proposition A1), and its very general member has Hodge group SU ( V , H ) for the F-hermitian form H of the polarisation.
A Mumford fourfold is a simple abelian fourfold X whose Hodge group is the Q -form of SL 2 3 given by the norm-one group D 1 of a quaternion algebra D over a totally real cubic field F, split at one real place and definite at the other two [87]. Its own Hodge ring is generated by the polarisation, and X × X carries exceptional Hodge classes of degree four that are not products of divisor classes and are not Weil classes of any quotient [90] (Proposition A20). We write T = D 0 ⊆ D for the trace-zero part, a nine-dimensional F-line on which Hg ( X ) acts by conjugation; T ( − 1 ) is a Hodge structure of K3 type with Hodge numbers ( 1 , 7 , 1 ) , and the exceptional classes of X × X are the real multiplication by F on the copy of T ( − 1 ) inside H 1 ( X ) ⊗ H 1 ( X ) .
For a K3 surface S the transcendental lattice T ( S ) is the orthogonal complement of NS ( S ) in H 2 ( S , Z ) ; T ( S ) Q is an irreducible Hodge structure, and E = End Hdg T ( S ) Q is a totally real field or a CM field, acting by self-adjoint endomorphisms in the first case [126]. The K3 lattice is Λ = U 3 ⊕ E 8 ( − 1 ) 2 , and Λ Q = Λ ⊗ Q . For a rational quadratic space W we use the discriminant d ( W ) ∈ Q × / Q × 2 and the Hasse invariants ε p ( W ) = ∏ i < j ( a i , a j ) p of a diagonalisation 〈 a 1 , ⋯ , a k 〉 , with ( , ) p the Hilbert symbol; the classification of rational quadratic forms by rank, signature, discriminant and Hasse invariants, and the existence theorem for prescribed local invariants, are Serre’s [108, Chap. IV].
Perfect complexes on A are objects of the bounded derived category of coherent sheaves; ch ( E ) ∈ H ev ( A ) is the Chern character, additive on triangles, multiplicative on exterior products, with ch ( E [ 1 ] ) = − ch ( E ) . The semiregularity map σ E : Ext 2 ( E , E ) → ⨁ q H q + 2 ( A , Ω q ) is that of Buchweitz and Flenner [11], after Bloch [20]; its injectivity implies that the Hodge locus of ch ( E ) coincides, near the given point, with the locus where E deforms [11,102].
Throughout, an assertion attributed to a reference is cited and not reproved; everything else is proved in the text. The phrase “very general” means outside a countable union of proper closed analytic subsets of the parameter space.

3. The Reduction to Abelian Varieties

3.1. The Statement (F3′) and the Class A

Definition 1.
For a variety X and p ≥ 0 let Ab p ( X ) ⊆ H 2 p ( X ) be the span of the classes Γ * γ of degree 2 p , where B runs over the complex abelian varieties, the point included, γ over Hdg * ( B ) , and Γ over the algebraic cycles on B × X with rational coefficients. The statement(F3′) is: Hdg p ( X ) = Ab p ( X ) for every X and every p.
Each Γ * γ is a Hodge class, since cup product with an algebraic class and push-forward along a projection are morphisms of Hodge structures of the right bidegree; so Ab p ( X ) ⊆ Hdg p ( X ) . Taking B a point shows that Ab p ( X ) contains every algebraic class, and taking X abelian, B = X and Γ the diagonal shows that (F3′) holds on every abelian variety.
Proposition 1.
HC holds if and only if AB and (F3′) both hold.
Proof. 
If HC holds then AB holds, and every Hodge class is algebraic, hence in Ab p ( X ) . Conversely let γ ∈ Hdg p ( X ) . By (F3′), γ = ∑ j Γ j * γ j with γ j a Hodge class on an abelian variety B j , algebraic by AB ; then pr B j * γ j ∪ [ Γ j ] is algebraic on B j × X and so is its push-forward to X.    □
Proposition 2.
Let A be the class of varieties X such that H * ( X ) is spanned by classes Γ * y with B abelian, Γ an algebraic cycle on B × X and y ∈ H * ( B ) .
(i)
Every X ∈ A satisfies (F3′).
(ii)
A contains the abelian varieties and the curves, and is closed under products and under images of surjective morphisms. It contains every product of curves, every symmetric product of a curve, and every variety dominated by a product of curves or by an abelian variety.
(iii)
Every X of dimension d satisfies (F3′) in the degrees 2 p with p ∈ { 0 , 1 , d − 1 , d } . In particular every variety of dimension at most three satisfies it in every degree.
Proof. 
(i) Let γ ∈ Hdg p ( X ) and write γ = ∑ j Γ j * y j with y j ∈ H * ( B j ) . After replacing y j by its component in the degree k j that Γ j * sends to H 2 p ( X ) , each Γ j * is a morphism of rational Hodge structures H k j ( B j ) ( m j ) → H 2 p ( X ) for a suitable Tate twist, and the sum is a morphism
Φ : W = ⨁ j H k j ( B j ) ( m j ) ⟶ H 2 p ( X ) , Φ ( y j ) j = γ .
A correspondence shifts degrees by an even number, so every k j is even and W is pure of weight 2 p . Polarisable rational Hodge structures form a semisimple abelian category, so W = ker Φ ⊕ W ′ with Φ | W ′ : W ′ → im Φ an isomorphism of Hodge structures. The Hodge class γ of im Φ is therefore Φ ( y ′ ) for a Hodge class y ′ = ( y j ′ ) j of W ′ ⊆ W , and each component y j ′ is a Hodge class on B j , since the projections of W onto its summands are morphisms of Hodge structures. Hence γ = ∑ j Γ j * y j ′ ∈ Ab p ( X ) .
(ii) The point, abelian varieties and curves are in A : for a curve C, H 0 and H 2 come from the point and H 1 ( C ) = H 1 ( J ( C ) ) by the Abel–Jacobi map. If X , Y ∈ A then H * ( X × Y ) = H * ( X ) ⊗ H * ( Y ) is spanned by exterior products of classes Γ * y ⊠ Γ * ′ y ′ = ( Γ × Γ ′ ) * ( y ⊠ y ′ ) with B × B ′ abelian. If f : X → X ′ is surjective, then f * f * = deg ( f ) on H * ( X ′ ) after cutting X by hyperplanes to the dimension of X ′ if needed, so every class of X ′ is f * of a class of X, and f * Γ * y = ( Γ f ∘ Γ ) * y with Γ f the graph of f. The listed varieties follow: a symmetric product is the image of a product of curves, and a variety dominated by a product of curves or an abelian variety is such an image.
(iii) Degrees 0 and 2 d are spanned by the class of X and the class of a point. In degree 2 every Hodge class is a divisor class by Lefschetz [70]. In degree 2 d − 2 , let θ be a hyperplane class; the hard Lefschetz theorem makes θ d − 2 ∪ : H 2 ( X ) → H 2 d − 2 ( X ) an isomorphism of Hodge structures, so every Hodge class of degree 2 d − 2 is θ d − 2 ∪ α with α a Hodge class of degree 2, algebraic by Lefschetz; a product of algebraic classes is algebraic.    □
Part (iii) covers every variety of dimension at most three whatever its Hodge numbers, the Calabi–Yau threefolds included, for which the Hodge numbers of the toric hypersurface families are tabulated in [14]; the first open degree is the middle degree of a fourfold.
Proposition 3
(Fourfolds with few zero-cycles). Let X be a smooth projective fourfold whose group of zero-cycles is supported on a threefold: CH 0 ( Z ) Q → CH 0 ( X ) Q is surjective for some closed Z ⊂ X of dimension at most three. Every uniruled fourfold is such: a general point lies on a rational curve, which meets an ample divisor D, so the point is rationally equivalent to a point of D, and the points of a dense open set generate CH 0 ( X ) . Then the Hodge conjecture holds for X in degree four, and X satisfies (F3′) in every degree.
Proof. 
Bloch and Srinivas prove, by the decomposition of the diagonal that the hypothesis provides, that a fourfold whose zero-cycles are supported on a threefold satisfies the Hodge conjecture for cycles of codimension two [24, Thm. 1]. So Hdg 2 ( X ) = Alg 2 ( X ) ⊆ Ab 2 ( X ) , the class of a cycle Z being Γ * 1 for the point and Γ = Z ; the other degrees are Proposition 2(iii).    □
For uniruled fourfolds the conclusion in degree four is the theorem of Conte and Murre [30]; the proposition isolates the hypothesis on zero-cycles behind it.

3.2. The Statement (F2)

For an abelian variety Y let R ( Y ) ⊆ H * ( Y ) be the Q -subalgebra generated by the divisor classes of Y and by the pull-backs f * w of Weil classes w of abelian varieties B of Weil type along homomorphisms f : Y → B . Every element of R ( Y ) is a Hodge class, and every Hodge class on an abelian variety of dimension at most five lies in R ( Y )  [91].
Definition 2.
The statement (F2) is: for every complex abelian variety Y, every Hodge class on Y that does not lie in R ( Y ) is algebraic.
The classes in R ( Y ) are the ones the known theorems reach when they reach anything: divisor classes by Lefschetz, Weil classes by Markman on fourfolds of Weil type of discriminant one [76], and on all abelian fourfolds of Weil type and on abelian sixfolds of split Weil type by Markman’s secant-sheaf construction [78,79], and by the propagation results of Appendix A (Theorem A3) wherever they propagate. So (F2) is the conjecture for abelian varieties with the known classes struck out. The next proposition says that striking them out removes nothing.
Setting 1.
Let X be a Mumford fourfold, V = H 1 ( X ) with its polarisation ψ, G = Hg ( X ) ⊆ Sp ( V , ψ ) , and e 0 ∈ Hdg 2 ( X × X ) an exceptional class, a Hodge class of degree four that is not a polynomial in divisor classes. Such classes exist [90];1 X is simple with End 0 ( X ) = Q  [87].
Lemma 1.
Let Y be an abelian variety. Let Y X ⊆ Y be the largest abelian subvariety isogenous to a power of X and Y ′ the identity component of the intersection of the kernels of the homomorphisms Y → X . Then Y X × Y ′ → Y is an isogeny, Hom ( X , Y ′ ) = 0 , and
H 1 ( Y ) = V ⊗ Q ⊕ H 1 ( Y ′ ) , Q = Hom 0 ( Y X , X ) ,
the first summand embedded by v ⊗ f ↦ f * v . Let Sp ( V , ψ ) act on H * ( Y ) = ⋀ * H 1 ( Y ) through the factor V. Then every class of R ( Y ) is invariant under Sp ( V , ψ ) , and for Y = X m the invariants of Sp ( V , ψ ) in H * ( X m ) are polynomials in divisor classes.
Proof. 
Poincaré reducibility gives Y ∼ Y X × Y ′ with Y X ∼ X k and no factor X in Y ′ ; since X is simple, Hom ( X , Y ′ ) = Hom ( Y ′ , X ) = 0 , and since End 0 ( X ) = Q , Q has dimension k and the displayed map is an isomorphism. A homomorphism h : Y → Z of abelian varieties decomposes up to isogeny as h X × h ′ , and h * acts on H 1 as ( 1 ⊗ ( f ↦ f ∘ h X ) ) ⊕ h ′ * , which commutes with the action of GL ( V ) on the factor V. So pull-backs along homomorphisms, cup products and exterior products are GL ( V ) -equivariant, and the classes invariant under Sp ( V , ψ ) form a subalgebra stable under pull-back. It remains to check the generators of R ( Y ) .
A divisor class on Y has Künneth components in ⋀ 2 ( V ⊗ Q ) , ( V ⊗ Q ) ⊗ H 1 ( Y ′ ) and ⋀ 2 H 1 ( Y ′ ) , each a Hodge class. The middle one is a homomorphism from Y ′ to the dual of Y X , which is zero. The first is a Hodge class on Y X , hence a G-invariant of ⋀ 2 ( V ⊗ Q ) = S 2 V ⊗ ⋀ 2 Q ⊕ ⋀ 2 V ⊗ S 2 Q . Since G contains the Hodge group of a simple abelian fourfold whose only Hodge classes of degree two are the divisor classes, ( S 2 V ) G = 0 and ( ⋀ 2 V ) G = Q ψ ; so the component lies in ψ ⊗ S 2 Q , which Sp ( V , ψ ) fixes. The third component is fixed because GL ( V ) acts trivially on H 1 ( Y ′ ) .
For a Weil class, let B be of Weil type for K and f : Y → B . The field K acts on B ∼ B X × B ′ through End 0 ( B X ) × End 0 ( B ′ ) , because Hom ( B X , B ′ ) = 0 ; so it acts on Q B = Hom 0 ( B X , X ) and on H 1 ( B ′ ) , and the eigenspace of H 1 ( B , Q ¯ ) for an embedding σ of K is V ⊗ Q B , σ ⊕ H 1 ( B ′ , Q ¯ ) σ . A Weil class lies in the sum over σ of the top exterior powers of these eigenspaces, and ⋀ top ( V ⊗ Q B , σ ⊕ W σ ) = ( det V ) ⊗ dim Q B , σ ⊗ ( det Q B , σ ) ⊗ 8 ⊗ det W σ , on which SL ( V ) ⊇ Sp ( V , ψ ) acts trivially. Pulling back along f preserves invariance.
For Y = X m , H * ( X m ) = ⋀ * ( V ⊗ Q m ) , and the invariants of Sp ( V , ψ ) in the tensor algebra of V are spanned by complete contractions with ψ by the first fundamental theorem for the symplectic group [51, Chap. 5]; a contraction of two factors of X m is the pull-back of the class ψ ∈ H 2 ( X × X ) along the projection to those factors, a divisor class, and a contraction within one factor is ψ ∈ H 2 ( X ) . So the invariants are polynomials in divisor classes.    □
Theorem 8. 
(F2) implies AB , and is therefore equivalent to it.
Proof. 
AB contains (F2). Assume (F2), let A be an abelian variety and 0 ≠ β ∈ Hdg p ( A ) . If β ∉ R ( A ) it is algebraic by (F2), so suppose β ∈ R ( A ) .
The line Q e 0 is not stable under Sp ( V , ψ ) : a stable line would be a one-dimensional representation of the connected semisimple group Sp ( V , ψ ) , hence trivial, and e 0 , which is not a polynomial in divisor classes, would be invariant, against Lemma 1. The stabiliser of the line is therefore a proper Zariski closed subgroup, and since Sp ( V , ψ ) ( Q ) is Zariski dense in Sp ( V , ψ )  [22, Cor. 18.3] there is g ∈ Sp ( V , ψ ) ( Q ) with g e 0 ∉ Q e 0 .
Put B = X × X , Y = A × B , and let b be an ample class on B. The decomposition of Lemma 1 for Y is the sum of those for A and B, so Sp ( V , ψ ) acts on H * ( Y ) = H * ( A ) ⊗ H * ( B ) by g ( x ⊗ y ) = g x ⊗ g y , and by the lemma it fixes R ( Y ) pointwise. If β ⊗ e 0 = pr A * β ∪ pr B * e 0 were in R ( Y ) , then g β ⊗ g e 0 = β ⊗ e 0 with β ≠ 0 , which forces g e 0 ∈ Q e 0 ; so β ⊗ e 0 ∉ R ( Y ) . On the other hand β ⊗ b 2 ∈ R ( Y ) , because the pull-back of a generator of R ( A ) or R ( B ) along a projection is a generator of R ( Y ) and R ( Y ) is an algebra. Hence also
β ⊗ ( e 0 + b 2 ) = β ⊗ e 0 + β ⊗ b 2 ∉ R ( Y ) .
The classes β ⊗ e 0 and β ⊗ ( e 0 + b 2 ) are Hodge classes of degree 2 p + 4 on Y outside R ( Y ) , so (F2) makes both algebraic, and their difference β ⊗ b 2 is algebraic. Then
pr A * ( β ⊗ b 2 ) ∪ pr B * b 6 = ∫ B b 8 β ,
a push-forward of a product of algebraic classes, is algebraic, and ∫ B b 8 > 0 because b is ample on the eightfold B. So β is algebraic.    □
Theorem 1 is Propositions 1, 2 and Theorem 8 together.

3.3. The Closure Graph

Appendix A.8 records the implications between the conjecture and the statements that have been used to approach it: the Lefschetz standard conjecture (L) for all varieties, the conjecture (M) that every Hodge class is motivated in André’s sense [6], the variational Hodge conjecture (V), the statement (F3) that the conjecture holds for all varieties that are not abelian, and the two statements of this paper.2 The rules are listed in Table 2, each with the place where it is proved; the first two rows are Propositions 1 and Theorem 8.
Proposition 4.
Under the rules of Table 2 the closure of a set of statements contains HC if and only if the set contains one of
{ F 3 } , { L , M } , { L , F 3 ′ } , { V , F 3 ′ } , { F 2 , F 3 ′ } , { AB , F 3 ′ } ,
and these six sets are the minimal ones with this property.
Proof. 
Each listed set derives HC : { F 3 } by the second row; { L , M } by the sixth; the other four through AB , which (F2), (L) and (V) each give, and the fifth row. For the converse, call a set closed if it is stable under the rules. A closed set without HC contains no (F3), does not contain (L) and (M) together, and, if it contains (F3′), contains none of AB , (F2), (L), (V), since each of these yields AB and then HC . The maximal closed sets without HC are therefore
M 1 = { F 2 , AB , L , V } , M 2 = { F 2 , AB , V , M } , M 3 = { F 3 ′ , M } ,
each of which is closed by inspection of the table. The closure of a set Σ contains HC exactly when Σ lies in none of the M i . If Σ contains (F3) this holds. Otherwise Σ needs an element outside M 1 , that is (F3′) or (M); one outside M 2 , that is (F3′) or (L); and one outside M 3 , that is one of (F2), AB , (L), (V). The minimal choices are (F3′) with one of (F2), AB , (L), (V), and (M) with (L). This gives the six sets, and each is minimal because removing any element lands in some M i . The same computation is carried out mechanically from the rule list in the Lean file of Appendix D, which returns these six sets and no others.    □
Figure 2 draws the graph. The two statements of this paper are the filled nodes; every route to HC other than (F3), which is HC itself, and (L)+(M) passes through (F3′).
Figure 2. The closure graph of Table 2. Grey arrows are rules with one premise; a black fork joins the two premises of a rule that needs both. The filled nodes are (F2) and (F3′).
Figure 2. The closure graph of Table 2. Grey arrows are rules with one premise; a black fork joins the two premises of a rule that needs both. The filled nodes are (F2) and (F3′).
Preprints 244139 g002

4. The Algebraic Locus in a Family

Definition 3.
Let π : X → S be a smooth projective morphism to a smooth connected complex quasi-projective variety, and let ξ be a global section of R 2 p π * Q that is a Hodge class at one point. The algebraic locus of ξ is the set Σ ( ξ ) of s ∈ S at which ξ s ∈ H 2 p ( X s ) is algebraic.
By the theorem of the fixed part [4,33], the flat sections of a polarisable variation form a sub-Hodge structure, constant along S; so ξ s is a Hodge class at every point once it is one somewhere.3
Lemma 2
(Dichotomy).
(i)
Σ ( ξ ) is a countable union of Zariski closed subsets of S.
(ii)
Either Σ ( ξ ) = S or Σ ( ξ ) is meagre in S an .
(iii)
The following are equivalent: Σ ( ξ ) = S ; Σ ( ξ ) has an interior point; Σ ( ξ ) is not meagre; ξ s is algebraic for the very general s. If Σ ( ξ ) is dense they are equivalent to Σ ( ξ ) being closed.
Proof. 
(i) Fix a relatively ample line bundle on X . For a Hilbert polynomial P the relative Hilbert scheme Hilb P ( X / S ) is projective over S [49], and the class of the fibre of the universal family in H 2 p ( X s ) is locally constant on each connected component of it, since a flat family of subschemes over a connected base has locally constant cycle class. A rational combination ∑ i a i [ Z i ] equal to ξ s is a point of a fibre product ∏ S Hilb P i ( X / S ) , projective over S and with countably many connected components; on each component the condition ∑ a i [ Z i ] = ξ is either empty or holds throughout, because both sides are flat sections. So Σ ( ξ ) is the union, over the countably many tuples ( P i , a i ) and components, of images of proper morphisms to S, each Zariski closed.
(ii) A countable union of proper Zariski closed subsets of an irreducible S has empty interior in S an , and each is closed, so the union is meagre; if one of the closed sets is S itself, Σ ( ξ ) = S .
(iii) By the Baire category theorem4 S an is not meagre in itself, so Σ ( ξ ) = S excludes meagreness, and a set with an interior point is not meagre; by (ii) not meagre forces Σ ( ξ ) = S . The complement of a countable union of proper closed subvarieties is the very general locus, so algebraicity at the very general point means that one of the closed sets is all of S. A dense set that is also closed is S.    □
Lemma 3
(Points of CM type are dense). Let G be a connected reductive group over Q , u 0 : U ( 1 ) → G R a homomorphism and D the orbit of u 0 under conjugation by G ( R ) + . The points u ∈ D for which u ( U ( 1 ) ) lies in the real points of a maximal torus of G defined over Q are dense in D .
Proof. 
Let u 1 ∈ D and W a neighbourhood of u 1 . A maximal torus T 1 of G R contains u 1 ( U ( 1 ) ) , since the centraliser of that compact torus is reductive and connected. Pick a regular element Y 1 in the Lie algebra t 1 of T 1 . The map ( g , Y ) ↦ Ad ( g ) Y from G ( R ) + × t 1 to g R is submersive at ( 1 , Y 1 ) because g R = t 1 ⊕ [ g R , Y 1 ] , so the image of a small neighbourhood N 1 × N 2 , with N 1 u 1 N 1 − 1 ⊆ W and N 2 made of regular elements, contains an open set of g R , hence a rational point Y = Ad ( g ) Y ′ . The identity component T of the centraliser of the rational regular semisimple Y is a maximal torus over Q with T R = g T 1 g − 1 , and u = g u 1 g − 1 ∈ W has u ( U ( 1 ) ) ⊆ T R .    □
When G = Hg ( A ) and u is the Hodge cocharacter of an abelian variety A, the condition of the lemma says that Hg ( A ) lies in a Q -torus, and then A is of CM type: a maximal Q -torus of GL ( H 1 ( A ) ) containing it is the unit group of a commutative semisimple algebra L of dimension 2 dim A commuting with Hg ( A ) , and the commutant of Hg ( A ) is End 0 ( A ) .
Proposition 5
(A family with a CM member through every Hodge class). Let A 0 be an abelian variety and γ a Hodge class on it. There are a polarised abelian scheme π : A → S over a smooth connected quasi-projective base, a flat section ξ of R 2 p π * Q , a point s 0 with A s 0 ≅ A 0 and ξ s 0 = γ , and a point c with A c of CM type; the points of CM type are dense in S.
Proof. 
Let G = Hg ( A 0 ) , let u 0 be the Hodge cocharacter of A 0 and D its G ( R ) + -orbit. Every u ∈ D defines a polarisable Hodge structure of weight one on H 1 ( A 0 ) with Hodge group contained in G, so an abelian variety A u up to isogeny; fixing the lattice H 1 ( A 0 , Z ) , the polarisation of A 0 and a level structure of level N ≥ 3 , the quotient S = Γ ∖ D by a neat congruence subgroup Γ ⊆ G ( Q ) is a smooth quasi-projective variety carrying a polarised abelian scheme with these fibres [33,87]. The class γ is invariant under G, so it is a flat section ξ of R 2 p π * Q , and ξ s 0 = γ at the image s 0 of u 0 . By Lemma 3 the points of CM type are dense in D , hence in S.    □
Theorem 9
(F2) is propagation from CM points). The following are equivalent.
(i)
(F2).
(ii)
AB .
(iii)
For every polarised abelian scheme π : A → S and every flat Hodge class ξ, algebraicity of ξ s at one point s forces Σ ( ξ ) = S .
(iv)
The same for the points s of CM type.
(v)
For every π, ξ with ξ c algebraic at a point c of CM type, Σ ( ξ ) has an interior point.
Proof. 
(i) and (ii) are Theorem 8. If (ii) holds, ξ t is a Hodge class at every t by the theorem of the fixed part and algebraic by (ii), so (iii) holds; (iv) is a case of (iii), and (iv) and (v) are equivalent by Lemma 2(iii).
Assume (iv). Let F be a CM field and n ≥ 2 . The family of abelian varieties of split Weil type for F over its Shimura variety is a polarised abelian scheme on which the Weil classes are flat sections, and by Theorem A2 they are algebraic at a member isogenous to a product of abelian varieties with complex multiplication by F, a member of CM type; so by (iv) they are algebraic on every member, and by pull-back along an F-linear isogeny on every abelian variety of split Weil type for F. By André’s theorem [5], every Hodge class on an abelian variety of CM type is a sum of pull-backs, along homomorphisms, of Weil classes of abelian varieties of split Weil type; so AB holds for every abelian variety of CM type. Now let A 0 be any abelian variety and γ a Hodge class on it, and take π , ξ , s 0 , c from Proposition 5. Then ξ c is algebraic, (iv) gives Σ ( ξ ) = S , and γ = ξ s 0 is algebraic. This is (ii).    □
Proposition 6
(The Weil classes). For every CM field F, every n ≥ 2 and every discriminant, the algebraic locus of the Weil classes in the polarised period domain D F contains an explicit point of CM type and is dense, and it is either D F or meagre.
Proof. 
This is Theorem A1, items (i) to (iii). The density is the Hecke orbit of the base point: the group G F ( Q ) of the unitary group of the hermitian form acts on D F through F-linear isogenies that preserve the polarisation up to a scalar, hence preserve the algebraicity of the Weil classes, and its closure contains G F ( R ) + , which acts transitively. The dichotomy is Lemma 2.    □
Proposition 7
((F3′) and the abelian part). Let H ab 2 p ( X ) ⊆ H 2 p ( X ) be the sum of the images of all maps Γ * : H k ( B ) → H 2 p ( X ) with B abelian and Γ an algebraic cycle on B × X . Then H ab 2 p ( X ) is a sub-Hodge structure, Ab p ( X ) = Hdg p ( X ) ∩ H ab 2 p ( X ) , and (F3′) holds for X in degree 2 p if and only if a Hodge complement T ab 2 p ( X ) of H ab 2 p ( X ) carries no nonzero Hodge class.
Proof. 
Each Γ * is a morphism of Hodge structures, so its image and the sum of finitely many images, which is all of H ab 2 p ( X ) by dimension, are sub-Hodge structures. A Hodge class in H ab 2 p ( X ) is ∑ Γ j * y j , and the argument of Proposition 2(i) replaces each y j by a Hodge class; so Ab p ( X ) = Hdg p ( X ) ∩ H ab 2 p ( X ) . A Hodge complement exists because the Hodge group is reductive, and the Hodge classes of H 2 p ( X ) are the direct sum of those of the two summands, which gives the last assertion.    □
Theorem 2 is Lemma 2, Theorem 9, Proposition 6 and Proposition 7. Figure 3 draws the locus of Proposition 6.
Figure 3. The period domain D n , n of a Weil family as a surface. The star is the CM base point, the dots and curves are strata of the algebraic locus, dense by the Hecke orbit. The locus is a countable union of closed strata; it is either the whole surface or meagre.
Figure 3. The period domain D n , n of a Weil family as a surface. The star is the CM base point, the dots and curves are strata of the algebraic locus, dense by the Hecke orbit. The locus is a countable union of closed strata; it is either the whole surface or meagre.
Preprints 244139 g003
Both statements are now of one shape. There is a dense set of points at which the class is algebraic by an explicit cycle; the locus is a countable union of closed strata; the statement is that this union is not meagre, that is, that algebraicity propagates from the dense set. The rest of the paper goes to the first point where the propagation is not known and builds the object that would carry it.

5. Constructions at (F3′): Real Multiplication on K3 Surfaces

5.1. Where the First Class Is

Let S be a projective K3 surface, T = T ( S ) Q and E = End Hdg T . The Hodge classes of S × S are H 0 ⊗ H 4 , H 4 ⊗ H 0 and End Hdg H 2 ( S ) = End ( NS Q ) ⊕ E , since T is irreducible with no class of type ( 1 , 1 ) and Hom Hdg ( T , NS Q ) = 0 . The first two pieces and End ( NS Q ) are spanned by products of divisor classes. So the Hodge conjecture for S × S is the algebraicity of E, and by Proposition 7 the same classes decide (F3′) for S × S once T is not reached from abelian varieties. When E is a CM field it is spanned by its elements of norm one, which are Hodge isometries of T and algebraic by Buskin’s theorem [26]; see [115, Rem. 2.16]. When E is totally real, the Rosati involution is the identity on it [126], every e ∈ E is self-adjoint for the intersection form q, and q ( e x , e y ) = q ( e 2 x , y ) . The first class is a real multiplication.

5.2. Chains of Correspondences

Lemma 4.
Let f : S ⤏ S ′ be a dominant rational map of K3 surfaces of degree d and Γ f the closure of its graph. Then f * = Γ f * and f * = Γ f * t are morphisms of Hodge structures, f * maps T ( S ′ ) Q isomorphically onto T ( S ) Q with q S ( f * a , f * b ) = d q S ′ ( a , b ) , and f * = d ( f * ) − 1 on T ( S ) Q .
Proof. 
The graph is an algebraic cycle, so both maps are induced by algebraic classes and are morphisms of Hodge structures. Pull-back of a nowhere vanishing 2-form is a nonzero 2-form, so f * ( T ( S ′ ) Q ) is a sub-Hodge structure of H 2 ( S ) that meets H 2 , 0 ( S ) ; the smallest sub-Hodge structure of H 2 ( S ) containing H 2 , 0 is T ( S ) Q , so f * T ( S ′ ) Q ⊇ T ( S ) Q , and since T ( S ′ ) Q is irreducible and f * is injective on it ( f * f * = d ), equality of dimensions gives f * T ( S ′ ) Q = T ( S ) Q . The projection formula gives q S ( f * a , f * b ) = q S ′ ( f * f * a , b ) = d q S ′ ( a , b ) and f * f * = d .    □
Lemma 5.
A Hodge isometry φ : T ( S ) Q → T ( S ′ ) Q is induced by an algebraic class on S × S ′ .
Proof. 
The spaces NS ( S ) Q and NS ( S ′ ) Q are the orthogonal complements of T ( S ) Q and T ( S ′ ) Q in the two copies of Λ Q , isometric to each other through φ ; by Witt’s cancellation theorem the complements are isometric, so φ extends to a Hodge isometry H 2 ( S ) → H 2 ( S ′ ) , which is algebraic by Buskin’s theorem [26] (see [59, Thm. 0.1]). The projector of H 2 ( S ) onto T ( S ) Q along NS ( S ) Q is algebraic, being 1 minus the projector onto NS , which is a sum ∑ i , j c i j D i ⊗ D j over a basis of divisor classes with ( c i j ) the inverse Gram matrix; composing gives φ on T and zero on NS .    □
Theorem 10
(Chains). Let S be a projective K3 surface with E totally real and e ∈ E with e 2 = n ∈ Q > 0 . Suppose there are projective K3 surfaces S = S 0 , S 1 , ⋯ , S r and, for each i, either a dominant rational map S i − 1 ⤏ S i of degree n i , a dominant rational map S i ⤏ S i − 1 of degree n i , or a Hodge isometry T ( S i − 1 ) Q ≅ T ( S i ) Q (and then n i = 1 ), such that
∏ i n i ± 1 ≡ n ( mod Q × 2 ) ,
the sign being + for maps leaving S i − 1 and − for maps arriving at it. Then e is algebraic on S × S .
Proof. 
Composing along the chain the maps f * for maps leaving S i − 1 , f * for maps arriving, and the classes of Lemma 5 for the isometries, gives an algebraic isomorphism ϕ : T ( S r ) Q → T ( S ) Q of Hodge structures with q S ( ϕ a , ϕ b ) = μ q S r ( a , b ) , where μ = ∏ n i ± 1 by Lemma 4 (a map leaving contributes f * = n i ( f * ) − 1 , of multiplier n i ; a map arriving contributes f * , of multiplier n i in the other direction, that is n i − 1 for the composite in the direction of ϕ ). Write μ = n c 2 with c ∈ Q × . Since e is self-adjoint with e 2 = n , it is a Hodge similarity of T ( S ) Q of multiplier n. So ψ = ( c − 1 ϕ ) − 1 ∘ e : T ( S ) Q → T ( S r ) Q is a Hodge isometry, algebraic by Lemma 5, and e = ( c − 1 ϕ ) ∘ ψ is a composite of algebraic correspondences, hence algebraic on S × S .    □
The case r = 1 with S ⤏ S 1 the quotient by a symplectic automorphism of prime order p is Varesco’s theorem [115, Thm. 2.1], which gives e = p and is used for p = 2 , 3 . What a chain adds is a composite multiplier.

5.3. Real Multiplication by Q ( 6 )

A symplectic automorphism of finite order on a K3 surface acts trivially on T ( S ) , its quotient has a K3 resolution Y [94], and T ( Y ) Q ≅ T ( S ) Q ( 1 / ord ) by Lemma 4. The invariant lattice of a symplectic involution is U 3 ⊕ E 8 ( − 2 ) , and a K3 surface has a symplectic involution exactly when E 8 ( − 2 ) embeds primitively in its Néron–Severi lattice [81,120]. The invariant lattice of a symplectic automorphism of order three is U ⊕ U ( 3 ) 2 ⊕ A 2 2 , and such an automorphism exists exactly when the Coxeter–Todd lattice K 12 ( − 2 ) embeds primitively in the Néron–Severi lattice [50, Thm. 4.1]. Put
V 2 = U 3 ⊕ E 8 ( − 2 ) Q , V 3 = U ⊕ U ( 3 ) 2 ⊕ A 2 2 Q ,
of ranks 14 and 10 and signatures ( 3 , 11 ) and ( 3 , 7 ) ; over Q , U ( 3 ) ≅ U and E 8 ( − 2 ) ≅ 〈 − 1 〉 8 ≅ 〈 − 2 〉 8 . Varesco [115, Props. 2.5 and 2.11] shows that a projective K3 surface whose transcendental space embeds isometrically in V 2 (resp. V 3 ) is Hodge isometric to one with a symplectic involution (resp. a symplectic automorphism of order three).
Lemma 6
(Embedding criterion). Let W and V be rational quadratic spaces with rk V − rk W ≥ 3 . Then W embeds isometrically in V if and only if the signature of V dominates that of W in both signs. In that case a complement W ′ exists with V ≅ W ⊕ W ′ .
Proof. 
Necessity is clear. For sufficiency put k = rk V − rk W ≥ 3 and let ( s + , s − ) be the difference of signatures, s + + s − = k . A quadratic space W ′ of rank k and signature ( s + , s − ) with prescribed discriminant d ′ = d ( V ) d ( W ) and prescribed Hasse invariants ε p ′ exists provided the ε p ′ equal 1 for almost all p, ∏ p ≤ ∞ ε p ′ = 1 , and ε ∞ ′ , d ′ match the signature [108, Chap. IV, Prop. 7], with no further condition when k ≥ 3 . Take ε p ′ = ε p ( V ) ε p ( W ) ( d ( W ) , d ′ ) p , the value forced by the formula ε p ( W ⊕ W ′ ) = ε p ( W ) ε p ( W ′ ) ( d ( W ) , d ( W ′ ) ) p . The product formula for these holds because it holds for V, for W and for the Hilbert symbol. Then W ⊕ W ′ and V have the same rank, signature, discriminant and Hasse invariants at every place, so they are isometric by the Hasse–Minkowski theorem [108, Chap. IV, Thm. 9], and W embeds in V.    □
Theorem 11
(Two-dimensional families with algebraic real multiplication by Q ( 6 ) ). Let S be a projective K3 surface with Q ( 6 ) ⊆ E whose transcendental space T = T ( S ) Q embeds isometrically in V 3 .
(i)
6 ∈ E is algebraic on S × S .
(ii)
The Hodge conjecture, and (F3′) , hold for S × S and for the Hilbert square S [ 2 ] .
(iii)
Such surfaces exist in two-dimensional families, whose very general member has rk T = 8 and E = Q ( 6 ) exactly. One such family is the one whose transcendental space is E 4 with the trace form of q E = 〈 − 2 , − 2 , − 2 + 6 , − 1 + 6 〉 over E = Q ( 6 ) .
Proof. 
(i) By [115, Prop. 2.11] S is Hodge isometric to a K3 surface S ′ with a symplectic automorphism σ of order three, and by Lemma 5 the isometry is algebraic, so we may replace S by S ′ . Let Y be the minimal resolution of S ′ / σ and f 1 : S ′ ⤏ Y the quotient map, of degree three, so that T ( Y ) Q ≅ T ( 1 / 3 ) , of rank rk T ≤ 8 and signature ( 2 , rk T − 2 ) : the rank is at most 8 because T embeds in V 3 of rank 10 and signature ( 3 , 7 ) with a complement that must contain the remaining positive direction, so the complement has rank at least two. Now T ( Y ) Q has rank at most 8 and signature dominated by ( 3 , 11 ) , and 14 − 8 ≥ 3 , so by Lemma 6 it embeds in V 2 . By [115, Prop. 2.5] Y is Hodge isometric to a K3 surface Y ′ with a symplectic involution ι ; let Z be the resolution of Y ′ / ι and f 2 : Y ′ ⤏ Z the quotient, of degree two. The chain S ′ ⤏ Y ≅ Y ′ ⤏ Z has multiplier 3 · 1 · 2 = 6 , and Theorem 10 applies to e = 6 (Figure 4).
(ii) By the first paragraph of this section it remains to make E algebraic. E contains Q ( 6 ) , and either E = Q ( 6 ) or E is a CM field: a totally real E has dim E T ≥ 3 , because the complex line H 2 , 0 lies in one eigenspace T σ ⊗ C and H 2 , 0 ¯ in the same one, so that T σ ⊗ R has signature ( 2 , * ) and dim E T ≥ 3 follows from q being negative definite on the other eigenspaces together with T ⊗ R having exactly two positive directions; and [ E : Q ] · dim E T = rk T ≤ 8 then forces [ E : Q ] ≤ 2 . If E = Q ( 6 ) , (i) suffices. If E is CM, it is spanned by its norm-one elements, Hodge isometries, algebraic by Buskin’s theorem. So every Hodge class on S × S is algebraic, and (F3′) follows. For S [ 2 ] the Chow motive is a direct summand of twists of h ( S × S ) , h ( S ) and 1  [32], so every Hodge class on S [ 2 ] is the image under an algebraic correspondence of a Hodge class on S × S or S, and both statements follow.
(iii) Let E = Q ( 6 ) with real embeddings σ 1 , σ 2 ( 6 ↦ ± 6 ), and let q E = 〈 a 1 , a 2 , a 3 , a 4 〉 with a 1 , a 2 negative at both places and a 3 , a 4 positive at σ 1 , negative at σ 2 . The Q -space T = E 4 with the form Tr E / Q ∘ q E has rank 8 and signature ( 2 , 2 ) + ( 0 , 4 ) = ( 2 , 6 ) , and multiplication by 6 is self-adjoint of square 6. For q E = 〈 − 2 , − 2 , − 2 + 6 , − 1 + 6 〉 one computes d ( T ) = 10 , ε 2 = 1 , ε 3 = − 1 , ε 5 = 1 and ε p = 1 otherwise, while V 3 has d = − 1 , ε 2 = 1 , ε 3 = − 1 , ε 5 = 1 ; and T ⊕ 〈 1 , − 10 〉 has the invariants of V 3 at every place, so V 3 ≅ T ⊕ 〈 1 , − 10 〉 by Hasse–Minkowski, with a complement containing a positive vector (Table 3; the Hilbert symbols are recomputed in the programs of Appendix D). Embed V 3 in Λ Q as the invariant lattice of an order-three symplectic action, with orthogonal complement K 12 ( − 2 ) , and let N ⊆ Λ be the primitive sublattice with N Q = T ⊥ ; it contains K 12 ( − 2 ) primitively and a positive class. The Hodge structures of K3 type on T for which 6 is a Hodge endomorphism are the periods ω ∈ P ( T σ 1 ⊗ C ) with q ( ω ) = 0 , q ( ω , ω ¯ ) > 0 , where T σ 1 is the σ 1 -eigenspace, of signature ( 2 , 2 ) ; this is a domain of dimension two [115, Prop. 2.3]. For each such ω the surjectivity of the period map gives a K3 surface S ω with T ( S ω ) Q ⊆ T , projective because N contains a positive class, carrying an order-three symplectic automorphism because K 12 ( − 2 ) ⊆ N primitively. For ω outside a countable union of proper closed analytic subsets, T ( S ω ) Q = T and E = Q ( 6 ) , by the argument of Theorem A23, which needs only dim E T ≥ 3 .    □
In Varesco’s families E is Q ( 2 ) or Q ( 3 )  [115, Thms. 2.9, 2.15]; van Geemen and Schütt obtain 2 , 3 , 7 from rational self-maps of elliptic K3 surfaces [121, §§6–7]. The lattice condition of Theorem 11 is a genuine restriction: among the 165 sign-admissible diagonal forms q E with small coefficients, 98 embed in V 3 and 67 do not (Appendix D).
For Q ( 2 ) the lattice condition is empty in low rank, and a single Nikulin quotient settles every surface of transcendental rank at most ten.
Theorem 12
(Real multiplication by Q ( 2 ) in transcendental rank at most ten). Let S be a projective K3 surface with Q ( 2 ) ⊆ E and rk T ( S ) ≤ 10 .
(i)
2 ∈ E is algebraic on S × S .
(ii)
The Hodge conjecture, and (F3′) , hold for S × S and for S [ 2 ] .
(iii)
For E = Q ( 2 ) and m ∈ { 4 , 5 } , the real multiplication is algebraic at every member of every family D W of Theorem 18, the very general one included, and (ii) holds there.
Proof. 
(i) T = T ( S ) Q has rank t ≤ 10 and signature ( 2 , t − 2 ) , and V 2 has rank 14 and signature ( 3 , 11 ) . Since 14 − t ≥ 4 , Lemma 6 embeds T in V 2 . By [115, Prop. 2.5]S is Hodge isometric to a K3 surface with a symplectic involution, and [115, Thm. 2.1] with p = 2 makes 2 algebraic; this is Theorem 10 for the chain formed by the Hodge isometry and the quotient by the involution.
(ii) As in the proof of Theorem 11(ii): a totally real E has dim E T ≥ 3 , so [ E : Q ] ≤ 3 , and [ E : Q ] is even because E ⊇ Q ( 2 ) , so E = Q ( 2 ) and (i) applies; a CM field E is spanned by its norm-one elements, which are algebraic by Buskin’s theorem [26]. So every Hodge class on S × S is algebraic, (F3′) follows, and S [ 2 ] follows by [32].
(iii) For [ ω ] ∈ D W the surface S ω is projective and T ( S ω ) Q ⊆ Tr W is preserved by 2 , as in the proof of Theorem 18; its rank is at most 2 m ≤ 10 , so (i) and (ii) apply.    □
Corollary 1
(Transcendental rank six for e ∈ { 2 , 3 , 6 } ). Let S be a projective K3 surface with rk T ( S ) ≤ 6 and Q ( e ) ⊆ E for some e ∈ { 2 , 3 , 6 } . Then e is algebraic on S × S , and the Hodge conjecture and (F3′) hold for S × S and S [ 2 ] . This holds at every member of the families of Witt index one of Corollary 6 for e ∈ { 2 , 3 , 6 } , which Floccari’s theorem does not reach, in particular for the first, second and fourth rows of Table 4.
Proof. 
The space T ( S ) Q has rank at most 6 and signature ( 2 , * ) , so it embeds in V 3 , of rank 10 and signature ( 3 , 7 ) , by Lemma 6. For e = 2 apply Theorem 12, for e = 3  [115, Prop. 2.11, Thm. 2.1], and for e = 6 Theorem 11. A totally real E has 3 [ E : Q ] ≤ 6 , so E = Q ( e ) or E is a CM field, and the rest is the proof of Theorem 11(ii).    □
In transcendental rank at least eight the quotients by finite symplectic groups stop at a lattice condition, which we determine.
Proposition 8
(The reach of symplectic quotients). Let S be a projective K3 surface with E = Q ( e ) and rk T ( S ) = 2 m ≥ 8 , and put T = T ( S ) Q .
(i)
Some chain as in Theorem 10 whose rational maps are quotients by finite groups of symplectic automorphisms makes e algebraic if and only if either e = 2 and T embeds in V 2 , or e ∈ { 3 , 6 } , m = 4 and T embeds in V 3 .
(ii)
For e = 2 the embedding exists for m ≤ 5 , never for m ≥ 7 , and for m = 6 exactly when ε p ( T ) = ε p ( V 2 ) at every prime p at which d ( T ) is a square in Q p . For e ∈ { 3 , 6 } and m = 4 it exists exactly when ε p ( T ) = ε p ( V 3 ) at every such prime, and never for m ≥ 5 .
(iii)
If m ≥ 7 , no dominant rational map of degree at least two from S to a K3 surface is birational to the quotient by a finite group.
Proof. 
(ii) V 2 and V 3 have discriminant − 1 (Table 3). For m ≤ 5 , resp. m ≤ 3 , the codimension is at least 3 and Lemma 6 applies. A complement of T must contain the third positive direction, so the codimension is at least 1, and at least 2 because both ranks are even; this excludes m ≥ 7 in V 2 and m ≥ 5 in V 3 . In codimension two the complement Y is binary of signature ( 1 , 1 ) , with d ( Y ) = − d ( T ) and, by the formula in the proof of Lemma 6, ε p ( Y ) = ε p ( V ) ε p ( T ) ( d ( T ) , − d ( T ) ) p = ε p ( V ) ε p ( T ) . A binary form with these local invariants exists at p unless d ( Y ) ≡ − 1 , that is d ( T ) is a square in Q p , and ε p ( Y ) = − 1  [108, Chap. IV, Prop. 6]; at the real place both V and T have ε ∞ = − 1 , so ε ∞ ( Y ) = 1 as signature ( 1 , 1 ) requires. The global form exists by [108, Chap. IV, Prop. 7], and V ≅ T ⊕ Y by the Hasse–Minkowski theorem.
(i) If the condition holds, 2 is algebraic by Theorem 12(i) and its proof, 3 by [115, Prop. 2.11, Thm. 2.1], and 6 by Theorem 11(i). Conversely, let a chain make e algebraic. By Proposition 20 every degree is of the form 2 a 3 b , and the multiplier is e modulo squares. For e ∈ { 2 , 6 } some degree is even, and for e ∈ { 3 , 6 } some degree is divisible by 3; the group of that quotient then contains a symplectic involution, resp. a symplectic automorphism of order three, on a surface X of the chain, and T ( X ) Q lies in its rational invariant lattice, V 2 resp. V 3 . By Lemmas 4 and 5, T ( X ) Q ≅ T ( k ) for some k ∈ Q × . For a rational quadratic space of rank n ≡ 0 ( mod 4 ) and discriminant d one has d ( T ( k ) ) = d and ε p ( T ( k ) ) = ( k , d ) p ε p ( T ) , since the exponents n ( n − 1 ) / 2 of ( k , k ) p and n − 1 of ( k , d ) p in the expansion of ∏ i < j ( k a i , k a j ) p are even and odd. So at the primes where d is a square, T ( k ) and T have the same Hasse invariants. By (ii), T ( k ) embeds in V 3 only for m = 4 , where 2 m = 8 , and in V 2 only for m ≤ 6 , where the criterion for m = 6 , with 2 m = 12 , does not see k; for m ≤ 5 there is nothing to prove. So T embeds, which is the condition.
(iii) If S ⤏ S ′ is birational to S → S / G with G ≠ 1 finite, G acts trivially on H 2 , 0 ( S ) , since the resolution of S / G has p g = 1 ; so G is symplectic and contains an element of prime order, whose coinvariant lattice has rank at least 8 [94, § 5]. It is negative definite and lies in NS ( S ) , which also contains an ample class orthogonal to it, so rk T ( S ) ≤ 22 − 8 − 1 = 13 .    □
For e = 2 and m = 6 both cases occur: for W = 〈 2 , 2 2 , − 4 − 2 2 , − 4 + 2 2 , − 4 + 2 2 , − 3 + 2 〉 one finds d ( T ) = 14 and the criterion holds, and for W = 〈 1 + 2 2 , 3 + 3 2 , − 3 − 2 2 , − 3 − 2 2 , − 3 − 2 , − 3 − 2 〉 one finds d ( T ) = 7 , a square in Q 3 , with ε 3 ( T ) = − 1 ≠ 1 = ε 3 ( V 2 ) . So the Hecke family of Section 5.9 remains the only route at the very general member for Q ( 2 ) with m ≥ 6 outside the criterion, for Q ( 3 ) and Q ( 6 ) with m ≥ 5 , and for every other real quadratic field with m ≥ 4 ; by (iii), when m ≥ 7 every rational map realising the real multiplication is non-Galois.

5.4. Isogenies of Elliptic Surfaces

Varesco’s quotients and the chains of Theorem 10 produce the square roots of the orders of symplectic automorphisms, 2 and 3, and their products. A different source of similarities is a rational self-map of a K3 surface, and the self-maps that exist in families come from isogenies of elliptic fibrations: van Geemen and Schütt obtain 2 and 3 from isogenies of degree 2 and 3 in this way, and for degrees 5 and 7 they give one surface each [121, §§6–7]. We show that the degrees 5 and 7 come in one-dimensional families, that in the families the isogeny is a complex multiplication by Q ( − 5 ) or Q ( − 7 ) and not a real one, the sign of its square being decided by how the twisting function behaves under the Fricke involution, and that the families satisfy the Hodge conjecture on all their powers.
Let N ∈ { 5 , 7 } and write N k for 125 when N = 5 and 49 when N = 7 . The modular curve X 0 ( N ) is rational, with Fricke’s coordinate t in which the involution ( E , C ) ↦ ( E / C , E [ N ] / C ) is w ( t ) = N k / t  [121, §7.6].5 The universal curve with a cyclic subgroup C t of order N over P t 1 is
E t : y 2 = x 3 − 27 c 4 ( t ) x − 54 c 6 ( t ) ,
with, for N = 7 ,
c 4 = ( t 2 + 13 t + 49 ) ( t 2 + 5 t + 1 ) , c 6 = ( t 2 + 13 t + 49 ) ( t 4 + 14 t 3 + 63 t 2 + 70 t − 7 ) ,
and for N = 5 ,
c 4 = ( t 2 + 10 t + 5 ) ( t 2 + 22 t + 125 ) , c 6 = ( t 2 + 4 t − 1 ) ( t 2 + 22 t + 125 ) 2 .
The discriminant is 1728 t ( t 2 + 13 t + 49 ) 2 for N = 7 and 1728 t ( t 2 + 22 t + 125 ) 3 for N = 5 , of degree 5 and 7. So the relatively minimal model R N of (2) over P t 1 is a rational elliptic surface with fibres of type I 1 at t = 0 , I N at t = ∞ , and at the two roots e 1 , e 2 of the quadratic factor, where c 4 vanishes, of type II for N = 7 and III for N = 5 ; the Euler numbers add up to 1 + N + 2 · 2 = 12 and 1 + N + 2 · 3 = 12 . The roots satisfy e 1 e 2 = N k , so w exchanges them and exchanges 0 with ∞.
Lemma 7
(The Fricke isogeny over the function field). E t / C t ≅ E w ( t ) over C ( t ) .
Proof. 
The two curves have the same j-invariant, because w is the Fricke involution of X 0 ( N ) . Two elliptic curves over C ( t ) with the same non-constant j-invariant differ by a quadratic twist χ ∈ C ( t ) × / C ( t ) × 2 , and a twist ramified at a point of P 1 changes the Kodaira type of the fibre there, exchanging I n with I n * , II with IV * and III with III * . So it suffices to show that E t / C t has the same fibre types as E w ( t ) , which are I N at 0, I 1 at ∞ and II or III at e 1 , e 2 : then χ is unramified on P 1 , hence constant, hence trivial over C .
At t = 0 the fibre of E is I 1 , the Tate curve G m / q Z over C ( ( t ) ) , whose only cyclic subgroup of order N rational over C ( ( t ) ) is μ N ; the quotient G m / q N Z is I N . At e i the local monodromy M ∈ SL 2 ( Z ) of E has order 6 or 4 and irreducible characteristic polynomial; the isogeny is given on the period lattices by a matrix A ∈ M 2 ( Z ) with det A = N , and the monodromy of the quotient is M ′ = A M A − 1 . The SL 2 ( Z ) -classes with this characteristic polynomial are those of M and M − 1 , which are the types II and II * , or III and III * . If M ′ = g M − 1 g − 1 with g ∈ SL 2 ( Z ) , then B = g − 1 A conjugates M to M − 1 . All rational matrices doing so form the coset B 0 Q ( M ) × , where B 0 = 0 1 1 0 for M = 0 − 1 1 tr M has determinant − 1 and the elements of the imaginary quadratic field Q ( M ) have positive norm; so det B < 0 , against det B = N . Hence the quotient has the type of M at e 1 , e 2 . At ∞ the j-invariant of the quotient has a simple pole, so the fibre is I 1 or I 1 * , and the Euler numbers of the other fibres add up to 11; since the total is 12 the fibre is I 1 .    □
For p ≠ q in P t 1 let u p , q = ( t − p ) ( t − q ) , and let X p , q be the elliptic K3 surface whose generic fibre is the quadratic twist E t ⊗ u p , q : the twist multiplies c 4 by u 2 and c 6 by u 3 , giving Weierstrass coefficients of degrees 8 and 12, and adds two fibres of type I 0 * at p and q when these are not among 0 , ∞ , e 1 , e 2 , by Tate’s algorithm; the same bookkeeping of Weierstrass coefficients, fibre types and their resolutions in higher dimension is carried out in [15]. For a double cover ψ : P s 1 → P t 1 let X ψ be the relatively minimal model of ψ * E .
Lemma 8. 
(i)
X p , q has Picard number at least 16: its Néron–Severi lattice contains U ⊕ A 6 ⊕ D 4 2 for N = 7 and U ⊕ A 4 ⊕ A 1 2 ⊕ D 4 2 for N = 5 . The same bound holds for X ψ , whose fibres 2 I 1 + 2 I N , together with four fibres of type III when N = 5 , and the pulled-back Mordell–Weil group of R N , of rank 2, give rank at least 16 by the Shioda–Tate formula.
(ii)
The fibrewise isogeny by C t is a dominant rational map of degree N
f : X p , q ⤏ X w ( p ) , w ( q ) , f : X ψ ⤏ X w ∘ ψ .
(iii)
The deck involution of ψ times [ − 1 ] on the fibres is a symplectic involution ι of X ψ , and the minimal resolution of X ψ / ι is X p , q with { p , q } the branch locus of ψ. Hence T ( X ψ ) Q ≅ T ( X p , q ) Q ( 1 / 2 ) as in Lemma 4, and the two surfaces have the same Picard number.
Proof. 
(i) Shioda–Tate with the fibres listed, and for X ψ the Mordell–Weil rank 8 − 6 = 2 of R N . (ii) By Lemma 7 the quotient of the generic fibre is E w ( t ) ⊗ u p , q ( t ) , and
u p , q ( w ( t ) ) = N k t − p N k t − q = p q t 2 u w ( p ) , w ( q ) ( t ) ,
a constant times a square times u w ( p ) , w ( q ) . So E w ( t ) ⊗ u p , q ( t ) ≅ E w ( t ) ⊗ u w ( p ) , w ( q ) ( w ( t ) ) , which is the pull-back along the automorphism w of the base of the generic fibre of X w ( p ) , w ( q ) ; pulling back along an automorphism of the base gives an isomorphic surface. For X ψ , ( ψ * E ) / C = ψ * ( E / C ) ≅ ψ * w * E . (iii) The quadratic twist by the function defining the double cover is the quotient of the base change by the deck involution composed with [ − 1 ] on the fibres; the involution is symplectic with eight fixed points, the 2-torsion of the two fibres over the ramification points, and the quotient’s resolution is the twist.    □
Lemma 9
(The sign of the Fricke isogeny). Let φ t : E t → E w ( t ) be an isogeny over C ( t ) with kernel C t , which exists by Lemma 7 and is unique up to sign, and let φ t * ( d x / y ) = λ ( t ) d x / y in the models (2).
(i)
φ w ( t ) ∘ φ t = [ − N ] on E t , and λ ( t ) λ ( w ( t ) ) = − N .
(ii)
Let u ∈ C ( t ) satisfy u ( w ( t ) ) = c ( t ) 2 u ( t ) with c ∈ C ( t ) , let X be the elliptic K3 surface with generic fibre E t ⊗ u , and let f : X ⤏ X be the fibrewise isogeny by C t followed by the isomorphism ( x , y ) ↦ ( c 2 x , c 3 y ) from E w ( t ) ⊗ u ( t ) to E w ( t ) ⊗ u ( w ( t ) ) , the fibre of X over w ( t ) . Then f * acts on H 2 , 0 ( X ) by a constant Λ with
Λ 2 = − N c ( t ) c ( w ( t ) ) .
For u = ( t − p ) ( t − N k / p ) one has c ( t ) c ( w ( t ) ) = 1 and Λ 2 = − N ; for u = t 2 − N k one has c ( t ) c ( w ( t ) ) = − 1 and Λ 2 = N .
Proof. 
(i) The composite is an endomorphism of E t over C ( t ) of degree N 2 . Since j is not constant, End C ( t ) ( E t ) = Z , so the composite is [ N ] or [ − N ] , and changing φ to − φ does not change it. We read off the sign on the upper half plane H . Let τ ↦ t ( τ ) be the quotient map H → X 0 ( N ) . By the moduli interpretation of X 0 ( N ) the pair ( E t ( τ ) , C t ( τ ) ) is ( C / ( Z τ + Z ) , 〈 1 / N 〉 ) , analytically in τ , and the Fricke involution lifts to τ ↦ w ( τ ) = − 1 / N τ . The map f τ ( z ) = z / τ carries Z τ + Z into Z w ( τ ) + Z = Z / N τ + Z , so it is an isogeny C / ( Z τ + Z ) → C / ( Z w ( τ ) + Z ) , and its kernel τ ( Z w ( τ ) + Z ) / ( Z τ + Z ) = ( Z / N + Z τ ) / ( Z τ + Z ) is 〈 1 / N 〉 . Its composite with the map of the same form at w ( τ ) is f w ( τ ) ( f τ ( z ) ) = z / τ w ( τ ) = − N z . On the set H ′ of τ with j ( τ ) ∉ { 0 , 1728 } the automorphisms of E τ are ± 1 , so φ t ( τ ) = ε ( τ ) f τ with ε ( τ ) ∈ { ± 1 } ; ε is continuous because φ is, hence constant on H ′ , which is H minus a discrete set and connected. So φ w ( t ) ∘ φ t = ε 2 f w ( τ ) ∘ f τ = [ − N ] on H ′ , hence identically. Pulling back d x / y along the composite gives λ ( t ) λ ( w ( t ) ) d x / y on one side and − N d x / y on the other.
(ii) Over C ( t ) ( u ) the twist by u is the same change of coordinates ( x , y ) ↦ ( u x , u 3 / 2 y ) on E t and on E w ( t ) , which multiplies d x / y by u − 1 / 2 on both, so the twisted isogeny pulls d x / y back to λ ( t ) d x / y . The isomorphism ( x , y ) ↦ ( c 2 x , c 3 y ) pulls d x / y back to c ( t ) − 1 d x / y , and the holomorphic two-form of X, which is d x / y ∧ d t on the Weierstrass model, restricts to the fibre over w ( t ) as d x / y ∧ d ( w ( t ) ) = w ′ ( t ) d x / y ∧ d t . Hence f * ( d x / y ∧ d t ) = Λ ( t ) d x / y ∧ d t with Λ ( t ) = w ′ ( t ) λ ( t ) / c ( t ) . The pull-back of a holomorphic two-form along a dominant rational map of K3 surfaces is holomorphic off the finitely many points of indeterminacy, hence everywhere, so f * preserves the line H 2 , 0 ( X ) and Λ is a constant. Then Λ 2 = Λ ( t ) Λ ( w ( t ) ) = w ′ ( t ) w ′ ( w ( t ) ) λ ( t ) λ ( w ( t ) ) / c ( t ) c ( w ( t ) ) , where w ′ ( t ) w ′ ( w ( t ) ) = ( w ∘ w ) ′ ( t ) = 1 and λ ( t ) λ ( w ( t ) ) = − N by (i). For u = ( t − p ) ( t − N k / p ) equation (3) with q = N k / p gives u ( w ( t ) ) = ( N k / t 2 ) u ( t ) , so c ( t ) = N k / 2 / t for a fixed square root N k / 2 , and c ( t ) c ( w ( t ) ) = N k / t w ( t ) = 1 . For u = t 2 − N k , u ( w ( t ) ) = N 2 k / t 2 − N k = − ( N k / t 2 ) u ( t ) , so c ( t ) = i N k / 2 / t and c ( t ) c ( w ( t ) ) = − 1 .    □
The repository [9] computes λ itself: the kernel polynomial of φ t is the factor of degree ( N − 1 ) / 2 of the N-th division polynomial over Z [ t , x ] , Vélu’s formulas give the quotient curve, and the isomorphism to (2) at w ( t ) has λ ( t ) 2 = − t 2 / 25 for N = 5 and − t 2 / 7 for N = 7 , so that λ ( t ) 2 λ ( w ( t ) ) 2 = N 2 and Λ 2 = − N on the pair twists directly; the program julia/frickesign.jl certifies the kernel polynomial (it divides the division polynomial, and it and its cofactor are irreducible modulo a prime) and carries out the rest, and the Lean file checks the polynomial identity it reduces to.
Theorem 13
(Fricke families). Let N ∈ { 5 , 7 } and K = Q ( − N ) . Let A N be the family of surfaces X p = X p , N k / p for p ∉ { 0 , ∞ , e 1 , e 2 , ± N k } , and B N the family of surfaces X ψ with ψ branched over a pair { b , N k / b } .
(i)
Every member X has a dominant rational self-map f of degree N with f * ∈ E = End Hdg T ( X ) Q , f * f * = N , ( f * ) 2 = − N and f * = − f * . So K ⊆ E , with − N = ± f * algebraic on X × X , and E is a CM field.
(ii)
The families are not isotrivial, and every member has rk T ( X ) ∈ { 2 , 4 , 6 } . At all but countably many members E = K and the Hodge group of T ( X ) is the unitary group of the K-hermitian form h on T ( X ) with Tr K / Q h = q X ; at the others T ( X ) is of CM type.
(iii)
For every member X and every n ≥ 1 the Hodge conjecture holds for X n and for the Hilbert scheme X [ n ] , and (F3′) holds there.
Van Geemen and Schütt state that the Fricke parametrisations give one-dimensional families with complex multiplication by Q ( − 5 ) and Q ( − 7 ) , and give the details for the two surfaces at the fixed points of w only [121, §7.6]; (i) and (ii) make their statement explicit, and (iii) is what the complex multiplication gives.
Proof. 
(i) w exchanges p and N k / p , so Lemma 8(ii) gives f : X p ⤏ X p , the map of Lemma 9(ii) for u = ( t − p ) ( t − N k / p ) ; for B N , w ∘ ψ has the same branch locus as ψ , so w ∘ ψ = ψ ∘ σ for a Möbius transformation σ and X w ∘ ψ ≅ X ψ . The correspondence of a dominant rational map preserves NS and its transpose does too, so it preserves T = NS ⊥ and acts there as a morphism of Hodge structures, nonzero since f * ω = Λ ω with Λ ≠ 0 ; on a resolution S of the graph with p 1 : S → X birational and p 2 of degree N, both p 1 * T and p 2 * T are irreducible sub-Hodge structures of H 2 ( S ) containing its unique ( 2 , 0 ) -line, so they coincide, and f * f * = p 2 * p 1 * p 1 * p 2 * = N on T. The field E acts faithfully on the line H 2 , 0 ( X ) , and f * acts there by Λ with Λ 2 = − N by Lemma 9(ii); so ( f * ) 2 = − N in E, and f * = N ( f * ) − 1 = − f * . For X ψ the generic fibre is E ψ ( s ) over C ( s ) , f is the fibrewise isogeny over σ , and σ is an involution: in the coordinate s with s 2 = ( t − b ) / ( t − N k / b ) one computes ( w ( t ) − b ) / ( w ( t ) − N k / b ) = b 2 / N k s 2 , so σ ( s ) = κ / s with κ 2 = b 2 / N k ; the argument of Lemma 9(ii) with c = 1 and σ in place of w gives Λ 2 = λ ( t ) λ ( w ( t ) ) σ ′ ( s ) σ ′ ( σ ( s ) ) = − N . The adjoint of f * for q X is f * = − f * , so the involution of E induced by q X is not the identity; by Zarhin’s theorem E is totally real or a CM field [126, Thm. 1.5.1], so it is a CM field, and K = Q + Q f * ⊆ E . The class of the graph of f is algebraic, its action on H 2 ( X ) is f * on T and an endomorphism of NS ( X ) Q , which is a sum of algebraic classes of NS ⊗ NS ; so f * , extended by zero on NS , is an algebraic class on X × X .
(ii) The elliptic fibration of X p has labelled singular fibres I 1 at 0, I N at ∞, II or III at e 1 , e 2 and I 0 * at p and N k / p . An isomorphism of elliptic fibrations induces a Möbius transformation preserving this configuration; the transformations preserving { 0 , ∞ } and { e 1 , e 2 } are t ↦ λ t and t ↦ λ / t with λ e 1 ∈ { e 1 , e 2 } , and since e 1 + e 2 ≠ 0 and e 1 2 ≠ e 2 2 they are the identity and w, which preserve { p , N k / p } . So X p and X p ′ have isomorphic fibrations only when p ′ ∈ { p , N k / p } . A K3 surface has finitely many elliptic fibrations up to automorphisms [112], so the map p ↦ [ X p ] to the moduli space has finite fibres and the family is not isotrivial; by local Torelli its period map has uncountable image. By (i) T ( X ) Q is a K-vector space, so its rank is even, and it is at most 6 by Lemma 8(i). It is also an E-vector space, so [ E : Q ] divides rk T ( X ) ; if E ≠ K then [ E : Q ] ∈ { 4 , 6 } , dim E T ( X ) = 1 and T ( X ) is of CM type. A Hodge structure of CM type on a sublattice of M ⊥ is determined by its endomorphism field, a subfield of the endomorphism algebra of the sublattice, and by an eigenline of that field, so the periods of such members form a countable subset of the period domain of M ⊥ ; the period map of the family is a non-constant holomorphic map from a connected curve, its fibres over a countable set are countable, so all but countably many members have E = K . There the Hodge group of T ( X ) is the unitary group of the K-hermitian form h with Tr K / Q h = q X by Zarhin’s theorem for CM fields [126, Thm. 2.3.1]. The same holds for B N by Lemma 8(iii).
(iii) By (i) E is a CM field containing K. If E = K then E = Q + Q f * is algebraic. Otherwise every element of E of norm one to its maximal real subfield is a Hodge isometry of T, extended by the identity on NS to a Hodge isometry of H 2 ( X , Q ) , algebraic by Buskin’s theorem [26]. The algebraic classes form a Q -subalgebra E ′ ⊆ E containing these elements; for x ∈ E with x ¯ ≠ x and t 1 ≠ t 2 in Q it contains ( t i + x ) / ( t i + x ¯ ) , hence their quotient 1 + ( t 2 − t 1 ) / ( t 1 + x ¯ ) , hence x ¯ ; so E ′ = E . In both cases End Hdg H 2 ( X ) = End ( NS ) ⊕ E is algebraic, and Hdg 2 ( X × X ) = End Hdg H 2 ( X ) ⊕ H 0 ⊗ H 4 ⊕ H 4 ⊗ H 0 since H 1 = H 3 = 0 , so the Hodge conjecture holds for X × X . For a K3 surface whose endomorphism field is a CM field the Hodge classes in the tensor algebra of T are generated by those of T ⊗ T , and the Hodge conjecture holds for every power [116, Thm. 2.2, Cor. 2.3], [104, Prop. 5.2, Thm. 5.4]. It passes to the Hilbert schemes X [ n ] , whose motives are direct summands of motives of powers of X [32], and (F3′) on a variety follows from the Hodge conjecture on it.    □
Figure 5. The family A N of Theorem 13. Over the Fricke coordinate t of X 0 ( N ) the involution w ( t ) = N k / t exchanges the cusps 0, ∞ and the elliptic points e 1 , e 2 ; the twist X p at a pair { p , N k / p } exchanged by w is carried to itself by the fibrewise isogeny f of degree N, and f * is the complex multiplication by − N on the transcendental lattice.
Figure 5. The family A N of Theorem 13. Over the Fricke coordinate t of X 0 ( N ) the involution w ( t ) = N k / t exchanges the cusps 0, ∞ and the elliptic points e 1 , e 2 ; the twist X p at a pair { p , N k / p } exchanged by w is carried to itself by the fibrewise isogeny f of degree N, and f * is the complex multiplication by − N on the transcendental lattice.
Preprints 244139 g005
The surface of [121, §7.6] with real multiplication by 7 is the twist at { 7 , − 7 } , the two fixed points of w, and the surface of [121, Prop. 7.7] is the twist at the fixed points ± 5 5 for N = 5 , which has Picard number 18 and complex multiplication as well: there u = t 2 − N k , c ( t ) c ( w ( t ) ) = − 1 and Λ 2 = N by Lemma 9(ii), so f * is the real multiplication by N . The families A N are the twists at the pairs exchanged by w, where c ( t ) c ( w ( t ) ) = 1 and Λ 2 = − N ; the two signs are the two values of the constant by which the twisting function changes under w once the square factor N k / t 2 is removed. The chains of Theorem 10 apply to every member, and a member whose field E contained 2 , 3 or 6 would have that element algebraic as well; by Theorem 13(ii) such a member is of CM type, and we know none.
Remark 1
(The Picard number). Theorem 13 does not say which of the three ranks the very general member has. The programs of the repository decide it by counting points. The surface X 1 , the twist at { 1 , N k } , has a Weierstrass model over Z with good reduction at ℓ = 11 for N = 7 and at ℓ = 7 for N = 5 , primes split in K; the program julia/frickepicard.jl counts its points over F ℓ i , i ≤ 3 , fibre by fibre, subtracts the trace of Frobenius on the trivial lattice, and completes the characteristic polynomial of Frobenius on the rank-six complement of the trivial lattice by the functional equation, whose sign is fixed by the nonzero middle coefficient. The polynomials are x 6 − 726 x 3 + 11 6 for N = 7 and x 6 + 6 x 5 + 56 x 4 + 392 x 3 + 2744 x 2 + 14406 x + 7 6 for N = 5 ; in the variable x / ℓ neither has a root of unity among its roots, each is irreducible over Q and splits over K into a cubic and its conjugate, as the action of f * requires. By the Tate conjecture for K3 surfaces over finite fields, used in the same way in [121, §7], ρ ( X 1 ) = 16 , so the very general member of A N and of B N has ρ = 16 , rk T = 6 , E = K with dim K T = 3 , and its Hodge group is the unitary group of a hermitian form of signature ( 1 , 2 ) on K 3 . At the primes inert in K, ℓ = 5 for N = 7 and ℓ = 11 for N = 5 , the odd traces vanish, as they must when Frobenius exchanges the two eigenspaces of f * . No statement of this section rests on this computation.

5.5. The Reach of Similarities

Every correspondence used so far acts on transcendental spaces as a Hodge similarity. The next proposition says that this is not an accident of the examples but a property of every correspondence whose components are small, and it measures what a correspondence must look like to act otherwise.
Proposition 9
(Structure of the transcendental action). Let X, X ′ be projective K3 surfaces, Γ ∈ CH 2 ( X × X ′ ) Q , and γ T : T ( X ) Q → T ( X ′ ) Q the map induced on transcendental spaces by composing the action of Γ with the orthogonal projections.
(i)
γ T is zero or an isomorphism of Hodge structures.
(ii)
If γ T ≠ 0 then a = γ T † γ T lies in the maximal totally real subfield of E = End Hdg T ( X ) Q and is totally positive, and γ T is a similarity exactly when a ∈ Q .
(iii)
Let Γ be irreducible, S a resolution of its normalisation with projections p 1 , p 2 of degrees d 1 , d 2 to X, X ′ . If p 1 * T ( X ) Q = p 2 * T ( X ′ ) Q inside H 2 ( S , Q ) , then γ T is a similarity of multiplier d 1 d 2 . This holds whenever p g ( S ) ≤ 1 .
(iv)
If Γ is irreducible and γ T is not a similarity, then p 1 * T ( X ) Q ∩ p 2 * T ( X ′ ) Q = 0 , so p g ( S ) ≥ 2 and the transcendental rank of S is at least twice that of X.
(v)
The Q -algebra generated by the actions on T ( X ) Q of all correspondences from X to itself whose irreducible components either fail to dominate a factor or have p g ≤ 1 is the multiquadratic subfield E 2 ⊆ E generated by the square roots of the products d 1 d 2 that occur. When [ E : Q ] is odd it is Q .
Proof. 
(i) T ( X ) Q is an irreducible Hodge structure [126], and the image of γ T is a sub-Hodge structure of H 2 ( X ′ ) ; if nonzero it contains the ( 2 , 0 ) -part, hence T ( X ′ ) Q , and γ T is injective by irreducibility. (ii) a is a Hodge endomorphism of T ( X ) , so a ∈ E , and it is self-adjoint, so it lies in the fixed field of the adjoint involution, the maximal totally real subfield. The form q ′ ( γ x , γ y ) = q ( x , a y ) is a polarisation of the Hodge structure T ( X ) up to sign on each real eigenspace of E, and comparison with q gives σ ( a ) > 0 at every real place. A similarity of multiplier μ has a = μ . (iii) Both p i * T are irreducible sub-Hodge structures of H 2 ( S ) with a ( 2 , 0 ) -part; if p g ( S ) ≤ 1 they share the unique ( 2 , 0 ) -line, so they meet and coincide. When they coincide, γ T = p 2 * p 1 * = d 2 ( p 2 * ) − 1 p 1 * on T, so p 2 * γ x = d 2 p 1 * x ; since q S ( p 1 * x , p 1 * y ) = d 1 q ( x , y ) and q S ( p 2 * x ′ , p 2 * y ′ ) = d 2 q ′ ( x ′ , y ′ ) , this gives d 2 q ′ ( γ x , γ y ) = q S ( p 2 * γ x , p 2 * γ y ) = d 2 2 q S ( p 1 * x , p 1 * y ) = d 1 d 2 2 q ( x , y ) , that is q ′ ( γ x , γ y ) = d 1 d 2 q ( x , y ) . (iv) If the two subspaces met they would coincide by irreducibility and (iii) would apply; so they are independent, their ( 2 , 0 ) -lines are distinct, and their sum has rank 2 rk T ( X ) inside the transcendental part of H 2 ( S ) . (v) A component that does not dominate a factor acts through NS or through H 0 and H 4 , hence by zero on T. The others act by similarities by (iii), a composite of similarities returning to X is an element Φ ∈ E with Φ † Φ ∈ Q , and for totally real E this is Φ 2 ∈ Q ; such elements generate E 2 . For a CM field E the elements with Φ † Φ ∈ Q are the Q × -multiples of the norm-one elements, which generate E, so the statement is about the totally real case, which is the one that matters for (F3′). An odd-degree field has E 2 = Q .    □
The correspondences of Lemmas 4 and 5, the Fourier–Mukai kernels between a K3 surface and its moduli spaces of sheaves, the rational Hodge isometries of hyper-Kähler varieties of K3[n] type [77], the Fricke isogenies of Theorem 13 and all their composites fall under (iii) or (v). A correspondence that realises an element of a cubic field must therefore have a component whose resolution has p g ≥ 2 and transcendental rank at least twice that of the surface. Such correspondences exist, and the only ones we know come from finite groups.
Proposition 10
(Hecke correspondences). Let Z be a smooth projective surface with an action of a finite group G, H ≤ G a subgroup, and X a smooth projective surface with p g ( X ) = 1 birational to Z / H ; write π : Z ⤏ X .
(i)
For g ∈ G the correspondence Γ g = ( π × π ) * [ graph of g ] acts on H 2 ( X , Q ) as the Hecke operator of the double coset H g H , and these operators span a Q -algebra isomorphic to a quotient of End G Q [ G / H ] acting by Hodge endomorphisms.
(ii)
Exactly one simple factor of this algebra acts nontrivially on H 2 , 0 ( X ) ; it is a field, it preserves T ( X ) Q , and it embeds into E by algebraic correspondences. If H 2 , 0 ( Z ) G = 0 this factor is not the one on which G acts trivially.
(iii)
Every field obtained in this way is abelian over Q . In particular a totally real cubic field F whose Galois closure has group S 3 is realised by no correspondence built from a finite group acting on a cover of X.
Proof. 
(i) Let Z ˜ → Z be an equivariant resolution and π : Z ˜ → X the quotient by H followed by the resolution; then π * π * = | H | and π * h * = π * for h ∈ H , so Γ g acts as π * g * π * = ( π * ) − 1 ∘ ∑ h ∈ H g H / H h * ∘ π * on H 2 ( X , Q ) = H 2 ( Z ˜ , Q ) H , the Hecke operator of H g H . The Hecke operators of ( G , H ) span End G Q [ G / H ] , a semisimple algebra, and the action on H 2 ( Z ˜ , Q ) H factors through it. (ii) The idempotents of the simple factors are rational combinations of Hecke operators, so they are Hodge endomorphisms of H 2 ( X ) ; they sum to 1 and act on the line H 2 , 0 ( X ) by 0 or 1, so exactly one of them, e V , acts as 1. A simple factor with nonzero image in the commutative ring End Hdg H 2 , 0 ( X ) = C through a ring homomorphism to a field is commutative, hence a field H V . For f ∈ H V , f ( T ( X ) Q ) is a sub-Hodge structure of the transcendental part of H 2 ( X ) ; if f ω ≠ 0 it contains ω , hence T ( X ) Q , and if f ω = 0 it has no ( 2 , 0 ) -part and lies in NS ∩ T = 0 . So H V → End Hdg T ( X ) Q = E is a homomorphism from a field, injective. The factor on which G acts trivially fixes H 2 , 0 ( Z ˜ ) G = H 2 , 0 ( Z / G ) , which is zero by hypothesis, so it is not H V . (iii) The centres of the simple factors of End G Q [ G / H ] are the character fields Q ( χ i ) of the irreducible constituents of Q [ G / H ] , generated by sums of roots of unity, so each lies in Q ( ζ | G | ) and is abelian; a commutative simple factor is its centre. A non-abelian cubic field is not a subfield of an abelian field.    □
For the dihedral group D n with n odd, H = 〈 τ 〉 of order two and Z the elliptic surfaces E ˜ a of [121, §4.7], the algebra End D n Q [ D n / H ] is Q × ∏ d ∣ n , d > 1 Q ( ζ d ) + , and the factor acting on ω is Q ( ζ n ) + ; this is the content of [121, §4.8], where the cycle Γ 1 + Γ − 1 induces ζ n + ζ n − 1 . The components of these correspondences are birational to E ˜ a , which has p g ≥ 2 , as Proposition 9(iv) requires.
Corollary 2
(Squares of the dihedral families). Let X be an elliptic K3 surface of one of the families of [121, Thm. 1.1], with real multiplication by F = Q ( ζ n ) + for n ∈ { 5 , 7 , 9 , 11 } , and suppose E = F or E is a CM field; the very general member has E = F . Then the Hodge conjecture holds for X × X , and (F3′) holds for X × X and X [ 2 ] . This covers the cubic fields Q ( ζ 7 ) + and Q ( ζ 9 ) + on families of dimensions 3 and 2.
Proof. 
If E = F , E is algebraic by Proposition 10(ii) and [121, §4.8]; if E is CM, by Buskin’s theorem as in Theorem 13(iii). The rest is as there.    □
Figure 6. The proof of Theorem 14. The surface is the two-dimensional period domain of the twelve-dimensional space T 12 with real multiplication by F = Q ( ζ 7 ) + or Q ( ζ 9 ) + ; the shaded patch is the image of the two-dimensional dihedral family, on which the multiplication is algebraic. The curves are the loci on which an F-line N becomes algebraic, and the surface of Mumford type is a point on one of them. The rational isometries x ↦ a x g of the quaternion algebra are dense in the real group, so one of them moves the point into the patch; Buskin’s theorem turns that isometry into a cycle.
Figure 6. The proof of Theorem 14. The surface is the two-dimensional period domain of the twelve-dimensional space T 12 with real multiplication by F = Q ( ζ 7 ) + or Q ( ζ 9 ) + ; the shaded patch is the image of the two-dimensional dihedral family, on which the multiplication is algebraic. The curves are the loci on which an F-line N becomes algebraic, and the surface of Mumford type is a point on one of them. The rational isometries x ↦ a x g of the quaternion algebra are dense in the real group, so one of them moves the point into the patch; Buskin’s theorem turns that isometry into a cycle.
Preprints 244139 g006

5.6. The Cyclic Cubic Fields Q ( ζ 7 ) + and Q ( ζ 9 ) +

The dihedral families reach the cyclic cubic fields only on their own members. For each of the two cubic fields they contain, F = Q ( ζ n ) + with n ∈ { 7 , 9 } , there is a two-dimensional family on a transcendental lattice of rank twelve: for n = 9 the family of [121, §5.6], and for n = 7 the Dickson deformation of the one-parameter family of Artebani, Sarti and Taki with Picard lattice U ⊕ E 8  [7, Ex. 6.1], which [121] do not use, their family for n = 7 being the three-dimensional one of Picard number four. Two is the largest dimension a family with this real multiplication on a transcendental space of rank twelve can have, and this maximality lets the cycle travel: every K3 surface of Mumford type for F and for the quaternion algebra ramified only at infinity is a Hodge isometric image of a member of the family, and Buskin’s theorem carries the cycle along the isometry. The result settles the frontier case of Section 5.10 for these two fields and this algebra.
Write θ = ζ n + ζ n − 1 , so that F = Q ( θ ) , K = Q ( ζ n ) = F ( − δ ) with − δ = ( ζ n − ζ n − 1 ) 2 = θ 2 − 4 , and δ = 4 − θ 2 is totally positive. For n = 7 one has θ 3 + θ 2 − 2 θ − 1 = 0 and the three real embeddings send θ to − 1.802 , − 0.445 and 1.247 ; for n = 9 one has θ 3 − 3 θ + 1 = 0 and the embeddings send θ to − 1.879 , 0.347 and 1.532 . Let p n be the prime dividing n; it is totally ramified in K, so it lies under a single place of F, and K / Q is unramified outside p n . The prime 2 has order three in ( Z / n ) × / { ± 1 } = Gal ( F / Q ) , so it is inert in F. Let D ∞ be the quaternion algebra over F ramified exactly at two real places and at no finite place. Put ε = θ for n = 7 and ε = − θ for n = 9 ; then ε is a unit positive at exactly one real place, and D ∞ = ( − 1 , ε ) F : the symbol is ramified at the two real places where ε < 0 , split at every odd finite place because both entries are units there, and split at the one place above 2 by the parity of the number of ramified places. Its corestriction to Q is unramified everywhere and split at infinity, hence is M 8 ( Q ) , so D ∞ is one of Mumford’s algebras and the fourfolds of [87] exist for it. For μ ∈ F totally positive let q μ be the quadratic form Tr F / Q ( μ trd ( y 2 ) ) on the nine-dimensional Q -space D ∞ 0 of elements of reduced trace zero, of signature ( 2 , 7 ) ; the surfaces S λ of Theorem A23 for the fourfolds of ( F , D ∞ ) have T ( S λ ) Q ≅ ( D ∞ 0 , q λ ) .
Theorem 14
(Real multiplication by Q ( ζ 7 ) + and Q ( ζ 9 ) + on the surfaces of Mumford type). Let n ∈ { 7 , 9 } , F = Q ( ζ n ) + and D ∞ as above. Let S be a projective K3 surface such that T ( S ) Q with its intersection form is isometric to ( D ∞ 0 , q μ ) for some totally positive μ ∈ F , by an isometry under which the multiplication by F on D ∞ 0 acts on T ( S ) Q by Hodge endomorphisms. Then E = F , the real multiplication by F is induced by algebraic cycles on S × S , and the Hodge conjecture and (F3′) hold for S × S and S [ 2 ] . This applies to every surface S λ of a Mumford fourfold of ( F , D ∞ ) , at every point of its Shimura curve.
Proof. 
Step 1, the family for n = 9 . Let X → B be the two-dimensional family of [121, §5.6]: elliptic K3 surfaces with very general Picard lattice U ⊕ A 2 4 ,6 deformations of the surfaces y 2 = x 3 + b x + c 1 t 9 + c 0 with a non-symplectic automorphism σ of order nine, obtained by replacing t 9 by the Dickson polynomial p 9 , a ( t ) . The transcendental lattice T 0 of a very general member has rank twelve and discriminant group ( Z / 3 ) 4 , is a Z [ ζ 9 ] -module of rank two through σ * , and carries the flat action of F through σ * + σ * − 1 , which on every member of the family is induced by the cycle Γ 1 + Γ − 1 of [121, §4.8] on a dihedral cover, as in Proposition 10. The period map of B has two-dimensional image [121, §5.6].
Step 1, the family for n = 7 . The surfaces X b : y 2 = x 3 + b x + c 1 t 7 + c 0 are the K3 surfaces of [7] with a = 0 there: σ ( x , y , t ) = ( x , y , ζ 7 t ) is a non-symplectic automorphism of order seven, the fibration has a fibre of type II * at t = ∞ and fourteen fibres of type I 1 , and for very general b the Picard lattice is the invariant lattice U ⊕ E 8 of σ , with T 0 = U 2 ⊕ E 8 ( − 1 )  [7, Table 3]. Replacing t 7 by the Dickson polynomial p 7 , a ( t ) = t 7 − 7 a t 5 + 14 a 2 t 3 − 7 a 3 t gives the surfaces7
X a : y 2 = x 3 + b x + c 1 p 7 , a ( t ) + c 0 ,
where p 7 , a is the Dickson polynomial with p 7 , a ( v + a v − 1 ) = v 7 + a 7 v − 7 ;8 these are the deformations (4.4) of [121, Prop. 4.6] with n = 7 , α = b and β ( u ) = c 1 u + c 0 ; by that proposition each X a is a K3 surface on whose transcendental space F acts by Hodge endomorphisms, induced by the cycle Γ 1 + Γ − 1 on the D 7 -cover as before. The polynomial c 1 p 7 , a + c 0 has degree seven, so at t = ∞ the coefficients of the Weierstrass equation vanish to orders 8 and 5 and the fibre of type II * survives; hence Pic ( X a ) contains the unimodular lattice U ⊕ E 8 of rank ten, and since the Picard number of a very general member of a family is at most that of a very general member of any subfamily, here the locus a = 0 , it equals U ⊕ E 8 ( − 1 ) , with T 0 = U 2 ⊕ E 8 ( − 1 ) , unimodular of rank twelve. The parameters ( b , c 1 , c 0 , a ) modulo the scalings ( x , y , t ) ↦ ( λ 2 x , λ 3 y , μ t ) , which send a to a / μ 2 , leave two, and the image of the period map is two-dimensional: otherwise it is the closure of the one-dimensional image of the locus a = 0 , and a very general X a is isomorphic by the Torelli theorem to some X b ′ , so carries a non-symplectic automorphism σ ′ of order seven. Such a σ ′ has no invariant vector in T 0 ⊗ Q , so its invariant lattice is Pic ( X a ) ; it fixes the class of a fibre and the class of the zero section O, hence preserves the fibration and O, the only effective curve in its class, and it acts on the base with order seven, since an automorphism over the identity of the base that fixes O is a group automorphism of the fibres, of order dividing six. It fixes t = ∞ , the only reducible fibre, so it acts by t ↦ ζ t + β with ζ a primitive seventh root of unity, and since the Weierstrass equation of a fibration with a chosen section is unique up to ( x , y ) ↦ ( λ 2 x , λ 3 y ) , the polynomial c = c 1 p 7 , a + c 0 satisfies c ( ζ t + β ) = λ 6 c ( t ) . The coefficients of t 7 , t 6 and t 5 give λ 6 = 1 , β = 0 and 7 a c 1 ( ζ 5 − 1 ) = 0 , so a = 0 . Thus the direction a is not isotrivial and the image is two-dimensional. If E ⊋ F then E has degree six or twelve and dim E T 0 ⊗ Q ≤ 2 , which allows at most one modulus by the count of [121, §5.4]; so E = F on the very general member.
In both cases the family is maximal: dim F T 12 = 4 for T 12 = T 0 ⊗ Q , the period domain of Hodge structures of K3 type on T 12 with F acting by Hodge endomorphisms has dimension 4 − 2 = 2 , and the period map has two-dimensional image in its quotient by the arithmetic group, so by the open mapping theorem the image contains a nonempty open set U of one of the two components of that domain.
Step 2, the quadratic structure. Since the intersection form q of T 12 is σ -invariant there is a unique hermitian form h on the two-dimensional K-space T 12 with q = Tr K / Q h , and a unique F-bilinear form Φ = Tr K / F h with q = Tr F / Q Φ . Diagonalise h = 〈 h 1 , h 2 〉 with h i ∈ F × . For x = u + v − δ one has Tr K / F ( h 1 x x ¯ ) = 2 h 1 ( u 2 + δ v 2 ) , so
Φ ≅ 2 h 1 〈 1 , δ , h , h δ 〉 , h = h 2 / h 1 ,
and 〈 1 , δ , h , h δ 〉 is the reduced norm form of the quaternion algebra Q = ( − δ , − h 1 h 2 ) F ; thus ( T 12 , Φ ) ≅ ( Q , 2 h 1 nrd ) as F-quadratic spaces. The signature of q is ( 2 , 10 ) and splits over the three real places of F as ( 2 , 2 ) at the place σ 1 whose eigenspace carries H 2 , 0 and ( 0 , 4 ) at the two others. From (5), signature ( 2 , 2 ) at σ 1 forces h 1 h 2 < 0 there, and ( 0 , 4 ) at σ 2 , σ 3 forces h 1 , h 2 < 0 there; so Q is split at σ 1 and ramified at σ 2 and σ 3 , and after exchanging h 1 and h 2 we may take h 1 > 0 at σ 1 and h 1 < 0 at the other two places.
Step 3, Q = D ∞ . The lattice T 0 is unimodular at every prime p ≠ p n , and K / Q is ramified only at p n , so for a place v of F not above p n the ring O K ⊗ Z p is a product of unramified extensions whose trace forms are perfect, and the hermitian O K , v -lattice T 0 ⊗ Z p is unimodular for h; its determinant h 1 h 2 is therefore a unit of F v times a norm from K v . Units are norms from an unramified extension, so ( − δ , − h 1 h 2 ) v = 1 at every finite v ∤ p n , including the place above 2. The ramification set of Q has even cardinality, contains σ 2 , σ 3 and is otherwise contained in the single place above p n ; so that place is unramified as well, and Q ≅ D ∞ . As a check independent of this argument, the rational invariants of T 0 ⊗ Q , computed from the Picard lattice by Witt cancellation in the K3 lattice, and those of Tr F / Q ( ε nrd D ∞ ) agree for both n: signature ( 2 , 10 ) , discriminant 1, Hasse invariant − 1 at 2 and + 1 at every odd prime, while the trace forms of the algebras over F with the same real ramification and ramification at finite places differ from them at 7 or 43 for n = 7 and at 3 or 19 for n = 9 .
Step 4, the surfaces of Mumford type inside T 12 . For n 0 ∈ Q × the F-line N = F n 0 has Φ ( n 0 , n 0 ) = 2 h 1 nrd ( n 0 ) , and its orthogonal complement is n 0 Q 0 , with Φ ( n 0 y , n 0 y ′ ) = − h 1 nrd ( n 0 ) trd ( y y ′ ) for y , y ′ ∈ Q 0 ; so ( N ⊥ , Φ ) ≅ ( D ∞ 0 , μ trd ) with μ = − h 1 nrd ( n 0 ) , and N is negative definite exactly when μ is totally positive. By the Hasse–Schilling norm theorem [103, Thm. 33.15] the reduced norms of Q × are the elements of F × positive at σ 2 and σ 3 , and − h 1 is negative at σ 1 and positive at the other two places, so every totally positive μ occurs. Given S as in the statement, choose n 0 with μ = − h 1 nrd ( n 0 ) and an F-linear isometry T ( S ) Q ≅ n 0 ⊥ ; the period ω 0 of S becomes a point of the period domain of T 12 at which F n 0 is algebraic and F acts by Hodge endomorphisms, because ω 0 lies in the σ 1 -eigenspace. Replacing S by its complex conjugate, whose period is ω ¯ 0 and for which the conclusion is equivalent, we may take ω 0 on the component containing U.
Step 5, moving the point. The group SO ( Φ ) ( R ) at the place σ 1 is ( SL 2 ( R ) × SL 2 ( R ) ) / { ± 1 } acting on Q ⊗ F , σ 1 R ≅ M 2 ( R ) by x ↦ a x g , and it acts transitively on each component of the two-dimensional period domain. The group Q 1 × Q 1 of pairs of elements of reduced norm one is simply connected, so by weak approximation [101, Thm. 7.8] its rational points are dense in its real points; hence there are a , g ∈ Q 1 ( Q ) with x ↦ a x g carrying ω 0 into U. This map is an F-linear isometry of ( T 12 , Φ ) , it carries n 0 ⊥ onto ( a n 0 g ) ⊥ with nrd ( a n 0 g ) = nrd ( n 0 ) , and it carries the Hodge structure of T ( S ) to the Hodge structure on ( a n 0 g ) ⊥ determined by the image point, which is the period of a member X b of the family. The rational classes of type ( 1 , 1 ) in T 12 at that period are exactly F · a n 0 g , since a rational class orthogonal to the period decomposes into a component in the line and a component in ( a n 0 g ) ⊥ that is a rational class of type ( 1 , 1 ) in a copy of T ( S ) , which has none. So T ( X b ) Q = ( a n 0 g ) ⊥ with the restricted action of F, and x ↦ a x g is a Hodge isometry φ : T ( S ) Q → T ( X b ) Q that commutes with F.
Step 6, the cycle. H 2 ( S , Q ) and H 2 ( X b , Q ) are both isometric to Λ Q , so by Witt cancellation NS ( S ) Q ≅ NS ( X b ) Q , and φ extends to a rational Hodge isometry H 2 ( S , Q ) → H 2 ( X b , Q ) , which is algebraic by Buskin’s theorem [26]. The action of e ∈ F on T ( X b ) is induced by the cycle of Step 1, so m e = φ − 1 ∘ m e ∘ φ on T ( S ) is induced by a cycle on S × S . Since T ( S ) has rank nine and F ⊆ E , E is F or a totally real field of degree nine with dim E T = 1 , which [118] excludes; so E = F is algebraic, and the Hodge conjecture for S × S and the rest follow as in Theorem 13(iii).    □
The next proposition says exactly which pairs ( F , D ) the mechanism reaches and where it stops.
Proposition 11
(What the dihedral mechanism reaches). Let n ∈ { 7 , 9 } and F = Q ( ζ n ) + .
(i)
Let X be a K3 surface with a purely non-symplectic automorphism of order n ∈ { 7 , 9 } whose transcendental lattice has rank twelve and is unimodular at every prime other than p n . Then ( T ( X ) Q , Φ ) is isometric to ( D ∞ , c nrd ) for some c ∈ F × , and every surface of Mumford type obtained from X or from a rational isometric image of X as in Step 4 has the algebra D ∞ . The Mumford algebras over F ramified at two finite places are reached by no family of rank twelve with such an automorphism.
(ii)
Let F = Q ( ζ 7 ) + and let T 18 = U 2 ⊕ A 6 ⊕ E 8 ( − 1 ) be the transcendental lattice of the three-dimensional family of [121, §5.4], with its F-bilinear form Φ 6 of dimension six. For every Mumford algebra D over F and every totally positive λ the space ( D 0 , λ trd ) embeds F-isometrically into ( T 18 ⊗ Q , Φ 6 ) , so the four-dimensional period domain of this family contains the Shimura curve of the surfaces S λ of every pair ( F , D ) . The image of the family is a threefold in it, and Step 5 would need a rational isometry carrying the period of S λ into that threefold.
(iii)
A finite group acting on the base P 1 of an elliptic surface is cyclic, dihedral, tetrahedral, octahedral or icosahedral, and the centres of the simple factors of its rational group algebra are the fields Q ( ζ m ) , Q ( ζ m ) + , Q ( − 3 ) , Q and Q ( 5 ) . The only cubic fields among them are Q ( ζ 7 ) + and Q ( ζ 9 ) + , so the base change of elliptic surfaces along covers of P 1 reaches no cyclic cubic field of conductor at least 13.
Proof. 
(i) Steps 2 and 3 of the theorem use only the order of the automorphism, the rank, the signature and the unimodularity away from p n , so they apply to X and give Φ ≅ 2 h 1 nrd D ∞ . A rational isometry preserves ( T , Φ ) up to the F-similarity class, which determines the algebra, and Step 4 exhibits the orthogonal complement of every negative definite F-line as ( D ∞ 0 , μ trd ) ; an algebra D ≠ D ∞ has ( D 0 , λ trd ) ≅ ( D ∞ 0 , μ trd ) for all λ , μ , since the even Clifford algebra of the ternary form λ trd is D.
(ii) T 18 ⊗ Q is a three-dimensional hermitian K-space and Φ 6 = 2 〈 h 1 , h 1 δ , h 2 , h 2 δ , h 3 , h 3 δ 〉 has signature ( 2 , 4 ) at σ 1 and ( 0 , 6 ) at σ 2 , σ 3 , by the same count as in Step 2. The form q 3 = λ trd on D 0 has signature ( 2 , 1 ) at σ 1 , where D is split, and ( 0 , 3 ) at the two ramified places. The Witt class of Φ 6 − q 3 is represented at every real place by a negative definite ternary form and at every finite place by a ternary form, because a class of odd dimension has an anisotropic representative of dimension one or three; these local ternary forms have the global discriminant d ( Φ 6 ) d ( q 3 ) , Hasse invariants + 1 at almost every place and satisfying the product formula, since they are products of the invariants of Φ 6 and q 3 and of Hilbert symbols of their discriminants; so by [95, Thm. 72:1] there is a global ternary form q 3 ′ with these localisations, and Φ 6 ≅ q 3 ⊥ q 3 ′ by the Hasse–Minkowski theorem. The image N of q 3 ′ is a negative definite F-subspace of dimension three with N ⊥ ≅ ( D 0 , λ trd ) , and the Hodge structures of K3 type on N ⊥ with F acting form the Shimura curve of the S λ of ( F , D ) . The family has three moduli [121, §5.4] in a domain of dimension 6 − 2 = 4 .
(iii) The list of finite subgroups of PGL 2 ( C ) is Klein’s. The simple factors of Q [ G ] have centres the fields Q ( χ ) generated by the values of the irreducible characters: Q ( ζ m ) for the cyclic group of order m, Q ( ζ m ) + for the dihedral group of order 2 m (its two-dimensional characters take the values ζ m k + ζ m − k ), Q and Q ( − 3 ) for A 4 , Q for S 4 , and Q and Q ( 5 ) for A 5 . By Proposition 10(iii) the correspondences cut out of the action act through these fields; Q ( ζ m ) is never cubic, and Q ( ζ m ) + is cubic exactly for m ∈ { 7 , 9 , 14 , 18 } , where it is Q ( ζ 7 ) + or Q ( ζ 9 ) + .    □
Part (ii) leaves open whether the threefold meets the Shimura curves of the other algebras. The next two results turn this into a question about subfamilies of the rank-eighteen family, and show where such a subfamily would have to be special.
Lemm 10
(Unimodular places split the algebra). Let F be a totally real cubic field with O F = Z [ θ ] , and let T be a lattice of rank nine whose form is b = Tr F / Q ∘ φ for an F-bilinear form φ on T Q , with θ self-adjoint. Let D = C 0 ( T Q , φ ) . If p is odd, θ preserves T p = T ⊗ Z p and T p is unimodular, then D is split at every place of F above p.
Proof. 
The idempotents e v of O F ⊗ Z p = ∏ v ∣ p O v are polynomials in θ , hence self-adjoint, so T p = ⨁ v L v is an orthogonal sum, with L v free of rank three over O v and b = Tr F v / Q p ∘ φ v on it. Since φ v ( x , L v ) is an O v -module, x lies in the b-dual of L v exactly when φ v ( x , L v ) ⊆ d v − 1 , d v the local different; so unimodularity gives L v # φ v = d v L v , and ψ = d v φ v , with d v a generator of d v , is a unimodular ternary form over O v . Its reduction is a nondegenerate ternary form over a finite field of odd characteristic, which has a nonzero zero by the Chevalley–Warning theorem; the zero is nonsingular because the gradient of the form there is twice a nonzero linear form, so by Hensel’s lemma ψ is isotropic, ψ ≅ H ⊥ 〈 c 〉 . Then C 0 ( ψ ) ≅ M 2 ( F v ) , and since the even Clifford algebra of a ternary form does not change under scaling, D ⊗ F v = C 0 ( φ v ) ≅ C 0 ( ψ ) is split.    □
Proposition 12
(Transport within a family). Let X → B be an irreducible family of projective K3 surfaces with a totally real cubic field F acting on the transcendental spaces by Hodge endomorphisms induced by a flat family of algebraic cycles. Let ( V , φ ) be an F-quadratic space of dimension three with Tr F / Q φ of signature ( 2 , 7 ) , let Ω be a component of { ω ∈ P ( V ⊗ F , σ 1 C ) : φ ( ω , ω ) = 0 , φ ( ω , ω ¯ ) > 0 } , let Ω sp ⊂ Ω be the set of periods for which V contains a rational class of type ( 1 , 1 ) , and let R ⊆ Ω be the set of ω for which ( V , ω ) is F-linearly Hodge isometric to T ( X b ) Q for some b ∈ B .
(i)
Either R is countable or R ⊇ Ω ∖ Ω sp . In the second case, every projective K3 surface S with T ( S ) Q F-linearly Hodge isometric to ( V , ω ) for some ω ∈ Ω ∖ Ω sp has its multiplication by F induced by algebraic cycles on S × S .
(ii)
The second case holds for some ( V , φ ) with even Clifford algebra D if and only if X contains a non-isotrivial analytic subfamily of positive dimension whose very general member has transcendental rank nine and algebra D.
(iii)
Let F = Q ( ζ n ) + with n ∈ { 7 , 9 } , and let X be a member of a family of [121, Thm. 1.1] with real multiplication by F, transcendental rank nine and algebra D ≠ D ∞ . Then T ( X ) ⊗ Z p is not unimodular for some odd prime p, and D ramifies at a place above p. In particular, for the algebra D p of Figure 7, T ( X ) ⊗ Z p is not unimodular.
Proof. 
(i) Ω sp is the union over the countably many nonzero rational x ∈ V of Ω ∩ x ⊥ , each a proper analytic subset, so it is a countable union of nowhere dense sets. Cover B by countably many simply connected open sets U with marked period maps p U to Λ C , and for each of the countably many F-linear isometric embeddings j : V → Λ Q compatible with the action of F on the transcendental spaces let A U , j be the analytic set of b ∈ U with p U ( b ) ∈ j ( V ) C and j − 1 p U ( b ) ∈ Ω . If T ( X b ) has rank nine then T ( X b ) Q = j ( V ) for such a j, so R ⊆ ⋃ ϕ U , j ( A U , j ) ⊆ R ∪ Ω sp with ϕ U , j = j − 1 ∘ p U . On each of the countably many irreducible components of the A U , j the map ϕ U , j is constant or, by the open mapping theorem, has image containing a nonempty open set. In the first case for all components R is countable. In the second, R contains a nonempty open set U 0 ∖ Ω sp . The group D 1 of elements of reduced norm one acts on V ≅ μ ( D 0 , nrd ) by conjugation, through F-linear isometries, so R and Ω sp are stable under D 1 ( F ) . It is simply connected, so by weak approximation [101, Thm. 7.8] D 1 ( F ) is dense in D 1 ( F ⊗ R ) and its image is dense in SL 2 ( R ) = D 1 ( F σ 1 ) , which acts transitively on Ω . For ω ′ ∈ Ω ∖ Ω sp choose g ∈ D 1 ( F ) with g − 1 ω ′ ∈ U 0 ∖ Ω sp ; then ω ′ ∈ g R = R . For the last statement, compose the given isometry with one onto T ( X b ) Q to get an F-linear Hodge isometry ψ : T ( S ) Q → T ( X b ) Q ; by Witt cancellation it extends to a Hodge isometry ϕ of H 2 , which is algebraic by Buskin’s theorem [26], and t Z ϕ ∘ Γ e ∘ Z ϕ induces ψ − 1 e ψ = e on T ( S ) for each e ∈ F .
(ii) If the second case holds, some component of some A U , j has nonconstant ϕ U , j ; it is a positive-dimensional analytic subfamily on which T ( X b ) Q ⊆ j ( V ) , with equality off the preimage of Ω sp , which is a countable union of proper analytic subsets, and it is non-isotrivial because its periods vary. Conversely, on such a subfamily T ( X b ) Q runs through the countably many F-isometry classes of three-dimensional spaces with algebra D, so one class ( V , φ ) is attained on an uncountable set, on which the periods are not constant by the Torelli theorem; so some ϕ U , j for this V is nonconstant and R is uncountable.
(iii) In the construction of [121, §4.8] the multiplication by θ is 1 2 ( π × π ) * ( Γ σ + Γ σ − 1 ) , where π is a double cover and σ an automorphism of order n upstairs: the integral cycle ( π × π ) * ( Γ σ + Γ σ − 1 ) acts as π * ( σ * + σ − 1 * ) π * , and since π * π * = 2 and σ * + σ − 1 * acts as θ on π * T ( X ) Q , it acts on T ( X ) as 2 θ . Hence θ preserves T ( X ) ⊗ Z p for every odd p, and O F = Z [ θ ] for both fields. If T ( X ) ⊗ Z p were unimodular at every odd p, Lemma 10 would make D split at every finite place not above 2. As in Step 2 of the proof of Theorem 14, the signature makes D ramified exactly at the real places σ 2 , σ 3 ; the prime 2 is inert in F, since 2 has order 3 modulo 7 and generates ( Z / 9 ) × ; and the number of ramified places is even. So D would be split above 2 as well, and D = D ∞ . Hence some odd p has T ( X ) ⊗ Z p non-unimodular, and Lemma 10 shows that every place where D ramifies away from 2 and ∞ lies over such a p; for D p these places lie over p.    □
So the rank-eighteen family reaches the surfaces S λ of ( F , D ) , D ≠ D ∞ , if and only if it has a non-isotrivial subfamily of transcendental rank nine with that algebra, and along such a subfamily the transcendental lattice is non-unimodular above the ramified primes of D. We do not know such a subfamily; this is where part (ii) stops.
For the fourfolds X of ( F , D ∞ ) the theorem leaves (F2) where it was: by the argument of Step 2 of the proof of Theorem 38, with T Z replaced by T ( S λ ) , on which the real multiplication is now algebraic, their exceptional classes are algebraic as soon as one algebraic cycle on S λ × X r acts nontrivially on T ( S λ ) , and the real multiplication supplies no such cycle.

5.7. Powers of K3 Surfaces and Kummer Type

The theorems of this section decide S × S and S [ 2 ] . On the higher powers the Hodge classes are described by Varesco [116, Thm. 2.16]. Let E be totally real, m = dim E T , and for each real embedding σ of E let V σ = T ⊗ E , σ C , so that T C = ⨁ σ V σ . The line ⋀ m V σ ⊂ T C ⊗ m is spanned by a class det V σ of type ( m , m ) , and the span of the det V σ over all σ is defined over Q : it is ⋀ E m T ⊂ ⋀ m T ⊂ T ⊗ m , a space of Hodge classes of dimension [ E : Q ] in H 2 m ( S m ) , the exceptional classes. Varesco’s theorem is that every Hodge class in the tensor algebra of T is obtained from the Hodge classes of T ⊗ T , which are the elements of E, and from the exceptional classes by tensor products, permutations of the factors, linear combinations and contractions with q S . Each of these operations carries algebraic classes on powers of S to algebraic classes: a tensor product is an exterior product, a permutation is the pull-back along an automorphism of S k , and the contraction of two factors is the push-forward along the projection S k → S k − 2 after cupping with the pull-back of the class of the diagonal of S × S , since q S ( a , b ) = ∫ S × S ( a ⊗ b ) ∪ [ Δ S ] .
Proposition 13
(Hodge classes on the powers). Let S be a projective K3 surface with E totally real and m = dim E T , so that m ≥ 3 as in the proof of Theorem 11(ii).
(i)
The Hodge conjecture holds for every power S n if and only if the elements of E are algebraic on S × S and the exceptional classes ⋀ E m T ⊂ H 2 m ( S m ) are algebraic. When E = Q , m = r = rk T and the exceptional class is the determinant det T , which spans ⋀ r T ⊂ H 2 r ( S r ) .
(ii)
For a surface of Theorems 11, 14 or Corollary 2 with E totally real, E is algebraic, and the Hodge conjecture for all powers of S is equivalent to the algebraicity of ⋀ E m T , m = rk T / [ E : Q ] : m = 3 for the surfaces S λ of Theorem 14, with rk T = 9 ; m = 4 for the very general members of the two-dimensional family of Theorem 14, with rk T = 12 .
Proof. 
(i) If the Hodge conjecture holds on all powers, both kinds of classes are algebraic, being Hodge classes on S × S and on S m . Conversely, H 2 ( S ) = NS ( S ) Q ⊕ T , and H 0 ( S ) , NS ( S ) Q and H 4 ( S ) are spanned by algebraic classes and are Tate twists of trivial Hodge structures. So a Hodge class of H * ( S n ) = ⨂ n H 0 ⊕ NS Q ⊕ T ⊕ H 4 is a sum of exterior products of algebraic classes with Hodge classes of the tensor powers T ⊗ k , k ≤ n , pulled back along the projections of S n onto k of its factors. By Varesco’s theorem and the remark preceding the proposition, the Hodge classes of T ⊗ k are algebraic as soon as E and the exceptional classes are. When E = Q there is one embedding, V σ = T C and ⋀ E m T = ⋀ r T .
(ii) The theorems cited prove that E is algebraic; m = rk T / [ E : Q ] since T is an E-vector space. The ranks are those of the nine-dimensional space D ∞ 0 and of the lattice T 0 in Step 1 of the proof of Theorem 14.    □
Proposition 14
(K3 surfaces of Kummer type over Q ). Let S be a projective K3 surface, r = rk T , let 〈 a 1 , ⋯ , a r 〉 be a diagonalisation of q S on T, d ( T ) = ∏ i a i ∈ Q × / Q × 2 its discriminant and ε p ( T ) = ∏ i < j ( a i , a j ) p its Hasse invariant at the prime p. The following are equivalent.
(a)
T embeds isometrically in U Q 3 , the quadratic space H 2 ( A , Q ) of an abelian surface.
(b)
There is a dominant rational map from S to a Kummer surface.
(c)
r ≤ 3 ; or r = 4 and ε p ( T ) = ( − 1 , − 1 ) p at every prime p at which d ( T ) is a square in Q p ; or r = 5 and ε p ( T ) = ( − 1 , − 1 ) p at every prime p.
Here ( − 1 , − 1 ) p = 1 for p odd and ( − 1 , − 1 ) 2 = − 1 , so (c) asks for ε p ( T ) = 1 at the odd primes concerned and ε 2 ( T ) = − 1 . If (a)–(c) hold, then S lies in the class A of Proposition 2, (F3′) holds for every power S n , and the Hodge conjecture holds for every S n and every Hilbert scheme S [ n ] .
The equivalence of (a) and (b) is Ma’s theorem [74, Prop. 3.2], where (a) is also shown equivalent to an embedding of lattices T ( S ) ↪ U 3 , and for a Kummer surface itself the Hodge conjecture on all powers follows from that on the powers of the abelian surface [116, §0.2]. What is added here is the form (c) of the condition, which reads it off the invariants of T, and the passage from Ma’s rational map to the class A .
Proof. 
(a)⇔(c). The space V = U Q 3 ≅ 〈 1 , − 1 〉 3 has signature ( 3 , 3 ) , discriminant − 1 and Hasse invariants ε p ( V ) = ( − 1 , − 1 ) p 3 = ( − 1 , − 1 ) p . By Witt’s theorem T embeds in V if and only if there is a quadratic space W over Q with T ⊕ W ≅ V , and by the Hasse–Minkowski theorem [108, Ch. IV, Thm. 9] this is a condition on invariants: W must have dimension 6 − r , signature ( 1 , 5 − r ) , discriminant d ( W ) = d ( V ) d ( T ) − 1 = − d ( T ) and, from ε p ( T ⊕ W ) = ε p ( T ) ε p ( W ) ( d ( T ) , d ( W ) ) p with ( d ( T ) , − d ( T ) ) p = 1 , Hasse invariants ε p ( W ) = ( − 1 , − 1 ) p ε p ( T ) . The signature forces r − 2 ≤ 3 . A quadratic space over Q with prescribed dimension k, signature, discriminant and Hasse invariants exists if and only if the invariants satisfy the product formula, the real invariants are those of the signature, and at every prime the local existence condition holds [108, Ch. IV, §2.3, Prop. 6, and §3.3, Prop. 7]; the first two conditions hold here because they hold for T and for V and T ⊗ R ⊕ W ⊗ R ≅ V ⊗ R , and the local condition is empty for k ≥ 3 , is ε p ( W ) = 1 at every p at which d ( W ) = − 1 in Q p × / Q p × 2 for k = 2 , and is ε p ( W ) = 1 at every p for k = 1 . With k = 6 − r , d ( W ) = − d ( T ) and ε p ( W ) = ( − 1 , − 1 ) p ε p ( T ) these are the three cases of (c).
(b) and the class A . Let f : S ⤏ Km ( A ) be dominant. By Lemma 4, f * = Γ f * t maps T ( Km ( A ) ) Q onto T. Let β : A ˜ → A be the blow-up of the sixteen points of order two and π : A ˜ → Km ( A ) the quotient by the involution; Γ A = ( β × π ) * [ A ˜ ] is an algebraic cycle on A × Km ( A ) with Γ A * = π * β * , and H 2 ( Km ( A ) , Q ) = π * β * H 2 ( A , Q ) ⊕ ⨁ i = 1 16 Q [ E i ] with E i the exceptional curves. Hence T ⊆ ( Γ f t ∘ Γ A ) * H 2 ( A , Q ) , and H 2 ( S ) = NS ( S ) Q ⊕ T is spanned by this image and by divisor classes, while H 0 ( S ) and H 4 ( S ) are spanned by the class of S and the class of a point. So S ∈ A , the point being the abelian variety of dimension zero, S n ∈ A and (F3′) holds on S n by Proposition 2. For the Hodge conjecture, the proof of Proposition 2(i) writes a Hodge class of S n as ∑ j Γ j * y j with y j a Hodge class on one of the abelian varieties spanning H * ( S n ) , here the powers A k with k ≤ n . The Hodge conjecture holds for every product of surfaces with p g = 1 and q = 2  [104], in particular for A k , so every y j , and with it every Hodge class of S n , is algebraic. The Chow motive of S [ n ] is a direct summand of a sum of Tate twists of the motives h ( S k ) , k ≤ n  [32], so every Hodge class of S [ n ] is the image under an algebraic correspondence of a Hodge class of some S k .    □
Remark 2
(The powers left open). (i) At ρ ( S ) ≥ 19 the condition (c) is empty and S has a Shioda–Inose structure [81]. At ρ ( S ) = 17 or 18 it is a condition on the Hasse invariants, met by U Q 2 ⊕ 〈 − 2 〉 , by U Q ⊕ 〈 2 , − 2 , − 2 〉 and by 〈 2 , 2 , − 2 , − 2 〉 , and failed at p = 2 and p = 3 by 〈 2 , 2 , − 6 , − 6 〉 , by 〈 2 , 2 , − 2 , − 6 , − 6 〉 and by 〈 2 , 6 , − 2 , − 2 , − 2 〉 , the primes of failure coming in pairs by the product formula; all six are transcendental lattices of K3 surfaces of Picard number 18 or 17 by [93] and the surjectivity of the period map. The Kummer condition is not needed for the Hodge conjecture on the powers: by Corollary 3 it holds for all powers and Hilbert schemes of every projective K3 surface of Picard number at least 17, the last three included, which embed in V 1 . For ρ ≤ 16 no surface is of Kummer type, and at ρ = 16 Corollary 4 decides the surfaces whose transcendental space has Witt index two.
(ii) The Fricke surfaces have complex multiplication by K = Q ( − N ) at every member, and Theorem 13(iii) settles the Hodge conjecture on all their powers. Their transcendental spaces U Q 2 ⊕ 〈 − 1 〉 ⊕ 〈 − N 〉 (Proposition 15) have discriminant N, so they are neither of Kummer type nor in the two families of Picard number 16 for which Varesco proves the Hodge conjecture on all powers from the algebraicity of the Kuga–Satake correspondence [116, Thm. 4.3, Exs. 4.4, 4.5], whose transcendental spaces U Q 2 ⊕ 〈 − 2 〉 2 and U Q 2 ⊕ 〈 − 6 〉 ⊕ 〈 − 2 〉 have discriminants 1 and 3 modulo squares; they are settled by the complex multiplication instead. For each surface of Proposition 13(ii) the exceptional classes are Hodge classes of degree 2 m on S m reached by no correspondence of this section, all of which act on T ⊗ T ; whether they are algebraic, and with it whether (F3′) and the Hodge conjecture hold on S m , is open. The one construction known to produce such classes is the Kuga–Satake correspondence, through which Varesco proves [116, Thm. 4.3].
Proposition 15
(The transcendental lattice and the Kuga–Satake variety of the Fricke surfaces). Let N ∈ { 5 , 7 } , K = Q ( − N ) , let M be the trivial lattice U ⊕ A 4 ⊕ A 1 2 ⊕ D 4 2 or U ⊕ A 6 ⊕ D 4 2 of the elliptic fibration of the members of A N , and let T = M Q ⊥ ⊂ H 2 ( X , Q ) , which is T ( X ) Q when ρ ( X ) = 16 .
(i)
T ≅ U Q 2 ⊕ 〈 − 1 〉 ⊕ 〈 − N 〉 as a quadratic space over Q , so d ( T ) = N and the centre of the even Clifford algebra C + ( T ) is K.
(ii)
The Kuga–Satake variety of T is isogenous to B 4 for an abelian fourfold B of K-Weil type with discriminant one. If ρ ( X ) = 16 and End Hdg T = K , then B is isogenous to E K × B 3 , where E K is an elliptic curve with complex multiplication by K and B 3 is a simple abelian threefold with End ( B 3 ) Q = K ; in particular B is not simple.
Proof. 
(i) The root lattices are rationally diagonal: D n = { x ∈ Z n : ∑ x i ∈ 2 Z } has index two in Z n , so D n ⊗ Q ≅ 〈 1 〉 n , and E 8 ⊃ D 8 with index two gives E 8 ⊗ Q ≅ 〈 1 〉 8 ; A n = { x ∈ Z n + 1 : ∑ x i = 0 } is the orthogonal complement of the vector ( 1 , ⋯ , 1 ) of norm n + 1 , so ( A n ⊕ 〈 n + 1 〉 ) ⊗ Q ≅ 〈 1 〉 n + 1 ; and A 1 2 = 〈 2 , 2 〉 ≅ 〈 1 , 1 〉 by the vectors ( 1 , 1 ) and ( 1 , − 1 ) of Z 2 . Hence, with all forms taken positive definite, A 4 ⊕ A 1 2 ⊕ D 4 2 ⊕ 〈 1 , 5 〉 ≅ 〈 1 〉 5 ⊕ 〈 1 〉 2 ⊕ 〈 1 〉 8 ⊕ 〈 1 〉 = 〈 1 〉 16 and A 6 ⊕ D 4 2 ⊕ 〈 1 , 7 〉 ≅ 〈 1 〉 7 ⊕ 〈 1 〉 8 ⊕ 〈 1 〉 = 〈 1 〉 16 over Q . The K3 lattice is Λ = U 3 ⊕ E 8 ( − 1 ) 2 , so Λ Q ≅ U Q 3 ⊕ 〈 − 1 〉 16 , while M Q ⊕ U Q 2 ⊕ 〈 − 1 , − N 〉 ≅ U Q 3 ⊕ 〈 − 1 〉 16 by the two isometries just written. Thus M Q ⊕ T ≅ Λ Q ≅ M Q ⊕ ( U Q 2 ⊕ 〈 − 1 , − N 〉 ) , and Witt’s cancellation theorem gives T ≅ U Q 2 ⊕ 〈 − 1 〉 ⊕ 〈 − N 〉 . Its discriminant is ( − 1 ) 2 ( − 1 ) ( − N ) = N , and the centre of C + ( T ) is Q ( ( − 1 ) 3 d ( T ) ) = Q ( − N ) . The program julia/frickelattice.jl recomputes these isometries by the invariants of Hasse–Minkowski.
(ii) By Lombardo’s theorem [73, Thms. 6.2, 6.4], in the form of [116, Thm. 4.1], the Kuga–Satake variety of a polarised Hodge structure of K3 type on U Q 2 ⊕ 〈 a 〉 ⊕ 〈 b 〉 with a , b < 0 is isogenous to B 4 for an abelian fourfold B of Q ( − a b ) -Weil type with discriminant one; here ( a , b ) = ( − 1 , − N ) , Q ( − a b ) = K and C + ( T ) ≅ M 4 ( K ) . Suppose ρ ( X ) = 16 and End Hdg T = K . By Zarhin’s theorem [126, Thm. 2.3.1] the Hodge group of T is G = U ( W , h ) , the unitary group of the hermitian form h on the three-dimensional K-space W = T with Tr K / Q h = q . Over C the two eigenspaces W σ , W σ ¯ of K on T C have dimension three and are isotropic for q, since q ( a x , y ) = q ( x , a ¯ y ) , so T C = W σ ⊕ W σ * and G C = GL ( W σ ) acting on W σ * by the dual representation. The Kuga–Satake Hodge structure on H 1 ( B , Q ) is the lift of that of T along Spin ( T ) → SO ( T )  [118, §5], so the Hodge group of B is the preimage G ˜ of G in Spin ( T ) , and H 1 ( B , C ) = S + ⊕ S − is the sum of the two half-spin representations of Spin ( T C ) ≅ SL 4 , on which the two eigenspaces of the centre K of C + ( T ) are S + and S − . For the isotropic decomposition T C = W σ ⊕ W σ * the spin representation is ⋀ • W σ ⊗ ( det W σ ) − 1 / 2 , with S + the even and S − the odd part. Restricted to G ˜ C , S + = ( ⋀ 0 W σ ⊕ ⋀ 2 W σ ) ⊗ det − 1 / 2 is the sum of the character χ = det − 1 / 2 and of the irreducible representation W σ * ⊗ det 1 / 2 of dimension three, and S − = ( W σ ⊕ ⋀ 3 W σ ) ⊗ det − 1 / 2 is the sum of W σ ⊗ det − 1 / 2 and of χ − 1 . The four summands are pairwise non-isomorphic, the two characters being distinct and the two three-dimensional representations having different weights, so they are the isotypic components of H 1 ( B , C ) under the connected group G ˜ C . An automorphism of C over Q carries an isotypic component to one of the same dimension, since the representation is defined over Q ; so U C = χ ⊕ χ − 1 and V C = S + ⊕ S − minus it are stable under all of them and are the complexifications of Q -subspaces U, V with H 1 ( B , Q ) = U ⊕ V . Both are stable under G ˜ , hence sub-Hodge structures, and under K, which acts by scalars on S + and on S − . By Poincaré reducibility B is isogenous to E K × B 3 with H 1 ( E K , Q ) = U and H 1 ( B 3 , Q ) = V . The algebra End Hdg ( U ) ⊗ C = End ( χ ) ⊕ End ( χ − 1 ) has dimension two and contains K, so End ( E K ) Q = K ; likewise End Hdg ( V ) ⊗ C has dimension two, so End ( B 3 ) Q = K , a field, and B 3 is simple; and Hom ( E K , B 3 ) = 0 since U and V share no isotypic component.    □
Remark 3
(Real multiplication at the fixed points). Theorem 13(iii) decides every power of every Fricke surface, so the Kuga–Satake variety of Proposition 15(ii) is not needed for them; the proposition records what the splitting of T into the two eigenspaces of K does to the Kuga–Satake construction. For a very general Hodge structure of K3 type on U Q 2 ⊕ 〈 − 1 〉 ⊕ 〈 − N 〉 , with endomorphism field Q , the fourfold B is simple with a definite quaternion algebra as endomorphism algebra [118, Prop. 5.7]; the Fricke families lie in the locus of the four-dimensional period domain where the endomorphism field is K, and there B splits. Real multiplication by N appears on the two twists at the fixed points of w, the surfaces of [121, §7.6, Prop. 7.7], where Λ 2 = N by Lemma 9(ii). The fixed points of w are two, so the fixed-point twists are two surfaces; one-dimensional families of transcendental rank six with algebraic real multiplication by 5 , 7 and every other real quadratic field are given by Theorem 16, through a correspondence that is not a chain.

5.8. Floccari’s Theorem and Real Multiplication of Degree Two

Floccari attached to every hyper-Kähler variety of generalised Kummer type a K3 surface and proved that its Kuga–Satake correspondence is algebraic [42,43]. The statement depends only on the rational transcendental space. Write Λ Kum n = U 3 ⊕ 〈 − 2 n − 2 〉 for the second cohomology lattice of a variety of generalised Kummer type of dimension 2 n , and for a positive integer m put
V m = U Q 3 ⊕ 〈 − m 〉 ,
a quadratic space of rank 7, signature ( 3 , 4 ) and discriminant m, which depends only on the class of m modulo squares; Λ Kum n ⊗ Q = V 2 n + 2 . We also write V F = V 1 ; it is Λ Kum 3 ( 2 ) ⊗ Q = U ( 2 ) Q 3 ⊕ 〈 − 16 〉 , since U ( 2 ) Q ≅ U Q by the basis e , 2 f and 〈 − 16 〉 ≅ 〈 − 1 〉 .
Theorem 15
(Floccari [43, Thm. 5.11]). Let S be a projective K3 surface such that T ( S ) Q embeds isometrically in V m for some positive integer m. Then the Kuga–Satake correspondence of S is algebraic, and the Hodge conjecture holds for S n for every n ≥ 1 .
The theorem numbers are those of the version arXiv:2501.02315 of [43]; the case m = 1 is [42, Thm. 1.3, Cor. 5.8]. Which transcendental spaces it reaches is decided by the following lemma. For a rational quadratic space T of even rank, C + ( T ) denotes its even Clifford algebra, a central simple algebra over the quadratic étale algebra Q [ x ] / ( x 2 − ( − 1 ) rk T / 2 d ( T ) )  [68, Ch. V, §2].
Lemma 11
(Transcendental spaces that embed in some V m ). Let T be a rational quadratic space of signature ( 2 , r − 2 ) with r ≥ 2 , and d = d ( T ) ∈ Q × / Q × 2 .
(i)
If r ≤ 5 , then T embeds isometrically in V m for some positive integer m.
(ii)
If r = 6 , then d > 0 ; let K = Q ( − d ) . The following are equivalent.
(a)
T embeds isometrically in V m for some positive integer m.
(b)
T has Witt index two: T ≅ U Q 2 ⊕ 〈 − x , − x d 〉 for some x ∈ Q > 0 .
(c)
T ≅ λ U Q 2 ⊕ 〈 − 1 , − d 〉 for some λ ∈ Q > 0 ; then (a) holds with m ≡ λ d modulo squares.
(d)
T ⊗ Q p ≅ U Q p 3 for every prime p such that − d ∈ Q p × 2 .
(e)
C + ( T ) ≅ M 4 ( K ) .
(iii)
If r ≥ 7 , then T embeds in no V m .
Proof. 
(i) Let r = 5 . An indefinite rational quadratic space of rank at least five is isotropic [108, Ch. IV, §3.2], so T ≅ U Q ⊕ 〈 x , y , z 〉 with 〈 x , y , z 〉 of signature ( 1 , 2 ) ; order the entries so that x > 0 > y and z < 0 . Then
T ⊕ 〈 − x , − y 〉 ≅ U Q ⊕ 〈 x , − x 〉 ⊕ 〈 y , − y 〉 ⊕ 〈 z 〉 ≅ U Q 3 ⊕ 〈 z 〉 ,
because a binary form 〈 a , − a 〉 represents zero and is a hyperbolic plane, and 〈 z 〉 ≅ 〈 − m 〉 for the positive integer m = − z s 2 with s a common denominator. If r < 5 , then T ⊕ 〈 − 1 〉 5 − r has signature ( 2 , 3 ) and the case r = 5 applies.
(ii) The discriminant of a space of signature ( 2 , 4 ) has the sign ( − 1 ) 4 , so d > 0 . (a)⇒(b): V m ⊇ U Q 3 contains a totally isotropic subspace N of dimension three, and T has codimension one in V m , so dim ( T ∩ N ) ≥ 2 and the Witt index of T is at least two; by the signature it is at most two. Hence T ≅ U Q 2 ⊕ 〈 − x , − y 〉 with x , y > 0 and x y ≡ d , that is y ≡ x d modulo squares. (b)⇒(c): λ U Q ≅ U Q , so U Q 2 ⊕ 〈 − x , − x d 〉 = x U Q 2 ⊕ 〈 − 1 , − d 〉 . (c)⇒(a): T ⊕ 〈 λ 〉 ≅ U Q 2 ⊕ 〈 λ , − λ 〉 ⊕ 〈 − λ d 〉 ≅ U Q 3 ⊕ 〈 − λ d 〉 = V m with m ≡ λ d .
(b)⇔(d). By (i) applied to a rank-five subspace, or directly by [108, Ch. IV, §3.2], T is isotropic, so T ≅ U Q ⊕ Q 4 with Q 4 of signature ( 1 , 3 ) and d ( Q 4 ) = − d . By Witt cancellation T has Witt index two if and only if Q 4 is isotropic, and by the Hasse–Minkowski theorem this holds if and only if Q 4 is isotropic over R and over every Q p . Over R it is indefinite. Over Q p with − d ∉ Q p × 2 it is isotropic, because an anisotropic quaternary form over Q p has square discriminant [108, Ch. IV, §2.2, Thm. 6]. Over Q p with − d ∈ Q p × 2 the form Q 4 has discriminant one; if it is isotropic it is U ⊕ B with B binary of discriminant − 1 , hence B ≅ U ; so Q 4 ⊗ Q p is isotropic if and only if it is U 2 , if and only if T ⊗ Q p ≅ U 3 .
(d)⇔(e). Let c ( T ) ∈ Br ( Q ) be the class of the Clifford algebra C ( T ) , central simple over Q since the rank is even. The centre of C + ( T ) is K, and C ( T ) ⊗ Q K ≅ M 2 ( C + ( T ) )  [68, Ch. V, §2], so [ C + ( T ) ] = res K / Q c ( T ) in Br ( K ) . As C + ( T ) has dimension 16 over K, (e) says that this class vanishes. The class c ( T ) has order at most two; restriction to a place w of K over a place v of Q multiplies local invariants by [ K w : Q v ] , so res K / Q c ( T ) = 0 if and only if c p ( T ) = 0 at every prime p that splits in K, the real place being complex in K. A prime p splits in K exactly when − d ∈ Q p × 2 ; then d ( T ⊗ Q p ) = − 1 = d ( U 3 ) , and two quadratic spaces over Q p of the same rank and discriminant are isometric if and only if their Clifford invariants agree; c ( U 3 ) = 0 . So (e) holds if and only if T ⊗ Q p ≅ U 3 at every such p.
(iii) V m has three positive and four negative directions, and T has r − 2 ≥ 5 negative ones.    □
Corollary 3
(Floccari’s theorem at Picard number at least seventeen). Let S be a projective K3 surface with ρ ( S ) ≥ 17 . Then the Kuga–Satake correspondence of S is algebraic, and the Hodge conjecture and (F3′) hold for every power S n and every Hilbert scheme S [ n ] .
Proof. 
T ( S ) Q has signature ( 2 , 20 − ρ ) and rank 22 − ρ ≤ 5 , so it embeds in some V m by Lemma 11(i), and Theorem 15 gives the Kuga–Satake correspondence and the Hodge conjecture for every S n . The Chow motive of S [ n ] is a sum of Tate twists of motives of symmetric products of S [32], so the Hodge conjecture for S [ n ] follows. When every Hodge class is algebraic, (F3′) holds, since the algebraic classes are the images of the class of a point under correspondences.    □
For ρ = 17 the embedding in V F = V 1 alone needs a condition on Hasse invariants, ε p ( T ) = ( − 1 , − 1 ) p at every prime p at which − d ( T ) is a square, and the theorem removes it. For instance 〈 1 , 1 , − 1 , − 23 , − 23 〉 , of discriminant − 1 , has ε 23 = ( − 23 , − 1 ) 23 = − 1 ≠ ( − 1 , − 1 ) 23 , so it does not embed in V 1 ; with x = 1 , y = z = − 23 in the proof of Lemma 11(i) it embeds in V 23 . Figure 8 summarises the K3 surfaces that Floccari’s theorem reaches.
Corollary 4
(Floccari’s theorem at Picard number sixteen). Let S be a projective K3 surface with ρ ( S ) = 16 , T = T ( S ) Q , d = d ( T ) and K = Q ( − d ) . If T satisfies the equivalent conditions of Lemma 11(ii), the Kuga–Satake correspondence of S is algebraic and the Hodge conjecture and (F3′) hold for every S n and every S [ n ] . If T has Witt index one, T embeds in no V m , and Theorem 15does not apply to S.
Proof. 
The first statement is Theorem 15 with Lemma 11(ii) and [32], as in Corollary 3; the second is (a)⇒(b) of the lemma.    □
The spaces 〈 1 , 1 , − 1 , − 1 , − 7 , − 7 〉 , 〈 1 , 1 , − 1 , − 1 , − 5 , − 23 〉 and 〈 1 , 1 , − 1 , − 1 , − 7 , − 23 〉 have Witt index two, with ( λ , m ) = ( 7 , 7 ) , ( 5 , 23 ) and ( 7 , 23 ) in Lemma 11(ii)(c); the first does not embed in V 1 , since that would force 〈 7 , 7 〉 ≅ 〈 1 , 1 〉 by Witt cancellation and 7 is not a sum of two rational squares. The spaces 〈 1 , 1 , − 1 , − 2 , − 3 , − 3 〉 and 〈 1 , 1 , − 1 , − 2 , − 5 , − 10 〉 have Witt index one. Each of these spaces contains an even lattice of rank six with a discriminant group on at most six generators, which embeds primitively in the K3 lattice by [93, Cor. 1.12.3], so each is T ( S ) Q for the very general member of a four-dimensional family of K3 surfaces. Of the 1980 diagonal forms 〈 a , b , − x 1 , − x 2 , − x 3 , − x 4 〉 with 1 ≤ a ≤ b ≤ 3 and x 1 ≤ x 2 ≤ x 3 ≤ x 4 in { 1 , 2 , 3 , 5 , 6 , 7 , 10 , 23 } , 1484 have Witt index two and 496 have Witt index one (Appendix C).
Lemma 12
(The space T d ). Let d > 1 be squarefree, E = Q ( d ) with real embeddings σ 1 ( d ) = d , σ 2 ( d ) = − d , let W be the three-dimensional E-quadratic space 〈 d , d , − 1 2 〉 , and let T d = Tr E / Q W , the Q -vector space W with the form Tr E / Q ∘ q W .
(i)
T d ≅ U Q 2 ⊕ 〈 − 1 〉 ⊕ 〈 − d 〉 ; it has signature ( 2 , 4 ) and discriminant d.
(ii)
Multiplication M by d is self-adjoint for the form of T d and M 2 = d ; the σ 1 -eigenspace W σ 1 ⊂ T d ⊗ R has signature ( 2 , 1 ) and the σ 2 -eigenspace is negative definite of rank 3.
(iii)
T d ⊕ 〈 d 〉 ≅ V F ; in particular T d embeds isometrically in V F .
(iv)
The even lattice L d = U 2 ⊕ 〈 − 4 〉 ⊕ 〈 − 4 d 〉 satisfies L d ⊗ Q ≅ T d and admits a primitive embedding in the K3 lattice Λ, whose orthogonal complement N d has signature ( 1 , 15 ) .
Proof. 
(i) For α ∈ E × the binary form ( x , y ) ↦ Tr E / Q ( α x y ) on E has, in the basis 1 , d , the Gram matrix
Tr α Tr α d Tr α d d Tr α .
For α = d it is 0 2 d 2 d 0 , a hyperbolic plane, and for α = − 1 2 it is diag ( − 1 , − d ) . The signature and the discriminant follow.
(ii) q W is E-bilinear, so Tr ( q W ( e x , y ) ) = Tr ( q W ( x , e y ) ) for e ∈ E . Over R , T d ⊗ R = W σ 1 ⊕ W σ 2 orthogonally, with the forms 〈 σ i d , σ i d , − 1 2 〉 , of signatures ( 2 , 1 ) and ( 0 , 3 ) .
(iii) 〈 − d 〉 ⊕ 〈 d 〉 ≅ U Q , since it represents zero; so T d ⊕ 〈 d 〉 ≅ U Q 3 ⊕ 〈 − 1 〉 .
(iv) Over Q , 〈 − 4 , − 4 d 〉 ≅ 〈 − 1 , − d 〉 . The lattice L d is even of signature ( 2 , 4 ) with discriminant group Z / 4 ⊕ Z / 4 d , generated by two elements. Since 3 − 19 ≡ 0 ( mod 8 ) , 2 ≤ 3 , 4 ≤ 19 and 22 − 6 > 2 , it embeds primitively in the even unimodular lattice Λ of signature ( 3 , 19 ) by Nikulin’s criterion [93, Cor. 1.12.3].    □
Theorem 16
(Real multiplication by every real quadratic field). Let d > 1 be squarefree and keep the notation of Lemma 12. Fix a marking Λ ⊃ L d and let
D d = { [ ω ] ∈ P ( W σ 1 ⊗ R C ) : q ( ω , ω ) = 0 , q ( ω , ω ¯ ) > 0 } ,
the complement of the real points in a smooth conic, a union of two discs.
(i)
For every [ ω ] ∈ D d there is a projective K3 surface S ω with H 2 , 0 ( S ω ) = C ω , N d ⊆ NS ( S ω ) and T ( S ω ) ⊆ L d ; each isomorphism class occurs for at most countably many ω, so the surfaces form a one-dimensional family.
(ii)
T ( S ω ) Q is an E-subspace of T d , and M restricts to a Hodge endomorphism of it. Outside a countable subset of D d one has T ( S ω ) Q = T d , ρ ( S ω ) = 16 and End Hdg T ( S ω ) Q = Q ( d ) exactly; at the other points the transcendental rank is 2 or 4 and the endomorphism field is a CM field.
(iii)
For every [ ω ] ∈ D d the Hodge conjecture holds for S ω n and for the Hilbert schemes S ω [ n ] , for every n ≥ 1 . In particular the real multiplication by d is algebraic on S ω × S ω , and the classes ⋀ E 3 T of Proposition 13on S ω 3 are algebraic.
(iv)
A rational quadratic space T of signature ( 2 , 4 ) embeds in V F if and only if T ≅ U Q 2 ⊕ 〈 − 1 〉 ⊕ 〈 − d ( T ) 〉 , if and only if ε p ( T ) = ( − 1 , d ( T ) ) p for every prime p.
Proof. 
(i) The point ω is a period of Λ , so the surjectivity of the period map [58, Ch. 7, Thm. 4.1] gives a marked K3 surface S ω with H 2 , 0 = C ω . Its Néron–Severi group ω ⊥ ∩ Λ contains L d ⊥ = N d , which contains a class of positive square, so S ω is projective [17, Thm. IV.6.2], and T ( S ω ) = NS ( S ω ) ⊥ ⊆ L d . Two markings of one surface differ by an element of the countable group O ( Λ ) .
(ii) The space V = T ( S ω ) Q is the smallest Q -subspace of T d whose complexification contains ω . As M ω = d ω , the complexification of M ( V ) contains ω too, so V ⊆ M ( V ) and M ( V ) = V : V is an E-subspace of E-dimension 1, 2 or 3. The real operator M is self-adjoint with eigenvectors ω , ω ¯ , so it preserves H 2 , 0 , H 0 , 2 and their orthogonal complement in V C , and M | V is a Hodge endomorphism. There are countably many proper E-subspaces V ′ ⊂ T d , and each V ′ ⊗ E , σ 1 C is a proper linear subspace of the plane P ( W σ 1 ⊗ C ) , which meets the smooth conic in at most two points; outside these points V = T d . Then ρ = 22 − 6 = 16 , and the endomorphism field E ′ contains Q ( d ) and is totally real or CM by Zarhin’s theorem [126, Thm. 1.6], of degree dividing 6. If E ′ is totally real then dim E ′ T ≥ 3 by the first paragraph of the proof of Proposition 20, so E ′ = Q ( d ) . If E ′ were CM, of degree 2 it would equal the real field Q ( d ) , and of degree 6 its cubic totally real subfield would contain Q ( d ) ; both are impossible. At a special point dim E V ≤ 2 ; a totally real field would need dim ≥ 3 over itself, so the endomorphism field is CM.
(iii) T ( S ω ) Q ⊆ T d embeds in V F = V 1 by Lemma 12(iii), so Theorem 15 applies. The classes named are Hodge classes on S ω 2 and S ω 3 . For the Hilbert schemes, the motive of S [ n ] is a sum of Tate twists of motives of symmetric products of S [32], so the Hodge conjecture for S [ n ] follows from that for the powers of S, as in Theorem 11. The real multiplication also has a second proof that uses only the Kuga–Satake part of Theorem 15: q ( M x , M y ) = q ( M 2 x , y ) = d q ( x , y ) , so M is a Hodge similarity, and a Hodge similarity between K3 surfaces satisfying the Kuga–Satake Hodge conjecture is algebraic by Varesco’s theorem [115, Thm. 4.5, Cor. 4.6], the Lefschetz standard conjecture in degree two being trivial for surfaces.
(iv) If T ⊂ V F then V F ≅ T ⊕ 〈 c 〉 with c ≡ d ( V F ) d ( T ) = d ( T ) modulo squares, and comparing signatures c > 0 . Since also V F ≅ U Q 2 ⊕ 〈 − 1 〉 ⊕ 〈 − c 〉 ⊕ 〈 c 〉 , Witt cancellation gives T ≅ U Q 2 ⊕ 〈 − 1 〉 ⊕ 〈 − c 〉 ; the converse is Lemma 12(iii) with d replaced by c. By the Hasse–Minkowski theorem two rational quadratic spaces with the same rank, discriminant and signature are isometric if and only if their Hasse invariants agree at every prime [108, Ch. IV, Thm. 9], and ε p ( 〈 1 , − 1 , 1 , − 1 , − 1 , − c 〉 ) = ( − 1 , − 1 ) p 3 ( − 1 , − c ) p 3 = ( − 1 , c ) p , using ( − 1 , − 1 ) p ( − 1 , − c ) p = ( − 1 , c ) p . Comparing with Lemma 11(ii)(c), these are the spaces in which λ can be taken to be 1; the spaces of Witt index two in which it cannot embed in V λ d and not in V 1 .    □
The surfaces of Theorem 16 are one family for each real quadratic field. Lemma 11 decides every family of transcendental rank six with real multiplication by a real quadratic field, through the quaternion algebra C + ( W ) .
Setting 2
(Real multiplication of degree two in rank six). Let e > 1 be squarefree, E = Q ( e ) with real embeddings s 1 ( e ) = e , s 2 ( e ) = − e , and let W be a three-dimensional quadratic space over E of signature ( 2 , 1 ) at s 1 and ( 0 , 3 ) at s 2 , with discriminant δ = d ( W ) ∈ E × / E × 2 and even Clifford algebra Q W = C + ( W ) , a quaternion algebra over E. Put T = Tr E / Q W , of signature ( 2 , 4 ) , and let D W ⊂ P ( W s 1 ⊗ R C ) , W s 1 = W ⊗ E , s 1 R , be the union of two discs cut out by q ( ω , ω ) = 0 , q ( ω , ω ¯ ) > 0 . If L ⊂ Λ is a primitive even sublattice with L ⊗ Q ≅ T , every [ ω ] ∈ D W is the period of a projective K3 surface S ω with T ( S ω ) Q ⊆ T an E-subspace on which e acts as a Hodge endomorphism, and outside a countable subset of D W one has T ( S ω ) Q = T , ρ ( S ω ) = 16 and End Hdg T ( S ω ) Q = E ; the proof of Theorem 16(i), (ii) applies word for word with T d replaced by T. Such a lattice exists for every W: a lattice in T scaled by 2 is even, of rank 6, with a discriminant group on at most six generators, and 6 + 6 < 22 in Nikulin’s criterion [93, Cor. 1.12.3].
Theorem 17
(Witt index of a trace form). In Setting 2, d ( T ) ≡ e N E / Q ( δ ) modulo squares, d = e N E / Q ( δ ) > 0 , and T has Witt index two if and only if
∑ P ∣ p inv P ( Q W ) = 0 in 1 2 Z / Z for every prime p with − d ∈ Q p × 2 .
In that case every surface S ω , [ ω ] ∈ D W , has an algebraic Kuga–Satake correspondence, the Hodge conjecture and (F3′) hold for all its powers and Hilbert schemes of points, and the real multiplication by e is algebraic on S ω × S ω .
Proof. 
For α = u + v e ∈ E × the binary form Tr E / Q 〈 α 〉 has, in the basis 1 , e , the Gram matrix
2 u 2 v e 2 v e 2 u e , det = 4 e ( u 2 − e v 2 ) = 4 e N E / Q ( α ) .
Diagonalising W = 〈 α 1 , α 2 , α 3 〉 gives d ( T ) ≡ e 3 N ( α 1 α 2 α 3 ) ≡ e N ( δ ) . The sign of N ( δ ) = s 1 ( δ ) s 2 ( δ ) is ( − 1 ) ( − 1 ) 3 = 1 , so d > 0 . By Lemma 11(ii), T has Witt index two if and only if T p = T ⊗ Q p is hyperbolic at every prime p with − d ∈ Q p × 2 ; fix such a p. Then T p = ⨁ P ∣ p Tr E P / Q p W P with W P = W ⊗ E E P , and d ( T p ) = d ≡ − 1 .
p splits in E. Then E ⊗ Q p = Q p × Q p and T p = W 1 ⊕ W 2 with W i = W P i ternary over Q p . From d ( W 1 ) d ( W 2 ) ≡ − 1 we get d ( W 2 ) ≡ − d ( W 1 ) = d ( − W 1 ) . Since W 1 ⊕ ( − W 1 ) is hyperbolic, Witt cancellation shows that T p is hyperbolic if and only if W 2 ≅ − W 1 . Ternary forms over Q p of the same discriminant are isometric if and only if their even Clifford algebras are isomorphic, and C + ( − W 1 ) ≅ C + ( W 1 ) because the rank is odd. So T p is hyperbolic if and only if inv P 1 ( Q W ) = inv P 2 ( Q W ) , that is, if and only if their sum vanishes.
p does not split in E and W P is isotropic. Then P is the only place over p, Q W is split at P , and W P ≅ H ⊕ 〈 − δ 〉 with H a hyperbolic plane over E P . The trace of an isotropic E P -line is a totally isotropic Q p -plane, so Tr H ≅ U 2 ; and Tr 〈 − δ 〉 is a binary form of discriminant e N ( − δ ) = d ≡ − 1 , hence hyperbolic. So T p ≅ U 3 , in accordance with inv P ( Q W ) = 0 .
p does not split in E and W P is anisotropic. Then inv P ( Q W ) = 1 2 and we show that T p is not hyperbolic. Let D be the quaternion division algebra over E P and n D its reduced norm form. The form δ W P is ternary, anisotropic and of discriminant one; over a local field there is exactly one such form up to isometry, the restriction of n D to the pure quaternions, so W P ⊕ 〈 δ 〉 ≅ δ n D ≅ n D , the last because the reduced norm of D is surjective onto E P × . In the Witt group W ( Q p ) the transfer is additive and Tr 〈 δ 〉 is hyperbolic, its discriminant being e N ( δ ) ≡ − 1 ; so T p = Tr n D in W ( Q p ) , and it suffices to show Tr n D ≠ 0 . The norm group N ( E P × ) has index two in Q p × , which contains Q p × 2 with index at least four; choose n ∈ N ( E P × ) ∖ Q p × 2 , b ∈ E P × with N ( b ) = n , and a ∈ Q p × with ( a , n ) p = − 1 . By the projection formula ( a , b ) P = ( a , N ( b ) ) p = − 1 , so the quaternion algebra ( a , b ) over E P is D and n D ≅ 〈 1 , − a 〉 ⊗ 〈 1 , − b 〉 . By Frobenius reciprocity for the transfer [106, Ch. 2, §5], Tr n D = 〈 1 , − a 〉 ⊗ φ with φ = Tr 〈 1 , − b 〉 , a quaternary form of discriminant e · e N ( b ) ≡ n ; so φ ≡ 〈 1 , − n 〉 modulo I 2 ( Q p ) and Tr n D ≡ 〈 1 , − a 〉 ⊗ 〈 1 , − n 〉 modulo I 3 ( Q p ) . Over a p-adic field I 3 = 0 : I 3 is additively generated by the three-fold Pfister forms, of dimension eight; every form of dimension at least five is isotropic [108, Ch. IV, §2.2, Thm. 6] and an isotropic Pfister form is hyperbolic [68, Ch. X, §1]; so Tr n D equals the Pfister form 〈 〈 a , n 〉 〉 in W ( Q p ) , the norm form of the quaternion algebra ( a , n ) p , which is a division algebra; that form is anisotropic and nonzero in W ( Q p ) .
In all three cases T p is hyperbolic if and only if ∑ P ∣ p inv P ( Q W ) = 0 ; no case uses p ≠ 2 . The rest follows from Lemma 11(ii), Theorem 15 and [32] as in Corollary 3, the class of e being a Hodge class on S ω × S ω .    □
Corollary 5
(Families of Witt index two for every real quadratic field). In Setting 2let W = 〈 s e , s ′ e , c 〉 with s , s ′ ∈ Q > 0 and c ∈ E totally negative. Then T = U Q 2 ⊕ Tr 〈 c 〉 has Witt index two, and the conclusions of Theorem 17hold on the whole family D W .
Proof. 
The signatures are as required, and the Gram matrix of Tr 〈 s e 〉 is 0 2 s e 2 s e 0 , a hyperbolic plane. Apply Lemma 11(ii)(b).    □
Lemma 12 is the case s = s ′ = 1 , c = − 1 2 ; the families of Corollary 5 with c ranging over the totally negative elements up to squares give infinitely many pairwise non-isometric transcendental spaces for every E, since d ( T ) ≡ e N ( c ) .
Corollary 6
(Families of Witt index one for every real quadratic field). For every real quadratic field E there are spaces W as in Setting 2for which T has Witt index one; for them Theorem 15applies to no surface of D W whose transcendental space is T, and in particular to no very general member.Table 4 gives one for e = 2 , 3 , 5 , 6 , 7 , with α = a + b e written ( a , b ) .
Proof. 
Let p ∤ 2 e be a prime that splits in E, with places P ≠ P ′ over it, and let Q be the quaternion algebra over E ramified exactly at s 2 and P ; it exists because the set of ramified places is finite of even cardinality [103, §32]. If Q = ( a , b ) , the restriction of the reduced norm to the pure quaternions is 〈 − a , − b , a b 〉 , of discriminant one and with even Clifford algebra ( − a b , a ) ≅ ( a , b ) = Q , since C + 〈 α , β , γ 〉 ≅ ( − α β , − β γ )  [68, Ch. V, §2]. At s 1 , where Q is split, this form is the determinant on trace-zero real 2 × 2 matrices, of signature ( 1 , 2 ) ; at s 2 it is positive definite. By weak approximation choose λ ∈ E × negative at s 1 and s 2 , in − e E P × 2 and in E P ′ × 2 . Then W = λ 〈 − a , − b , a b 〉 has signature ( 2 , 1 ) at s 1 and ( 0 , 3 ) at s 2 , Q W = Q because scaling a ternary form does not change its even Clifford algebra, and δ ≡ λ 3 ≡ λ . In Q p = E P = E P ′ one has N ( λ ) = λ P λ P ′ ≡ − e , so − d = − e N ( δ ) ≡ e 2 is a square in Q p , while inv P ( Q ) + inv P ′ ( Q ) = 1 2 . By Theorem 17, T has Witt index one, and Lemma 11(ii) says it embeds in no V m . For the table, the signatures, d and the Hasse invariants of T at every prime are computed in Appendix C; at the prime listed T p ≅ U 3 .    □
Table 4. Spaces W = 〈 α 1 , α 2 , α 3 〉 over E = Q ( e ) as in Corollary 6: T = Tr W has Witt index one, and T ⊗ Q p ≅ U 3 at the prime p listed, where − d is a square.
Table 4. Spaces W = 〈 α 1 , α 2 , α 3 〉 over E = Q ( e ) as in Corollary 6: T = Tr W has Witt index one, and T ⊗ Q p ≅ U 3 at the prime p listed, where − d is a square.
e α 1 α 2 α 3 d mod Q × 2 p
2 ( 3 , 3 ) ( 2 , 4 ) ( − 3 , 2 ) 14 3
3 ( 2 , 4 ) ( 0 , 3 ) ( − 1 , 0 ) 11 3
5 ( − 1 , 1 ) ( 2 , 1 ) ( − 4 , 1 ) 55 2
6 ( − 3 , 2 ) ( 0 , 3 ) ( − 4 , 0 ) 15 2
7 ( 0 , 2 ) ( 3 , 3 ) ( − 1 , 0 ) 6 7
Families of higher dimension are reached on a dense set of curves.
Theorem 18
(Dense curves in every family with real multiplication of degree two). Let E = Q ( e ) be real quadratic, m ≥ 4 , and W an m-dimensional quadratic space over E of signature ( 2 , m − 2 ) at s 1 and negative definite at s 2 ; let L ⊂ Λ be a primitive sublattice with L ⊗ Q ≅ Tr E / Q W , and let D W ⊂ P ( W s 1 ⊗ C ) be the period domain of dimension m − 2 cut out by q ( ω , ω ) = 0 , q ( ω , ω ¯ ) > 0 . There are countably many three-dimensional E-subspaces W ′ ⊂ W such that
(i)
D W ′ = D W ∩ P ( W s 1 ′ ⊗ C ) is a nonempty curve, whose very general member has T ( S ω ) Q = Tr W ′ , ρ = 16 and End Hdg T ( S ω ) Q = E ;
(ii)
for every [ ω ] ∈ D W ′ the Kuga–Satake correspondence of S ω is algebraic, the Hodge conjecture and (F3′) hold for all powers and Hilbert schemes of points of S ω , and the real multiplication by E is algebraic on S ω × S ω ;
(iii)
the union of the curves D W ′ is dense in D W .
Proof. 
Write q for the E-form of W, q i for the real form on W s i , and Tr q for the rational form of Tr W , so that Tr W ⊗ R = W s 1 ⊕ W s 2 with Tr q = q 1 + q 2 . Since Tr q ( v ) = 2 u for q ( v ) = u + v ′ e , the rational isotropic vectors of Tr q are the v with q ( v ) ∈ Q e ; for such v ≠ 0 write q ( v ) = s e , so that s > 0 exactly when q 1 ( v ) > 0 , and then Tr 〈 q ( v ) 〉 is a hyperbolic plane. If Y ⊆ W is an E-subspace of signature ( k , dim Y − k ) at s 1 , k ≥ 1 , and negative definite at s 2 , with dim E Y ≥ 3 , then Tr q | Y is indefinite of rank at least six, hence isotropic over Q  [108, Ch. IV, §3.2]; its projective quadric is smooth with a rational point, hence rational over Q , so its rational points are dense in its real points. The real isotropic cone of Tr q | Y is { ( y 1 , y 2 ) : q 1 ( y 1 ) = − q 2 ( y 2 ) } , and as q 2 is negative definite its projection to Y s 1 is { q 1 ≥ 0 } . Hence: for every y 1 ∈ Y s 1 with q 1 ( y 1 ) > 0 there are rational v ∈ Y with q ( v ) ∈ Q > 0 e whose s 1 -components are arbitrarily close to positive multiples of y 1 .
Let [ ω ] ∈ D W and ω = x + i y , so that x , y ∈ W s 1 are orthogonal with q 1 ( x ) = q 1 ( y ) > 0 ; choose z ∈ W s 1 orthogonal to both with q 1 ( z ) < 0 . Apply the last sentence with Y = W to get v with q ( v ) = s e , s > 0 , v s 1 close to a multiple of x; then with Y = v ⊥ , of dimension m − 1 ≥ 3 and signature ( 1 , m − 2 ) at s 1 , to get w ⊥ v with q ( w ) = s ′ e , s ′ > 0 , w s 1 close to a multiple of the projection of y to v s 1 ⊥ . The space 〈 v , w 〉 ⊥ is negative definite at both real places, since W s 1 has only two positive directions; choose a rational u ≠ 0 in it with u s 1 close to a multiple of the projection of z. Then q ( u ) is totally negative, and W ′ = E v ⊕ E w ⊕ E u has the diagonal form 〈 s e , s ′ e , q ( u ) 〉 of Corollary 5: Tr W ′ = U Q 2 ⊕ Tr 〈 q ( u ) 〉 has Witt index two. The three-space W s 1 ′ , of signature ( 2 , 1 ) , is close to 〈 x , y , z 〉 ∋ ω , so the conic D W ′ contains points close to [ ω ] ; this proves (iii), and W ′ ranges over the countably many E-subspaces of W.
For [ ω ′ ] ∈ D W ′ the surface S ω ′ is projective, since L ⊥ has signature ( 1 , 21 − 2 m ) and contains a class of positive square, and T ( S ω ′ ) Q , the smallest rational subspace whose complexification contains ω ′ , lies in Tr W ′ . So it embeds in some V m ′ by Lemma 11(ii), and (ii) follows from Theorem 15 and [32] as in Corollary 3; the multiplication by e preserves T ( S ω ′ ) Q because ω ′ is an eigenvector, as in the proof of Theorem 16(ii). Statement (i) is that proof with T d replaced by Tr W ′ .    □
The union of the curves of Theorem 18 is a countable union of proper closed analytic subsets of D W , hence meagre; whether the real multiplication is algebraic at a very general member of D W is not decided by it, and by Lemma 2 its algebraic locus is either all of D W or meagre. Figure 9 draws the curves. Such families exist for 4 ≤ m ≤ 10 by Bayer-Fluckiger, van Geemen and Schütt [27], who construct families of dimension m − 2 with real multiplication by a totally real field E whenever m [ E : Q ] ≤ 21 .
Table 5. Families of projective K3 surfaces with real multiplication by Q ( d ) on which it is known to be algebraic.
Table 5. Families of projective K3 surfaces with real multiplication by Q ( d ) on which it is known to be algebraic.
d source dimension rk T all powers S n
2  [115]; [121, Prop. 6.2] 4 12 open
2, every family Theorem 12 2; 3 8; 10 open
2, 3, 6 in rank six Corollary 1 1 6 open
3  [115]; [121, Prop. 7.2] 2; 3 8; 12 open
5  [121, Thm. 1.1] 7 20 open
6 Theorem 11 2 8 open
sums of two squares  [105], with [118] 1 6 proved [116]
every squarefree d > 1 Theorem 16 1 6 proved
every d, T of Witt index two Theorem 17 1 6 proved
every d, curves dense in each family Theorem 18 1 ≥ 8 proved
For d = N ∈ { 5 , 7 } the space T N = U Q 2 ⊕ 〈 − 1 〉 ⊕ 〈 − N 〉 is exactly the transcendental space of the Fricke families of Proposition 15(i). So the families A N , with complex multiplication by Q ( − N ) , and the curve D N of Theorem 16, with real multiplication by Q ( N ) , lie in the same four-dimensional period domain, and every K3 surface whose transcendental space lies in T N satisfies the Hodge conjecture on all its powers by Theorem 15. The correspondences realising d in this subsection are built from the Kuga–Satake cycle, not from chains of symplectic quotients or fibrewise isogenies, so Proposition 20 is not contradicted, and Proposition 9 is consistent with them, since multiplication by d is a similarity of multiplier d. What they do not give is a family of dimension at least two on which the real multiplication is algebraic at every member: Theorem 15 needs an embedding in a space of rank seven, so the transcendental rank is at most six, a totally real field of degree f acting on a space of rank six needs 3 f ≤ 6 , and in rank six Lemma 11(ii) requires Witt index two.

5.9. The Hecke Graph of a Family with Real Multiplication

The very general member of a family of Theorem 18 is reached by a class that lives on pairs of different surfaces. Its Hodge locus is the graph of a Hecke correspondence, the class is algebraic on a dense set of that graph, and the semiregularity count on a product of two K3 surfaces says what an object at one point of the graph has to achieve.
Setting 3
(The Hecke family). Let E = Q ( e ) with e > 1 squarefree, 3 ≤ m ≤ 10 , W and L ⊂ Λ as in Theorem 18 (such L exist by [27]), and N = L ⊥ , of rank 22 − 2 m and signature ( 1 , 21 − 2 m ) . Let φ ∈ End ( Λ Q ) be multiplication by e on L Q = Tr E / Q W and 0 on N Q ; it is self-adjoint, and q ( φ x , φ y ) = e q ( x , y ) on L Q . Put D L = { [ w ] ∈ P ( L C ) : q ( w ) = 0 , q ( w , w ¯ ) > 0 } , of dimension 2 m − 2 ; it contains D W , of dimension m − 2 . For [ w ] ∈ D L also [ φ w ] ∈ D L , since q ( φ w ) = e q ( w ) = 0 and q ( φ w , φ w ¯ ) = e q ( w , w ¯ ) ; so g φ [ w ] = [ φ w ] is an automorphism of D L , the action of a similitude of multiplier e. Fix h ∈ N with h 2 > 0 orthogonal to no ( − 2 ) -vector of N, a neat subgroup Γ ⊆ O ( Λ ) of finite index fixing N pointwise and preserving L, and let Γ ′ ⊆ Γ be the subgroup of finite index of those g for which φ g φ − 1 on L Q , extended by the identity on N Q , lies in Γ. Let B ⊆ Γ ′ ∖ D L be the complement of the images of the hyperplanes δ ⊥ and ( φ δ ) ⊥ , δ ∈ Λ , δ 2 = − 2 , δ · h = 0 ; they form a locally finite family, so B is open and dense, and Zariski open by Borel’s extension theorem [21]. The two maps [ w ] ↦ [ w ] and [ w ] ↦ [ φ w ] from B to the moduli space of N-polarised K3 surfaces with level-Γ structure are holomorphic, hence algebraic [21], and pulling back the universal family along them gives a smooth projective family of pairs ( S 1 , S 2 ) over B with markings μ i : H 2 ( S i , Z ) → Λ , h ample on both, and [ μ 2 ( H 2 , 0 ( S 2 ) ) ] = g φ [ μ 1 ( H 2 , 0 ( S 1 ) ) ] . Let Φ W ∈ H 2 ( S 1 , Q ) ⊗ H 2 ( S 2 , Q ) be the class corresponding under Poincaré duality on S 1 to μ 2 − 1 φ μ 1 . The monodromy acts through g ↦ ( g , φ g φ − 1 ) , so Φ W is a flat global section; it is of Hodge type at every point, as is every element of theadmissible span
A = Q 1 ⊗ 1 + ( N ⊗ 1 + 1 ⊗ N ) + ( Q pt ⊗ 1 + Q 1 ⊗ pt + Q Φ W + N ⊗ N ) + ( pt ⊗ N + N ⊗ pt ) + Q pt ⊗ pt ,
where N stands for μ i − 1 ( N ) . We write γ ∈ A as γ = r + ( D 1 ⊗ 1 + 1 ⊗ D 2 ) + ( α pt ⊗ 1 + β 1 ⊗ pt + c Φ W + Ξ ) + ( pt ⊗ D 3 + D 4 ⊗ pt ) + d pt ⊗ pt and call c itsRM coefficient.
On a K3 surface S with period w the group H 1 ( S , T S ) acts on H * ( S ) by contraction. The action is a derivation of the cup product killing H 0 and H 4 , so on H 2 it is skew for q; with a = κ ⌟ w ∈ H 1 , 1 ( S ) it is
κ ⌟ x = a q ( w ¯ , x ) − w ¯ q ( a , x ) q ( w ¯ , w ) ,
and κ ↦ a is the isomorphism H 1 ( T S ) ≅ H 1 , 1 ( S ) under which a is the derivative of the period.
Theorem 19
(The Hodge locus is the Hecke graph). In Setting 3, let ( S 1 , S 2 ) be marked K3 surfaces with periods w 1 , w 2 ∈ Λ C and Φ the class of μ 2 − 1 φ μ 1 in H 4 ( S 1 × S 2 , Q ) .
(i)
Φ is a Hodge class if and only if w 1 ∈ L C and [ w 2 ] = [ φ w 1 ] .
(ii)
At such a point the first-order Hodge locus of Φ in H 1 ( T S 1 ) ⊕ H 1 ( T S 2 ) has dimension 2 m − 2 and is the tangent space of the graph of g φ . The Hodge locus of Φ in the forty-dimensional germ Def ( S 1 × S 2 ) is the germ of the graph of g φ , smooth of codimension 42 − 2 m .
(iii)
D W is the fixed locus of g φ . At [ w ] ∈ D W the map μ 2 − 1 μ 1 is a rational Hodge isometry, algebraic by Buskin’s theorem [26], and its composite with Φ differs from the real multiplication by e on S 1 × S 1 by an element of NS ⊗ NS .
(iv)
If Φ W is algebraic on a nonempty open subset of a connected component of B , it is algebraic on every fibre of B , and the real multiplication by e is algebraic on S × S for every K3 surface S with period in the image of D W in B , which is the complement of a locally finite union of hyperplane sections of D W .
Proof. 
(i) By Künneth and Poincaré duality, H 2 ( S 1 ) ⊗ H 2 ( S 2 ) = Hom ( H 2 ( S 1 ) , H 2 ( S 2 ) ) as Hodge structures up to a Tate twist, so Φ is Hodge if and only if φ maps H 1 2 , 0 into H 2 2 , 0 and F 1 1 = w 1 ⊥ into F 2 1 = w 2 ⊥ ; as φ is self-adjoint the second condition is φ w 2 ∈ C w 1 . If φ w 1 = 0 then w 1 ∈ N C , which is impossible because Re w 1 and Im w 1 span a positive plane and N R has one positive direction. So w 2 ∈ C φ w 1 ⊆ L C , and φ w 2 ∈ C φ 2 w 1 = C e w 1 L , where w 1 = w 1 L + w 1 N ; this is proportional to w 1 only if w 1 N = 0 . The converse holds because φ ( φ w 1 ) = e w 1 .
(ii) Normalise w 2 = φ w 1 , so q 2 = q ( w 2 , w ¯ 2 ) = e q 1 , q 1 = q ( w 1 , w ¯ 1 ) , and φ maps H 1 , 1 ( S 1 ) ∩ L C onto H 1 , 1 ( S 2 ) ∩ L C . With dual bases, Φ = 1 q 1 ( w ¯ 1 ⊗ φ w 1 + w 1 ⊗ φ w ¯ 1 ) + ∑ k e k * ⊗ φ ( e k ) , where ( e k ) is a basis of H 1 , 1 ( S 1 ) and ( e k * ) its q-dual basis. For κ = ( κ 1 , κ 2 ) with a i = κ i ⌟ w i , (6) and ∑ k q ( a 1 , e k * ) e k = a 1 give
κ ⌟ Φ = 1 q 1 a 1 − φ ( a 2 ) e ⊗ w ¯ 2 + 1 q 1 w ¯ 1 ⊗ a 2 − φ ( a 1 ) ,
the two terms lying in H 1 , 1 ⊗ H 0 , 2 and H 0 , 2 ⊗ H 1 , 1 . It vanishes if and only if a 2 = φ ( a 1 ) and a 1 = φ 2 ( a 1 ) / e , that is a 1 ∈ L C ∩ H 1 , 1 ( S 1 ) and a 2 = φ a 1 : a space of dimension 2 m − 2 , which is the tangent space at ( [ w 1 ] , [ φ w 1 ] ) of the graph { ( [ w ] , [ φ w ] ) } of g φ , the derivative of the period being a. The Hodge locus is an analytic germ containing the smooth ( 2 m − 2 ) -dimensional germ of the graph and with Zariski tangent space of the same dimension, so it is that germ; Def ( S 1 × S 2 ) = H 1 ( T S 1 ) ⊕ H 1 ( T S 2 ) has dimension 40, since H 0 ( T S ) = H 1 ( O S ) = 0 .
(iii) If φ w = λ w then λ ≠ 0 by the argument of (i), w ∈ L C , λ = ± e , and w lies in an eigenspace; the − e -eigenspace W s 2 ⊗ C is negative definite, so w ∈ W s 1 ⊗ C and [ w ] ∈ D W , and conversely. Then S 1 and S 2 have the same period line, so μ 2 − 1 μ 1 is a Hodge isometry, algebraic by [26, Thm. 1.1]. Its composite with Φ is the class of μ 1 − 1 φ μ 1 , which is e on T ( S 1 ) Q ⊆ L Q and maps NS ( S 1 ) Q to itself; the real multiplication is e on T and 0 on NS , and the difference is a Hodge class in NS ⊗ NS , a combination of products of divisors.
(iv) The connected components of B are smooth and quasi-projective and the family is smooth and projective, so by Lemma 2 the algebraic locus of the flat class Φ W is all of the component as soon as it has an interior point. The two components of D L are exchanged by complex conjugation, which carries a marked K3 surface with period w to its conjugate variety with period w ¯ and algebraic classes to algebraic classes, and fixes Φ W because φ is rational; so Φ W is algebraic on every fibre of B . At a point of D W in B we have S 2 ≅ S 1 , and (iii) gives the real multiplication. A hyperplane δ ⊥ with δ 2 = − 2 and δ · h = 0 contains D W only if the L-component x of δ is orthogonal to W s 1 , that is x ∈ W s 2 ; then φ x = − e x with x and φ x rational, so x = 0 and δ ∈ N , which the choice of h excludes. The hyperplane ( φ δ ) ⊥ contains D W only if φ δ = φ x is orthogonal to W s 1 , which again forces x = 0 .    □
So real multiplication on the ( m − 2 ) -dimensional family D W follows from the algebraicity of the multiplier-e Hecke correspondence on the ( 2 m − 2 ) -dimensional family of N-polarised K3 surfaces; for m = 10 the latter have Picard number two and transcendental rank twenty. On a dense set of the Hecke family the class is algebraic.
Corollary 7
(Algebraic points of the Hecke family, after Buskin). The set of [ w ] ∈ D L such that T ( S w ) Q is a positive definite rational plane P ⊂ L Q of discriminant e modulo squares is dense in D L , and at each such point Φ W is algebraic on S 1 × S 2 .
Proof. 
Density. L Q is indefinite of rank 2 m ≥ 6 , of signature ( 2 , 2 m − 2 ) . Let p 1 , p 2 be an orthonormal basis of a positive plane of L R . Rational positive vectors are dense among the positive directions, so choose v ∈ L Q with q ( v ) > 0 close to a multiple of p 1 . The space v ⊥ ⊂ L Q is indefinite of rank 2 m − 1 ≥ 5 , hence isotropic [108, Ch. IV, §3.2] and universal, so the quadric { x ∈ v ⊥ : q ( x ) = e q ( v ) } has a rational point; projection from it makes the quadric rational over Q , so its rational points are dense in its real points, which meet every positive direction of v R ⊥ . Choose x on it close to a multiple of the projection of p 2 . Then P = 〈 v , x 〉 is positive definite with Gram matrix diag ( q ( v ) , e q ( v ) ) , of discriminant e q ( v ) 2 ≡ e , close to 〈 p 1 , p 2 〉 ; one of the isotropic lines of P C is a point [ w ] ∈ D L close to [ p 1 + i p 2 ] , and T ( S w ) Q = P because the rational lines of P are anisotropic.
Algebraicity. T ( S 1 ) Q = P has rank two, so End Hdg T ( S 1 ) Q = K = Q ( − e ) , and T ( S 2 ) Q = φ ( P ) . The element k = 1 / − e ∈ K acts on T ( S 1 ) Q as a similitude of multiplier N K / Q ( k ) = 1 / e , so ψ = φ ∘ k : T ( S 1 ) Q → T ( S 2 ) Q is a rational Hodge isometry. The complements NS ( S i ) Q of P and φ ( P ) in Λ Q are isometric by Witt cancellation, so ψ extends to a Hodge isometry H 2 ( S 1 , Q ) → H 2 ( S 2 , Q ) , algebraic by [26]. The field K is algebraic on S 1 × S 1 by the argument in the proof of Theorem 13(iii), the projector onto T ( S 1 ) Q is algebraic, so φ | T = ψ ∘ k − 1 is algebraic, and the rest of Φ W is a Hodge class in NS ( S 1 ) ⊗ NS ( S 2 ) .    □
The surfaces of Corollary 7 are singular K3 surfaces, and the class has a representative built from Shioda–Inose structures.
Corollary 8
(An explicit representative, after Morrison and Shioda–Mitani). At a point of Corollary 7, let K = Q ( − e ) . Each S i has a Shioda–Inose structure [81, Thm. 6.3, Cor. 6.4]: a rational map π i : S i ⤏ Km ( A i ) of degree two and the quotient q i : A i → Km ( A i ) , with Hodge isometries π i * : T ( S i ) ( 2 ) → T ( Km A i ) and q i * : T ( A i ) ( 2 ) → T ( Km A i )  [81, Thm. 5.7]. The abelian surface A i is a product E i × E i ′ of elliptic curves with complex multiplication by orders of K [111]. For an isogeny f = f 1 × f 2 : A 1 → A 2 and λ ∈ End ( E 1 ) , the correspondence
Θ λ = Γ π 2 t ∘ Γ q 2 ∘ Γ f ∘ Γ λ × 1 ∘ Γ q 1 t ∘ Γ π 1
acts on T ( S 1 ) Q as a Hodge similitude onto T ( S 2 ) Q of multiplier 16 N ( λ ) deg f , and for suitable λ it equals κ φ | T with κ ∈ Q × . Then
Φ W = κ − 1 [ Θ λ ] 22 − κ − 1 [ Θ λ ] NS ⊗ NS + ( Φ W ) NS ⊗ NS ,
where [ ] 22 is the Künneth component in H 2 ⊗ H 2 ; so Φ W is a rational combination of graphs of rational maps, their transposes and products of divisors, obtained without Buskin’s theorem or the Kuga–Satake construction.
Proof. 
Singular K3 surfaces have Shioda–Inose structures by [81, Cor. 6.4]. A i has Picard number four, so it is isomorphic to a product of two isogenous elliptic curves with complex multiplication [111], by orders of the field of T ( A i ) Q ≅ T ( S i ) Q ( 2 ) ; this field is K for i = 1 , 2 , since φ ( P ) has discriminant e 2 · e . Elliptic curves with complex multiplication by orders of the same field are isogenous, which gives f. On transcendental lattices π 1 * , q 1 * , q 2 * and π 2 * = 2 ( π 2 * ) − 1 have multiplier 2, f * has multiplier deg f , and ( λ × 1 ) * acts on T ( E 1 × E 1 ′ ) ⊆ H 1 ( E 1 ) ⊗ H 1 ( E 1 ′ ) through λ , with multiplier N ( λ ) ; each is a morphism of Hodge structures, which gives the multiplier. The Hodge maps T ( S 1 ) Q → T ( S 2 ) Q form a one-dimensional K-vector space containing φ | T , with K = End Hdg T ( S 1 ) Q acting by precomposition. Transported to T ( A 1 ) Q by the Hodge isomorphism 1 2 q 1 * π 1 * , this action is that of ( λ × 1 ) * up to complex conjugation, both K-structures on the rank-two space being determined by their action on H 2 , 0 . So Θ λ = Θ 1 ∘ λ ′ with λ ′ ∈ { λ , λ ¯ } , and Θ 1 = φ | T ∘ k 0 for some k 0 ∈ K × ; choosing λ in the order End ( E 1 ) and in Q k 0 − 1 , or in Q k ¯ 0 − 1 , gives Θ λ = κ φ | T . The Künneth projectors of a K3 surface are algebraic, so [ Θ λ ] 22 is algebraic; as a Hodge map H 2 ( S 1 ) → H 2 ( S 2 ) it maps T to T and NS to NS , because T ( S i ) Q is irreducible with h 2 , 0 = 1 and contains no rational class of type ( 1 , 1 ) . Its NS ⊗ NS part and that of Φ W are Hodge classes in NS ⊗ NS , hence products of divisors.    □
No Fourier–Mukai kernel produces Φ W .
Corollary 9
(Kernels of equivalences, after Mukai). At a point of B :
(i)
no kernel P of an equivalence D b ( S 1 ) → D b ( S 2 ) has ch ( P ) ∈ A with RM coefficient c ≠ 0 ;
(ii)
if F = P 1 ⊕ P 2 with P 1 , P 2 kernels of equivalences, then dim Ext 2 ( F , F ) ≥ 44 .
Proof. 
(i) The cohomological transform of P is a Hodge isometry of Mukai lattices H ˜ ( S 1 , Q ) → H ˜ ( S 2 , Q )  [85], given by the Mukai vector ch ( P ) td ; its restriction to H 2 ⊗ H 2 is that of ch ( P ) . For ch ( P ) ∈ A the transform preserves L Q , on which it is ± c φ , a similitude of multiplier c 2 e ≠ 1 since e is not a rational square. (ii) Ext 2 ( P i , P i ) ≅ H H 2 ( S 1 ) has dimension h 0 ( ⋀ 2 T ) + h 1 ( T ) + h 2 ( O ) = 22 for each i.    □
The semiregularity count on Y = S 1 × S 2 differs from the abelian one of Theorem 20 (Remark 5). Write H ˜ ( S ) = H 0 ⊕ H 2 ⊕ H 4 with the Mukai pairing 〈 ( r , D , s ) , ( r ′ , D ′ , s ′ ) 〉 = D D ′ − r s ′ − r ′ s and U = H 0 ⊕ H 4 , so that H ˜ ( S i , Q ) = L Q ⊕ ( U ⊕ N ) Q at every point of B . For a perfect complex F on Y let v = ch ( F ) td Y , td Y = ( 1 + pt ) ⊗ ( 1 + pt ) , and φ ˜ v : H ˜ ( S 1 ) → H ˜ ( S 2 ) the map obtained by contracting the first factor of v with the Mukai pairing. If ch ( F ) ∈ A with RM coefficient c, then φ ˜ v preserves the decomposition, its L-block is c φ and its ( U ⊕ N ) -block is a map M; let M * be its Mukai adjoint, μ = c 2 e and V v = ker ( M * M − μ ) ⊆ ( U ⊕ N ) C . In the basis ( 1 , pt ) of U write v’s U ⊗ U coefficients as ρ 0 (of 1 ⊗ 1 ), a, b (of pt ⊗ 1 , 1 ⊗ pt ) and δ (of pt ⊗ pt ); so ρ 0 = r , a = α + r , b = β + r , δ = d + α + β + r .
Setting 4
(Hochschild input). For a perfect complex F on a smooth projective variety Y let χ F : H H 2 ( Y ) → Ext 2 ( F , F ) , ξ ↦ ξ ∘ id F , and let ι F : Ext * ( F , F ) → H H * ( Y ) be the boundary–bulk map, with ι F ( id F ) = ch H H ( F )  [29, Sect. 6, Prop. 11]. Let I K : H H 2 ( Y ) → H T 2 ( Y ) and I K : H H * ( Y ) → H * ( Y , C ) be the Hochschild–Kostant–Rosenberg isomorphisms twisted by td Y 1 / 2 . We say that Y satisfies (HH) if for every perfect complex F
(a)
I K ι F ( χ F ( ξ ) ) = I K ( ξ ) ⌟ v ( F ) for every ξ ∈ H H 2 ( Y ) , and
(b)
for κ ∈ H 1 ( Y , T Y ) the obstruction At ( F ) ∘ κ to deforming F to first order along κ equals χ F ( ( I K ) − 1 κ ) ,
and that Y satisfies (HH′) if moreover ι F agrees on Ext 2 ( F , F ) with the semiregularity map σ F of Buchweitz and Flenner up to an isomorphism of the target. The ingredients of (a) are the module identity ι F ( ξ ∘ ϕ ) = ξ · ι F ( ϕ ) , which follows from the definitions of [29]; the compatibility I K ( ξ · x ) = I K ( ξ ) ⌟ I K ( x ) , conjectured by Căldăraru and proved by Calaque, Rossi and Van den Bergh [31]; and I K ( ch H H F ) = v ( F )  [28]. The same compatibility is used in [56, proof of Lemma 4.2]. Part (b) uses c 1 ( Y ) = 0 , so that td Y 1 / 2 has no term of degree two. We have not checked (HH′) in the literature, and we use (HH) and (HH′) only as stated hypotheses.
Proposition 16
(The semiregularity count on a product of two K3 surfaces). Let F be a perfect complex on Y = S 1 × S 2 at a point of B with ch ( F ) ∈ A and RM coefficient c ≠ 0 , and v = v ( F ) . Let R ( v ) be the rank of H T 2 ( Y ) → H * ( Y , C ) , η ↦ η ⌟ v , on the forty-four-dimensional space H T 2 ( Y ) , and r ( v ) its rank on H 2 ( O Y ) ⊕ H 1 ( T Y ) .
(a)
R ( v ) = 46 − 2 m − dim V v and r ( v ) = 44 − 2 m − dim ( V v ∩ pt ⊥ ∩ ( M * pt ) ⊥ ) . Hence 22 ≤ R ( v ) ≤ 46 − 2 m , R ( v ) − r ( v ) ∈ { 0 , 1 , 2 } , and the first-order Hodge locus of v in H 1 ( T Y ) has dimension 2 m − 2 + dim ( V v ∩ N C ∩ ( M * 1 ) ⊥ ∩ ( M * pt ) ⊥ ) . If Ξ = 0 and D 1 = ⋯ = D 4 = 0 , the eigenvalues of M * M on U are ( a b ± ρ 0 δ ) 2 , and V v = 0 , R = 46 − 2 m = r + 2 , unless μ is one of them; for ρ 0 = 0 , a b = μ , a δ ≠ 0 one has R = 45 − 2 m , r = 43 − 2 m , and for a = b = 0 , ρ 0 δ = μ one has R = r = 44 − 2 m .
(b)
If Y satisfies (HH), then dim Ext 2 ( F , F ) ≥ R ( v ) .
(c)
If Y satisfies (HH) and dim Ext 2 ( F , F ) = R ( v ) , then ι F is injective on Ext 2 ( F , F ) and F extends to first order along every first-order deformation of Y tangent to the Hodge locus of v, in particular along B . If moreover Y satisfies (HH′), then F is semiregular, Φ W is algebraic on every fibre of B , and the real multiplication by e is algebraic on S × S for every S as in Theorem 19(iv).
(d)
For the structure sheaf of the diagonal of S × S , where φ ˜ v is an isometry, r = 21 and R = 22 = dim Ext 2 ( O Δ , O Δ ) ; the bivector part of H T 2 is needed already for Fourier–Mukai kernels.
Proof. 
(a) By Künneth, H T 2 ( Y ) = H T 2 ( S 1 ) ⊕ H T 2 ( S 2 ) , because h 1 ( O S ) = h 0 ( T S ) = 0 , acting by η 1 ⊗ 1 + 1 ⊗ η 2 . On a K3 surface with period s, for a ∈ s ⊥ ∩ s ¯ ⊥ = C 1 ⊕ H 1 , 1 ⊕ C pt put
ξ a ( x ) = a 〈 s ¯ , x 〉 − s ¯ 〈 a , x 〉 〈 s ¯ , s 〉 .
Then ξ 1 is contraction with the bivector dual to s (it maps s ↦ 1 and pt ↦ s ¯ / 〈 s ¯ , s 〉 and kills 1, s ¯ and H 1 , 1 ); for a ∈ H 1 , 1 , ξ a is (6); and ξ q pt , q = 〈 s ¯ , s 〉 , is cup product with s ¯ . So H T 2 ( S ) = { ξ a } , the map a ↦ ξ a is bijective, and every ξ a is skew for the Mukai pairing. For skew operators, ( ξ a 1 ⊗ 1 + 1 ⊗ ξ a 2 ) v = 0 if and only if ξ a 2 φ ˜ = φ ˜ ξ a 1 , φ ˜ = φ ˜ v . Normalise s 2 = φ ˜ s 1 = c φ w 1 ; then s ¯ 2 = φ ˜ s ¯ 1 and 〈 φ ˜ x , φ ˜ y 〉 = μ 〈 x , y 〉 for x , y ∈ L C , and
ξ a 2 φ ˜ ( x ) = a 2 〈 s ¯ 1 , x 〉 − s ¯ 2 〈 φ ˜ * a 2 / μ , x 〉 〈 s ¯ 1 , s 1 〉 , φ ˜ ξ a 1 ( x ) = φ ˜ a 1 〈 s ¯ 1 , x 〉 − s ¯ 2 〈 a 1 , x 〉 〈 s ¯ 1 , s 1 〉 .
Equality for all x says ( a 2 − φ ˜ a 1 ) 〈 s ¯ 1 , x 〉 = s ¯ 2 〈 φ ˜ * a 2 / μ − a 1 , x 〉 . The vector a 2 − φ ˜ a 1 is orthogonal to s 2 while 〈 s ¯ 2 , s 2 〉 ≠ 0 , so both sides vanish: a 2 = φ ˜ a 1 and φ ˜ * φ ˜ a 1 = μ a 1 . Conversely these give equality, with a 2 ∈ s 2 ⊥ ∩ s ¯ 2 ⊥ . Since φ ˜ * φ ˜ = μ on L and φ ˜ is block diagonal, the kernel is ( L C ∩ s 1 ⊥ ∩ s ¯ 1 ⊥ ) ⊕ V v , of dimension 2 m − 2 + dim V v , and R = 44 − ( 2 m − 2 ) − dim V v . Removing the bivectors means that a 1 and a 2 have no component along 1, that is a i ⊥ pt ; for a 2 = φ ˜ a 1 this is 〈 a 1 , M * pt 〉 = 0 , automatic on the L-part, which gives r. The commutative part is a i ∈ H 1 , 1 , that is a i ⊥ 1 , pt , which on the ( U ⊕ N ) -part means a 1 ∈ N C and 〈 a 1 , M * 1 〉 = 〈 a 1 , M * pt 〉 = 0 . As dim V v ≤ rk ( U ⊕ N ) = 24 − 2 m , R ≥ 22 ; two linear conditions cut V v in codimension at most two, so R − r ∈ { 0 , 1 , 2 } . In the basic shape M ( 1 ) = − ( a 1 + δ pt ) and M ( pt ) = − ( ρ 0 1 + b pt ) , and M * M on U is a b + ρ 0 δ 2 ρ 0 b 2 a δ a b + ρ 0 δ , with the stated eigenvalues, while M = 0 on N; for ρ 0 = 0 , a b = μ , a δ ≠ 0 the eigenspace is C pt , inside pt ⊥ ∩ ( M * pt ) ⊥ because M * pt = − a pt ; for a = b = 0 , ρ 0 δ = μ , V v = U C , V v ∩ pt ⊥ = C pt and 〈 pt , M * pt 〉 = ρ 0 ≠ 0 .
(b) By (HH)(a), { η ⌟ v } = I K ( ι F ( χ F ( H H 2 ( Y ) ) ) ) , of dimension at most dim Ext 2 ( F , F ) .
(c) Equality makes ι F ∘ χ F of rank dim Ext 2 , so ι F is injective on Ext 2 . By (HH)(b) the obstruction along κ is o = χ F ( ( I K ) − 1 κ ) , and I K ι F ( o ) = κ ⌟ v , which vanishes when κ is tangent to the Hodge locus of v; so o = 0 . The tangent space of B lies in that locus by (a). Under (HH′), σ F is injective, and by [11, Thm. 5.1] and [102] F deforms over the germ of the Hodge locus of ch ( F ) , which contains the germ of B ; the nearby fibres are projective, so the deformed complexes are algebraic, and Φ W = c − 1 [ ch 2 ] 22 − Ξ is algebraic near the point, the Künneth projectors being algebraic. Theorem 19(iv) finishes.
(d) For O Δ the map φ ˜ v is an isometry of H ˜ ( S ) preserving C pt (the cohomological transform of O Δ is the identity [85]), and the computation of (a) with φ ˜ * φ ˜ = 1 on all of H ˜ gives the kernel { ( ξ , φ ˜ ξ φ ˜ − 1 ) } , of dimension 22, and of dimension 21 on H 1 ( T ) ⊕ H 2 ( O ) ; so R = 22 and r = 21 . By Hochschild–Kostant–Rosenberg, Ext 2 ( O Δ , O Δ ) = H H 2 ( S ) has dimension h 0 ( ⋀ 2 T S ) + h 1 ( T S ) + h 2 ( O S ) = 22 .    □
So the target on S 1 × S 2 is R ( v ) , which is 46 − 2 m for the generic admissible shape: 26 for the maximal families m = 10 , 38 for m = 4 . Contracting into ch instead of v, as on an abelian variety, gives a different number for some admissible shapes (Appendix C.6). The smallest possible target, 22, is attained exactly when M is a similitude, and such shapes exist.
Proposition 17
(Similitude kernels).
(i)
There is a rational similitude M 0 of ( U ⊕ N ) Q of multiplier e. The map Φ ˜ = φ ⊕ M 0 is a Hodge similitude H ˜ ( S 1 , Q ) → H ˜ ( S 2 , Q ) of multiplier e at every point of B , flat over B , and its dual class divided by td Y lies in A with c = 1 .
(ii)
If M * M = μ on U ⊕ N , then R ( v ) = 22 , r ( v ) = 22 , or 21 when M * pt ∈ Q pt , and the commutative first-order Hodge locus of v is { a 1 ∈ H 1 , 1 ( S 1 ) : 〈 a 1 , n 1 〉 = 〈 a 1 , n 2 〉 = 0 } , where n 1 , n 2 are the N-components of M * pt and M * 1 ; it has dimension 18 when n 1 , n 2 are independent.
Proof. 
(i) H ˜ ( S , Q ) ≅ U Q 4 ⊕ E 8 ( − 1 ) Q 2 . The plane U ( e ) is isometric to U (basis e 1 , e 2 / e ). E 8 ( − 1 ) Q ≅ 〈 − 1 〉 8 by Hasse–Minkowski: same rank, discriminant and signature, both unimodular at odd primes, and the product formula fixes the invariant at 2. The form 〈 1 , 1 , 1 , 1 〉 is a two-fold Pfister form, hence round [68, Ch. X], and represents e by Lagrange’s four-square theorem, so 〈 − e 〉 4 ≅ 〈 − 1 〉 4 . Hence H ˜ ( S , Q ) ( e ) ≅ H ˜ ( S , Q ) , and also L Q ( e ) ≅ L Q via φ ; by Witt cancellation ( U ⊕ N ) Q ( e ) ≅ ( U ⊕ N ) Q , and an isometry between them is M 0 . The classes of U ⊕ N are of type ( 1 , 1 ) in the Mukai Hodge structure on both factors at every point, φ maps w 1 to w 2 , and the monodromy fixes U and N pointwise; so Φ ˜ is a flat Hodge similitude. Multiplication by ( 1 − pt ) ⊗ ( 1 − pt ) preserves ( U ⊕ N ) ⊗ ( U ⊕ N ) + Q Φ W , which is A written in Mukai form.
(ii) This is (a) of Proposition 16 with V v = ( U ⊕ N ) C : then dim V v = 24 − 2 m , the two conditions are independent unless M * pt is proportional to pt , and on the commutative part the L-component of a 1 is free and the N-component is subject to the two conditions, which gives ( 2 m − 2 ) + ( 22 − 2 m ) − 2 = 18 .    □
A semiregular similitude kernel would therefore make a multiplier-e Hodge similitude algebraic on an eighteen-dimensional family of K3 surfaces of transcendental rank twenty containing B ; the target 22 solves the Hecke problem of Theorem 19 in its largest form and is no shortcut. The Euler form constrains every candidate.
Lemma  13
(The RM coefficient is an integer). For a perfect complex F at a point of B with ch ( F ) ∈ A and c ≠ 0 , c ∈ Z ; so μ = c 2 e ≥ e ≥ 2 .
Proof. 
ch 2 = c 1 2 / 2 − c 2 is integral, since the Künneth components of c 1 2 are D 1 2 , D 2 2 , which are even, and 2 D 1 ⊗ D 2 . Its component in H 2 ⊗ H 2 is c Φ W + Ξ , an integral class, hence by unimodularity an integral map f : Λ → Λ . On L it is c φ , as Ξ kills L, with values in L Q ∩ Λ = L ; it is injective and multiplies the form by μ , so μ 2 m disc L = [ L : c φ L ] 2 disc L and μ m ∈ Z . So the rational number μ is an integer, and c 2 e ∈ Z with e squarefree forces c ∈ Z .    □
Proposition  18
(Euler-form constraints). Let F be as in Lemma 13.
(i)
χ ( F , F ) = 〈 v , v 〉 = 2 m μ + τ with τ = tr ( M * M ) = 2 ρ 0 δ + 2 a b − 2 q ( D 1 , D 1 + D 4 ) − 2 q ( D 2 , D 2 + D 3 ) + tr ( Ξ * Ξ ) .
(ii)
If Ext < 0 ( F , F ) = 0 , then dim Ext 2 ( F , F ) = 〈 v , v 〉 − 2 hom ( F , F ) + 2 ext 1 ( F , F ) ; if moreover dim Ext 2 = R ( v ) , then hom − ext 1 = ( 〈 v , v 〉 − R ( v ) ) / 2 . For a similitude kernel with dim Ext 2 = 22 this gives hom ( F , F ) = 12 μ − 11 + ext 1 ≥ 12 e − 11 ≥ 13 , so F is not simple.
(iii)
If F is a simple pure sheaf of dimension two with tr ( Ξ * Ξ ) ≥ 0 , then dim Ext 2 ( F , F ) ≥ 2 α β + 2 m μ + tr ( Ξ * Ξ ) − 2 ≥ 4 m with α , β ≥ 1 , and dim Ext 2 ( F , F ) ≤ 46 − 2 m forces m ( e + 1 ) ≤ 23 : m ≤ 7 for e = 2 , m ≤ 5 for e = 3 , m = 3 for e ∈ { 5 , 6 } , and never for e ≥ 7 .
Proof. 
(i) Hirzebruch–Riemann–Roch on Y gives χ ( F , G ) = ∫ Y v ( F ) ∨ v ( G ) , as td Y is self-dual. On a K3 surface ∫ x ∨ y = − 〈 x , y 〉 , and ( ) ∨ is multiplicative for the Künneth decomposition, so the signs cancel and χ = 〈 , 〉 ⊗ 〈 , 〉 (for O Y both sides are 4). Writing v = ∑ V k l b k ⊗ b l , 〈 v , v 〉 = tr ( φ ˜ v * φ ˜ v ) , which is μ on the 2 m -dimensional L-block plus τ ; expanding gives the formula: the coefficients of 1 ⊗ 1 , pt ⊗ 1 , 1 ⊗ pt , pt ⊗ pt in v are ρ 0 , a , b , δ , those of 1 ⊗ N and pt ⊗ N are D 2 and D 2 + D 3 , those of N ⊗ 1 and N ⊗ pt are D 1 and D 1 + D 4 , and 〈 1 , pt 〉 = − 1 . (ii) K Y = 0 , so Serre duality gives ext i = ext 4 − i and χ = 2 hom − 2 ext 1 + ext 2 . For a similitude kernel 〈 v , v 〉 = 2 m μ + ( 24 − 2 m ) μ = 24 μ , and Lemma 13 gives μ ≥ e ≥ 2 . (iii) For a pure sheaf of dimension two ch 0 = ch 1 = 0 , so ρ 0 = 0 , D 1 = D 2 = 0 , a = α , b = β , τ = 2 α β + tr ( Ξ * Ξ ) , and ch 2 = ∑ n i [ Z i ] with n i > 0 and surfaces Z i . Then α = ∑ n i deg ( Z i → S 2 ) ≥ 0 , and likewise β , the degree being 0 for a non-dominant projection. If a projection of Z i is not dominant, the transpose of [ Z i ] kills w 2 : either w 2 pulls back to 0 on Z i , which then maps to a curve, or its push-forward to S 1 is a class of type ( 2 , 0 ) supported on a curve. The transpose of c Φ W + Ξ maps w 2 to c e w 1 ≠ 0 , so some Z i dominates both factors and α , β ≥ 1 . With hom = 1 , (ii) and μ ≥ 2 give the bound, and 2 + 2 m μ − 2 ≤ 46 − 2 m with μ ≥ e gives m ( e + 1 ) ≤ 23 .    □
Proposition  19
(The graph of a Kummer isogeny). Let g : A 1 → A 2 be an isogeny of degree e of abelian surfaces. It induces a rational map g ¯ : Km ( A 1 ) ⤏ Km ( A 2 ) of degree e, whose indeterminacy consists of the k = 8 ( e − 1 ) images of the 16 e − 16 points of g − 1 ( A 2 [ 2 ] ) ∖ A 1 [ 2 ] ; its graph is Z ≅ Bl k Km ( A 1 ) , embedded in Y = Km ( A 1 ) × Km ( A 2 ) , acting on the image of H 2 ( A 1 , Q ) through g * , a similitude of multiplier e. Moreover Z · Z = 8 e + 16 and dim Ext 2 ( O Z , O Z ) = 8 e + 14 + 2 ext 1 ( O Z , O Z ) .
Proof. 
Near a point x ∉ A 1 [ 2 ] with g ( x ) ∈ A 2 [ 2 ] the map is locally C 2 → C 2 / ± 1 , which lifts to the minimal resolution after one blow-up of the origin, the exceptional curve mapping isomorphically onto the ( − 2 ) -curve; near A 1 [ 2 ] , g is a local isomorphism commuting with − 1 and lifts. So ( b , g ˜ ) : Z → Y , with b the blow-up, is injective and immersive, a closed embedding. Then c ( N Z ) = c ( T Y | Z ) / c ( T Z ) with c 1 ( T Y | Z ) = 0 , c 2 ( T Y | Z ) = 24 + 24 e (the degrees of the two projections are 1 and e), c 2 ( Z ) = 24 + k and K Z 2 = − k , so Z · Z = c 2 ( N Z ) = 24 + 24 e − c 2 ( Z ) + K Z 2 = 24 e − 2 k = 8 e + 16 . By Grothendieck–Riemann–Roch only [ Z ] 2 survives in χ ( O Z , O Z ) , and hom ( O Z , O Z ) = 1 , so Proposition 18(ii) gives the last formula.    □
Whether such a graph is admissible for some ( W , L ) is not decided here; for m = 3 the target 46 − 2 m = 40 exceeds 8 e + 14 for e = 2 , 3 , so Proposition 18 does not exclude it.

5.10. Where This Stops

The frontier case of (F3′) in Theorem A23 and Remark A6 is a K3 surface S λ of Picard number 13 with T ( S λ ) Q ≅ T , the nine-dimensional representation of a Mumford fourfold, and E = F a totally real cubic field, which may have Galois group S 3 . Theorem 14 settles it for the two cyclic fields F = Q ( ζ 7 ) + and Q ( ζ 9 ) + with the algebra D ∞ ; the frontier is S λ for every other pair ( F , D ) : for these two fields with the algebras ramified at two finite places, for the cyclic cubic fields of conductor at least 13, and for every field that is not cyclic. There, by Proposition 9(v), no composite of similarities reaches F,9 so Theorem 10, the Fricke isogenies and every correspondence of the kinds listed are blind to it; by Proposition 9(iv) a correspondence that does reach it has an irreducible component whose resolution has p g ≥ 2 , transcendental rank at least 18 and b 2 ≥ 31 ; and by Proposition 10(iii) that component is not the graph of an element of a finite group acting on a cover of S λ when F is not cyclic. When F is cyclic the dihedral families are the natural source, as van Geemen observed for Q ( ζ 9 ) +  [118, Rem. 7.9]; Theorem 14 is the statement that for the two fields of conductor 7 and 9 and the algebra D ∞ the source actually reaches S λ , and Proposition 11 says that it reaches nothing else. The only correspondences of this size that we know to act on a transcendental space of K3 type are the dihedral ones, which give abelian fields, and the Kuga–Satake correspondence into the cohomology of an abelian variety of dimension 2 7 , whose algebraicity is open [34,117]; in Section 6 we replace its abelian variety by a sixfold and show what the replacement does and does not give. Within quadratic fields the method reaches d for d ∈ { 2 , 3 , 6 } in families, 2 on every surface of transcendental rank at most ten (Theorem 12), with Proposition 8 marking where it stops in higher rank, and the Fricke isogenies give complex multiplication by Q ( − 5 ) and Q ( − 7 ) on one-dimensional families, where Theorem 13 settles the Hodge conjecture on every power; real multiplication by 5 and 7 is algebraic on the two surfaces of [121] and on no family known to us (Remark 3); for Q ( 10 ) , Q ( 14 ) and the other composite discriminants no family is known on which the chain of Theorem 10 can be started. On the powers beyond the square, Proposition 13 reduces the Hodge conjecture for every surface of this section with totally real endomorphism field to its exceptional classes, Proposition 14 settles it with (F3′) for the surfaces of Kummer type over Q , Theorem 13 settles it for every member of the Fricke families, and Remark 2 records what is left.
Proposition  20
(Chains of symplectic quotients reach exactly 2 , 3 , 6 ). Let S be a projective K3 surface whose endomorphism field E is totally real and different from Q , and let S = S 0 , ⋯ , S r be a chain as in Theorem 10 in which every rational map is the quotient by a finite group of symplectic automorphisms. Then no degree n i is divisible by 5 or by 7, every n i divides a power of 24, and the multiplier ∏ n i ± 1 lies in { 1 , 2 , 3 , 6 } modulo squares. In particular no such chain makes 10 , 14 , 15 or 21 algebraic, and the surfaces with real multiplication by these fields are not reached by Theorem 10 in this form.
Proof. 
Along the chain the transcendental spaces are isomorphic up to scaling of the form, by Lemma 4 and Lemma 5, so every S i has the same transcendental rank t and the same endomorphism field E as S. By Zarhin [126, Thms. 2.2.1 and 2.3.1] the Hodge group of T ( S ) Q is the group of E-linear special isometries of the E-quadratic space W with T ( S ) Q = Tr E / Q W , of dimension m = t / [ E : Q ] over E. If m ≤ 2 this group is commutative (trivial for m = 1 , a torus of dimension [ E : Q ] for m = 2 ). Then every g ∈ Hdg ( Q ) commutes with the Hodge group and is an endomorphism of the Hodge structure, so g ∈ E . The elements of the totally real field E are self-adjoint for the form on T ( S ) Q and g is an isometry, so g = g * = g − 1 and g 2 = 1 . As Q -points are Zariski dense in a connected linear algebraic group over Q , the Hodge group would be trivial, which is impossible because h 2 , 0 ( S ) = 1 . So m ≥ 3 and t ≥ 3 [ E : Q ] ≥ 6 .
Let G be a finite group of symplectic automorphisms of a K3 surface X and Ω G ⊂ NS ( X ) its coinvariant lattice, negative definite because G fixes the transcendental lattice and an ample class. If 5 divides | G | , then G contains an element of order five, whose coinvariant lattice has rank 16 [94, § 5] and lies in Ω G , so rk Ω G ≥ 16 ; if 7 divides | G | the same argument with the rank 18 of an element of order seven gives rk Ω G ≥ 18 . The transcendental lattice T ( X ) lies in Ω G ⊥ ⊂ H 2 ( X , Z ) , which has rank 22 − rk Ω G ≤ 6 and signature ( 3 , 19 − rk Ω G ) , and T ( X ) has signature ( 2 , rk T ( X ) − 2 ) ; so rk T ( X ) ≤ 5 when 5 divides | G | and rk T ( X ) ≤ 3 when 7 does. Every symplectic automorphism of finite order has order at most eight [94], so if some n i = | G i | is not of the form 2 a 3 b then 5 or 7 divides it, the source of that quotient has transcendental rank at most five, and so has S; this contradicts t ≥ 6 . Hence every n i is of the form 2 a 3 b , and the multiplier is 2 a 3 b modulo squares, that is 1, 2, 3 or 6. A multiplier n makes only e with e 2 ≡ n algebraic, which gives the list.    □
Remark 4.
The proposition settles the lattice question of whether a chain S → Y ← S ′ through a symplectic involution and a symplectic automorphism of order five, of multiplier 2 / 5 ≡ 10 , can start on a surface with real multiplication by Q ( 10 ) : it cannot, since S ′ would have transcendental rank at most five and real multiplication by a real quadratic field needs rank at least six. The same bound shows that the surfaces of transcendental rank at most five with a symplectic automorphism of order five or seven have endomorphism field Q or a CM field; in the CM case the Hodge group is a torus and the chain closes only through the complex multiplication itself. Real multiplication by 10 , 14 or any field outside { Q ( 2 ) , Q ( 3 ) , Q ( 6 ) } therefore needs, in transcendental rank at least eight, a correspondence that is not a composite of symplectic quotients, Hodge isometries and fibrewise isogenies (Corollary 10 below), fibrewise isogenies of degree five or seven requiring Picard number at least sixteen (Proposition 21); in transcendental rank six with Witt index two the Kuga–Satake correspondence of Theorems 16 and 17 is one, and in higher rank one is known only on the curves of Theorem 18.
Fibrewise isogenies do not enlarge the list of fields of Proposition 8. By a fibrewise isogeny of degree N we mean a dominant rational map of K3 surfaces commuting with genus-one fibrations over P 1 and inducing on generic fibres a morphism of genus-one curves of degree N.
Proposition  21
(Fibrewise isogenies). Let S be a projective K3 surface whose endomorphism field is totally real, with rk T ( S ) ≥ 6 , and let S ⤏ S ′ or S ′ ⤏ S be a fibrewise isogeny of prime degree ℓ. Then ℓ ∈ { 2 , 3 , 5 , 7 , 13 } , and ρ ( S ) ≥ 9 if ℓ = 2 , ρ ( S ) ≥ 10 if ℓ = 3 , and ρ ( S ) ≥ 16 if ℓ ∈ { 5 , 7 , 13 } .
Proof. 
By Lemma 4 T ( S ) Q and T ( S ′ ) Q are similar, so we may take the map to leave S. The relative Jacobian J of the fibration of S has T ( J ) Q ≅ T ( S ) Q  [58, Ch. 11, Cor. 4.7], hence the same Picard number, and a morphism of torsors of degree ℓ induces an isogeny of the Jacobians of degree ℓ; so we may assume that the fibration has a section and that its generic fibre E over K = C ( t ) has a K-rational subgroup C of order ℓ.
If j ( E ) is constant, E becomes constant, equal to E 0 , after a finite Galois base change B → P 1 with group G, and T ( S ) Q is a sub-Hodge structure of ( H 1 ( E 0 ) ⊗ H 1 ( B ) ) G . If E 0 has complex multiplication by K ′ , then K ′ acts on this space by Hodge endomorphisms commuting with G and preserves T ( S ) Q , the smallest sub-Hodge structure containing H 2 , 0 , so the endomorphism field contains a CM field and is not totally real [126]. Otherwise G acts on H 1 ( E 0 ) through ± 1 , and h 2 , 0 ( S ) = 1 forces the part of H 1 ( B ) on which G acts through the same sign to have dimension 2, so rk T ( S ) ≤ 4 . So j is not constant, ( E , C ) defines a non-constant map g : P 1 → X 0 ( ℓ ) of some degree k with j = j ℓ ∘ g , and X 0 ( ℓ ) has genus 0, which for a prime ℓ means ℓ ∈ { 2 , 3 , 5 , 7 , 13 } .
For ℓ = 2 , C is a section of order two, translation by it is a symplectic involution and the isogeny is the quotient by it [120, §4]; so E 8 ( − 2 ) ⊆ NS ( S ) and ρ ≥ 9  [81]. Let ℓ be odd. By the Shioda–Tate formula [109, Cor. 6.13], ρ = 26 − μ − 2 a + rk MW , where μ and a are the numbers of multiplicative and additive fibres. The poles of j, of total order k ( ℓ + 1 ) , lie over the two cusps of X 0 ( ℓ ) ; there are P ≤ 2 k of them, with fibres I n or I n * , and we let s be the number of the latter. Every other singular fibre is additive. It has Euler number at least 6 unless it is of type II , III or IV , which requires j = 0 with order prime to 3 or j = 1728 with odd order, that is, a preimage under g of an elliptic point of X 0 ( ℓ ) of order 3 or 2, since j ℓ is ramified to order 3, resp. 2, at the other points over 0, resp. 1728. X 0 ( 3 ) has one elliptic point of order 3, X 0 ( 5 ) two of order 2, X 0 ( 7 ) two of order 3 and X 0 ( 13 ) two of each order. The Euler numbers add up to 24, so the a ′ = a − s additive fibres off the poles have Euler numbers summing to B = 24 − k ( ℓ + 1 ) − 6 s , and
ρ ≥ 26 − P − s − 2 a ′ + rk MW ≥ 26 − 2 k − s − 2 a ′ + rk MW .
For given ( ℓ , k , s ) , a ′ is largest when the cheapest fibres are used: type II (Euler number 2) or III (Euler number 3) at most k times per elliptic point, then I 0 * . Raising s by one lowers B by 6 and the bound for 2 a ′ by at least 2, so s = 0 is the worst case. With s = 0 and k ≤ 24 / ( ℓ + 1 ) this gives: for ℓ = 13 , k = 1 , a ′ ≤ 4 ( II , II , III , III ) and ρ ≥ 16 ; for ℓ = 7 , ( k , a ′ ) = ( 1 , 4 ) , ( 2 , 4 ) , ( 3 , 0 ) and ρ ≥ 16 , 14 , 20 ; for ℓ = 5 , ( k , a ′ ) = ( 1 , 4 ) , ( 2 , 4 ) , ( 3 , 2 ) , ( 4 , 0 ) and ρ ≥ 16 , 14 , 16 , 18 ; for ℓ = 3 , ( k , a ′ ) = ( 1 , 4 ) , ( 2 , 4 ) , ( 3 , 4 ) , ( 4 , 4 ) , ( 5 , 2 ) , ( 6 , 0 ) and ρ ≥ 16 , 14 , 12 , 10 , 12 , 14 . The two cases with bound 14 have four fibres of type II , resp. III , and 2 k = 4 multiplicative fibres; no fibre is starred, so the surface is the pull-back by g, without twist, of the rational elliptic surface R ℓ over X 0 ( ℓ ) with fibres I 1 , I 7 , II , II , resp. I 1 , I 5 , III , III , whose Mordell–Weil rank is 2, and pull-back multiplies the height pairing by k [109, Prop. 11.14]; so ρ ≥ 16 there too.    □
Corollary 10
(Chains of symplectic quotients and fibrewise isogenies). Let S be a projective K3 surface with endomorphism field Q ( e ) , e squarefree, and rk T ( S ) = 2 m ≥ 8 . Suppose a chain as in Theorem 10, whose rational maps are quotients by finite groups of symplectic automorphisms and fibrewise isogenies, makes e algebraic. Then e ∈ { 2 , 3 , 6 } and m ≤ 6 , and if e = 2 and m = 6 , then T ( S ) Q ( c ) embeds in V 2 for c = 1 or c = 3 .
Proof. 
Along the chain every S i has the same endomorphism field and transcendental rank as S. A fibrewise isogeny factors over K into multiplication by an integer and isogenies of prime degree, and the intermediate surfaces are K3 surfaces, since a dominant rational map and its dual preserve p g and q; a finite group of symplectic automorphisms of order divisible by a prime p contains one of order p, which requires p ≤ 7 and ρ ≥ 17 for p = 5 , 7 and ρ ≥ 13 for p = 3  [94]. So by Proposition 21 and ρ ≤ 14 , no step has degree divisible to an odd power by a prime p ≥ 5 , and ∏ n i ± 1 ≡ e forces e ∈ { 2 , 3 , 6 } . A prime factor 2 of e needs a step of degree 2 up to squares, hence a symplectic involution on some S i (for a fibrewise 2-isogeny by [120, §4]), so E 8 ( − 2 ) ⊆ NS ( S i ) , T ( S i ) Q embeds in V 2 and m ≤ 6 ; a prime factor 3 needs ρ ≥ 10 , so again m ≤ 6 . For e = 2 , T ( S i ) Q ≅ T ( S ) Q ( c ) with c ∈ { 1 , 2 , 3 , 6 } modulo squares, and T ( S ) Q ( 2 ) ≅ T ( S ) Q because 2 is a similitude of multiplier 2.    □
So the composites of symplectic quotients, fibrewise isogenies and Hodge isometries reach no real quadratic field other than Q ( 2 ) , Q ( 3 ) , Q ( 6 ) in transcendental rank at least eight, and nothing in rank at least fourteen; what they leave for these three fields is rank twelve for Q ( 2 ) through the embedding of T ( S ) Q ( 3 ) , and ranks ten and twelve, and rank eight outside V 3 , for Q ( 3 ) and Q ( 6 ) through fibrewise isogenies of degree three, which need ρ ≥ 10 .
For real multiplication of degree two the frontier has two parts. (a) Transcendental rank six with Witt index one: the families of Corollary 6, which exist for every real quadratic field; for e ∈ { 2 , 3 , 6 } the real multiplication is algebraic on them by Corollary 1, so this part concerns the other fields, such as e = 5 , 7 in Table 4. There C + ( T ) ≅ M 2 ( D ) with D a quaternion division algebra over K by Lemma 11(ii)(e); right multiplication by C + ( T ) on H 1 of the Kuga–Satake variety commutes with its Hodge structure [117], so that variety is isogenous to B 2 for an abelian eightfold B on which D op acts, outside every case in which the Kuga–Satake Hodge conjecture is known, and Markman’s theorems on Weil classes [78,79] give cycles on the Kuga–Satake variety but no cycle on S × KS ( S ) that would transport them to S. (b) Transcendental rank at least seven: the very general member of every family of Theorem 18 outside the cases of Proposition 8, so Q ( 2 ) from rank twelve on, Q ( 3 ) and Q ( 6 ) from rank eight on outside the lattice condition, and every other real quadratic field, and every family with real multiplication by a field of degree at least three; by Corollary 10 adding fibrewise isogenies to the symplectic quotients reaches no other real quadratic field and nothing from rank fourteen on. Theorem 18 puts a dense countable union of curves into the algebraic locus, and by Lemma 2 the locus is then all of D W or meagre; which one holds is open. By Theorem 19 it is all of D W as soon as Φ W is algebraic on one open subset of the Hecke family B , on which it is algebraic at a dense set of points (Corollaries 7 and 8). The open step is an object F at one point of B with ch ( F ) ∈ A , RM coefficient c ≠ 0 and dim Ext 2 ( F , F ) = R ( v ( F ) ) , which is 46 − 2 m for the generic shape (26 for m = 10 ) and 22 for similitude kernels; under the Hochschild inputs (HH) and (HH′) such an object suffices (Proposition 16(c)). By Corollary 9 and Proposition 18 it is not a kernel of an equivalence or a sum of two, not a simple similitude kernel, and, when m ( e + 1 ) > 23 , not a simple pure sheaf of dimension two with tr ( Ξ * Ξ ) ≥ 0 . In both parts the obstacle is the same: the Kuga–Satake Hodge conjecture is known for K3 surfaces only through varieties of generalised Kummer type, whose second Betti number is seven, and the hyper-Kähler sixfolds of OG6 type used by Floccari and Fu [39], with b 2 = 8 , and of K3[3] type used by van Geemen and Rapagnetta [119] are attached to abelian fourfolds of Weil type of discriminant one and give no Kuga–Satake statement for K3 surfaces of transcendental rank seven.

6. Constructions at (F2)

The first classes (F2) has to reach are the Weil classes of a very general abelian 2 n -fold of Weil type with n ≥ 3 , and the exceptional classes of a Mumford fourfold. We build an object at each.

6.1. The Polarised Semiregularity Criterion

Let A be a very general member of the polarised family D n , n of abelian 2 n -folds of Weil type for an imaginary quadratic field K, with polarisation class θ and Weil plane W ( A ) . Then Hdg * ( A ) = Q [ θ ] ⊕ W ( A ) and W ( A ) · θ = 0 Theorem A1. A class γ ∈ H ev ( A ) has the polarised shape if
γ = N ω + ∑ k = 0 2 n c k θ k , N ≠ 0 , 0 ≠ ω ∈ W ( A ) , c k ∈ Q .
Its Hankel rank ρ ∈ { 0 , 1 , 2 , 3 } is the rank of the ( 2 n − 1 ) × 3 Hankel matrix [ μ i + j ] 0 ≤ i ≤ 2 n − 2 , 0 ≤ j ≤ 2 , μ k = k ! c k (Setting 10); it is the number of distinct exponentials e a θ needed to write the polynomial part when that number is at most two, and 3 otherwise.10
Theorem  20
(The polarised criterion). Let n ≥ 3 and let E be a perfect complex on A with ch ( E ) = γ of polarised shape and Hankel rank ρ. Then
dim Ext 2 ( E , E ) ≥ r ( γ ) : = ( 4 + ρ ) n 2 − 2 n ,
and if equality holds the semiregularity map11 of E is injective, E deforms with A along the whole of D n , n , and the Weil classes are algebraic on every member of the family with full Hodge group. If some c k ≠ 0 with 1 ≤ k ≤ 2 n − 1 , the first-order Hodge locus of γ in H 1 ( A , T A ) is the tangent space of D n , n , of dimension n 2 . A perfect complex whose Chern character is a nonzero multiple of a Weil class plus classes in degrees 0 and 4 n never has injective semiregularity map (Theorem A7), so the polynomial part is forced.
The number r ( γ ) is the rank of the contraction H T 2 ( A ) → H * ( A ) , ξ ↦ ξ ⌟ γ , on the Hochschild cohomology H T 2 = ⨁ p + q = 2 H q ( A , ⋀ p T A ) , of dimension 4 n 2 ; its kernel is the annihilator Ann ( γ ) , which meets H 1 ( A , T A ) in the first-order Hodge locus. The semiregularity map σ E of a perfect complex composes with the characteristic map ev E : H T 2 → Ext 2 ( E , E ) , ξ ↦ ξ ⌟ At ( E ) , to the contraction into ch ( E ) ; this is how the lower bound arises, and it is the form in which the criterion is used below. Figure 10 shows r ( γ ) ; the two values that matter below are r = 57 at n = 3 , ρ = 3 and r = 104 at n = 4 , ρ = 3 . On the products E 0 6 and E 0 8 of copies of the elliptic curve E 0 = C / Z [ i ] no convolution of the line bundles of Proposition A4 meets the criterion, in any order, with any shifts and any placement of the multiples of the polarisation: dim Ext 2 ≥ r ( γ ) + 1 (Theorem A9), and dim Ext 2 ≥ 69 > 57 at E 0 6 (Theorem A12).
Remark  5
(The Todd class). The identity σ E ∘ ev E = ( · ) ⌟ ch ( E ) behind Theorem 20 holds as stated because the Todd class of an abelian variety is 1. On a general smooth projective variety Y the Hochschild–Kostant–Rosenberg comparison has to be twisted by td Y 1 / 2 , and the boundary–bulk map of a perfect complex sends ξ · id to the contraction of the twisted polyvector field ξ into the Mukai vector v = ch · td Y 1 / 2  [29, Sect. 6], [28,31]. On a product of two K3 surfaces this changes the count (Proposition 16).
The criterion has an equivariant form. A finite subgroup G of A × A ∨ acts on D perf ( A ) by g = t a * ( − ) ⊗ P ξ up to the central extension of Mukai [84]; if G is isotropic for its commutator pairing, the restricted extension of G by C × is abelian and splits because C × is divisible, so G acts genuinely. Perry calls E G-semiregular if it carries a G-equivariant structure and σ E is injective on Ext 2 ( E , E ) G , and proves that this suffices for the conclusions of semiregularity on abelian varieties [98, Thm. 1.2, Lemma 2.7].
Corollary  11
(Equivariant criterion). Let X → S be a family of polarised abelian 2 n -folds of Weil type over a connected base along which θ and the Weil plane stay of Hodge type, 0 ∈ S , and G ⊆ X 0 × X 0 ∨ a finite group acting on D perf ( X 0 ) . Let E be a G-equivariant perfect complex on X 0 with Ext < 0 ( E , E ) = 0 and ch ( E ) = γ of polarised shape, and r 0 the rank of H T 2 ( X 0 ) → H * ( X 0 ) , ξ ↦ ξ ⌟ γ . Then dim Ext 2 ( E , E ) G ≥ r 0 , and if equality holds the Weil classes are algebraic on every member of the family. At a member with Hodge group SU ( V , H ) , r 0 = r ( γ ) .
Proof. 
For g = t a * ( − ) ⊗ P ξ the Atiyah class satisfies At ( g E ) = g * At ( E ) , since translations commute with the formation of Atiyah classes and the Atiyah class of P ξ ∈ Pic 0 is c 1 ( P ξ ) = 0 ; translations act trivially on H T 2 ( X 0 ) , which consists of translation-invariant polyvector fields. Transporting along the equivariant structure g E ≅ E therefore fixes ev E ( ξ ) = ξ ⌟ exp At ( E ) , so U = ev E ( H T 2 ) lies in Ext 2 ( E , E ) G . By (A4), σ E ( U ) is the image of the contraction with γ , of dimension r 0 , so dim Ext 2 ( E , E ) G ≥ r 0 , and equality makes σ E injective on Ext 2 ( E , E ) G : E is G-semiregular. By [98, Lemma 2.7] every equivariant lift of E is semiregular in the invariant category, and [98, Thm. 1.2], applied with B 0 = 0 and w 0 = γ , which stays of Hodge type along S, shows that γ stays algebraic along S; so N ω = γ − ∑ k c k θ k is algebraic at every member, and W = K · ω is algebraic because K acts by algebraic correspondences. The last statement is Theorem A6.    □
For the line bundles of Proposition A4 on E 0 2 n the largest such group is ( Z / 2 ) n , when all t i − p are odd, and it divides the off-diagonal groups between pieces by 2 | S | (Proposition A14); the bounds of the appendix that rest on those groups therefore say nothing about Ext 2 ( E , E ) G . The exclusion of Theorem A9 rests on the diagonal and holds for Ext 2 ( E , E ) G as well: no G-equivariant convolution of these pieces meets the equivariant criterion.

6.2. The Weil Plane of a Product

Lemma  14.
Let A, A ′ be of Weil type for the same field K, of dimensions 2 m , 2 m ′ , with the diagonal K-action on A × A ′ . Then W ( A × A ′ ) is a two-dimensional rational subspace of the four-dimensional space W ( A ) ⊗ W ( A ′ ) ⊆ H 2 n ( A × A ′ ) , n = m + m ′ , and every rational w ∈ W ( A × A ′ ) is ∑ i , j a i j ω i ⊗ ω j ′ with a i j ∈ Q for rational bases ( ω i ) , ( ω j ′ ) of W ( A ) , W ( A ′ ) .
Proof. 
Over C , H 1 ( A × A ′ ) = V ⊕ V ′ splits into eigenspaces V ± ⊕ V ± ′ of K, and ⋀ K 2 n of the sum is spanned by det V + ⊗ det V + ′ and det V − ⊗ det V − ′ ; these lie in the span of the four products ω ± ⊗ ω ± ′ of the eigenvectors of W ( A ) ⊗ C and W ( A ′ ) ⊗ C . Both W ( A × A ′ ) and W ( A ) ⊗ W ( A ′ ) are defined over Q , so the inclusion holds over Q . The explicit formula is (A3): ω 1 ( A × A ′ ) = ω 1 ⊗ ω 1 ′ − d ω 2 ⊗ ω 2 ′ for K = Q ( − d ) .    □

6.3. An Object with the Polarised Chern Character

Theorem  21
(The product object). Let A, A ′ be very general in their polarised Weil families, with ample classes θ A , θ A ′ , θ = θ A + θ A ′ , and suppose that W ( A ) and W ( A ′ ) are algebraic. Then there is a perfect complex E on A × A ′ , a finite direct sum with shifts of exterior products
F i ⊠ G j ′ , F i ⊠ L ′ k , L l ⊠ G j ′ , L l ⊠ L ′ k ,
with F i , G j ′ coherent sheaves on A, A ′ and L, L ′ the ample bundles, such that
ch ( E ) = N w + ∑ k = 0 2 n c k θ k , N ≠ 0 , 0 ≠ w ∈ W ( A × A ′ ) ,
has the polarised shape with Hankel rank ρ = 3 , so that r ( ch E ) = 7 n 2 − 2 n , and the first-order Hodge locus of ch ( E ) in H 1 ( A × A ′ , T ) is the tangent space of D n , n .
The hypotheses hold in the two open cases nearest to the known ones: A, A ′ very general Weil fourfolds of trivial discriminant, where the Weil classes are algebraic by Markman’s theorem for discriminant one [76] and the specialisation of algebraic classes, so that A × A ′ is an eightfold with n = 4 ; and A such a fourfold and A ′ = B a Weil surface of any discriminant d, whose Weil plane is of type ( 1 , 1 ) and algebraic by Lefschetz, so that A × B is a sixfold of discriminant d with n = 3 . For every d the Weil classes of A × B itself are algebraic by Lemma 14, and for d = − 1 those of every member of its family are, by Markman’s secant-sheaf theorem [78]. For d ≠ − 1 in Q × / N ( k × ) the very general member of the family of discriminant d is open; on the six-dimensional subdomains of Proposition 34 and at every point of complex multiplication (Corollary 15) the Weil classes are algebraic.
Proof. 
Since Hdg * ( A ) = Q [ θ A ] ⊕ W ( A ) , algebraic classes on A are rational combinations of Chern characters of coherent sheaves, and the Chern character of a coherent sheaf is a Hodge class, the algebraicity of W ( A ) gives coherent sheaves F 1 , F 2 with ch ( F i ) = N i ω i + p i ( θ A ) , N i ≠ 0 , for a basis ( ω 1 , ω 2 ) of W ( A ) ; likewise G 1 ′ , G 2 ′ on A ′ with ch ( G j ′ ) = N j ′ ω j ′ + p j ′ ( θ A ′ ) . Fix 0 ≠ w ∈ W ( A × A ′ ) rational, w = ∑ a i j ω i ⊗ ω j ′ by Lemma 14, and a common denominator M of the numbers M a i j / ( N i N j ′ ) , so that these are integers. Negative multiplicities are realised by the shift [ 1 ] , which changes the sign of ch .
Step 1. E 0 = ⨁ i , j ( F i ⊠ G j ′ ) ⊕ M a i j / ( N i N j ′ ) has, by multiplicativity,
ch ( E 0 ) = M w + ∑ i ω i ⊗ q i ′ ( θ A ′ ) + ∑ j q j ( θ A ) ⊗ ω j ′ + s ,
with q i ′ , q j polynomials and s ∈ Q [ θ A ] ⊗ Q [ θ A ′ ] .
Step 2. The classes e k θ A ′ = ch ( L ′ k ) , k = 0 , ⋯ , 2 m ′ , form a basis of Q [ θ A ′ ] / ( θ A ′ 2 m ′ + 1 ) : the matrix ( k j / j ! ) j , k has determinant ∏ j ( 1 / j ! ) times the Vandermonde determinant of 0 , 1 , ⋯ , 2 m ′ , which is nonzero. Write q i ′ = ∑ k b i k e k θ A ′ and add ( F i ⊠ L ′ k ) ⊕ − M b i k / N i , after enlarging M to clear denominators; this removes ω i ⊗ q i ′ and adds only terms p i ( θ A ) ⊗ e k θ A ′ , which lie in Q [ θ A ] ⊗ Q [ θ A ′ ] . Symmetrically remove q j ⊗ ω j ′ with summands L l ⊠ G j ′ .
Step 3. The classes e l θ A + k θ A ′ = ch ( L l ⊠ L ′ k ) form a basis of Q [ θ A ] / ( θ A 2 m + 1 ) ⊗ Q [ θ A ′ ] / ( θ A ′ 2 m ′ + 1 ) by the same determinant, so the remaining part in Q [ θ A ] ⊗ Q [ θ A ′ ] is cancelled by line bundles with integer multiplicities and shifts. Finally add ( L ⊠ L ′ ) ⊗ k , k = 0 , 1 , 2 , with generic nonzero multiplicities μ k . The result is
ch ( E ) = M w + ∑ k = 0 2 μ k e k θ ,
a polarised shape whose polynomial part is a sum of three distinct exponentials with nonzero coefficients. Three exponentials are not one of the shapes of Hankel rank at most two, so ρ = 3 ; and some c k with 1 ≤ k ≤ 2 n − 1 is nonzero, so the first-order Hodge locus is the tangent space of D n , n by Theorem 20.    □
The construction was carried out in the exterior-algebra model of H * ( A × A ′ ) for ( m , m ′ ) = ( 2 , 1 ) and ( 2 , 2 ) : the cross terms of ch ( E 0 ) vanish, nothing is left outside W ( A × A ′ ) ⊕ Q [ θ ] after Steps 2 and 3, the Hankel rank is 3, and the first-order Hodge locus has dimension 9, resp. 16, equal to n 2 (Appendix D).

6.4. What a Direct Sum Cannot Do

The tangent space of D n , n at a product point contains the tangent space of the product locus D m , m × D m ′ , m ′ , of dimension m 2 + m ′ 2 , and n 2 − m 2 − m ′ 2 = 2 m m ′ mixed directions. In the same model, each exterior product F ⊠ G ′ of sheaves of polarised shape has first-order Hodge locus inside T D n , n equal to the tangent space of the product locus: all 2 m m ′ mixed directions obstruct it, four for A × B and eight for A × A ′ (Figure 11). The object E of Theorem 21 has the whole tangent space as its first-order locus because its Chern character is a single polarised class and not a product. Semiregularity is a different matter.
Proposition  22
(Direct sums are never semiregular).
(i)
Let E = ⨁ k E k be a direct sum of at least two nonzero complexes. Then the semiregularity map σ E vanishes on ⨁ k ≠ l Ext 2 ( E k , E l ) . So σ E is injective only if Ext 2 ( E k , E l ) = 0 for all k ≠ l .
(ii)
For every direct-sum realisation E of Theorem 21, σ E is not injective and dim Ext 2 ( E , E ) > r ( ch E ) .
Proof. 
(i) The semiregularity map of Buchweitz and Flenner is built from the traces ξ ↦ tr ξ ∘ At ( E ) p of compositions with powers of the Atiyah class [11, Sect. 4]. The Atiyah class of a direct sum is block diagonal, so for ξ in an off-diagonal block Ext 2 ( E k , E l ) the composition ξ ∘ At ( E ) p is off-diagonal, and the trace of an off-diagonal block is zero.
(ii) The summands of E include F 1 ⊠ L ′ k and F 1 ⊠ L ′ l for two exponents k < l , because Step 2 of the proof uses every power L ′ k with 0 ≤ k ≤ 2 m ′ (the exponentials form a basis and q 1 ′ is a nonzero polynomial of positive degree for a sheaf of polarised shape, which has p 1 of positive degree by Theorem 20). By the Künneth formula
Ext 2 ( F 1 ⊠ L ′ k , F 1 ⊠ L ′ l ) ⊇ Ext 2 ( F 1 , F 1 ) ⊗ H 0 ( A ′ , L ′ l − k ) ,
and dim Ext 2 ( F 1 , F 1 ) ≥ r ( ch F 1 ) ≥ 12 by Theorem 20 on the fourfold A ( n = 2 , ρ ≥ 0 ), while H 0 of an ample bundle is nonzero. So an off-diagonal block is nonzero, σ E has a kernel by (i), and dim Ext 2 ( E , E ) > r because equality would force injectivity.    □
Proposition  23
(Convolutions). Let E δ be a one-sided twisted complex on pieces P 1 , ⋯ , P N with shifts s a , ordered so that the differential δ = ( δ a b ) a < b , δ a b ∈ Ext 1 + s a − s b ( P a , P b ) , is strictly upper triangular; let γ = ch ( E δ ) = ∑ a ( − 1 ) s a ch ( P a ) .
(i)
The semiregularity map σ E δ vanishes on every class of Ext 2 ( E δ , E δ ) represented by a strictly upper triangular cocycle. So the criterion holds for E δ only if every strictly upper triangular cocycle of degree two is a coboundary.
(ii)
Suppose no term of the twisted differential carries a class of degree one into a diagonal block Ext * ( P c , P c ) . Let v ∈ T D n , n have a nonzero mixed component. If some piece P = X ⊠ Y has rk Y c 1 ( X ) / θ A ≠ rk X c 1 ( Y ) / θ A ′ , then ev E δ ( v ) ≠ 0 and σ E δ ( ev E δ ( v ) ) = 0 : the criterion fails.
(iii)
Suppose the pieces split into two nonempty groups with Ext * ( P a , P b ) = 0 in both directions for a, b in different groups. Then E δ = E ′ ⊕ E ′ ′ is the direct sum of the convolutions of the groups, and the criterion holds for E δ only if every v ∈ T D n , n annihilates ch ( E ′ ) and ch ( E ′ ′ ) separately.
Proof. 
(i) The components of σ E δ ( ξ ) are the traces tr ( ξ ∘ At ( E δ ) p ) / p !  [11, Sect. 4]. In the harmonic model of a twisted complex the Atiyah class is represented by [ ∇ , ∂ ¯ + δ ] with ∇ the direct sum of the Chern connections of the pieces; it is block upper triangular with diagonal blocks At ( P a ) and off-diagonal blocks ∇ δ a b . The homotopy transfer ξ ↦ ∑ k ( − h [ δ , · ] ) k ξ that produces the chain-level representative of ξ raises or preserves the block position b − a , so a strictly upper triangular cocycle has a strictly upper triangular representative. The product of an upper triangular block matrix with a strictly upper triangular one is strictly upper triangular, and its trace, the sum of the traces of its diagonal blocks, is zero. The trace is a chain map for ∂ ¯ + [ δ , · ] by cyclicity, so the vanishing holds in cohomology. Since σ E δ ∘ ev E δ is contraction into γ , of rank r ( γ ) , the kernel of σ E δ contains the image of the strictly upper triangular classes and dim Ext 2 ≥ r ( γ ) + dim ( that image ) .
(ii) The Atiyah class of X ⊠ Y is At ( X ) ⊗ 1 + 1 ⊗ At ( Y ) , so the diagonal component of ev E δ ( v ) at P is v ⌟ At ( P ) , with trace v ⌟ ch 1 ( P ) . Write v = v A + v A ′ + z ⊗ d ′ + z ′ ⊗ d with v A ∈ T D m , m , v A ′ ∈ T D m ′ , m ′ , z ∈ H 0 , 1 ( A ) , d ′ ∈ H 0 ( T A ′ ) , z ′ ∈ H 0 , 1 ( A ′ ) , d ∈ H 0 ( T A ) . Since v kills θ = θ A + θ A ′ and v A ⌟ θ A = v A ′ θ A ′ ⌟ = 0 , we have v ⌟ θ A ′ = − v ⌟ θ A = − z ′ ∧ ( d ⌟ θ A ) , nonzero when d ≠ 0 because θ A is nondegenerate and z ′ ∧ is injective on H 0 , 1 ( A ) ; the case d = 0 , d ′ ≠ 0 is symmetric. On a very general member c 1 ( X ) and c 1 ( Y ) are multiples of the polarisations, so v ⌟ ch 1 ( P ) = rk Y c 1 ( X ) / θ A − rk X c 1 ( Y ) / θ A ′ v ⌟ θ A ≠ 0 under the hypothesis on P. Under the assumption on the differential, the diagonal components of a class in Ext 2 ( E δ , E δ ) are well defined, so ev E δ ( v ) ≠ 0 , while σ E δ ( ev E δ ( v ) ) = v ⌟ γ = 0 because v is tangent to the Hodge locus of γ .
(iii) With no Ext groups between the groups there are no components of δ between them, so the twisted complex is block diagonal,
Ext 2 ( E δ , E δ ) = Ext 2 ( E ′ , E ′ ) ⊕ Ext 2 ( E ′ ′ , E ′ ′ ) , dim Ext 2 ( E δ , E δ ) ≥ dim H T 2 − dim Ann ( ch E ′ ) ∩ Ann ( ch E ′ ′ ) ,
with σ E δ = σ E ′ ⊕ σ E ′ ′ ; the inequality holds because for each summand the contraction into its Chern character factors through ev and σ , so that dim Ext 2 ( E ′ , E ′ ) is at least the rank of ξ ↦ ξ ⌟ ch E ′ , and likewise for E ′ ′ . Here r ( γ ) = dim H T 2 − dim Ann ( γ ) with Ann ( γ ) = T D n , n . Equality dim Ext 2 = r ( γ ) forces the common annihilator to be all of T D n , n .    □
For the pieces of Theorem 21 the splitting in (iii) that one would like to use separates the pieces F i ⊠ Y carrying a secant sheaf on A from the pieces L l ⊠ Y ′ . Two statements settle what that splitting gives. The first computes the annihilator of the secant pieces by hand; the second shows that the splitting is never available.
Theorem  22
(The annihilator of the secant pieces). Let E ′ be the sum, with the multiplicities and shifts of Theorem 21, of the pieces F i ⊠ G j ′ and F i ⊠ L ′ k . Then
ch ( E ′ ) = M w + ∑ i M N i p i ( θ A ) ⊗ R i , R i = ∑ j a i j ω j ′ ,
where ( a i j ) is the matrix of w in the bases ( ω i ) , ( ω j ′ ) and is invertible, and the annihilator of ch ( E ′ ) in T D n , n is the tangent space T D m , m ⊕ T D m ′ , m ′ of the product locus, of dimension m 2 + m ′ 2 . If Ext * ( F i , L l ) = 0 for every power L l among the pieces, every convolution E δ of the pieces has dim Ext 2 ( E δ , E δ ) ≥ r ( γ ) + 2 m m ′ , which is 61 at n = 3 and 120 at n = 4 .
Proof. 
Steps 1 and 2 of Theorem 21 add and subtract, for each i, the classes
∑ j M a i j N i N j ′ ch ( F i ) ch ( G j ′ ) and M N i ch ( F i ) ⊗ q i ′ ,
where q i ′ = ∑ j ( a i j / N j ′ ) p j ′ ; the difference is ch ( F i ) ⊗ R i , and summing over i with ch ( F i ) = N i ω i + p i ( θ A ) gives (11), the terms ω i ⊗ R i adding up to M w by the definition of a. By Lemma 14 w = x ω + ⊗ ω + ′ + x ¯ ω − ⊗ ω − ′ with x ≠ 0 , a tensor of rank two, so a is invertible and R 1 , R 2 form a basis of the Weil plane of A ′ .
Write V = H 1 ( A , C ) = V + ⊕ V − for the eigenspaces of K, ω ± = det V ± , and V ± ′ , ω ± ′ for their analogues on A ′ . A tangent vector v ∈ H 1 ( T A × A ′ ) acts on cohomology as a derivation and decomposes as v = v A + v A ′ + μ + μ ′ with v A ∈ H 0 , 1 ( A ) ⊗ H 0 ( T A ) , v A ′ its analogue on A ′ , and the mixed parts μ ∈ H 0 , 1 ( A ) ⊗ H 0 ( T A ′ ) , μ ′ ∈ H 0 , 1 ( A ′ ) ⊗ H 0 ( T A ) . Let v ∈ T D n , n . The three Künneth components of v ⌟ θ vanish separately: v A ⌟ θ A = 0 , v A ′ ⌟ θ ′ = 0 and μ ⌟ θ ′ + μ ′ ⌟ θ A = 0 . Contraction into a nondegenerate form is an isomorphism from H 0 ( T ) onto H 0 , 1 , so μ ↦ μ θ ′ and μ ′ ↦ μ ′ θ A are injective into H 0 , 1 ( A ) ⊗ H 0 , 1 ( A ′ ) , and μ = 0 exactly when μ ′ = 0 . Next, v ⌟ ( ω + ⊗ ω + ′ ) = 0 . The term ( v A ⌟ ω + ) ⊗ ω + ′ has no component in Λ 2 m V + ⊗ Λ 2 m ′ V + ′ , because that component is the trace of v A on V + and v A sends V + 1 , 0 into V + 0 , 1 and kills V + 0 , 1 ; it lies in Λ 2 m − 1 V + ⊗ V − ⊗ Λ 2 m ′ V + ′ , while the other three terms lie in Λ 2 m V + ⊗ Λ 2 m ′ − 1 V + ′ ⊗ V − ′ , in Λ 2 m − 1 V + ⊗ Λ 2 m ′ V + ′ ⊗ V − ′ and in Λ 2 m V + ⊗ V − ⊗ Λ 2 m ′ − 1 V + ′ , four distinct weight spaces for G L ( V + ) × G L ( V − ) × G L ( V + ′ ) × G L ( V − ′ ) . Hence v A ⌟ ω + = 0 , and in the same way v A ⌟ ω − = 0 and v A ′ ⌟ ω ± ′ = 0 . A vector of T D n , n with vanishing mixed part therefore lies in T D m , m ⊕ T D m ′ , m ′ , and conversely every such sum kills θ and the Weil plane; the dimension m 2 + m ′ 2 is that of the two Weil families [125].
Now let v ∈ T D n , n have μ ≠ 0 . It kills M w ; v A kills p i ( θ A ) and v A ′ kills R i ; so v ⌟ ch ( E ′ ) = ( μ + μ ′ ) ⌟ ∑ k θ A k ⊗ R ( k ) , where p i = ∑ k c i k θ A k and R ( k ) = ∑ i ( M c i k / N i ) R i . Write μ = ∑ s z s ⊗ d s ′ and μ ′ = ∑ s z s ′ ⊗ d s . The contraction of θ A k ⊗ R ( k ) against μ ′ is k ∑ s θ A k − 1 ( d s ⌟ θ A ) ⊗ ( z s ′ ∧ R ( k ) ) , of Künneth bidegree ( 2 k − 1 , 2 m ′ + 1 ) , and against μ it is ∑ s ( z s ∧ θ A k ) ⊗ ( d s ′ ⌟ R ( k ) ) , of bidegree ( 2 k + 1 , 2 m ′ − 1 ) . These bidegrees are pairwise distinct over all k and both types, so every term vanishes on its own. Some R ( k ) is nonzero, since p 1 , p 2 are not both zero and R 1 , R 2 are a basis. For R = σ ω + ′ + σ ¯ ω − ′ with σ ≠ 0 the map z ′ ↦ z ′ ∧ R is injective on H 0 , 1 ( A ′ ) , because z ′ ∧ ω + ′ = z − ′ ∧ ω + ′ and z ′ ∧ ω − ′ = z + ′ ∧ ω − ′ lie in independent weight spaces and vanish only with z ∓ ′ ; dually d ′ ↦ d ′ ⌟ R is injective on H 0 ( T A ′ ) . If the nonzero R ( k ) has k ≥ 1 , hard Lefschetz makes θ A k − 1 injective on H 1 ( A ) , since k − 1 ≤ 2 m − 1 , so the μ ′ -term vanishes only if ∑ s ( d s ⌟ θ A ) ⊗ z s ′ = 0 , that is μ ′ = 0 , and then μ = 0 . If k = 0 the μ -term is ( id ⊗ ( · ⌟ R ( 0 ) ) ) ( μ ) , which vanishes only if μ = 0 . Either way the mixed part is zero, against the assumption.
The bound is Proposition 23(iii): the common annihilator of ch E ′ and ch E ′ ′ in T D n , n lies in that of ch E ′ , so dim Ext 2 ≥ dim H T 2 − m 2 − m ′ 2 = r ( γ ) + n 2 − m 2 − m ′ 2 , and n = m + m ′ .    □
Proposition  24
(Secant and line-bundle pieces are never orthogonal). For every realisation of Theorem 21 there are an index i and a power L l among the pieces with χ ( F i , L l ) ≠ 0 . Hence Ext * ( F i , L l ) ≠ 0 and Ext * ( L l , F i ) ≠ 0 , and the hypothesis of the last sentence of Theorem 22 fails.
Proof. 
The Todd class of A is 1, ω i θ A = 0 and ∫ θ A 4 = 24 , so Riemann–Roch gives χ ( F i , L l ) = ∫ p i ( − θ A ) e l θ A = P i ( l ) with
P i ( l ) = c i 0 l 4 − 4 c i 1 l 3 + 12 c i 2 l 2 − 24 c i 3 l + 24 c i 4 ,
a nonzero polynomial because p i ≠ 0 . Expand p i = ∑ l b i l e l θ A , 0 ≤ l ≤ 4 , modulo θ A 5 , which is possible since the five exponentials span Q [ θ A ] / ( θ A 5 ) , and call the set S i of l with b i l ≠ 0 the support of p i . Then P i ( s ) = ∑ l ∈ S i b i l ( s − l ) 4 . If P i vanished on all of S i with | S i | ≥ 2 , the vector ( b i l ) l ∈ S i would lie in the kernel of the symmetric matrix ( s − l ) 4 s , l ∈ S i . That matrix is nonsingular: for two elements its determinant is − ( s − l ) 8 , for three it is 2 ∏ ( s − l ) 4 , for the five four-element subsets of { 0 , ⋯ , 4 } it is 29952, 1032192, 6054912, 1032192, 29952, and for the whole set it is 7962624. So a support with at least two elements contains some l with P i ( l ) ≠ 0 .
Step 2 of Theorem 21 uses the pieces L l ⊠ G j ′ for every l in the support of q j = ∑ i ( M a i j / N i ) p i , and the last step uses L k ⊠ L ′ k for three values of k whose net multiplicities are nonzero. If p 1 and p 2 are independent, q 1 and q 2 span the same plane because a is invertible, so the union of their supports is S 1 ∪ S 2 ; if p 2 is a multiple of p 1 , both q j are multiples of p 1 , not both zero, with the same support. In either case every l ∈ S 1 ∪ S 2 gives a piece L l ⊠ G j ′ . If | S 1 | ≥ 2 or | S 2 | ≥ 2 we are done. Otherwise p i = b i e l i θ A with b i ≠ 0 . If l 1 ≠ l 2 then L l 2 ⊠ G j ′ is a piece and P 1 ( l 2 ) = b 1 ( l 2 − l 1 ) 4 ≠ 0 . If l 1 = l 2 = l the residue of Steps 1 to 3 lies in e l θ A ⊗ Q [ θ ′ ] , so at least two of the three final exponents k differ from l, and P i ( k ) = b i ( k − l ) 4 ≠ 0 . The vanishing of Ext * ( F i , L l ) would force χ ( F i , L l ) = 0 ; Serre duality gives Ext * ( L l , F i ) ≠ 0 as well.    □
Theorem 22 is therefore a conditional statement whose condition never holds: Proposition 23(iii) cannot be applied to the splitting between secant pieces and line-bundle pieces, and no bound on dim Ext 2 of a convolution follows from the Chern characters alone. The hypothesis of part (ii) can fail with line bundles only. The cone of a section of L l ⊠ L ′ k ′ − k , viewed as a map L l ⊠ L ′ k → L l ⊠ L ′ k ′ with k ′ > k , is a twisted complex of two pieces in which the Serre dual of the section, a class of degree one, enters the diagonal block of the piece L l ⊠ L ′ k ; the only terms that can enter a diagonal block of a line-bundle piece are of this form, with the same power on A. Part (ii) applies to every piece L l ⊠ L ′ k with l ≠ k whose block receives no such term, which is how Step 3 of Theorem 21 uses it.

6.5. Where the Product Object Stops

Every test that depends on ch ( E ) alone is passed: polarised shape, N ≠ 0 , Hankel rank 3, first-order Hodge locus equal to T D n , n , lower bound 57 or 104. The criterion then asks for dim Ext 2 ( E , E ) = r , Proposition 22 shows that no direct sum achieves it, and Proposition 23 says what a convolution of the same pieces must do: kill every strictly upper triangular class of degree two, carry classes of degree one from the lower triangle into the diagonal blocks of every piece with mismatched slopes, and, since by Proposition 24 the secant pieces and the line-bundle pieces are never orthogonal, carry classes between the two groups in a way that cancels the mixed directions, which by Theorem 22 the secant pieces alone leave uncancelled. All three demands are statements about the groups Ext * ( F 1 , F 2 ) , Ext * ( F i , L l ) , Ext * ( L l , F i ) and their analogues on A ′ . Those groups are not computed in [78,79,80], where the only object with known Ext groups on a very general Weil fourfold of trivial discriminant is the twisted transform Φ ( F ⊠ F ∨ ) with dim Ext k = 1 , 8 , 18 , 8 , 1 , whose Brauer class is half the Poincaré class and which therefore cannot be mixed with untwisted line bundles; and we have not computed them. Riemann–Roch fixes χ ( F i , L l ) = P i ( l ) and the Künneth formula fixes everything on the line-bundle side, so the unknown is exactly the thirteen graded groups Ext * ( F i , F j ) and Ext * ( F i , L l ) . This is the exact stopping point: the polarised criterion at a product point is a finite question about the Ext algebra of the secant sheaves, and Proposition 23 lists what the answer must contain for the route to stay open. The analogous question on E 0 6 , for line bundles, closed negatively with dim Ext 2 ≥ 69 Theorem A12; that is an analogy and not a proof of the answer here.
At a point of complex multiplication the thirteen groups are known. We record them, since they turn the stopping point into a statement about one explicit variety.
Lemma  15
(The Ext groups at a point of complex multiplication). Let K = Q ( i ) , E 0 = C / Z [ i ] , and let A 0 = E 0 6 carry the K-action by i on three factors and by − i on the other three; it is a member of D 3 , 3 for every discriminant, since the K-hermitian forms ∑ j ≤ 3 a j z j w ¯ j − ∑ j > 3 a j z j w ¯ j with a j > 0 are polarisations of every discriminant class. Let L be ample on A 0 with class θ, let Z , Z ′ ⊆ A 0 be abelian subvarieties of dimension three with Z ∩ Z ′ finite, and put χ Z = ( θ | Z ) 3 / 6 , a positive integer.
(i)
Every Hodge class on A 0 is a polynomial in divisor classes, and every product of three divisor classes is a rational combination of classes [ Z ] of abelian threefolds. The Weil classes are such combinations, and no Weil class is a multiple of a single [ Z ] . For every Z, ch ( O Z ) = [ Z ] and ch ( O Z ⊗ L a ) = [ Z ] e a θ .
(ii)
Ext k ( O Z ⊗ L a , O Z ′ ⊗ L b ) = 0 for k ≠ 3 , and dim Ext 3 = | Z ∩ Z ′ | = ( Z · Z ′ ) , for all a , b .
(iii)
The groups Ext k ( O Z ⊗ L a , O Z ⊗ L b ) are given by (13) below, with N Z ≅ O Z 3 . For a = b it has dimension 6 k and the characteristic map H T 2 ( A 0 ) → Ext 2 is onto; for b > a it lives in degrees 0 to 3, of dimension 3 k ( b − a ) 3 χ Z ; for b < a in degrees 3 to 6, of dimension 3 k − 3 ( a − b ) 3 χ Z .
(iv)
For b > a , Ext k ( O Z ⊗ L a , L b ) is concentrated in degree 3 and Ext k ( L a , O Z ⊗ L b ) in degree 0, both of dimension ( b − a ) 3 χ Z ; for b < a the first is concentrated in degree 6 and the second in degree 3, both of dimension ( a − b ) 3 χ Z ; for a = b the dimensions are 3 k − 3 and 3 k .
(v)
For b > a , Ext k ( L a , L b ) = H k ( A 0 , L b − a ) is concentrated in degree 0, of dimension ( b − a ) 6 θ 6 / 720 ; for b < a in degree 6; for a = b it is H k ( O A 0 ) , of dimension 6 k .
Proof. 
(i) The Hodge group of A 0 is the norm-one torus K 1 acting by z on V + and z ¯ on V − , so over C the Hodge classes of degree 2 k are ⋀ k V + ⊗ ⋀ k V − , which is spanned by the products of k elements of V + ⊗ V − = Hdg 2 ( A 0 ) C ; since Hdg 2 = NS Q by Lefschetz, the Q -span of the products of k divisor classes has full dimension 6 k 2 . The classes f * [ 0 ] , f ∈ Hom ( E 0 6 , E 0 ) = Z [ i ] 6 , are the rank-one hermitian forms v v ¯ T and span NS Q . For f 1 , f 2 , f 3 independent over K the map f = ( f 1 , f 2 , f 3 ) : A 0 → E 0 3 is a surjective homomorphism, hence flat, and f 1 * [ 0 ] f 2 * [ 0 ] f 3 * [ 0 ] = f * [ 0 ] = [ f − 1 ( 0 ) ] is a positive multiple of the class of the abelian threefold ker ( f ) ∘ ; for dependent f i the product is zero. A Weil class w satisfies w θ = 0 , while [ Z ] θ = i * ( θ | Z ) ≠ 0 because ( θ | Z ) 3 > 0 . Since Z is a translate of a subgroup, N Z is trivial and td ( N Z ) = 1 , so Grothendieck–Riemann–Roch gives ch ( i * O Z ) = i * ( 1 ) = [ Z ] and ch ( O Z ⊗ L a ) = [ Z ] e a θ .
(ii) Since dim Z ∩ Z ′ = 0 and the tangent spaces are translation invariant, T Z ⊕ T Z ′ = T A 0 at every point of Z ∩ Z ′ , so Z and Z ′ are locally transverse complete intersections and Z ∩ Z ′ is reduced. In local coordinates with Z = { x 1 = x 2 = x 3 = 0 } and Z ′ = { y 1 = y 2 = y 3 = 0 } , the Koszul resolution of O Z gives E x t k ( O Z , O Z ′ ) as the cohomology of the Koszul complex of x 1 , x 2 , x 3 on O Z ′ , a regular sequence there; so E x t k = 0 for k ≠ 3 and E x t 3 = O Z ∩ Z ′ . The local-to-global spectral sequence has one row, H k − 3 ( Z ∩ Z ′ , O ) , and twisting by line bundles does not change a sheaf on a finite reduced scheme; the intersection number equals | Z ∩ Z ′ | by transversality.
(iii) Let π : A 0 → A 0 / Z be the quotient. The normal bundle N Z is the restriction of the trivial bundle π * T A 0 / Z , so it extends to all of A 0 , and by the theorem of Arinkin and Căldăraru [1] the derived restriction i * i * O Z is formal, i * i * O Z ≅ ⨁ p ⋀ p N Z ∨ [ p ] . Hence
Ext k ( O Z ⊗ L a , O Z ⊗ L b ) = Ext Z k i * i * O Z , L b − a | Z = ⨁ p + q = k H q Z , L b − a | Z ⊗ ⋀ p N Z .
For a = b the dimension is ∑ p + q = k 3 q 3 p = 6 k ; for b ≠ a the cohomology of L b − a | Z on the abelian threefold Z sits in degree 0 or 3 according to the sign of b − a , with dimension | b − a | 3 χ Z by Riemann–Roch, which gives the two ranges of degrees and their dimensions. The characteristic map sends H 2 ( O A 0 ) onto H 2 ( O Z ) by restriction, H 1 ( T A 0 ) = H 1 ( O ) ⊗ T 0 onto H 1 ( O Z ) ⊗ N Z , 0 through the quotient T 0 → N Z , 0 , and H 0 ( ⋀ 2 T A 0 ) = ⋀ 2 T 0 onto ⋀ 2 N Z , 0 ; all three are surjective, the first because H * ( A 0 ) → H * ( Z ) is onto for an abelian subvariety.
(iv) By Grothendieck duality for the regular embedding, whose relative dualising sheaf det N Z is trivial, and by adjunction,
Ext k ( O Z ⊗ L a , L b ) = H k − 3 Z , L b − a | Z , Ext k ( L a , O Z ⊗ L b ) = H k Z , L b − a | Z .
On Z the cohomology of an ample bundle is concentrated in degree 0 and that of an anti-ample one in degree 3, of dimension χ = c 1 3 / 6 by Riemann–Roch, and c 1 ( L m | Z ) 3 = m 3 ( θ | Z ) 3 ; for m = 0 one gets H * ( O Z ) .
(v) is the same statement on A 0 , with χ ( L m ) = m 6 θ 6 / 720 .    □
Lemma  16
(Threefolds that meet along a positive-dimensional set). Let Z , Z ′ ⊆ A 0 be distinct abelian threefolds, let W = Z ∩ Z ′ , of dimension d ∈ { 0 , 1 , 2 } , and put c = 3 − d and m = b − a .
(i)
H o m ( O Z , O Z ′ ) = 0 , and E x t q ( O Z , O Z ′ ) ≅ O W d q − c for c ≤ q ≤ 3 , zero otherwise.
(ii)
Ext k ( O Z ⊗ L a , O Z ′ ⊗ L b ) = 0 unless c ≤ k ≤ 3 when m > 0 , 3 ≤ k ≤ 3 + d when m < 0 , and c ≤ k ≤ 3 + d when m = 0 . In particular the group vanishes in degrees k ≤ 2 when m < 0 , and Ext 1 ≠ 0 only when d = 2 and m ≥ 0 , where Ext 1 = H 0 ( W , L m | W ) .
(iii)
For the threefolds Z ζ , Z ζ ′ of Appendix B, d is the number of indices j with ζ j = ζ j ′ , W is the product of the shared curves with ∏ ζ j ≠ ζ j ′ | ζ j − ζ j ′ | 2 points, and for d = 2 and m = 0 , dim Ext 1 = | ζ l − ζ l ′ | 2 ∈ { 2 , 4 } , l being the index at which they differ.
Proof. 
(i) W is a closed subgroup scheme of A 0 , reduced in characteristic zero, so a disjoint union of translates of an abelian variety of dimension d, and it is nonempty since Z and Z ′ contain the origin. At a point of W choose translation-invariant linear coordinates in which Z = { x 1 = x 2 = x 3 = 0 } . The restrictions of x 1 , x 2 , x 3 to Z ′ are linear forms cutting out W, of codimension c in Z ′ ; after a linear change x 1 , ⋯ , x c restrict to a regular sequence on Z ′ with zero locus W and x c + 1 , ⋯ , x 3 restrict to zero. The Koszul resolution of O Z computes E x t q ( O Z , O Z ′ ) as the cohomology of the Koszul complex of these restrictions on O Z ′ , the tensor product of the Koszul complex of the regular sequence, with cohomology O W in degree c, and of d complexes O Z ′ → 0 O Z ′ ; so E x t q = ⋀ q − c ( O W d ) . Since c ≥ 1 , H o m = 0 .
(ii) The local-to-global spectral sequence has E 2 p , q = H p ( W , E x t q ⊗ L m | W ) with 0 ≤ p ≤ d and c ≤ q ≤ 3 . On each component of W the bundle L m | W is ample for m > 0 , with cohomology in degree 0 only, anti-ample for m < 0 , with cohomology in degree d only, and trivial for m = 0 . This gives the three ranges. For k = 1 the only possible term is E 2 0 , 1 , nonzero only when c = 1 and m ≥ 0 , and it survives to E ∞ because E 2 1 , 0 = E 2 2 , 0 = 0 .
(iii) In the coordinates of Appendix B, Z ζ = { u j + ζ j v j = 0 , j = 1 , 2 , 3 } ; a shared index gives a form vanishing on Z ζ ′ , and an index with ζ j ≠ ζ j ′ gives the form ( ζ j − ζ j ′ ) v j on Z ζ ′ , whose zero locus in C ζ j ′ has | ζ j − ζ j ′ | 2 points by Lemma A28(ii). So W is as stated, and Ext 1 = H 0 ( W , O W ) has dimension the number of components of W.    □
Theorem 20 applies at A 0 with the same number r = 57 : the rank of the contraction of H T 2 into ch ( E ) is invariant under SU ( V , H ) ( R ) , which fixes the Weil classes and θ and acts transitively on D 3 , 3 , so it is the rank computed at a very general point, and the argument of Buchweitz and Flenner is local on the base. An object E with dim Ext 2 ( E , E ) = 57 at A 0 would therefore deform along all of D 3 , 3 and prove the Weil classes algebraic on every member with full Hodge group. By (i) every Chern character of polarised shape is realised at A 0 by a convolution of pieces O Z a ⊗ L k a [ s a ] and L k [ H ] , with at least two subvariety pieces, since w is not a multiple of one [ Z ] , and at least three powers of L, since the Hankel rank is three; by (ii) to (v) every group between two pieces is known. Appendix B writes the classes, the intersection numbers and every dimension of the lemma in coordinates for the threefolds that are products of elliptic curves of the three Weil surfaces, and exhibits the signed sum of thirty-two of them equal to 32 ( α + α ¯ ) . What the lemma decides is how the diagonal classes can disappear.
Proposition  25
(No cancellation in two steps). Let E be a convolution of pieces F a [ s a ] of the types of Lemma 15, whose threefolds meet pairwise in finite sets, and no two of which are isomorphic sheaves placed at shifts differing by at most one. A class of Ext 2 ( F a , F a ) is cancelled in the spectral sequence of the convolution only through a closed chain a = a 0 , a 1 , ⋯ , a m = a of pieces with Ext k i ( F a i − 1 , F a i ) ≠ 0 and ∑ i k i = m , and no such chain has m = 2 . The chains with m = 3 are exactly those through three twists k < k ′ < k ′ ′ of one threefold, O Z ⊗ L k [ s ] → O Z ⊗ L k ′ [ s + 1 ] → O Z ⊗ L k ′ ′ [ s + 2 ] → O Z ⊗ L k [ s ] , with degrees 0 , 0 , 3 . Hence every cancellation among the diagonal classes of degree two is a Massey product of length at least three.
Proof. 
The Maurer–Cartan components of the twisted complex and the class η whose differential is computed all have total degree one, that is, a component in Ext k ( F a , F b ) has k + s b − s a = 1 . A product of m such components along a closed chain lands in Ext ∑ k i − m + 2 ( F a , F a ) , since the shifts cancel around the chain and the m-fold product lowers the degree by m − 2 ; so it reaches degree two exactly when ∑ k i = m . For m = 2 both components have degree at most two. Between two pieces that are not isomorphic sheaves, Lemma 15 gives the degrees: 3 and 3 between O Z and O Z ′ ; 0 to 3 one way and 3 to 6 the other between two twists of one O Z ; 3 and 3 between O Z ⊗ L a and L b with a < b , 6 and 0 with a > b , and 3 to 6 against 0 to 3 for a = b ; 0 and 6 between two powers of L. In every case the two degrees add up to at least three. Between two copies of one sheaf at shifts s, s ′ the two degrees are 1 − s ′ + s and 1 − s + s ′ , which are both at most two only when | s − s ′ | ≤ 1 , the case excluded. For m = 3 the edges of degree at most two are, by the same list, the edges that raise the twist and the edges L a → O Z ⊗ L a ; a closed chain cannot consist of them alone, so one edge has degree at least three and the other two have degree zero. Two edges of degree zero from a line bundle lead to a power of L or to a subvariety piece of strictly larger twist, from which every edge back to the line bundle has degree six; from O Z ⊗ L k they lead only to O Z ⊗ L k ′ ′ with k ′ ′ > k through an intermediate twist k ′ , and the edge back has degree three, through H 3 ( Z , L k − k ′ ′ | Z ) ⊗ ⋀ 0 N Z . These are the chains displayed, and their three groups are nonzero.
   □
For line-bundle pieces alone the same bookkeeping holds on E 0 6 , and there the Massey products of length four are the ones that act, with the result dim Ext 2 ≥ 69 Theorem A12. With subvariety pieces and three powers of L no closed chain of any length exists, and the criterion fails for a reason that involves no product at all.
Proposition  26
(Convolutions with three line bundles fail). Let E be a convolution at A 0 of pieces O Z a ⊗ L k a [ s a ] and L l b [ t b ] in which no sheaf occurs twice, the threefolds Z a are pairwise distinct, there are at most three line bundles, and Ext 1 ( O Z a ⊗ L k a , O Z b ⊗ L k b ) = 0 whenever s a = s b . Suppose ch ( E ) has the polarised shape. Then no term of the twisted differential carries a class of total degree one into a diagonal block, and
dim Ext 2 ( E , E ) > r ( ch E ) :
the criterion fails for E . The hypothesis on Ext 1 holds when the threefolds meet pairwise in finite sets, and for the thirty-two threefolds Z ζ of Appendix B placed at one twist with shifts of the parity of ζ 1 ζ 2 ζ 3 .
Proof. 
Step 1, no closed chain. As in the proof of Proposition 25, a term of the differential landing in a diagonal block runs along a closed chain a 0 , a 1 , ⋯ , a m = a 0 of pieces with a nonzero component in Ext k i ( F a i − 1 , F a i ) , k i = 1 + s a i − 1 − s a i , so that ∑ i k i = m ; the pieces of the chain are distinct, because the differential is strictly upper triangular in the one-sided order. We show that no such chain exists. Between two distinct threefold pieces Hom = 0 by Lemma 16(i), and a component of degree one needs s a = s b , which the hypothesis excludes; so every edge between two threefold pieces has degree at least two. By Lemma 15(iv) and (v), an edge L l → O Z ⊗ L k has degree 0 when k > l , 3 when k < l and 0 to 3 when k = l , so a degree at most one forces k ≥ l ; an edge O Z ⊗ L k → L l has degree 3 when l > k , 6 when l < k and 3 to 6 when l = k , so a degree at most four forces l ≥ k ; and an edge between two line bundles has degree 0 when it raises the twist and 6 when it lowers it. A chain of threefold pieces alone has ∑ i k i ≥ 2 m > m . Otherwise the chain visits ℓ ≤ 3 line-bundle pieces; between consecutive visits it either passes directly, by an edge that raises the twist (u such edges in all) or lowers it (w in all), or runs through r ≥ 1 threefold pieces, entered with degree e and left with degree x; let R be the number of runs. Then R + u + w = ℓ , m = ℓ + ∑ runs r , a run contributes at least e + x + 2 ( r − 1 ) to ∑ i k i , and so
∑ i k i − m ≥ ∑ runs e + x + r − 3 + 5 w − u ,
where each bracket is at least 0 + 3 + 1 − 3 = 1 . If w ≥ 1 then u ≤ ℓ − R − w ≤ 2 − R and the right side is at least R + 5 − ( 2 − R ) > 0 . If w = 0 then u = ℓ − R ≤ 3 − R and the right side is at least 2 R − 3 , so R ≤ 1 ; R = 0 is a cycle of edges that all raise the twist, which cannot close; and for R = 1 the single bracket must equal u, so r ≤ 2 and e + x ≤ 5 − r , that is ( e , x ) ∈ { ( 0 , 3 ) , ( 0 , 4 ) , ( 1 , 3 ) } for r = 1 and ( e , x ) = ( 0 , 3 ) for r = 2 . In both cases the run enters the first threefold, of twist k, from L l with k ≥ l , leaves the last one, of twist k ′ , for L l ′ with l ′ ≥ k ′ , and returns from L l ′ to L l by u ≥ 1 edges that raise the twist, so l ′ < l and k ′ < k . For r = 1 this reads k < k . For r = 2 the internal edge goes from the twist k to the strictly smaller twist k ′ and has degree two, which Lemma 16(ii) forbids. So there is no closed chain.
Step 2, the class ev E ( v ) . By Step 1 the diagonal components of a class in Ext 2 ( E , E ) are well defined, as in Proposition 23(ii), and for v ∈ H 1 ( A 0 , T A 0 ) the diagonal components of ev E ( v ) are the classes ev F a ( v ) = v ⌟ At ( F a ) of the pieces, the Atiyah class of the twisted complex being block upper triangular with diagonal blocks At ( F a ) . For a piece F = O Z ⊗ L k the semiregularity map sends ev F ( v ) to v ⌟ ch ( F ) = v ⌟ [ Z ] + k v ⌟ ( [ Z ] θ ) + ⋯ , whose component of lowest degree is v ⌟ [ Z ] . There is at least one threefold piece, since the Weil component of ch ( E ) is nonzero and powers of L contribute none. Step 3 gives v ∈ T D 3 , 3 with v ⌟ [ Z a ] ≠ 0 for such a piece, so ev E ( v ) ≠ 0 ; and σ E ( ev E ( v ) ) = v ch ( E ) = 0 , because T D 3 , 3 annihilates θ and the Weil plane. So σ E has a nonzero kernel, while its rank is at least the rank r ( ch E ) of σ E ∘ ev E ; hence dim Ext 2 ( E , E ) ≥ r ( ch E ) + 1 .
Step 3, no [ Z ] is annihilated by T D 3 , 3 . The tangent space T D 3 , 3 ⊆ H 1 ( A 0 , T A 0 ) = Hom ( V 1 , 0 , V 0 , 1 ) consists of the K-linear maps that are skew for the polarisation; it is the part g − 1 , 1 of the Lie algebra g of SU ( V , H ) in the Hodge structure that A 0 induces on it, and the contraction v⌟ on H * ( A 0 ) = ⋀ * V C is the derivation by which g C acts. Suppose v ⌟ [ Z ] = 0 for all v ∈ g − 1 , 1 . Since [ Z ] is real it is annihilated by the conjugate space g 1 , − 1 as well, and g − 1 , 1 ⊕ g 1 , − 1 , the off-diagonal blocks of sl ( V + ⊕ V − ) , generates g C = sl 6 ; so [ Z ] is invariant under the connected group SU ( V , H ) , in particular under its rational points. Every g ∈ SU ( V , H ) ( Q ) ⊆ GL K ( V ) = End 0 ( A 0 ) × is a quasi-isogeny of A 0 , because E 0 has complex multiplication by K; for an integer n with φ = n g an isogeny, φ * [ Z ] = [ Z ] gives φ * [ Z ] = φ * φ * [ Z ] = deg ( φ ) [ Z ] , so the class of the abelian threefold φ ( Z ) , whose tangent space is g T Z , is a positive multiple of [ Z ] . The group SU ( V , H ) ( C ) ≅ SL 6 ( C ) acts transitively on the three-dimensional subspaces of V + , which determine the K-subspaces of V, the condition g T Z ∩ T Z = 0 is Zariski open and nonempty in g, and SU ( V , H ) ( Q ) is Zariski dense in the connected group SU ( V , H )  [22, Cor. 18.3]. So some g makes Z ∩ φ ( Z ) finite; it contains the origin, and ( Z · φ ( Z ) ) = | Z ∩ φ ( Z ) | ≥ 1 by Lemma 15(ii). But ( Z · φ ( Z ) ) is a positive multiple of [ Z ] 2 , and [ Z ] 2 is a multiple of f * ( [ 0 ] 2 ) = 0 for the map f : A 0 → E 0 3 of Lemma 15(i), since [ 0 ] 2 = 0 on E 0 3 . This contradiction proves the claim, and the proposition.
For the last sentence of the statement: when the threefolds meet in finite sets every group between two of them sits in degree three by Lemma 15(ii); and two of the Z ζ at the same shift have the same value of ζ 1 ζ 2 ζ 3 , hence differ in at least two coordinates, so d ≤ 1 and Ext 1 = 0 by Lemma 16(ii).    □
Proposition 26 settles the configuration that Appendix B writes down, the thirty-two threefolds Z ζ at one twist with three powers of L: no convolution of these thirty-five pieces meets the criterion. By Lemma 15(i) every object meeting the criterion at A 0 is a convolution of pieces O Z ⊗ L k [ s ] and L l [ H ] , so what remains are the convolutions with four or more line bundles, with a sheaf at two shifts, with a threefold at several twists, or with two threefold pieces at the same shift meeting along an abelian surface, as two of the Z ζ with two equal coordinates do. Closed chains exist among them. With the threefolds at twist 0 and the line bundles of positive twist, every edge from a line bundle into a threefold piece and every edge out of one has degree three, so the shortest closed chain through a threefold piece runs through five line bundles at five consecutive shifts, L l 1 [ σ − 2 ] → L l 2 [ σ − 1 ] → L l 3 [ σ ] → L l 4 [ σ + 1 ] → L l 5 [ σ + 2 ] → O Z [ σ ] → L l 1 [ σ − 2 ] with 0 < l 1 ≤ ⋯ ≤ l 5 and degrees 0 , 0 , 0 , 0 , 3 , 3 ; one threefold at three twists carries the chain of Proposition 25; and two pieces O Z ζ ⊗ L k [ s ] , O Z ζ ′ ⊗ L k [ s ] with two equal coordinates carry the chain of length two with degrees 1 , 1 , through the groups H 0 ( W , O W ) of Lemma 16(iii). The question at the stopping point is whether the diagonal components ev F a ( v ) a , v ∈ T D 3 , 3 , lie in the image of the products along such chains; the Hodge locus of the obstruction, the only other way the criterion could fail, is settled by Theorem 20 itself. The next subsection answers the question twice without computing a product: by counting the blocks that no chain reaches, and by observing that one threefold piece outside every closed chain decides it.

6.6. Counting the Remaining Configurations at E 0 6

The four kinds of convolution left open by Proposition 26 can be tested by a bound that uses no product at all. Fix pieces F a with shifts s a and a minimal A ∞ -model of their full subcategory, so that Hom • ( E , E ) for a convolution E of the F a is the direct sum of the blocks Ext k ( F a , F b ) placed in total degree k + s b − s a , with the differential x ↦ ∑ p , q m p + q + 1 ( δ , ⋯ , δ , x , δ , ⋯ , δ ) built from the twisting cochain δ , whose components lie in the blocks of total degree one. Call a block Ext k ( F a , F b ) of total degree two isolated when no block of total degree one maps to it and it maps to no block of total degree three along such a term: that is, there is no chain a → ⋯ → a ′ and b ′ → ⋯ → b of pieces with every step a nonzero block of total degree one and ( a ′ , b ′ ) ≠ ( a , b ) carrying a block of total degree one, and no such chain a ′ → ⋯ → a , b → ⋯ → b ′ with ( a ′ , b ′ ) carrying a block of total degree three.
Lemma  17
(Isolated blocks). For every twisting cochain δ on the pieces F a , the isolated blocks of total degree two inject into Ext 2 ( E , E ) ; so dim Ext 2 ( E , E ) is at least the sum of their dimensions.
Proof. 
An element of an isolated block is a cycle, since every term of its differential lands in a block of total degree three reached from it along a chain of components of δ , and there is none; and it is not a boundary, since every term of the differential of a class of total degree one lands in a block reached from that class along such a chain, and none reaches an isolated block. The blocks are direct summands, so the cycles of the isolated blocks meet the boundaries only in zero.    □
The bound depends only on the Ext groups of Lemmas 15 and 16 and on the shifts, not on δ , so it can be computed over all configurations of a family at once. We did this for the configurations of the four kinds within a window of twists 1 to 4 and shifts − 3 to 3 around the thirty-five pieces, keeping those whose Chern character has the polarised shape, and recorded for each the sums C 1 , C 2 , C 3 of the block dimensions in total degrees one to three, the Euler characteristic C 2 − C 1 − C 3 and the isolated blocks (Table 6). The programs are in the repository named in the Data availability section.
Proposition  27
(What counting decides). At A 0 = E 0 6 , with the pieces of Appendix B:
(i)
For the thirty-five pieces of Proposition 26 with the three line bundles at any shifts in the window, dim Ext 2 ( E , E ) ≥ 2605 for every twisting cochain, by isolated blocks alone.
(ii)
Adding to them a threefold Z ζ ′ meeting Z ζ along an abelian surface, at the shift of Z ζ and at a neighbouring shift so that the Chern character keeps the polarised shape, gives dim Ext 2 ≥ 5325 : the configuration of two threefold pieces at the same shift meeting along a surface fails for every twisting cochain.
(iii)
Adding one further copy of a threefold at a second shift within the window never gives the polarised shape.
(iv)
For four line bundles and for one threefold at three twists the isolated blocks are 0 at the configurations of Table 6 and the Euler characteristic is negative throughout the window: the bound decides nothing there.
Proof. 
(i), (ii) and (iv) are Lemma 17 applied to the block dimensions of Lemmas 15 and 16, enumerated over the window; (iii) is the enumeration of Chern characters, since a second copy of O Z ζ at a shift of the other parity changes the Weil component by ± 2 [ Z ζ ] , which no choice of line bundles compensates.    □
Proposition 27 closes the configuration (d) of two threefolds at one shift and every placement of the thirty-five pieces. For the configurations (a) and (c) the bound is silent, and a different argument decides them: it suffices that one threefold piece lies outside every closed chain. The configuration (a) with more line bundles is decided in Section 6.7, by hand up to five and by the enumeration in the repository up to seven.
Let Γ be the directed graph whose vertices are the pieces F a and which has an edge a → b , for a ≠ b , when the block of total degree one, Ext 1 + s a − s b ( F a , F b ) , is nonzero. The chains of Lemma 17 are its paths, a closed chain in the sense of Proposition 25 is one of its directed cycles, and the components of every twisting cochain δ sit on its edges. Call a piece chain-free when it lies on no directed cycle of Γ ; this is a property of the pieces and their shifts alone.
Proposition  28
(A piece outside every closed chain). Let E be a convolution at A 0 of pieces O Z a ⊗ L k a [ s a ] and L l b [ t b ] whose Chern character has the polarised shape, and suppose that some threefold piece F a = O Z a ⊗ L k a [ s a ] is chain-free. Then, for every twisting cochain δ,
dim Ext 2 ( E , E ) ≥ r ( ch E ) + 1 ,
and the criterion fails for E .
Proof. 
Let η ∈ Hom 1 ( E , E ) have components η b c in the blocks of total degree one, and consider the component of its differential in the block Ext 2 ( F a , F a ) . It is a sum of terms m p + q + 1 ( δ , ⋯ , δ , η b c , δ , ⋯ , δ ) with p + q ≥ 1 , and such a term is nonzero only if its arguments are composable and run from F a back to F a : the p components of δ on one side of η b c along edges of Γ from a to b, the q components on the other side along edges from c to a, and η b c ≠ 0 . If b = c , the p + q ≥ 1 components of δ form a closed walk of positive length, which is impossible because δ is strictly upper triangular in the one-sided order. If b ≠ c , then b → c is an edge of Γ , and a → ⋯ → b → c → ⋯ → a is a closed walk through a with at least two edges, whose consecutive vertices are distinct; its first return to a, with repeated vertices shortcut, is a directed cycle of Γ through a. Since F a is chain-free, no term survives: the differential of every element of total degree one has zero component in Ext 2 ( F a , F a ) . Hence the map π a : Ext 2 ( E , E ) → Ext 2 ( F a , F a ) that reads off the component of a cocycle in that block is well defined on cohomology.
As in the proof of Proposition 23(ii), the Atiyah class of the twisted complex is block upper triangular with diagonal blocks At ( F b ) , and the homotopy transfer to the minimal model only raises the block position, so it preserves the diagonal blocks of an upper triangular element; hence for v ∈ H 1 ( A 0 , T A 0 ) the class ev E ( v ) satisfies π a ( ev E ( v ) ) = ev F a ( v ) = v At ( F a ) . By Step 3 of the proof of Proposition 26 there is v ∈ T D 3 , 3 with v [ Z a ] ≠ 0 , and then ev F a ( v ) ≠ 0 , because the semiregularity map of F a sends it to v ch ( F a ) , whose component of lowest degree is v [ Z a ] . So ev E ( v ) ≠ 0 , while σ E ( ev E ( v ) ) = v ch ( E ) = 0 , since T D 3 , 3 annihilates θ and the Weil plane. The kernel of σ E is nonzero and its rank is at least r ( ch E ) , the rank of σ E ∘ ev E ; the bound follows, and equality in Theorem 20 is excluded.    □
Corollary  12
(Configurations with a chain-free piece). Let E be a convolution at A 0 whose Chern character has the polarised shape and whose pieces are the thirty-two threefolds O Z ζ of Appendix B at twist 0, at the shifts s and s + 1 of the parity of ζ 1 ζ 2 ζ 3 , together with further pieces, line bundles L l [ H ] and threefolds O Z ⊗ L k [ H ] , all of positive twist. The criterion fails for E , for every twisting cochain, in each of the following cases.
(i)
The further pieces are at most four line bundles, of pairwise distinct twists: then Γ has no directed cycle.
(ii)
Every further piece sits at a shift at most s + 2 : then no edge of Γ ends at a piece of shift s + 1 , and the sixteen pieces there are chain-free.
(iii)
The further pieces are at most five line bundles, at any shifts: then a cycle through a piece at shift σ needs five line bundles at the shifts σ − 2 , ⋯ , σ + 2 , and the sixteen pieces at one of the two shifts are chain-free.
Both bounds are sharp: L l [ s + 3 ] enters every piece at shift s + 1 , and six line bundles L l 1 [ s − 2 ] , ⋯ , L l 6 [ s + 3 ] with 0 < l 1 ≤ ⋯ ≤ l 6 put each of the thirty-two pieces on a cycle of length six.
Proof. 
Along a directed cycle a 0 , a 1 , ⋯ , a m = a 0 of Γ write k i = 1 + s a i − 1 − s a i for the degree of the block on the edge a i − 1 → a i ; the shifts cancel around the cycle, so ∑ i k i = m . The degrees available on the edges are read off from Lemmas 15 and 16. Two of the thirty-two pieces at one shift have the same value of ζ 1 ζ 2 ζ 3 , hence differ in at least two coordinates, so d ≤ 1 and the degrees between them are 2 and 3; as Hom = 0 between distinct threefolds and a degree one would need equal shifts, every edge between two of the thirty-two has degree at least 2. An edge from a line bundle of positive twist into a piece of twist 0 has degree 3, and so has an edge from a piece of twist 0 into a line bundle of positive twist; an edge into a piece of twist 0 from a threefold piece of positive twist has degree at least 3; an edge between two line bundles has degree 0 when it raises the twist, 6 when it lowers it, and 1 + t − t ′ ≥ 0 between two copies of one line bundle at shifts t and t ′ . Parts (i) and (iii) use of the thirty-two pieces only that they have one twist, pairwise distinct threefolds and no Ext 1 at equal shifts, as in Proposition 26.
(i) A cycle among pieces of twist 0 alone has ∑ i k i ≥ 2 m , so a cycle visits ℓ ′ ≥ 1 line-bundle pieces, ℓ ′ ≤ 4 . Between two consecutive visits it either runs through r ≥ 1 pieces of twist 0, entered and left with degree 3 and with r − 1 internal edges of degree at least 2, which contributes at least 2 r + 4 to ∑ i k i and r to m, or it passes directly along an edge that raises the twist, u such edges in all, or lowers it, w in all. With R runs, R + u + w = ℓ ′ and
0 = ∑ i k i − m ≥ ∑ runs ( r + 4 ) + 6 w − ℓ ′ ≥ 4 R + 5 w − u .
If R ≥ 1 this gives u ≥ 4 and ℓ ′ ≥ 5 . If R = 0 it gives w = 0 , and a cycle whose edges all raise the twist does not close. So Γ has no cycle; every threefold piece is chain-free, there is one, and Proposition 28 applies.
(ii) An edge u → v into a piece v at shift s + 1 has degree 1 + s u − ( s + 1 ) = s u − s , at most 2 when s u ≤ s + 2 . No line bundle and no threefold piece of positive twist has an edge of degree at most 2 into a piece of twist 0, and an edge into v from one of the thirty-two would have degree 0 or 1, which do not occur. So no edge ends at v.
(iii) Let a cycle pass through a piece v of twist 0 at shift σ , and let q count its direct edges between two copies of one line bundle, so that R + u + w + q = ℓ ′ ≤ 5 and, the q edges contributing at least 0, 0 ≥ 4 R + 5 w − u − q as in (i). Since R ≥ 1 , this forces R = 1 , w = 0 , u + q = 4 , ℓ ′ = 5 and equality throughout: the only run is v itself, and the four direct edges have degree 0, so each raises the shift by one, while the two edges of degree 3 at v lower it by two. The five line-bundle pieces therefore sit at the shifts σ − 2 , ⋯ , σ + 2 , one at each, with twists that do not decrease along the shifts. Five line bundles occupy at most five shifts, and the five shifts around s differ from the five around s + 1 ; so the pieces at one of the two shifts are chain-free.
For the sharpness, Ext 3 ( L l , O Z ζ ) ≠ 0 is the edge L l [ s + 3 ] → O Z ζ [ s + 1 ] ; and the six line bundles carry the cycle L l 1 [ s − 2 ] → ⋯ → L l 5 [ s + 2 ] → O Z ζ [ s ] → L l 1 [ s − 2 ] through every piece at shift s and the cycle L l 2 [ s − 1 ] → ⋯ → L l 6 [ s + 3 ] → O Z ζ [ s + 1 ] → L l 2 [ s − 1 ] through every piece at shift s + 1 , each edge being one of the nonzero groups listed.    □
Remark  6
(The configurations of Table 6). Corollary 12(i) covers the configuration (a) of the table, (ii) covers (c), whose further twists sit at shifts at most 2 = s + 2 , and (iii) covers every configuration of the thirty-two pieces with at most five line bundles. The programs of the repository rebuild Γ from the dimensions of Lemmas 15 and 16 and find, in the window of the table, no cycle in any of the 2401 configurations of (a); at least fifteen chain-free threefold pieces in each of the 3024 configurations of (c) with further twists 1 ≤ k < k ′ ≤ 3 , twenty-two at the configuration of the table; no edge into the sixteen pieces at shift 1 from the thirty-two or from the 1560 pieces of positive twist at shifts − 3 to 2, while L l [ 3 ] and O Z ( 1 , 1 , 1 ) ⊗ L 1 [ 3 ] do enter; in each of the 28 812 configurations of four or five line bundles of twists 1 to 5 at shifts − 3 to 3 one class of sixteen pieces chain-free, the other class, when it lies on cycles, having line bundles at its five shifts; and with L 1 [ − 2 ] , ⋯ , L 6 [ 3 ] every piece on a cycle. The Lean file checks the counts 32 and 22 of chain-free pieces at the configurations (a) and (c) of the table, the absence of entering edges from the line bundles L 1 to L 4 and the twists 1 to 3 of the sixty-four threefolds Z ζ at shifts − 3 to 2, the entering edges at shift 3 and the six-cycles. No statement above rests on these computations.

6.7. Chains Through Line Bundles

Corollary 12 needs the threefolds at one twist and the further pieces of positive twist, and it stops at five line bundles because six of them put every threefold piece on a cycle. This subsection removes both restrictions. It allows the threefolds any twists and any shifts, it works for every polarisation of A 0 and hence at every discriminant, and in place of the absence of cycles it uses a property of the cycles that the Maurer–Cartan equation of the twisted complex forces. By hand this excludes every convolution of distinct threefolds and at most five line bundles; the enumeration of Remark 9, a computation in the repository, extends the exclusion to six and seven line bundles and to all but two arrangements of eight. The next subsection turns from counting pieces to the support of any object meeting the criterion.
Setting  5
(Threefolds and line bundles at A 0 ). Throughout this subsection E = E δ is a convolution at A 0 = E 0 6 , in the sense of Proposition 23, of threefold pieces O Z ζ ⊗ L k ζ [ s ζ ] , ζ ∈ μ 4 3 , and line-bundle pieces L l i [ t i ] , where L now denotes a line bundle of class θ h (defined below), such that (S1) each threefold Z ζ occurs in at most one piece, (S2) the exponents l i are pairwise distinct, and (S3) two threefold pieces at one shift differ in at least two coordinates of ζ, so that Ext 1 = 0 between them by Lemma 16(ii) and (iii). The twists k ζ , l i and the shifts are arbitrary integers. For h ∈ Z > 0 6 let θ h = ∑ j ( h j a j + h 3 + j b j ) be the class of the hermitian form of Lemma 15 with coefficients h, so that θ ( 1 , ⋯ , 1 ) = θ is the class of L, let D h be the family D 3 , 3 of the discriminant of θ h , which contains ( A 0 , θ h ) , and suppose that ch ( E ) has the polarised shape with respect to θ h : ch ( E ) = N ω + ∑ k c k θ h k with N ≠ 0 and 0 ≠ ω ∈ W . The Weil plane W of (A10) does not depend on h. The class θ h is that of a positive hermitian form, so a line bundle L of class θ h is ample, and for h = ( 1 , ⋯ , 1 ) it is the line bundle of Lemma 15. The proofs of Lemmas 15 and 16 use only that L is ample and Serre duality, so their statements on vanishing and nonvanishing hold for L of class θ h word for word.
Remark  7
(Why the twist follows the polarisation). If the pieces were twisted by powers of the bundle of class θ while the polarised shape is taken with respect to θ h with h not constant, the shape would force c 1 = ⋯ = c 5 = 0 . Indeed, by Lemma 20 the part of weight zero of ch ( E ) in H 2 k is P k θ k + R k − 3 X 0 θ k − 3 with rational P k , R k − 3 ( R k − 3 = 0 for k < 3 ), and it must equal c k θ h k . In the basis of squarefree monomials u S = ∏ x ∈ S x in a 1 , a 2 , a 3 , b 1 , b 2 , b 3 one has θ h k = k ! ∑ | S | = k h S u S , h S = ∏ x ∈ S h x , and X 0 θ k − 3 = ( k − 3 ) ! ∑ | S | = k t ( S ) u S , with t ( S ) the number of three-element subsets of S meeting each pair { a j , b j } once. Suppose c k ≠ 0 . For k = 1 , 2 all h S are equal, so h is constant. For k = 3 the sets { a j , b j , x } have t = 0 , so h j h 3 + j h x does not depend on x ∉ { a j , b j } , and h is constant. For k = 4 the three sets { a j , b j , a l , b l } have t = 0 , so h 1 h 4 = h 2 h 5 = h 3 h 6 = : p ; the sets { a j , b j , x , y } with x , y from the two other pairs have t = 2 , so p h x h y is independent of the choice, which gives h l = h 3 + l for all l and then h l 2 = p . For k = 5 every S has t ( S ) = 4 , so h S = h 1 ⋯ h 6 / h x , x ∉ S , is independent of x. Then γ = N ω + c 0 + c 6 θ h 6 , the contraction has rank at most 48 and the first-order Hodge locus is larger than T D h . With L of class θ h the same weight-zero parts read P k θ h k + R k − 3 X 0 θ h k − 3 , and the shape no longer forces c k = 0 : it asks R k − 3 X 0 θ h k − 3 ∈ Q θ h k , and then c k = P k + λ R k − 3 with the constant of proportionality λ. For k = 4 the condition is R 1 = 0 , because the coefficient of a 1 b 1 a 2 b 2 is 0 in X 0 θ h and 24 h 1 h 4 h 2 h 5 ≠ 0 in θ h 4 . The program julia/chernh.jl of Appendix D checks both statements for h ∈ { 1 , 2 , 3 } 6 and computes the ranks 57, 48, 39, 30 of the contraction for γ of polarised shape, for c 1 = ⋯ = c 5 = 0 , and with in addition c 0 = 0 and c 6 = 0 .
Remark  8
(The discriminant of θ h ). The hermitian form of θ h is H h = diag ( h 1 , h 2 , h 3 , − h 4 , − h 5 , − h 6 ) up to a rational scalar, so the discriminant of D h is − h 1 ⋯ h 6 in Q × / N ( Q ( i ) × ) . When h 1 ⋯ h 6 is a sum of two rational squares, in particular for h = ( 1 , ⋯ , 1 ) , D h is the family of abelian sixfolds of Weil type for Q ( i ) of discriminant − 1 , on which Koike proved the Weil classes algebraic [64, Cor. 2.1]; an object at ( A 0 , θ h ) meeting the criterion would then give a new proof of a known case. Discriminant − 1 is exactly the split Weil type for sixfolds, so Markman’s theorem for abelian sixfolds of split Weil type [78,79] covers the same families as a theorem on polarised families; as a theorem on abelian varieties with an action of Q ( i ) it covers every member that carries some split Q ( i ) -compatible polarisation (Proposition 34(a)). For h 1 ⋯ h 6 not a sum of two squares, for instance h = ( 3 , 1 , 1 , 1 , 1 , 1 ) , the Weil classes are algebraic on the six-dimensional subdomains of Proposition 34(c) and at every point of complex multiplication of D h (Corollary 15), ( A 0 , θ h ) among them; at the very general member of D h they are open. Every exclusion of this subsection holds for every h.
Theorem 20 applies at ( A 0 , θ h ) with r = r ( γ ) for every h, by the argument given before Proposition 25 with SU ( V , H h ) in place of SU ( V , H ) : the group fixes W and θ h and acts transitively on D h . The tangent space T D h ⊆ H 1 ( A 0 , T A 0 ) consists of the K-linear maps skew for θ h ; it annihilates θ h and W, and Step 3 of the proof of Proposition 26 holds for it word for word, SU ( V , H h ) ( Q ) being again a group of quasi-isogenies of A 0 Zariski dense in SU ( V , H h ) : for every abelian threefold Z there is v ∈ T D h with v ⌟ [ Z ] ≠ 0 . Suppose now that the criterion holds for E , that is dim Ext 2 ( E , E ) = r ( γ ) ; for γ of Hankel rank three, r ( γ ) = 57 . Then σ E is injective, and for every v ∈ T D h the class ev E ( v ) ∈ Ext 2 ( E , E ) , which σ E sends to v ⌟ ch ( E ) = 0 , vanishes. Everything below draws consequences from this vanishing.
Let Γ be the graph of Section 6.6 on the pieces F a of E , and let Γ δ ⊆ Γ be the subgraph of the edges a → b on which the component δ a b is nonzero. Since δ is strictly upper triangular in the one-sided order, Γ δ has no directed cycle, and its walks are paths. The twisted complex lives over the minimal A ∞ -model of the pieces, whose underlying graded space is ⨁ a , b Ext * ( F a , F b ) with the block Ext k ( F a , F b ) in total degree k + s b − s a . The cochain δ satisfies the Maurer–Cartan equation ∑ k ≥ 2 m k ( δ , ⋯ , δ ) = 0 , and the differential of the twisted complex is D ( η ) = ∑ p , q ≥ 0 m p + q + 1 ( δ , ⋯ , δ , η , δ , ⋯ , δ ) with p components of δ before η and q after it. A term of either sum is nonzero only if its arguments are composable: along a path of Γ δ for the Maurer–Cartan equation, and, for a term of D ( η ) a d with the argument η b c , along a path of Γ δ from a to b, the edge b → c of Γ , and a path of Γ δ from c to d.
Lemma  18
(The cocycle of a tangent vector). For v ∈ H 1 ( A 0 , T A 0 ) the class ev E ( v ) is represented in the minimal model by a cocycle c ( v ) whose component in the diagonal block Ext 2 ( F a , F a ) is ev F a ( v ) = v ⌟ At ( F a ) and whose component in a block ( a , c ) with a ≠ c is a sum of terms each of which is a composite along a path of Γ δ from a to c. In particular c ( v ) a c = 0 when Γ δ has no path from a to c.
Proof. 
As in the proof of Proposition 23(i), the Atiyah class of the twisted complex is represented in the harmonic model by [ ∇ , ∂ ¯ + δ ] , block upper triangular with diagonal blocks At ( F a ) and off-diagonal blocks ∇ δ a b , which vanish off the edges of Γ δ ; so v ⌟ [ ∇ , ∂ ¯ + δ ] has diagonal blocks v ⌟ At ( F a ) and off-diagonal blocks on the edges of Γ δ only. The homotopy transfer ξ ↦ ∑ k ( − h [ δ , · ] ) k ξ to the minimal model composes each block of ξ with components of δ on either side, the homotopy h acting block by block; so every term of the component ( a , c ) of the transferred cocycle is a composite of one block of v ⌟ [ ∇ , ∂ ¯ + δ ] , sitting on a vertex or an edge of Γ δ , with components of δ along edges of Γ δ leading from a to it and from it to c. For a = c only the term with k = 0 occurs, since a composite of k ≥ 1 components of δ from a back to a would be a cycle of Γ δ .    □
Call a directed cycle of Γ effective for δ when exactly one of its edges lies outside Γ δ .
Proposition  29
(Every threefold piece lies on an effective cycle). Suppose E = E δ meets the criterion. Then for every v ∈ T D h there is η of total degree one with D ( η ) = c ( v ) , and for every piece F a
ev F a ( v ) = ∑ m p + q + 1 ( δ , ⋯ , δ , η b c , δ , ⋯ , δ ) ,
the sum running over the effective cycles a → ⋯ → b → c → ⋯ → a of Γ through a, with b → c the edge outside Γ δ and p, q the numbers of edges before and after it. In particular every threefold piece lies on an effective cycle whose edge outside Γ δ carries a nonzero component of η.
Proof. 
The class ev E ( v ) vanishes, so its representative c ( v ) of Lemma 18 is a coboundary D ( η ) . The diagonal component D ( η ) a a is the sum of the terms m p + q + 1 ( δ , ⋯ , η b c , ⋯ , δ ) whose arguments are composable from a to a: a path of Γ δ from a to b, the edge b → c with η b c ≠ 0 , and a path of Γ δ from c to a. If b = c the term is m 1 ( η a a ) = 0 when p = q = 0 , and otherwise its components of δ form a cycle of Γ δ ; so b ≠ c . The two paths are increasing in the one-sided order, so they meet only in a; the edge b → c is not in Γ δ , as otherwise Γ δ would contain a cycle; and the walk is a directed cycle of Γ with exactly one edge outside Γ δ . By Lemma 18 the left side is c ( v ) a a = ev F a ( v ) . For a threefold piece choose v ∈ T D h with v ⌟ [ Z a ] ≠ 0 ; then ev F a ( v ) ≠ 0 , as in the proof of Proposition 28, so some term of the sum is nonzero.    □
Lemma  19
(Two rules). Let a = L l a [ H ] , b = L l b [ t + 1 ] and c = L l c [ t + 2 ] be line-bundle pieces of E with l a < l b < l c , so that a → b → c is a path of Γ whose edges carry sections, δ a b ∈ H 0 ( L l b − l a ) and δ b c ∈ H 0 ( L l c − l b ) .
(i)
The path a → b → c is not the only path of Γ δ from a to c: either δ vanishes on one of its two edges, or Γ δ has another path from a to c.
(ii)
Suppose a → b → c is the only path of Γ from a to c, let v ∈ T D h and let η satisfy D ( η ) = c ( v ) . If δ b c = 0 ≠ δ a b then η b c = 0 , and if δ a b = 0 ≠ δ b c then η a b = 0 .
Proof. 
(i) The component ( a , c ) of the Maurer–Cartan equation is the sum, over the paths of Γ δ from a to c, of the products m k ( δ , ⋯ , δ ) along them. If a → b → c were the only such path it would read m 2 ( δ a b , δ b c ) = 0 , and m 2 on sections is their product, which is nonzero for two nonzero sections of line bundles on the integral variety A 0 .
(ii) A term of D ( η ) a c runs along a walk of Γ from a to c made of a path of Γ δ , one edge carrying η and a path of Γ δ . If the walk repeats a vertex, the repetition is a vertex common to the two paths of Γ δ , and shortcutting it leaves a path of Γ δ from a to c; under the hypothesis this path is a → b → c with both components nonzero, which is excluded in either case. So every term runs along a path of Γ from a to c, that is along a → b → c , and D ( η ) a c = ± m 2 ( δ a b , η b c ) ± m 2 ( η a b , δ b c ) . The right side c ( v ) a c vanishes by Lemma 18, since Γ δ has no path from a to c. When δ b c = 0 ≠ δ a b this gives δ a b η b c = 0 , the product of the nonzero section δ a b with η b c ∈ Ext 0 ( L l b , L l c ) = H 0 ( L l c − l b ) , so η b c = 0 ; the other case is the same with the roles exchanged.    □
Lemma  20
(Both parities occur). Grade H * ( A 0 , C ) by Z 3 , giving d z j and d z ¯ 3 + j the weight e j and d z ¯ j , d z 3 + j the weight − e j . Then a j , b j , θ and every θ h have weight 0, c j has weight 2 e j , c ¯ j has weight − 2 e j , W lies in the weights ± 2 ( e 1 + e 2 + e 3 ) , and the weight-zero part of [ Z ζ ] is X 0 = ∏ j ( a j + b j ) for every ζ. Consequently ∑ ζ ( − 1 ) s ζ = 0 over the threefold pieces of E : there are threefold pieces at even shifts and at odd shifts.
Proof. 
The weights are read off from (A7), (A8) and (A10), and in [ Z ζ ] = ∏ j ( a j + b j + ζ ¯ j c j + ζ j c ¯ j ) a product of one term from each factor has weight zero only if it takes a j or b j from every factor, the e j being independent. The part of ch ( E ) in H 6 is ∑ ζ ( − 1 ) s ζ [ Z ζ ] + P θ h 3 with P = ∑ i ( − 1 ) t i l i 3 / 6 , by Lemma 15(i) for L of class θ h , and it equals N ω + c 3 θ h 3 . Taking the parts of weight zero, with Q = ∑ ζ ( − 1 ) s ζ ,
Q X 0 + P θ h 3 = c 3 θ h 3 .
The squarefree cubic monomials in a 1 , b 1 , ⋯ , a 3 , b 3 are linearly independent. The coefficient of a 1 b 1 a 2 is 0 in X 0 and 6 h 1 h 4 h 2 ≠ 0 in θ h 3 , so P = c 3 . The coefficient of a 1 a 2 a 3 is 1 in X 0 and 6 h 1 h 2 h 3 in θ h 3 , so Q + 6 P h 1 h 2 h 3 = 6 c 3 h 1 h 2 h 3 . Hence Q = 0 . There is a threefold piece, since N ω ≠ 0 and the line bundles contribute nothing to the Weil component; so there are pieces of both parities.    □
The edges of Γ between line-bundle pieces are of two kinds, by Lemma 15(v): an up edge L l [ s ] → L l ′ [ s + 1 ] with l < l ′ , of degree 0, and a down edge L l [ s ] → L l ′ [ s − 5 ] with l > l ′ , of degree 6; in general an edge of degree k joins the shifts s and s − ( k − 1 ) . By Lemma 15(iv) an edge L l → O Z ⊗ L k has degree 0 when k > l , 3 when k < l and 0 to 3 when k = l , an edge O Z ⊗ L k → L l has degree 3 when l > k , 6 when l < k and 3 to 6 when l = k , and by Lemma 16(ii) and (S3) an edge between two threefold pieces has degree at least 2, and at least 3 when it lowers the twist. A run is a maximal path of Γ through threefold pieces from a line-bundle piece L l [ s ] to a line-bundle piece L l ′ [ s − β ] ; its excess is β , and it is decreasing when l ′ < l .
Lemma  21
(The shape of a cycle). Let C be a directed cycle of Γ through a threefold piece, with ℓ line-bundle pieces, R ≥ 1 runs, u up edges and w down edges, so that ℓ = R + u + w .
(i)
A run through r threefold pieces has excess at least r, with equality only when the twists do not decrease along it. A decreasing run has excess at least 3, with equality only for L l → O Z ⊗ L k → O Z ′ ⊗ L k ′ → L l ′ with k ′ ≤ l ′ < l ≤ k and degrees 0 , 3 , 3 , whose two threefold pieces have the parity of shift of L l ′ .
(ii)
u = ∑ runs β + 5 w . If ℓ ≤ 7 then w = 0 , exactly one run is decreasing, and ℓ ≥ 2 R + 2 .
(iii)
If ℓ ≤ 5 then R = 1 : the line-bundle pieces of C sit at ℓ consecutive shifts, joined by up edges in the order of their exponents, and the run leads from the highest back to the lowest. If ℓ = 4 the run is the one of (i) with excess 3, and both its threefold pieces have the parity of the lowest shift.
Proof. 
(i) An edge of degree k lowers the shift by k − 1 . Along a run the entering edge has degree e ≥ 0 , the r − 1 internal edges have degree at least 2 and the leaving edge has degree x ≥ 3 , so the excess is at least − 1 + ( r − 1 ) + 2 = r , with equality only for e = 0 , internal degrees 2 and x = 3 , that is l ≤ k 1 ≤ ⋯ ≤ k r ≤ l ′ for the twists. Let l ′ < l . With r = 1 the excess ( e − 1 ) + ( x − 1 ) is at least 4: e = 0 forces k ≥ l > l ′ , hence x = 6 ; e = 3 gives 2 + ( x − 1 ) ≥ 4 ; and e ∈ { 1 , 2 } forces k = l > l ′ , hence x = 6 . With r = 2 and degrees e , d , x the excess e + d + x − 3 equals 3 only for ( e , d , x ) = ( 0 , 3 , 3 ) : e = 0 forces k 1 ≥ l > l ′ , and then either k 2 ≥ k 1 > l ′ , so x = 6 , or k 2 < k 1 , so d ≥ 3 and x ≥ 3 , with x = 3 exactly when k 2 ≤ l ′ ; e = 3 gives at least 3 + 2 + 3 ; e = 1 forces k 1 = l , and then k 2 ≥ l > l ′ gives x = 6 while k 2 < k 1 gives d ≥ 3 , each at least 7; e = 2 gives at least 2 + 2 + 3 . With r ≥ 3 the excess is at least r + 1 ≥ 4 . Along the run with ( e , d , x ) = ( 0 , 3 , 3 ) from L l at shift s the shifts are s + 1 , s − 1 and s − 3 .
(ii) The shifts return to their value around C: each up edge adds 1, each down edge subtracts 5 and each run subtracts its excess. Each run has excess at least 1, so ℓ = R + u + w = R + ∑ β + 6 w ≥ 2 R + 6 w , and w ≥ 1 forces ℓ ≥ 8 . Let w = 0 . The exponents also return to their value, and the up edges raise them, so some run is decreasing. With D ≥ 1 decreasing runs, of excess at least 3, ℓ = R + ∑ β ≥ R + 3 D + ( R − D ) = 2 R + 2 D ; so ℓ ≤ 7 forces D = 1 , and ℓ ≥ 2 R + 2 .
(iii) ℓ ≤ 5 gives R = 1 by (ii), and u = ℓ − 1 : the line-bundle pieces form a chain of ℓ − 1 up edges, each raising the shift by one and the exponent, closed by one run from the last to the first, which is decreasing. For ℓ = 4 the run has excess u = 3 , so it is the run of (i), and the parity of its threefold pieces is that of its exit, the lowest line bundle.    □
Theorem  23
(Convolutions with at most five line bundles fail). In the Setting 5, if E has at most five line-bundle pieces then dim Ext 2 ( E , E ) > r ( γ ) : the criterion fails for E , for every twisting cochain, for every twist of the threefolds and for every polarisation θ h .
Proof. 
Suppose the criterion holds. By Lemma 20 there are threefold pieces of both parities of shift, and by Proposition 29 each lies on an effective cycle carrying a nonzero component of η , in particular on a directed cycle of Γ . A cycle through a threefold piece has at least four line-bundle pieces by Lemma 21(ii), so with at most three there is none. Let there be four or five, L 1 , ⋯ , L n in the order of their exponents.
With four, every cycle through a threefold piece uses all four, at consecutive shifts in the order of the exponents, and its threefold pieces have the parity of the lowest shift, by Lemma 21(iii); so the threefold pieces on cycles have one parity, and a threefold piece of the other parity lies on no cycle.
With five, a cycle through a threefold piece has four or five line bundles, at consecutive shifts in the order of the exponents. If the five are not at five consecutive shifts in that order, no cycle uses all five, and two cycles of four, omitting L m and L m ′ with m ≠ m ′ , have the same lowest shift: the three line bundles common to both occupy three of four consecutive shifts in each; if these are consecutive, x , x + 1 , x + 2 , the omitted bundles sit at x − 1 or at x + 3 , and at the same one of the two, since L m at x − 1 and L m ′ at x + 3 would put the five at five consecutive shifts in the order of the exponents; and if they leave a gap, both omitted bundles fill it. So again the threefold pieces on cycles have one parity, and a threefold piece of the other parity lies on no cycle.
It remains to treat L i at the shift σ − 3 + i , i = 1 , ⋯ , 5 . The cycles through threefold pieces are C 5 through all five, C 1234 and C 2345 , each a chain of up edges followed by one run. The only path of Γ from L i to L i + 2 is L i → L i + 1 → L i + 2 : a path leaving the chain of up edges enters a run, which ends at least three shifts below its start, at L 1 or L 2 , and from there the up edges lead only through line bundles already visited. By Lemma 19(i) applied to the three triples, δ vanishes on one of the edges L 1 → L 2 , L 2 → L 3 , on one of L 2 → L 3 , L 3 → L 4 , and on one of L 3 → L 4 , L 4 → L 5 . An effective cycle has exactly one edge outside Γ δ . So C 5 , with two such edges, is not effective; C 1234 is effective only when its edge outside Γ δ is L 2 → L 3 , and then δ 12 ≠ 0 , so that η 23 = 0 by Lemma 19(ii); and C 2345 is effective only when that edge is L 3 → L 4 , with δ 23 ≠ 0 , so that η 34 = 0 . No effective cycle carries a nonzero component of η on its edge outside Γ δ .
Each case contradicts Proposition 29. So the criterion fails, and since Theorem 20 gives dim Ext 2 ≥ r ( γ ) with equality excluded, dim Ext 2 > r ( γ ) .    □
Lemma  22
(The shifts along a cycle form an interval). The shifts of the line-bundle pieces along a directed cycle of Γ, or along a path of Γ between two line-bundle pieces the second of which is at the higher shift, take every value between their least and their greatest. Hence, if the shifts of the line-bundle pieces of E do not form an interval of integers, the line bundles split along a gap into two groups such that every cycle of Γ through a threefold piece, and every path a → b → c of Lemma 19, has its line bundles in one group.
Proof. 
Between consecutive line-bundle pieces of a cycle or a path the shift rises by exactly 1, along an up edge, or falls, along a down edge or a run; from the least shift to the greatest it therefore passes through every intermediate value. The three line bundles of Lemma 19 are at three consecutive shifts.    □
Lemma  23
(A sharper second rule). Let a = L l a [ H ] , b = L l b [ t + 1 ] , c = L l c [ t + 2 ] be line-bundle pieces with l a < l b < l c , suppose that exactly one of δ a b , δ b c vanishes, and let N be a set of edges of Γ on which every η with D ( η ) = c ( v ) , v ∈ T D h , vanishes. Suppose that(H1) Γ δ has no path from a to c, and(H2)every path of Γ from a to c other than a → b → c has at least two edges outside Γ δ , or exactly one, lying in N. Then η vanishes on the edge of a → b → c outside Γ δ .
Proof. 
By (H1) and Lemma 18, c ( v ) a c = 0 , so D ( η ) a c = 0 . A term of D ( η ) a c runs along a path of Γ δ from a to x, the component η x y , and a path of Γ δ from y to c. If x = y , or x → y lies in Γ δ , or the walk repeats a vertex, the walk contains a path of Γ δ from a to c, which (H1) excludes; Γ δ is acyclic, so a repeated vertex shortcuts to such a path. So every nonzero term runs along a path of Γ from a to c with exactly one edge outside Γ δ , carrying η , and by (H2) the terms along paths other than a → b → c vanish. What is left is ± m 2 ( δ a b , η b c ) if δ b c = 0 , or ± m 2 ( η a b , δ b c ) if δ a b = 0 , a product of two sections of line bundles on the integral variety A 0 one of which is nonzero; so the other vanishes.    □
Lemma 19(ii) is the case N = ∅ in which a → b → c is the only path of Γ . Edges found by Lemma 23 may be added to N and the lemma applied again.
Remark  9
(Six, seven and eight line bundles). The rules used in the proof of Theorem 23 depend on the shifts of the line bundles and the order of their exponents alone. The program of the repository described in Appendix D runs through every arrangement of n ≤ 8 line bundles whose shifts form an interval, builds the graph of up edges, down edges and runs, admitting every excess and every parity of threefold shift that the degrees of Lemmas 15 and 16 allow for a run, and, for every set of up edges on which δ is taken to vanish that Lemma 19(i) admits, records the parities of the threefold pieces on the cycles with at most one edge in that set whose edge in the set is not forced to carry η = 0 by Lemma 19(ii). Runs and down edges are taken to carry nonzero components, and the condition that the edge outside Γ δ point downwards in the one-sided order is dropped; both choices enlarge the set of cycles counted, so an arrangement for which no choice reaches both parities is one for which Proposition 29 and Lemma 20 exclude every twisting cochain. The counts are: 75 arrangements of four line bundles and 541 of five, none reaching a parity; 4683 of six and 47 293 of seven, of which 4 and 140 reach one parity and none both; and 545 835 of eight, of which 31 reach both. By Lemma 22 an arrangement whose shifts do not form an interval splits into arrangements of fewer line bundles, each contributing the parities it reaches, and with at most eight line bundles at most one part has six or more, so at most one parity is reached. Hence, in the Setting 5, every convolution with at most seven line-bundle pieces fails the criterion for every polarisation θ h , and so does every convolution with eight line bundles outside the thirty-one arrangements; those have their eight line bundles at four to seven consecutive shifts with at most three at one shift, the simplest being L l 1 [ 0 ] , L l 2 [ 1 ] , L l 3 [ 2 ] , L l 4 [ 3 ] , L l 5 [ 3 ] , L l 6 [ 4 ] , L l 7 [ 5 ] , L l 8 [ 6 ] with l 1 < ⋯ < l 8 . These two statements are the output of the enumeration and rest on it; Theorem 23 rests on nothing computed.
A finer enumeration reduces the thirty-one to two. It replaces Lemma 19(ii) by Lemma 23 and treats the support of δ on runs as a variable. Call an up edge, down edge or run between two line bundleslivewhen δ is nonzero on it, on every edge of the run in the last case, and let G be the set of live edges. If E meets the criterion then: (C1) G has no directed cycle, because Γ δ has none; (C2) for line bundles a, c two shifts apart the paths of G from a to c are not exactly one path a → b → c of up edges, by Lemma 19(i); (C3) the least set Z of dead up edges closed under Lemma 23, read on G, carries η = 0 ; and (C4) for each parity of shift some cycle through a run admitting that parity has exactly one edge outside G, and that edge is not in Z , by Lemma 20 and Proposition 29. Each condition follows from the lemma cited, since paths of Γ δ between line bundles project to paths of G and paths of G lift to walks of Γ δ . Running through all 2 m sets G, 15 ≤ m ≤ 22 , for each of the thirty-one arrangements, twenty-nine admit no G satisfying (C1) to (C4), and two remain:
L l 1 [ 0 ] , L l 2 [ 1 ] , L l 3 [ 1 ] , L l 4 [ 2 ] , L l 5 [ 1 ] , L l 6 [ 2 ] , L l 7 [ 2 ] , L l 8 [ 3 ] and L l 1 [ 0 ] , L l 2 [ 1 ] , L l 3 [ 1 ] , L l 4 [ 2 ] , L l 5 [ 2 ] , L l 6 [ 3 ] , L l 7 [ 3 ] , L l 8 [ 4 ] ,
l 1 < ⋯ < l 8 , with 165 and 269 admissible sets G. Both have rank ∑ ( − 1 ) t i = 0 , so the eight arrangements of rank ± 2 , where Proposition 32 below leaves only the products, are excluded, and so is the arrangement displayed above. The same procedure for nine line bundles, over 7 087 261 arrangements whose shifts form an interval, leaves 101, each one of the two above with a ninth line bundle added; by Lemma 22 a non-interval arrangement of nine leaves only one of the two with a ninth bundle across a gap. So the rules of this subsection cannot bound the number of line bundles: in both remaining arrangements every two line bundles two shifts apart are joined by two or more paths of up edges, Lemma 23 never applies, and a typical admissible G is a complex of sections L l 1 → L l 2 ⊕ L l 3 → L l 4 ⊕ L l 5 → L l 6 ⊕ L l 7 → L l 8 with the threefold pieces on runs L l 6 , L l 7 → O Z ⊗ L k → O Z ′ ⊗ L k ′ → L l 1 and L l 8 → O Z ⊗ L k → O Z ′ ⊗ L k ′ → L l 2 , L l 3 , each with one dead edge carrying η. Deciding them needs the products H 0 ( A 0 , L m ) ⊗ H 0 ( Z , L k ) → H 0 ( Z , L m + k ) and the higher products along these runs. Two independent implementations in julia/linebundlefine.jl, sharing only the table of runs, give the same counts.

6.8. The Support of an Object Meeting the Criterion

The results of this subsection do not count pieces. They start from one consequence of the criterion and read off where an object meeting it can be supported. Let E be a perfect complex on A 0 whose Chern character has the polarised shape with respect to θ h and which meets the criterion, dim Ext 2 ( E , E ) = 57 , so that σ E is injective. For v ∈ T D h the obstruction ob E ( v ) = v ⌟ At ( E ) ∈ Ext 2 ( E , E ) to deforming E along v is sent by σ E to v ⌟ ch ( E )  [11], which vanishes because T D h annihilates θ h and W; so
ob E ( v ) = 0 for every v ∈ T D h .
Everything below is drawn from (14). The open subsets U ⊆ A 0 are Zariski open; inside the proofs we pass to the analytic spaces, which changes no Ext group and no Atiyah class on the projective A 0  [107], and the cohomology with supports and the Chern characters with supports are those of the analytic spaces. For a closed subset Z of an open U we write H Z q ( U , F ) and H Z q ( F ) for the cohomology and the local cohomology sheaves with supports in Z [52]. Two facts are used throughout: if Z is a submanifold of codimension c and F is locally free then H Z q ( F ) = 0 for q ≠ c (depth below c, the Čech complex of c local equations above c), so that H Z c + p ( U , F ) = H p ( U , H Z c ( F ) ) ; and if C ⊆ Z is closed of codimension at least c + 1 then H C 0 ( H Z c ( F ) ) = 0 , because this sheaf embeds into H C c ( F ) , which vanishes. So a section of H Z c ( F ) that vanishes outside C vanishes.
For the threefolds we use the coordinates of Appendix B. Fix ζ ∈ μ 4 3 and T = Z ζ . The linear forms n m = z m + ζ m z 3 + m on the universal cover cut out the lift of T; their differentials d n m , and the d z m for m ≤ 3 , are translation-invariant forms on A 0 and together a frame of Ω A 0 1 . On a tubular neighbourhood N of T the n m themselves are single-valued, and π ( z ) = z − ∑ m ζ ¯ m n m ( z ) ∂ 3 + m retracts N onto T with z m ∘ π = z m ; the pair ( n , π ) identifies N with T × D 3 , and a holomorphic function on V × ( D * ) 3 , for V ⊆ T open and D * the punctured disc, is a Laurent series in n 1 , n 2 , n 3 whose coefficients are holomorphic on V. The local cohomology H T 3 ( O ) is computed by the Čech complex of the cover of N ∖ T by the three open sets { n m ≠ 0 } , all of whose intersections are Stein; so on T the sheaf H T 3 ( O ) is the sheaf of Laurent tails ∑ a , b , c ≥ 1 g a b c n 1 − a n 2 − b n 3 − c with g a b c ∈ O T , and H T 3 ( F ) = H T 3 ( O ) ⊗ F for F locally free. The part of H T 3 ( O ) annihilated by the ideal I T is O T · [ 1 / n 1 n 2 n 3 ] . We write ρ T : H T 3 ( O ) → O T for the map sending the class of ∑ g a b c n 1 − a n 2 − b n 3 − c to g 111 , a morphism of sheaves of vector spaces on T with ρ T ( g [ 1 / n 1 n 2 n 3 ] ) = g | T for g holomorphic near T, extended coefficientwise to H T 3 ( Ω q ) in the frame above. For a coherent sheaf G on U supported in T its multiplicity μ T ( G ) is the length of G at the generic point of T, the sum over k ≥ 0 of the ranks of the O T -modules I T k G / I T k + 1 G ; and μ T ( E | U ) = ∑ q ( − 1 ) q μ T ( H q ( E | U ) ) when every H q ( E | U ) is supported in T.
Lemma  24
(The Chern character with supports). Let U be open, Z ⊆ U closed, and F a perfect complex on U whose cohomology sheaves are supported in Z.
(i)
The trace R H o m ( F , F ) → O U factors through a unique morphism tr Z : R H o m ( F , F ) → R Γ Z O U . Setting ch k Z ( F ) = tr Z ( At ( F ) k ) / k ! ∈ H Z k ( U , Ω k ) defines classes that forget to ch k ( F ) , are compatible with restriction to open subsets and with enlarging Z, are additive in exact triangles, and satisfy
v ⌟ ch k + 1 Z ( F ) = 1 k ! tr Z ob F ( v ) ∘ At ( F ) k ∈ H Z k + 2 ( U , Ω k )
for v ∈ H 1 ( U , T U ) , where ob F ( v ) = v ⌟ At ( F ) . In particular ob F ( v ) = 0 implies v ⌟ ch k + 1 Z ( F ) = 0 for every k.
(ii)
For T = Z ζ and every open U meeting T the class [ H ] : = ch 3 T ( O T ∩ U ) in H 0 ( U , H T 3 ( Ω 3 ) ) is
[ H ] = c d n 1 ∧ d n 2 ∧ d n 3 n 1 n 2 n 3 ,
with one constant c ≠ 0 , the same for all sixty-four threefolds.
(iii)
If the cohomology sheaves of F are supported in T ∩ U then ch 3 T ( F ) = μ T ( F ) [ H ] . If they are supported in a union Z of threefolds T i of the sixty-four, any two meeting in dimension at most two, then ch 3 Z ( F ) = ∑ i μ T i ( F ) [ T i ] Z , where [ T i ] Z is the image of [ T i ] in H Z 3 ( U , Ω 3 ) .
Proof. 
(i) Let j : U ∖ Z → U . In the triangle R Γ Z O → O → R j * O U ∖ Z the third term receives no nonzero morphism from a complex K whose cohomology sheaves are supported in Z, since Hom ( K , R j * O U ∖ Z ) = Hom ( K | U ∖ Z , O U ∖ Z ) = 0 ; so Hom ( K , R Γ Z O ) → Hom ( K , O ) is bijective. The complex R H o m ( F , F ) has cohomology supported in Z, which gives tr Z and its uniqueness; the same bijection applied to the sources F ∨ ⊗ F ⊗ Ω k transports every identity of the trace to tr Z : its compatibility with restriction to open subsets, with the morphism R Γ Z → R Γ Z ′ for Z ⊆ Z ′ , its additivity in triangles, and its graded cyclicity. The Atiyah class At ( F ) ∈ Ext 1 ( F , F ⊗ Ω 1 ) is compatible with restriction and with triangles, so the classes ch k Z have the listed properties, and cyclicity applied to the k + 1 factors of tr Z ( At k + 1 ) gives v ⌟ tr Z ( At k + 1 ) = ( k + 1 ) tr Z ( ( v ⌟ At ) ∘ At k ) , which is the displayed identity.
(ii) The trace is O U -linear and At ( O T ) 3 is a morphism between O T -modules, so [ H ] is annihilated by I T on every open subset and is a section of O T ∩ U [ 1 / n 1 n 2 n 3 ] ⊗ Ω 3 . By the compatibility with restriction it suffices to take U = A 0 , where the coefficients are holomorphic functions on the compact connected T, hence constants: [ H ] = ∑ a g a [ 1 / n 1 n 2 n 3 ] ω a with g a ∈ C and ω a running through the wedge products of three elements of the frame d n 1 , d n 2 , d n 3 , d z 1 , d z 2 , d z 3 . For 0 < λ < 1 the map φ λ ( z ) = z − ( 1 − λ ) ∑ m ζ ¯ m n m ( z ) ∂ 3 + m is a biholomorphism of N onto a smaller tubular neighbourhood of T, fixes T pointwise, and satisfies n m ∘ φ λ = λ n m , z m ∘ φ λ = z m and φ λ * O T = O T . The Atiyah class, the trace and the local cohomology are natural under biholomorphisms onto open subsets, and a section of H T 3 ( Ω 3 ) is determined by its restriction to any neighbourhood of T; so φ λ * [ H ] = [ H ] . If ω a contains a of the d n m , then φ λ * multiplies g a [ 1 / n 1 n 2 n 3 ] ω a by λ a − 3 ; so g a = 0 unless a = 3 , and [ H ] = c [ d n 1 ∧ d n 2 ∧ d n 3 / n 1 n 2 n 3 ] . The automorphisms z 3 + m ↦ i e m z 3 + m of A 0 permute the sixty-four threefolds transitively and carry the forms n m , d n m of one to those of another, so c does not depend on ζ . Forgetting supports sends [ H ] to ch 3 ( O T ) ∈ H 3 ( A 0 , Ω 3 ) , which is the class of T by the Riemann–Roch theorem for the embedding of T [44, Ch. 15] (the Chern character defined by the Atiyah class is the usual one on the projective A 0  [11]); and [ Z ζ ] θ 3 = 48 ≠ 0 by Lemma A28, so c ≠ 0 .
(iii) Both identities are between sections of a sheaf, so they may be checked locally on U and, by the second fact recalled above, off a closed subset of codimension at least four. For the first, ch 3 T is additive and F is an iterated extension of its shifted cohomology sheaves, so let G be a coherent sheaf supported in T; near a point of T some I T k G vanishes, and G is an iterated extension of the O T -modules G k = I T k G / I T k + 1 G , of ranks r k with ∑ k r k = μ T ( G ) . Each G k is locally free of rank r k off a proper closed analytic subset of T, where it is locally O T r k and ch 3 T ( G k ) = r k [ H ] by additivity and (ii). For the second, let C be the union of the pairwise intersections of the T i , of codimension at least four. On U ∖ C the T i are disjoint, R Γ Z = ⨁ i R Γ T i by the Mayer–Vietoris triangle for supports, and F ≅ R Γ Z F = ⨁ i R Γ T i F , a sum of perfect complexes with cohomology supported in the T i , whose multiplicities along T i are those of F ; so there ch 3 Z ( F ) = ∑ i μ T i ( F ) [ T i ] Z by additivity and the first identity.    □
Theorem  24
(Locality). Let E be as above, let T = Z ζ , and let U ⊆ A 0 be open with T ∩ U ≠ ∅ , every cohomology sheaf of E | U supported in T, and μ T ( E | U ) ≠ 0 . Then the union S of the components of T ∖ U of codimension one in T is a nonempty ample divisor on T. In particular, if T is an irreducible component of the support of E along which E has nonzero multiplicity, then the other components of the support cut T in a set whose divisorial part is a nonempty ample divisor.
The case of a shifted line bundle on T ∩ U is the case μ T = ± 1 with one cohomology sheaf.
Proof. 
Let v ∈ T D h . Since T is a subgroup, its normal bundle is O T ⊗ N 0 with N 0 = T 0 A 0 / T 0 T , and we write κ T ( v ) ∈ H 1 ( T , O T ) ⊗ N 0 for the image of v under H 1 ( A 0 , T A 0 ) → H 1 ( T , T A 0 | T ) → H 1 ( T , N T ) , and κ T ( v ) m ∈ H 1 ( T , O T ) for its image under d n m : N 0 → C ; on a cover of A 0 by open sets U a with v represented by vector fields v a b , the class κ T ( v ) m is represented by the functions d n m ( v a b ) | T . By (14) and Lemma 24(i), (iii),
0 = v ⌟ ch 3 T ( E | U ) = μ T ( E | U ) v ⌟ [ H ]
in H T 4 ( U , Ω 2 ) = H 1 ( T ∩ U , H T 3 ( Ω 2 ) ) . The right-hand side is represented by the cocycle v a b ⌟ [ H ] = c [ ( ν 1 d n 2 ∧ d n 3 − ν 2 d n 1 ∧ d n 3 + ν 3 d n 1 ∧ d n 2 ) / n 1 n 2 n 3 ] with ν m = d n m ( v a b ) , and ρ T sends it to c ( κ T ( v ) 1 d n 2 ∧ d n 3 − κ T ( v ) 2 d n 1 ∧ d n 3 + κ T ( v ) 3 d n 1 ∧ d n 2 ) restricted to T ∩ U . As μ T ( E | U ) c ≠ 0 , every κ T ( v ) m restricts to zero on T ∩ U ; the d n m are a basis of N 0 ∨ , so κ T ( v ) lies in the kernel of the restriction H 1 ( T , O T ) ⊗ N 0 → H 1 ( T ∩ U , O T ) ⊗ N 0 , for every v ∈ T D h .
We compute that kernel. Let C be the union of the components of T ∖ U of codimension at least two in T, so that T ∩ U = ( T ∖ S ) ∖ C . Since T is smooth, O T has depth at least two along C, the local cohomology H C 1 ( T ∖ S , O ) vanishes, and H 1 ( T ∖ S , O ) → H 1 ( T ∩ U , O ) is injective; so the kernel is that of H 1 ( T , O ) → H 1 ( T ∖ S , O ) , which is zero when S = ∅ . Let S ≠ ∅ , let H ⊆ T be the connected component of the stabiliser { x ∈ T : S + x = S } , an abelian subvariety, and let q : T → T ′ = T / H . Each component of S is H-invariant, H being connected, so S = q − 1 ( S ′ ) for the divisor S ′ = q ( S ) of T ′ , whose stabiliser is the finite group of components of the stabiliser of S; hence S ′ is ample [88, §6, Application 1] and T ′ ∖ S ′ is affine. By Poincaré reducibility there is an abelian subvariety H ′ ⊆ T with H × H ′ → T an isogeny, and the composite H × H ′ → T → T ′ is the projection to H ′ followed by the isogeny H ′ → T ′ . So the preimage of T ∖ S in H × H ′ is H × U ′ with U ′ ⊆ H ′ the preimage of the affine T ′ ∖ S ′ under a finite map, which is affine. Pullback along the finite étale map H × U ′ → T ∖ S is injective on H 1 ( O ) , since composed with the trace it is multiplication by the degree, and by the Künneth formula and H 1 ( U ′ , O ) = 0 ,
H 1 ( H × U ′ , O ) = H 1 ( H , O ) ⊗ H 0 ( U ′ , O ) .
The composite H 1 ( T , O ) → H 1 ( H × U ′ , O ) factors through H 1 ( H × H ′ , O ) = H 1 ( H , O ) ⊕ H 1 ( H ′ , O ) , an isomorphism for the isogeny, followed by the restriction to H × U ′ , which is injective on the first summand and zero on the second. So the kernel is the second summand, which is q * H 1 ( T ′ , O ) : the classes of the invariant ( 0 , 1 ) -forms on T vanishing on the tangent space of H. Hence κ T ( v ) ( u ¯ ) = 0 for every v ∈ T D h and every u ∈ T 0 H , where H = T when S = ∅ .
Finally we compute κ T ( v ) in the coordinates of Appendix B. Write v = ∑ k , l v l k d z ¯ k ⊗ ∂ l , so that v replaces d z l by ∑ k v l k d z ¯ k . The K-linear maps have v l k ≠ 0 only when one of k , l is at most 3 and the other at least 4, and
v ⌟ θ h = i 2 ∑ k < l ( h l v l k − h k v k l ) d z ¯ k ∧ d z ¯ l
vanishes exactly when h l v l k = h k v k l for all k < l ; so T D h is the nine-dimensional space of the v with v 3 + m , j = Y m j and v j , 3 + m = ( h 3 + m / h j ) Y m j for j , m ≤ 3 , Y any complex 3 × 3 matrix, and the other entries zero. The threefold T = Z ζ is { z j + ζ j z 3 + j = 0 , j = 1 , 2 , 3 } , with tangent space spanned by w j = ζ j ∂ j − ∂ 3 + j and normal space N 0 with basis f m = [ ∂ 3 + m ] , so that [ ∂ m ] = ζ ¯ m f m . The class κ T ( v ) is that of the invariant form ∑ k , l v l k d z ¯ k | T ⊗ [ ∂ l ] , and on w ¯ j = ζ ¯ j ∂ ¯ j − ∂ ¯ 3 + j , where d z ¯ j ( w ¯ j ) = ζ ¯ j and d z ¯ 3 + j ( w ¯ j ) = − 1 ,
κ T ( v ) ( w ¯ j ) = ∑ m = 1 3 ζ ¯ j Y m j − h 3 + j h m ζ ¯ m Y j m f m .
Let 0 ≠ u = ∑ j u j w j ∈ T 0 T with u a ≠ 0 , choose b ≠ a , and let Y have the entry Y a b = 1 and all others 0. Then
κ T ( v ) ( u ¯ ) = ζ ¯ b u ¯ b f a − h 3 + a h b u ¯ a f b ≠ 0 .
So T 0 H = 0 : S is nonempty, H is trivial, S = S ′ has finite stabiliser, and S is ample.    □
Theorem 24 says where the rest of the support must cut a threefold of nonzero multiplicity; the next results say where it can lie. Fix ζ with ζ 1 ζ 2 ζ 3 = ± 1 , let T = Z ζ , fix j ∈ { 1 , 2 , 3 } with { j , k , l } = { 1 , 2 , 3 } , and let ζ ( j ) be ζ with the sign of the j-th coordinate changed, T ′ = Z ζ ( j ) . On the j-th factor B j the curves C ζ j and C − ζ j meet in the set P j of the four points of order two of C ζ j ( | 2 ζ j | 2 = 4 by Lemma A28), so
S : = T ∩ T ′ = P j × C ζ k × C ζ l
is a disjoint union of four surfaces, and Z = T ∪ T ′ is cut out near S by n j n j ′ = n k = n l = 0 , where n j ′ = z j − ζ j z 3 + j is the j-th linear form of T ′ . For y ∈ C ζ k × C ζ l let Γ y = C ζ j × { y } ⊆ T and Γ y ′ = C − ζ j × { y } ⊆ T ′ ; each meets S in the four points P j × { y } . On C ζ j one has z 3 + j = − ζ ¯ j z j , so n j ′ | Γ y = 2 z j is a coordinate vanishing at the four points and d n j ′ | Γ y = 2 d z j is a nonzero invariant form; on C − ζ j likewise n j | Γ y ′ = 2 z j and d n j | Γ y ′ = 2 d z j .
Lemma  25
(The curves through the four points). Let y = ( y k , y l ) ∈ C ζ k × C ζ l with y k ∉ C ξ for ξ ≠ ζ k and y l ∉ C ξ for ξ ≠ ζ l , which excludes finitely many values of each coordinate. The threefolds Z ξ , ξ ∈ μ 4 3 , meeting Γ y or Γ y ′ are those with ξ k = ζ k and ξ l = ζ l ; among the thirty-two with ξ 1 ξ 2 ξ 3 = ± 1 they are T and T ′ alone.
Proof. 
Z ξ ∩ Γ y = ( C ξ j ∩ C ζ j ) × ( { y k } ∩ C ξ k ) × ( { y l } ∩ C ξ l ) ; the first factor contains the origin, and y k ∈ C ξ k for ξ k ≠ ζ k only for the finitely many y k ∈ C ξ k ∩ C ζ k . The same holds for Γ y ′ . The condition ξ k = ζ k , ξ l = ζ l , ξ 1 ξ 2 ξ 3 = ± 1 leaves ξ j = ± ζ j .    □
Theorem  25
(Two threefolds meeting along a surface). Let E be as above and let U ⊆ A 0 be open such that every cohomology sheaf of E | U is supported in Z = T ∪ T ′ and Γ y ∪ Γ y ′ ⊆ U for general y ∈ C ζ k × C ζ l . Then μ T ( E | U ) = μ T ′ ( E | U ) = 0 .
Proof. 
Let μ = μ T ( E | U ) and μ ′ = μ T ′ ( E | U ) , so that by Lemma 24(iii)
ch 3 Z ( E | U ) = μ [ H ] Z + μ ′ [ T ′ ] Z .
Let v ∈ T D h be the vector whose matrix Y, in the notation of the proof of Theorem 24, is E j k , the matrix with the single entry Y j k = 1 . By (14) and Lemma 24(i) the class v ⌟ ch 3 Z ( E | U ) vanishes in H Z 4 ( U , Ω 2 ) . The Mayer–Vietoris sequence for supports
H S 4 ( U , Ω 2 ) → H T 4 ( U , Ω 2 ) ⊕ H T ′ 4 ( U , Ω 2 ) → H Z 4 ( U , Ω 2 )
is exact, so some s ∈ H S 4 ( U , Ω 2 ) maps to the pair
( μ v ⌟ [ H ] , ± μ ′ v ⌟ [ T ′ ] ) .
Near S the forms d n j , d n j ′ , d n k , d n l , d z k , d z l are a frame of Ω 1 , the functions n j , n j ′ , n k , n l are transverse coordinates to S, and the coordinates z k , z l of T restrict to coordinates on each component of S; as S is smooth of codimension four, H S 4 ( U , Ω 2 ) = H 0 ( S ∩ U , H S 4 ( Ω 2 ) ) , and s is a Laurent tail
s = ∑ α , β , γ , δ ≥ 1 c α β γ δ n j − α n j ′ − β n k − γ n l − δ
in each coefficient of the frame of Ω 2 , the c α β γ δ holomorphic functions on S ∩ U . The map H S 4 ( U , Ω 2 ) → H T 4 ( U , Ω 2 ) = H 1 ( T ∩ U , H T 3 ( Ω 2 ) ) forgets the support: since R Γ S = R Γ S R Γ T it is the connecting map of 0 → H T 3 ( Ω 2 ) → j * H T 3 ( Ω 2 ) | T ∖ S → H S 1 ( H T 3 ( Ω 2 ) ) → 0 , which sends the tail s, read as a section over ( T ∩ U ) ∖ S of H T 3 ( Ω 2 ) whose coefficients have poles along S, to its class. Applying ρ T to that class gives the class in H 1 ( T ∩ U , O T ) of the principal parts ∑ β c 1 β 11 n j ′ − β along S, and symmetrically for T ′ with ∑ α c α 111 n j − α .
We compare the coefficients of d n j ∧ d n l . In v ⌟ [ T ′ ] this coefficient is zero, because [ T ′ ] = c [ d n j ′ ∧ d n k ∧ d n l / n j ′ n k n l ] contains no d n j ; in v ⌟ [ H ] it is ± c [ d n k ( v ) / n j n k n l ] , and ρ T sends μ v ⌟ [ H ] to ± μ c κ T ( v ) k . Let c α β γ δ be the coefficients of d n j ∧ d n l in s. Restricting to the curve Γ y ′ ⊆ U , y general, the T ′ side gives that the class of the principal parts ∑ α c α 111 ( p ) n j − α at the four points p ∈ P j × { y } vanishes in H 1 ( Γ y ′ , O ) ; by Serre duality its pairing with the holomorphic form d n j | Γ y ′ , the sum of the residues ∑ p c 1111 ( p ) , is zero. Restricting to Γ y ⊆ U , the T side gives that the class of the principal parts ∑ β c 1 β 11 ( p ) n j ′ − β equals ± μ c κ T ( v ) k | Γ y in H 1 ( Γ y , O ) . Pair both with d n j ′ | Γ y . With the normalisation in which a class represented by a Dolbeault form ϕ pairs with ω to ( 2 π i ) − 1 ∫ ϕ ∧ ω , a class given by principal parts g p pairs to ∑ p Res p ( g p ω ) up to a sign independent of the class; so the left side pairs to ± ∑ p c 1111 ( p ) = 0 . The class κ T ( v ) k is represented by the invariant form ∑ b ( v k b + ζ k v 3 + k , b ) d z ¯ b | T , since d n k ( ∂ k ) = 1 , d n k ( ∂ 3 + k ) = ζ k and d n k kills the other ∂ m ; on Γ y only d z ¯ j and d z ¯ 3 + j = − ζ j d z ¯ j survive, and the entries of v for Y = E j k are v k , 3 + j = h 3 + j / h k and v 3 + k , j = 0 among those that occur, so
κ T ( v ) k | Γ y = − ζ j h 3 + j h k d z ¯ j | Γ y ,
1 2 π i ∫ Γ y κ T ( v ) k ∧ d n j ′ = − ζ j h 3 + j π i h k ∫ C ζ j d z ¯ j ∧ d z j ≠ 0 .
Hence μ = 0 . The coefficients of d n j ′ ∧ d n l give μ ′ = 0 in the same way, with the roles of T and T ′ exchanged: there v ⌟ [ H ] has coefficient zero, the restriction to Γ y paired with d n j ′ gives ∑ p c 1111 ′ ( p ) = 0 for the coefficients c ′ of that component, and on Γ y ′ , where d z ¯ 3 + j = ζ j d z ¯ j , the class κ T ′ ( v ) k | Γ y ′ = ζ j ( h 3 + j / h k ) d z ¯ j pairs with d n j to a nonzero number.    □
Proposition  30
(Where the rest of the support lies). Let E be as above, let ζ 1 ζ 2 ζ 3 = ± 1 , and suppose that T = Z ζ lies in the support of E with μ T ( E | U T ) ≠ 0 , where U T is the complement of the components of the support other than T. Then for each j ∈ { 1 , 2 , 3 } there is an irreducible component R j of the support of E of dimension at least three, not one of the thirty-two threefolds Z ξ with ξ 1 ξ 2 ξ 3 = ± 1 , which meets Γ y ∪ Γ y ′ for every y ∈ C ζ k × C ζ l ; and R j ∩ ( T ∪ T ′ ) has an irreducible component that maps onto C ζ k × C ζ l .
Proof. 
Let R be the union of the components of the support not contained in T ∪ T ′ and U = A 0 ∖ R . Then every cohomology sheaf of E | U is supported in T ∪ T ′ and μ T ( E | U ) = μ T ( E | U T ) ≠ 0 , so by Theorem 25 the curves Γ y ∪ Γ y ′ are not contained in U for general y: the closed set R ∩ ( T ∪ T ′ ) projects onto a dense, hence the whole, subset of C ζ k × C ζ l under the proper projection ( z k , z l ) . Some irreducible component R j of R then has R j ∩ ( T ∪ T ′ ) projecting onto C ζ k × C ζ l , so one of its components does, and R j meets every Γ y ∪ Γ y ′ . If dim R j ≤ 2 then R j = R j ∩ ( T ∪ T ′ ) ⊆ T ∪ T ′ , which is excluded; so dim R j ≥ 3 . By Lemma 25 no Z ξ with ξ 1 ξ 2 ξ 3 = ± 1 other than T and T ′ meets Γ y or Γ y ′ for general y.    □
Corollary  13
(Objects supported on the thirty-two threefolds). No perfect complex E on A 0 whose support lies in the union of the thirty-two threefolds Z ξ , ξ 1 ξ 2 ξ 3 = ± 1 , has a Chern character of polarised shape with respect to some θ h and meets the criterion. In particular no convolution of pieces O Z ξ ⊗ L k [ s ] over these threefolds, at any twists and shifts, does so; the same holds for the thirty-two threefolds with ξ 1 ξ 2 ξ 3 = ± i .
Proof. 
Suppose E has polarised shape and meets the criterion. For each ξ with ξ 1 ξ 2 ξ 3 = ± 1 and each j, let U be the complement of the thirty threefolds other than Z ξ and Z ξ ( j ) . Every cohomology sheaf of E | U is supported in Z ξ ∪ Z ξ ( j ) , and Γ y ∪ Γ y ′ ⊆ U for general y by Lemma 25; so μ Z ξ ( E ) = 0 by Theorem 25. Two distinct threefolds meet in dimension at most two, so Lemma 24(iii) with U = A 0 and Z the union of the thirty-two gives ch 3 Z ( E ) = ∑ ξ μ Z ξ ( E ) [ Z ξ ] Z = 0 , and forgetting supports, ch 3 ( E ) = 0 . But ch 3 ( E ) = N ω + c 3 θ h 3 with N ω ≠ 0 , and ω ∈ W and θ h 3 are linearly independent, W having the weights ± 2 ( e 1 + e 2 + e 3 ) and θ h 3 weight zero (Lemma 20). A convolution of pieces over the threefolds is supported in their union. The automorphism z 4 ↦ i z 4 of A 0 preserves θ , every θ h and W, and exchanges the two families of thirty-two threefolds.    □
Corollary 13 treats one family at a time; the threefolds of both families together need one more relation, read off from a line of four of them. Fix j ∈ { 1 , 2 , 3 } with { j , k , l } = { 1 , 2 , 3 } and ξ k , ξ l ∈ μ 4 . For a ∈ μ 4 let T a = Z x with x j = a , x k = ξ k , x l = ξ l ; two of the four lie in each family. Let C a = { z j + a z 3 + j = 0 } ⊆ B j , let Q = ⋃ a C a , a divisor on B j , and let Π = C ξ k × C ξ l ⊆ B ′ = B k × B l , a surface, so that A 0 = B j × B ′ and
Z : = ⋃ a ∈ μ 4 T a = Q × Π .
Two of the T a meet in ( C a ∩ C b ) × Π , a union of surfaces, so Lemma 24(iii) applies to Z. The divisor Q is ample: the connected component H of its stabiliser { x ∈ B j : Q + x = Q } is an abelian subvariety which, being connected, leaves each C a invariant, so H ⊆ C 1 ∩ C i , a finite set, and Q has finite stabiliser [88, §6, Application 1]. Hence B j ∖ Q is affine.
Lemma  26
(The four threefolds through a surface). Let E be as above and let U ⊆ A 0 be open such that every cohomology sheaf of E | U is supported in Z and Q × { y } ⊆ U for general y ∈ Π . Then the multiplicities μ a = μ T a ( E | U ) satisfy
∑ a ∈ μ 4 μ a = 0 , ∑ a ∈ μ 4 a μ a = 0 ,
that is, μ a = m a 2 for one integer m: μ 1 = μ − 1 = − μ i = − μ − i . In particular, if one of the four threefolds carries no cohomology sheaf of E | U , none of them has nonzero multiplicity.
Proof. 
Let n a = z j + a z 3 + j , n k = z k + ξ k z 3 + k , n l = z l + ξ l z 3 + l and f = ∏ a n a , so that near Z the set Z is cut out by f = n k = n l = 0 , a regular sequence, Q = { f = 0 } in B j and Π = { n k = n l = 0 } in B ′ . As for one threefold, the pair ( n k , n l ) and the retraction along ∂ 3 + k , ∂ 3 + l identify a tubular neighbourhood of Π in B ′ with Π × D 2 , and the local cohomology H Z 3 ( F ) of a locally free sheaf F, computed by the Čech complex of the cover of the complement of Z by { f ≠ 0 } , { n k ≠ 0 } , { n l ≠ 0 } , is the sheaf on Z of Laurent tails ∑ β , γ ≥ 1 g β γ n k − β n l − γ whose coefficients g β γ are sections of H Q 1 ( F ) : over W × V , with W ⊆ B j and V ⊆ Π open and Stein, a coefficient is a section of F on ( W ∖ Q ) × V modulo the sections on W × V . In particular H C a 1 ( F ) ⊆ H Q 1 ( F ) , and [ T a ] = c [ d n a ∧ d n k ∧ d n l / n a n k n l ] Z has the single tail n k − 1 n l − 1 with coefficient c [ d n a / n a ] ∧ d n k ∧ d n l , where [ d n a / n a ] ∈ H 0 ( C a , H C a 1 ( Ω B j 1 ) ) is the class of the form d n a / n a with a pole along C a .
Let v ∈ T D h have the matrix Y = E j k , as in the proof of Theorem 25: then v replaces d z 3 + j by d z ¯ k and d z k by ( h 3 + j / h k ) d z ¯ 3 + j and kills the other d z m , so d n a ( v ) = a d z ¯ k , d n k ( v ) = ( h 3 + j / h k ) d z ¯ 3 + j and d n l ( v ) = 0 . By (14) and Lemma 24(i), (iii),
0 = v ⌟ ch 3 Z ( E | U ) = ∑ a μ a v ⌟ [ T a ] Z = c ∑ a μ a a d z ¯ k d n k ∧ d n l − ( h 3 + j / h k ) d z ¯ 3 + j d n a ∧ d n l n a n k n l Z
in H Z 4 ( U , Ω 2 ) = H 1 ( Z ∩ U , H Z 3 ( Ω 2 ) ) , the classes of the invariant ( 0 , 1 ) -forms d z ¯ k , d z ¯ 3 + j in H 1 ( U , O ) acting by cup product. The decomposition Ω A 0 2 = Ω B j 2 ⊕ ( Ω B j 1 ⊠ Ω B ′ 1 ) ⊕ Ω B ′ 2 into locally free summands splits the local cohomology; d n k ∧ d n l lies in the third summand and d n a ∧ d n l in the second, so the second summand of the identity reads
[ d z ¯ 3 + j ] ∪ ∑ a μ a d n a ∧ d n l n a n k n l Z = 0
in H 1 Z ∩ U , H Z 3 ( Ω B j 1 ⊠ Ω B ′ 1 ) . Reading off the coefficient of d n l n k − 1 n l − 1 in the frame d n k , d n l , d z k , d z l of Ω B ′ 1 is a morphism of sheaves of vector spaces H Z 3 ( Ω B j 1 ⊠ Ω B ′ 1 ) → H Q 1 ( Ω B j 1 ) ⊠ O Π on Z, linear over the functions pulled back from B j ; restricting to Q × { y } ⊆ Z ∩ U for general y, where the ( 0 , 1 ) -form d z ¯ 3 + j restricts to the same form on B j , gives
[ d z ¯ 3 + j ] ∪ ∑ a μ a [ d n a / n a ] = 0
in H 1 ( Q , H Q 1 ( Ω B j 1 ) ) = H Q 2 ( B j , Ω B j 1 ) . Since B j ∖ Q is affine, H 1 ( B j ∖ Q , Ω 1 ) = 0 , and the exact sequence H 1 ( B j ∖ Q , Ω 1 ) → H Q 2 ( B j , Ω 1 ) → H 2 ( B j , Ω 1 ) shows that forgetting the support is injective on H Q 2 ( B j , Ω 1 ) ; it is compatible with the cup product by H 1 ( B j , O ) . The argument of Lemma 24(ii) in codimension one, with the scaling z ↦ z − ( 1 − λ ) a ¯ n a ( z ) ∂ 3 + j of B j , which fixes C a , multiplies n a by λ and preserves z j , shows that ch 1 C a ( O C a ) = c 1 [ d n a / n a ] in H 0 ( C a , H C a 1 ( Ω 1 ) ) with a constant c 1 , the same for the four curves since z 3 + j ↦ i z 3 + j permutes them and carries n a to n i a , and c 1 ≠ 0 since forgetting the support gives ch 1 ( O C a ) = [ C a ] ≠ 0 . So the image of [ d n a / n a ] in H 1 ( B j , Ω 1 ) is c 1 − 1 [ C a ] , and
[ d z ¯ 3 + j ] ∪ ∑ a μ a [ C a ] = 0 in H 2 ( B j , Ω 1 ) = H 1 , 2 ( B j ) .
By Lemma A28(i) with ( u , v ) = ( z j , z 3 + j ) , [ C a ] = i 2 ( d z j + a d z 3 + j ) ∧ ( d z ¯ j + a ¯ d z ¯ 3 + j ) , so
[ C a ] ∧ d z ¯ 3 + j = i 2 d z j + a d z 3 + j ∧ d z ¯ j ∧ d z ¯ 3 + j ,
and the invariant forms d z j ∧ d z ¯ j ∧ d z ¯ 3 + j and d z 3 + j ∧ d z ¯ j ∧ d z ¯ 3 + j are linearly independent in H 1 , 2 ( B j ) . Their coefficients ∑ a μ a and ∑ a a μ a vanish. Separating real and imaginary parts of the second relation gives μ 1 = μ − 1 and μ i = μ − i , and the first then gives μ i = − μ 1 ; so μ a = μ 1 a 2 , and if some μ a vanishes all do.    □
Theorem  26
(Objects supported on the sixty-four threefolds). No perfect complex E on A 0 whose support lies in the union of the sixty-four threefolds Z x , x ∈ μ 4 3 , has a Chern character of polarised shape with respect to some θ h and meets the criterion. In particular no convolution of pieces O Z x ⊗ L k [ s ] over threefolds of both families, at any twists and shifts, does so.
Proof. 
Suppose E has polarised shape and meets the criterion, and write μ x = μ Z x ( E ) . Fix j and ( ξ k , ξ l ) as above and let U be the complement of the sixty threefolds Z x ′ with ( x k ′ , x l ′ ) ≠ ( ξ k , ξ l ) . Every cohomology sheaf of E | U is supported in Z = ⋃ a T a , and μ T a ( E | U ) = μ x for x j = a , the generic point of T a lying in U. For y = ( y k , y l ) ∈ Π with y k ∉ C ξ for ξ ≠ ξ k and y l ∉ C ξ for ξ ≠ ξ l , which excludes finitely many values of each coordinate, Z x ′ ∩ ( Q × { y } ) = ( C x j ′ ∩ Q ) × ( { y k } ∩ C x k ′ ) × ( { y l } ∩ C x l ′ ) is empty unless ( x k ′ , x l ′ ) = ( ξ k , ξ l ) ; so Q × { y } ⊆ U for general y, and Lemma 26 gives that μ x / x j 2 does not depend on x j when x k , x l are fixed. This holds for j = 1 , 2 , 3 , so μ x / ( x 1 2 x 2 2 x 3 2 ) = μ x / σ ( x ) 2 is independent of each coordinate of x separately, hence constant: μ x = m σ ( x ) 2 for one integer m. Two distinct threefolds meet in dimension at most two, so Lemma 24(iii) with U = A 0 and Z the union of the sixty-four, followed by forgetting supports and (A11) with m = 2 there, gives
ch 3 ( E ) = ∑ x μ x [ Z x ] = m ∑ ω ∈ μ 4 ω 2 S ( ω ) = m ∏ j = 1 3 T 2 , j = 0 .
But ch 3 ( E ) = N ω + c 3 θ h 3 ≠ 0 , as in the proof of Corollary 13. A convolution of pieces over the threefolds is supported in their union.    □
The theorem contains Corollary 13, since a complex supported on one family has μ x = 0 on the other, hence m = 0 ; the corollary is kept for its proof, which localises to two threefolds and is what Proposition 30 uses. The relations of Lemma 26 hold along every line of four threefolds in the support, whatever else the support contains, as long as the rest of the support misses Q × { y } for general y.
Theorem 26 removes every convolution built from threefold pieces alone, of one family or of both, at any twists and shifts; it leaves the convolutions whose support reaches outside the sixty-four threefolds. For these the next two results say what the rest of the support can be. By (A10) the Weil class is ω = ( p + q i ) α + ( p − q i ) α ¯ with p , q ∈ Q ; for x ∈ μ 4 3 let σ ( x ) = x 1 x 2 x 3 and c ( x ) = ( x 1 2 , x 2 2 , x 3 2 ) ∈ { ± 1 } 3 . The fibres of c are the eight orbits of the group { ± 1 } 3 acting on μ 4 3 by sign changes; σ ( x ) 2 = ∏ j c ( x ) j , so the four orbits with ∏ j c j = 1 make up the family σ = ± 1 and the other four the family σ = ± i .
Lemma  27
(The four orbits). Let ν : μ 4 3 → Q satisfy ∑ x ν x [ Z x ] ≡ N ω modulo θ h H 4 ( A 0 , C ) . Then for every c ∈ { ± 1 } 3
∑ c ( x ) = c ν x σ ( x ) = N p / 4 , ∏ j c j = 1 , − i N q / 4 , ∏ j c j = − 1 .
So if N p ≠ 0 each orbit of the family σ = ± 1 carries an x with ν x ≠ 0 , and if N q ≠ 0 each orbit of the family σ = ± i does.
Proof. 
For ε ∈ { ± 1 } 3 let X ε = ∏ j ( c j if ε j = 1 , c ¯ j if ε j = − 1 ) , of weight 2 ε ; X ε spans the part of H 6 ( A 0 , C ) of that weight, H 4 ( A 0 , C ) has no class of weight ± 2 ε , and θ h has weight zero, so the component of weight 2 ε of θ h H 4 is zero and α = X ( 1 , 1 , 1 ) , α ¯ = X ( − 1 , − 1 , − 1 ) . In [ Z x ] = ∏ j ( a j + b j + x ¯ j c j + x j c ¯ j ) the coefficient of X ε is ∏ ε j = 1 x ¯ j ∏ ε j = − 1 x j = σ ( x ) ∏ ε j = 1 x ¯ j 2 = σ ( x ) ∏ ε j = 1 c ( x ) j , as x j 4 = 1 . Writing W c = ∑ c ( x ) = c ν x σ ( x ) , the coefficient of X ε in ∑ x ν x [ Z x ] is ∑ c W c χ ε ( c ) with χ ε ( c ) = ∏ ε j = 1 c j , and the χ ε are the eight characters of { ± 1 } 3 . The hypothesis says that this coefficient is N ( p + q i ) for ε = ( 1 , 1 , 1 ) , N ( p − q i ) for ε = ( − 1 , − 1 , − 1 ) and 0 otherwise. Fourier inversion gives 8 W c = N ( p + q i ) c 1 c 2 c 3 + N ( p − q i ) , which is the claim.    □
Proposition  31
(A smooth complete intersection cannot carry the rest). Let E be as above and suppose that its support lies in Y ∪ ⋃ σ ( x ) = ± 1 Z x , where Y ⊆ A 0 is a smooth complete intersection of dimension d ≤ 5 of divisors in the linear systems | L h m i | of powers of the polarisation L h . Then:
(i)
N p ≠ 0 , and every orbit of the family σ = ± 1 contains a threefold Z x ⊆ Y along which E has nonzero multiplicity.
(ii)
If Y meets every Z x in dimension at most one, then E does not exist: no such complex has polarised shape and meets the criterion.
Proof. 
(i) Let Y 1 , ⋯ , Y c be the divisors, c = 6 − d ≥ 1 . Each A 0 ∖ Y i is affine and smooth, as is every intersection of them, so has vanishing homology above degree six [2]; the Mayer–Vietoris spectral sequence of the cover of A 0 ∖ Y by these c open sets gives H q ( A 0 ∖ Y ) = 0 for q > 5 + c , and Lefschetz duality H q ( A 0 , Y ) ≅ H 12 − q ( A 0 ∖ Y ) gives H q ( A 0 , Y ) = 0 for q ≤ d . So H 2 d − 6 ( A 0 , Q ) → H 2 d − 6 ( Y , Q ) is onto, as 2 d − 6 < d . For η ∈ H 2 d − 6 ( A 0 ) the projection formula gives i * i * η = η ∪ [ Y ] = ∏ i m i θ h c η , so i * H 2 d − 6 ( Y ) ⊆ θ h H 4 ( A 0 ) . Now let Σ = Y ∪ ⋃ x ∈ I Z x , I the set of x with σ ( x ) = ± 1 and Z x ⊆ Supp E , Z x ⊆ Y . By the Riemann–Roch theorem for singular varieties [44, Ch. 18], applied to the cohomology sheaves of E , which are coherent sheaves on infinitesimal neighbourhoods of Σ , the class ch 3 ( E ) ∈ H 6 ( A 0 , Q ) is the image of a class τ ∈ A 3 ( Σ ) Q (the Todd class of A 0 being 1). The localisation sequence A 3 ( Y ) → A 3 ( Σ ) → A 3 ( Σ ∖ Y ) → 0 is exact, A 3 ( Σ ∖ Y ) is free on the classes of the three-dimensional components Z x ∖ Y , x ∈ I , and the component of τ there is ∑ x ∈ I μ Z x ( E ) [ Z x ∖ Y ] , the top-dimensional part of the Riemann–Roch class of a sheaf being its cycle. Hence
ch 3 ( E ) − ∑ x ∈ I μ Z x ( E ) [ Z x ] ∈ i * H 2 d − 6 ( Y ) ⊆ θ h H 4 ( A 0 ) ,
and ch 3 ( E ) = N ω + c 3 θ h 3 with θ h 3 ∈ θ h H 4 . Lemma 27 applies to ν x = μ Z x ( E ) for x ∈ I and ν x = 0 otherwise: since ν vanishes on the family σ = ± i , N q = 0 , so N p ≠ 0 as N ω ≠ 0 , and each orbit of the family σ = ± 1 contains x ∈ I with μ Z x ( E ) ≠ 0 .
(ii) Take x from (i), T = Z x . The multiplicity of E along T on the complement of the other components equals μ Z x ( E ) ≠ 0 , so Proposition 30 gives a component R 1 of the support, of dimension at least three and not among the thirty-two threefolds, hence contained in Y, such that R 1 ∩ ( Z x ∪ Z x ( 1 ) ) contains a surface. But Y ∩ ( Z x ∪ Z x ( 1 ) ) has dimension at most one.    □
The hypotheses of Proposition 31(ii) hold, for instance, when the rest of the support is a smooth curve or a smooth surface cut out by divisors of the polarisation and meeting the threefolds in finitely many points; the smooth divisors Y meet every Z x in a surface and are only constrained by (i) and by Theorem 24. When the support is large the localised obstruction says nothing at all:
Proposition  32
(Where the localised obstruction stops). Let E be as above and suppose that its support contains an ample effective divisor Y. Then v ⌟ ch k + 1 Z ( E | U ) = 0 for every v ∈ T D h , every k ≥ 0 , every open U ⊆ A 0 and every closed Z ⊆ U containing the support of E | U . In particular a convolution of line-bundle pieces L l i [ t i ] and threefold pieces with ∑ i ( − 1 ) t i ≠ 0 has support A 0 , and the criterion for it is decided by its products alone.
Proof. 
Let Σ be the support of E and X = A 0 ∖ Y , an affine variety, so that H 1 ( X , T X ) = 0 and H k + 1 ( X , Ω k ) = 0 for k ≥ 0 . The restriction of v to X vanishes, so the image of v ⌟ ch k + 1 Σ ( E ) under H Σ k + 2 ( A 0 , Ω k ) → H Σ ∖ Y k + 2 ( X , Ω k ) , which is v | X ⌟ ch k + 1 Σ ∖ Y ( E | X ) , vanishes; by the exact sequence
H Y k + 2 ( A 0 , Ω k ) → H Σ k + 2 ( A 0 , Ω k ) → H Σ ∖ Y k + 2 ( X , Ω k )
the class v ⌟ ch k + 1 Σ ( E ) comes from a class b ∈ H Y k + 2 ( A 0 , Ω k ) . The sequence
H k + 1 ( X , Ω k ) → H Y k + 2 ( A 0 , Ω k ) → H k + 2 ( A 0 , Ω k )
shows that b is determined by its image in H k + 2 ( A 0 , Ω k ) , which is v ⌟ ch k + 1 ( E ) = 0 . So v ⌟ ch k + 1 Σ ( E ) = 0 , and its restriction to U, followed by the enlargement of supports from Σ ∩ U to Z, is v ⌟ ch k + 1 Z ( E | U ) . A complex of rank ∑ i ( − 1 ) t i ≠ 0 has a cohomology sheaf of positive rank, whose support is A 0 .    □
What remains at E 0 6 . By Theorem 26 every convolution of threefold pieces fails, of one family or of both, at any twists and shifts, which covers the threefolds at several twists of Setting 5; Theorem 23 and Remark 9 cover the convolutions with at most seven line bundles, eight line bundles outside two arrangements, and nine outside the arrangements containing one of the two. Left are the convolutions in those arrangements and with ten or more line bundles, with or without threefold pieces at any twists. Both remaining eight-bundle arrangements have rank zero, so the support is a proper closed subset containing every threefold piece; Theorem 24 and Propositions 30 and 31 constrain it, and the products decide the rest. The products along the chains of Proposition 25 and along the cycles through the line bundles are computable in the theta-function model of Section A.4, where the fourfold products of line-bundle pieces at distance at least two were found to give dim Ext 2 ≥ 69 ; we have not computed them, and (F2) at E 0 6 stands or falls with them.

6.9. Weil Sixfolds with a Split Polarisation

Markman proved that the Weil classes of every polarised abelian sixfold of split Weil type are algebraic [78,79], for every imaginary quadratic field. The discriminant is an invariant of the polarisation, not of the abelian variety with its action of K, and the Weil classes depend only on the latter. This subsection uses that freedom. Throughout, K = Q ( − d ) is imaginary quadratic and N = N K / Q . For an abelian 2 n -fold ( A , η ) of Weil type for K, a polarisation h is K-compatible if η ( k ) * h = N ( k ) h ; it defines a K-hermitian form H h on H 1 ( A , Q ) of signature ( n , n ) , whose determinant det H h ∈ Q × / N ( K × ) is the discriminant of ( A , η , h ) and has the sign ( − 1 ) n . By Landherr’s theorem [69] hermitian forms over K are classified by dimension, signature and determinant, so ( A , η , h ) is of split Weil type, H h having an isotropic K-subspace of dimension n, if and only if det H h = ( − 1 ) n . The Weil plane W ( A ) = ⋀ K 2 n H 1 ( A , Q ) depends on ( A , η ) only. Every ( A , η ) of Weil type has K-compatible polarisations: if E is a Riemann form and a = η ( − d ) , then E ( x , y ) + d − 1 E ( a x , a y ) is a Riemann form, since a commutes with the complex structure, and it satisfies E ′ ( a x , y ) = − E ′ ( x , a y ) .
The split condition is a condition on the trace form. For a K-hermitian form H write q H ( x ) = H ( x , x ) , a quadratic form over Q of dimension 2 dim K V , and 〈 〈 x 〉 〉 = 〈 1 , − x 〉 .
Theorem  27
(Discriminant and trace form). Let H be a K-hermitian form of K-dimension 2 n and signature ( n , n ) , and B H = ( − d , ( − 1 ) n det H ) . The following are equivalent: (a) H is split; (b) det H = ( − 1 ) n in Q × / N ( K × ) ; (c) the quaternion algebra B H is split; (d) q H is hyperbolic. The algebra B H contains K and is the Clifford invariant e 2 ( q H ) ∈ Br 2 ( Q ) . If H is not split, its Witt index is n − 1 : H ≅ ( n − 1 ) H ⊥ H 2 with H the hyperbolic plane and H 2 binary, anisotropic, of signature ( 1 , 1 ) , with ( − d , − det H 2 ) = B H .
Proof. 
Hermitian forms over K are diagonalisable with entries in Q ; if H = 〈 a 1 , ⋯ , a 2 n 〉 , then q H = 〈 1 , d 〉 ⊗ 〈 a 1 , ⋯ , a 2 n 〉 , since N ( u + v − d ) = u 2 + d v 2 . A binary form 〈 b 1 , b 2 〉 is isotropic if and only if b 1 N ( x ) + b 2 N ( y ) = 0 has a nonzero solution, that is, if and only if − b 1 b 2 ∈ N ( K × ) . A hermitian form is isotropic exactly when its trace form is. While the anisotropic part has signature ( k , k ) with k ≥ 2 , its trace form is indefinite of dimension 4 k ≥ 8 , hence isotropic by Meyer’s theorem, and a hyperbolic plane splits off; so H ≅ ( n − 1 ) H ⊥ H 2 with H 2 binary of signature ( 1 , 1 ) , and det H = ( − 1 ) n − 1 det H 2 . By the binary criterion H 2 is isotropic if and only if det H = ( − 1 ) n modulo norms, and by Witt cancellation the Witt index is n in that case and n − 1 otherwise. This proves (a)⇔(b) and the statement on the Witt index. The equivalence (b)⇔(c) is the norm residue property: ( − d , b ) splits if and only if b ∈ N ( K × ) . In the Witt ring an even-dimensional form φ of dimension 2 m is congruent modulo I 2 to 〈 〈 δ ( φ ) 〉 〉 , δ ( φ ) = ( − 1 ) m det φ  [68, Ch. II]; for φ = 〈 a 1 , ⋯ , a 2 n 〉 this gives q H = 〈 〈 − d 〉 〉 ⊗ φ ≡ 〈 〈 − d , ( − 1 ) n det H 〉 〉 modulo I 3 , and since e 2 vanishes on I 3 and e 2 ( 〈 〈 a , b 〉 〉 ) = ( a , b )  [68, Ch. V], e 2 ( q H ) = B H ; the element − d generates K inside ( − d , b ) . For (c)⇔(d): q H lies in I 2 and has signature 0, so it is hyperbolic over R ; over Q p a form in I 2 with trivial Clifford invariant lies in I 3 = 0 and is hyperbolic. By the Hasse–Minkowski theorem q H is hyperbolic if and only if e 2 ( q H ) is locally trivial everywhere, and by the local–global principle for Br ( Q ) if and only if e 2 ( q H ) = B H is split. The last statement follows by applying the formula for e 2 to H 2 , with n = 1 .    □
Corollary  14
(Constructions with a rational Lagrangian are split). A K-hermitian form whose trace form has a Lagrangian subspace defined over Q is split. In particular, for an abelian variety X, V = H 1 ( X , Q ) ⊕ H 1 ( X , Q ) ∨ with its duality pairing ( , ) V and t ∈ K with t ¯ = − t , the hermitian forms H t = − t 2 ( , ) V + t Ξ t of Markman’s construction [79, Lemma 7.1(3)] are split, for every X and every t.
Proof. 
A quadratic form with a rational Lagrangian is hyperbolic; apply Theorem 27(d)⇒(a). For H t , Ξ t is antisymmetric [79, Lemma 7.1(2)], so q H t = − t 2 ( x , x ) V with − t 2 ∈ Q > 0 , and H 1 ( X , Q ) is a Lagrangian subspace of ( , ) V .    □
So every family reached by the secant-sheaf construction is split, whatever X, Θ , q and t; a non-split family needs a source of the action of K whose hermitian form is not a multiple of the pairing of X × X ∨ . By the Witt index of Theorem 27 the Q -rank of SU ( V , H ) drops from n to n − 1 in the non-split case, which matches Brosnan’s observation that maximal unipotent degenerations exist only for split forms [23].
Remark  10
(Half-spin representations). Since q H ∈ I 2 , the even Clifford algebra is C 0 ( q H ) ≅ A × A with A central simple of dimension 2 4 n − 2 in the Brauer class B H  [68, Ch. V]; for n = 3 , C 0 ( q H ) ≅ M 16 ( B H ) × M 16 ( B H ) , whose simple modules are B H 16 , of Q -dimension 64, when B H is a division algebra and Q 32 when it splits. So the half-spin representations of Spin ( q H ) are defined over Q exactly in the split case, and the spinor formalism of [78,79], in which the even cohomology of an abelian variety is a rational half-spin representation, has no rational model for a non-split form.
Proposition  33
(Split ( 2 n + 2 ) -folds give all 2 n -folds). Fix K and n ≥ 1 . If the Weil classes are algebraic on every polarised abelian ( 2 n + 2 ) -fold of split Weil type for K, they are algebraic on every abelian 2 n -fold of Weil type for K, of every discriminant. In particular, split eightfolds for K give all sixfolds for K.
Proof. 
Let ( A 1 , η 1 , h 1 ) have dimension 2 n and discriminant δ , and put b = ( − 1 ) n δ > 0 . Let E be an elliptic curve with complex multiplication by K, and A 2 = E × E with K acting by ( α , α ¯ ) and the polarisation h 2 = p 1 * θ + b p 2 * θ ; it is an abelian surface of Weil type with hermitian form diag ( 1 , − b ) . The product A 1 × A 2 with the diagonal action and h 1 + h 2 has determinant − b δ = ( − 1 ) n + 1 δ 2 , the split class, so W ( A 1 × A 2 ) is algebraic. In the notation of Proposition A2, ω ( A ) = ω 1 ( A ) + δ K ω 2 ( A ) spans ⋀ 2 n V − and ω ( A 1 × A 2 ) = ω ( A 1 ) ω ( A 2 ) . On the surface A 2 , ω ( A 2 ) 2 ∈ ⋀ 4 V − = 0 since dim V − = 2 , so ω 1 ( A 2 ) ω 2 ( A 2 ) = 0 and ω 1 ( A 2 ) 2 = d ω 2 ( A 2 ) 2 , while ω ( A 2 ) ω ( A 2 ) ¯ = 2 ω 1 ( A 2 ) 2 spans ⋀ 2 V − ⊗ ⋀ 2 V + = H 4 ( A 2 ) ; so c = ∫ A 2 ω 1 ( A 2 ) 2 ≠ 0 . The classes ω i ( A 2 ) are rational of type ( 1 , 1 ) , hence algebraic. By (A3),
pr 1 * ω 1 ( A 1 × A 2 ) · pr 2 * ω 1 ( A 2 ) = c ω 1 ( A 1 ) , pr 1 * ω 2 ( A 1 × A 2 ) · pr 2 * ω 1 ( A 2 ) = c ω 2 ( A 1 ) ,
and the left-hand sides are algebraic. So W ( A 1 ) is algebraic.    □
By Proposition 33 the non-split sixfolds for Q ( i ) , the families D h through E 0 6 of Remark 8 among them, would follow from the split family of eightfolds through E 0 8 of Section 6.10. For a discriminant δ let D K , δ be the nine-dimensional period domain of Setting 8 for F = K , n = 3 and discriminant δ , with group G = SU ( V , H ) , and Σ K , δ ⊆ D K , δ the algebraic locus of the Weil classes.
Proposition  34
(Odd pieces). Let ( A , η ) be an abelian sixfold of Weil type for K.
(a)
If A carries one K-compatible polarisation h ′ with det H h ′ = − 1 , then W ( A ) is algebraic.
(b)
If H 1 ( A , Q ) contains a K-stable sub-Hodge structure of odd K-dimension, then A carries K-compatible polarisations of every discriminant, and W ( A ) is algebraic. This applies to A ∼ B × E with E an elliptic curve with complex multiplication by K, and to A ∼ B 1 × B 2 with K-stable threefolds of signatures ( 2 , 1 ) and ( 1 , 2 ) .
(c)
For every δ and every v ∈ V with H ( v , v ) ≠ 0 , the subdomain D v ⊆ D K , δ of the points at which K v is a sub-Hodge structure has dimension six and lies in Σ K , δ , and the union of the D v is dense in D K , δ .
Proof. 
(a) W ( A ) depends only on ( A , η ) , and ( A , η , h ′ ) is of split Weil type, so Markman’s theorem [78, Thm. 1.5.1] applies.
(b) Take a K-compatible polarisation h, let V 1 be the given piece and V 2 its H h -orthogonal complement, a K-stable sub-Hodge structure; both have odd K-dimension. Write H h = H 1 ⊕ H 2 accordingly. For c ∈ Q > 0 the corresponding Riemann form is positive on each summand, so H 1 ⊕ c H 2 is, after multiplication by a positive integer, the hermitian form of a K-compatible polarisation of A, of signature ( 3 , 3 ) and determinant det ( H h ) c dim K V 2 ≡ det ( H h ) c modulo N ( K × ) , as c 2 = N ( c ) . Since det H h < 0 , the choice c = δ ′ / det H h , positive for every δ ′ < 0 , gives the discriminant δ ′ , and δ ′ = − 1 gives a split polarisation; apply (a). On B × E the piece H 1 ( E ) has K-dimension one, and on B 1 × B 2 the piece H 1 ( B 1 ) has K-dimension three.
(c) At a point of D v , K v is a K-stable sub-Hodge structure of K-dimension one, so (b) applies. The points of D v are the Hodge structures of Weil type on v ⊥ , of signature ( 3 , 2 ) or ( 2 , 3 ) , together with a choice of type on K v , so dim D v = 3 · 2 = 6 . For g ∈ G ( Q ) , g D v = D g v , and every G ( Q ) -orbit in D K , δ is dense by Theorem A3(ii).    □
The loci of products of a Weil fourfold and a Weil surface, of dimension five, are not reached in this way, since rescaling a piece of even K-dimension does not move the discriminant; there the Weil classes are algebraic by Lemma 14. A member of D K , δ with Hodge group SU ( V , H ) has Néron–Severi rank one and no second K-compatible polarisation, so Proposition 34 ends exactly at the very general member. At the points of complex multiplication the discriminant can always be chosen (Figure 12).
Quaternionic multiplication gives a second dense union of six-dimensional loci in every non-split component, on which the Weil classes are products of divisor classes. Let B = ( − d , β ) , β ∈ Q > 0 , be an indefinite quaternion algebra containing K = Q ( μ ) , μ 2 = − d , with basis 1 , μ , j , μ j , j 2 = β , j μ = − μ j , canonical involution x ↦ x ¯ , and the positive involution τ ( x ) = μ − 1 x ¯ μ . An abelian variety A of dimension 2 m with B ⊆ End 0 ( A ) and a polarisation E whose Rosati involution restricts to τ on B is of type II in Shimura’s classification [110], [19, Ch. 9].
Theorem  28
(Weil classes on quaternionic loci). Let ( A , E ) be as above.
(i)
( A , K , E ) is a polarised abelian 2 m -fold of Weil type for K.
(ii)
If B is a division algebra, det H E = ( − β ) m in Q × / N ( K × ) ; for odd m the algebra of Theorem 27 is B H E = B , and ( A , K , E ) is not split.
(iii)
The Weil plane W ( A ) lies in the Q -span of products of m divisor classes. In particular the Weil classes of A are algebraic.
(iv)
Let m = 3 and let δ < 0 be a non-split discriminant, − δ ∉ N ( K × ) . Then B δ = ( − d , − δ ) is an indefinite division algebra, the sixfolds of (i) with B = B δ form six-dimensional special subvarieties of the nine-dimensional domain D K , δ , the union of their translates by G ( Q ) is dense in D K , δ , and the Weil classes are algebraic on all of them.
(v)
The very general member of each of these subvarieties is simple, with End 0 ( A ) = B δ and with the rational multiples of E as its only K-compatible polarisations; Proposition 34 does not apply to it.
Proof. 
(i) B ⊗ C = M 2 ( C ) acts on H 1 , 0 ( A ) , which is therefore C 2 ⊗ U ′ with dim U ′ = m ; so μ has the eigenvalues ± − d on H 1 , 0 , each with multiplicity m. Since τ ( μ ) = μ ¯ = − μ , E ( μ x , μ y ) = E ( x , − μ 2 y ) = d E ( x , y ) = N ( μ ) E ( x , y ) , so E is K-compatible.
(ii) By the classification of type II forms [110, Sect. 4], [19, Sect. 9.2], H 1 ( A , Q ) = B m with B acting by left multiplication and E ( x , y ) = ∑ a , b Tr Q trd ( x a ν a b y ¯ b μ ) for a matrix ν hermitian for the canonical involution; one checks E ( b x , y ) = E ( x , τ ( b ) y ) from τ ( b ) ¯ = μ b μ − 1 . Hermitian matrices over a division algebra with involution are diagonalisable, with diagonal entries in the symmetric elements, which are Q ; so H 1 ( A , Q ) is an orthogonal sum of m lines B with E c ( x , y ) = c trd ( x y ¯ μ ) , c ∈ Q × . On one line take the K-basis 1 , j and H ( x , y ) = E ( μ x , y ) + μ E ( x , y ) . Then H ( 1 , 1 ) = trd ( μ 2 ) = − 2 d , H ( j , j ) = trd ( μ j j ¯ μ ) = − β trd ( μ 2 ) = 2 β d , and H ( 1 , j ) = trd ( μ j ¯ μ ) + μ trd ( j ¯ μ ) = 0 , both being reduced traces of pure quaternions. So H c = c diag ( − 2 d , 2 β d ) , of signature ( 1 , 1 ) and determinant − β ( 2 c d ) 2 . Summing over the lines, det H E = ( − β ) m modulo squares, and for odd m, B H E = ( − d , ( − 1 ) m ( − β ) m ) = ( − d , β ) = B , which is not split; apply Theorem 27.
(iii) Over C , V C = H 1 ( A , C ) ∨ = C 2 ⊗ U with dim U = 2 m and B ⊗ C = M 2 ( C ) acting on the first factor. Let a be the standard alternating form on C 2 , so that a ( x u , v ) = a ( u , x ¯ v ) , and s ( u , v ) = a ( u , μ v ) ; then s ( x u , v ) = s ( u , τ ( x ) v ) , and s is symmetric because μ ¯ = − μ . Every form with E ( b x , y ) = E ( x , τ ( b ) y ) is s ⊗ e for a unique form e on U, alternating and nondegenerate because E is. The centraliser L B of B in Sp ( V , E ) is a Q -group with L B ( C ) = Sp ( U , e ) , and the Hodge group of A, which commutes with End 0 ( A ) and preserves E, lies in L B . Let u ± be eigenvectors of μ on C 2 ; the eigenspaces of K on V C are V ± = u ± ⊗ U , of dimension 2 m , and W ( A ) ⊗ C is spanned by generators of ⋀ 2 m V + ∨ and ⋀ 2 m V − ∨ . For a symplectic basis f 1 , ⋯ , f 2 m of U put Ω ± = ∑ k ≤ m ( u ± f k ) ∨ ∧ ( u ± f k + m ) ∨ = ( u ± ∨ ⊗ u ± ∨ ) ⊗ e ∨ ∈ ⋀ 2 V C ∨ . It is invariant under Sp ( U , e ) , and
Ω ± m = m ! ( u ± f 1 ) ∨ ∧ ⋯ ∧ ( u ± f 2 m ) ∨
generates ⋀ 2 m V ± ∨ . The L B ( C ) -invariants of ⋀ 2 V C ∨ are the complexification of N = ( ⋀ 2 V ∨ ) L B , which is fixed by the Hodge group and so consists of rational classes of type ( 1 , 1 ) , that is, divisor classes. Let R be the Q -span of products of m elements of N. Then W ( A ) ⊗ C ⊆ R ⊗ C , and since W ( A ) is defined over Q , W ( A ) ⊆ ( R ⊗ C ) ∩ H 2 m ( A , Q ) = R .
(iv) B δ is indefinite because − δ > 0 , and it is a division algebra because − δ is not a norm from K. Take β = − δ and m = 3 . The complex structures on V R = R 2 ⊗ U R commuting with B δ , preserving E = s ⊗ e and positive for it are 1 ⊗ J with J in the Siegel space of ( U R , e ) , s being definite on R 2 ; this is a nonempty domain of dimension 6, and each point, with a lattice stable under an order of B δ , is a polarised abelian sixfold as in (i). By (ii) its hermitian form has signature ( 3 , 3 ) and determinant − β = δ , so it is isometric to H δ by Landherr’s theorem [69], and the domain embeds in D K , δ as the set of points commuting with an embedded copy of B δ , a special subvariety. For g ∈ G ( Q ) the translate by g is the locus of g B δ g − 1 , which contains K and on which the Rosati involution of E is again of the form τ ; the union of the translates contains a G ( Q ) -orbit, which is dense by Theorem A3(ii). By (iii) the Weil classes are algebraic at every point of these loci.
(v) A proper reductive Q -subgroup H ⊂ L B through which the Hodge cocharacter of the point J factors cuts out an analytic subset of the Siegel space; it is proper, since otherwise H ( R ) would contain the normal subgroup of L B ( R ) = Sp 6 ( R ) generated by the conjugates of the Hodge cocharacter, which is all of Sp 6 ( R ) , the group being simple. There are countably many such H, so at a very general point the Hodge group is L B , and End 0 ( A ) is its commutant in End ( V ) , which is B δ since Sp ( U , e ) acts irreducibly on U. An abelian variety whose endomorphism algebra is a division algebra is simple. Rational classes in NS ( A ) Q correspond to the Rosati-symmetric elements of End 0 ( A ) ; for x = a + b μ + c j + e ′ μ j , τ ( x ) = x means x ¯ μ = μ x , which holds if and only if b = 0 , a three-dimensional space. The polarisation E ( x · , · ) is K-compatible if and only if − μ x μ = d x , that is, x μ = μ x , so x ∈ K ∩ { τ ( x ) = x } = Q . A simple non-CM A has no K-stable sub-Hodge structure of odd K-dimension and no second K-compatible polarisation, so neither part of Proposition 34 applies.    □
The generic endomorphism algebra of the type II families is that of [110, Sect. 4]; the Hodge classes of abelian varieties with quaternionic multiplication are studied by Murty [89]. For m = 2 the argument of (iii) recovers the algebraicity of the Weil classes of fourfolds with quaternionic multiplication. The identity Ω m = m ! · ( generator ) is checked mechanically for m ≤ 4 in lean/HodgeClosure.lean, and the discriminants of (ii) by julia/qmdisc.jl.
Lemma  28
(A split polarisation on a sextic CM field). Let L be a totally real field of degree six with real embeddings s 1 , ⋯ , s 6 , M = K L , and ε ∈ { ± 1 } 6 with exactly three entries − 1 . For a ∈ L × let H a ( x , y ) = Tr M / K ( a x y ¯ ) on the six-dimensional K-vector space M. There is a ∈ L × with sign s i ( a ) = ε i for all i and H a split.
Proof. 
H a ( x , y ) = H 1 ( a x , y ) , and multiplication by a has K-determinant N M / K ( a ) = N L / Q ( a ) , so det H a = N L / Q ( a ) det H 1 , with det H 1 > 0 because H 1 is positive definite. Over C , H a is the sum of the 〈 s i ( a ) 〉 , of signature ( 3 , 3 ) . By Landherr’s theorem we need a with the signs ε and
N L / Q ( a ) / c ∈ N ( K × ) , c = − 1 / det H 1 < 0 .
A rational number r is a norm from K if and only if ( r , − d ) v = 1 at every place v of Q , by the Hasse–Minkowski theorem for x 2 + d y 2 = r z 2  [108, Ch. IV, §3].
Local step. Let p be a prime that does not split in K, so that K p = K ⊗ Q p is a field, and let w be a place of L over p. If K p ⊆ L w , then N ( L w × ) N ( K p × ) = Q p × by local class field theory [92, Ch. V]: N ( L w × ) is the norm group of the maximal abelian subextension of L w / Q p , which does not contain K p , and norm groups of abelian extensions reverse inclusions. If K p ⊆ L w , then L w ⊗ Q p K p ≅ L w × L w with the involution exchanging the factors, L w × 0 is an isotropic K p -subspace of half the dimension for H a , and the block of H a at w is hyperbolic for every a. Hence: if some w ∣ p is of the first kind there is a p ∈ ( L ⊗ Q p ) × with ( N ( a p ) / c , − d ) p = 1 ; if every w ∣ p is of the second kind, H a ⊗ Q p is hyperbolic for every a, so det H a = − 1 in Q p × / N ( K p × ) and ( N ( a ) / c , − d ) p = ( − det H a , − d ) p = 1 for every a. At primes that split in K every symbol ( · , − d ) p is 1.
Global step. Let S be the set of primes dividing 2 d or at which c is not a unit, and choose a p as above for p ∈ S , a p = 1 when no choice is needed. Let m be the product of all real places of L and of a power of each prime of L over S, so high that u ≡ 1 mod × m implies N ( u ) ∈ N ( K p × ) for p ∈ S . By weak approximation choose a 0 ∈ L × with the signs ε and so close to a p at p ∈ S that ( N ( a 0 ) / c , − d ) p = 1 . Write ( a 0 ) = a S b with a S supported over S and b prime to S. Every ray class modulo m contains prime ideals [92, Ch. VII, §13], so there are a prime ideal q of L not over S and u ≡ 1 mod × m with ( u ) = q b − 1 . Put a = a 0 u . It has the signs ε , satisfies ( N ( a ) / c , − d ) p = 1 for p ∈ S , and ( a ) = a S q . Let q 0 be the prime under q . At a prime p ∉ S ∪ { q 0 } both N ( a ) and c are p-units and p ∤ 2 d , so ( N ( a ) / c , − d ) p = 1 . At the real place N ( a ) < 0 and c < 0 . By the product formula ( N ( a ) / c , − d ) q 0 = 1 as well, so (15) holds, det H a = − 1 , and H a is split.    □
Lemma  29
(A CM algebra through K). Let A be an abelian sixfold of CM type and η : K → End 0 ( A ) of Weil type. Then End 0 ( A ) contains a commutative semisimple subalgebra M ⊇ η ( K ) of dimension 12, of the form M = K ⊗ Q L with L a product of totally real fields, [ L : Q ] = 6 , and H 1 ( A , Q ) is free of rank one over M.
Proof. 
Fix a K-compatible polarisation; its Rosati involution † is positive, preserves η ( K ) and is complex conjugation on it. End 0 ( A ) is semisimple; let Z ′ be the subalgebra generated by η ( K ) and the centre, which is commutative, semisimple and †-stable, and C ′ its commutant, †-stable. As A is of CM type, End 0 ( A ) = ∏ i M n i ( F i ) with CM fields F i and ∑ i n i [ F i : Q ] = 12 ; hence C ′ = ∏ k M m k ( Z k ′ ) over the field factors Z k ′ of Z ′ with ∑ k m k [ Z k ′ : Q ] = 12 , and † is of the second kind on each factor since it is nontrivial on K. The †-symmetric elements of C ′ ⊗ R are tuples of hermitian matrices, and those with distinct eigenvalues form a nonempty Zariski open subset of the rational vector space of symmetric elements of C ′ ; so there is a rational symmetric s whose components have distinct eigenvalues, and M = Z ′ [ s ] is étale of dimension ∑ k m k [ Z k ′ : Q ] = 12 . M is †-stable and † is positive on it, so † preserves each field factor, as tr ( e e † ) > 0 for a nonzero idempotent e, and each factor is a field with a positive involution nontrivial on K, a CM field containing K. Put L = M † = 1 . Each factor M j acts on e j H 1 ( A , Q ) ≠ 0 , so dim Q e j H 1 ≥ [ M j : Q ] , with equality in total, and H 1 is free of rank one over M.    □
Corollary  15
(Markman’s theorem at points of complex multiplication). Let A be an abelian sixfold of CM type and K ⊆ End 0 ( A ) an imaginary quadratic field for which A is of Weil type. Then W ( A ) is algebraic. In particular, for every K and every discriminant δ, the algebraic locus Σ K , δ contains every point of complex multiplication of D K , δ .
Proof. 
Let M = K ⊗ L be as in Lemma 29. The idempotents of L = ∏ j L j split A, up to K-isogeny, as ∏ j A j with complex multiplication by K L j , of K-dimension [ L j : Q ] and K-signature ( p j , q j ) , ∑ j ( p j , q j ) = ( 3 , 3 ) .
Some [ L j : Q ] is odd. Then H 1 ( A j ) is a K-stable sub-Hodge structure of odd K-dimension, and Proposition 34(b) applies.
L is a field. Let Φ be the CM type. The Q -bilinear forms E on M = H 1 ( A , Q ) with E ( α x , y ) = E ( x , α ¯ y ) for α ∈ M are the forms E ξ ( x , y ) = Tr M / Q ( ξ x y ¯ ) , ξ ∈ M , by nondegeneracy of the trace; E ξ is alternating if and only if ξ ¯ = − ξ , that is ξ = a − d with a ∈ L ; and E ξ is a Riemann form if and only if, for each σ ∈ Φ , Im σ ( ξ ) has the sign fixed by the convention for E ( x , J y ) , because M ⊗ R = ∏ σ ∈ Φ C with the complex structure acting as i on each factor. Writing σ ∈ Φ over s i , Im σ ( a − d ) = s i ( a ) Im σ ( − d ) , so the Riemann forms of this shape are those with sign s i ( a ) = ± ε i , where ε i = 1 when σ | K is the fixed embedding of K and − 1 otherwise; as A is of Weil type for K, exactly three ε i are − 1 . These forms are K-compatible, and Tr M / Q ( a − d x y ¯ ) = Tr K / Q − d H a ( x , y ) , so their hermitian form is H a up to a rational scalar. By Lemma 28, applied to ± ε , one of them is split, and Proposition 34(a) applies.
L has two or three factors, all of even degree. The degrees are ( 4 , 2 ) or ( 2 , 2 , 2 ) . If every piece is balanced, p j = q j , then each A j is of Weil type for K and W ( A ) lies in the tensor product of the W ( A j ) by Lemma 14; these are algebraic, by Markman’s theorem for all abelian fourfolds of Weil type [78] for the piece of K-dimension four and by the Lefschetz theorem for the pieces of K-dimension two, whose Weil classes lie in H 2 . Otherwise some piece of K-dimension two has signature ( 2 , 0 ) or ( 0 , 2 ) ; on it the complex structure is the action of an element of K ⊗ R , every K-line is a sub-Hodge structure, and Proposition 34(b) applies.    □
Theorem  29
(The Hodge conjecture for sixfolds with complex multiplication by K L ). Let M be a CM field of degree 12 that contains an imaginary quadratic field, and A an abelian sixfold with M ⊆ End 0 ( A ) . Then the ring of Hodge classes of A is generated by divisor classes and by the Weil classes W ( A , K ′ ) of the imaginary quadratic subfields K ′ ⊆ M for which ( A , K ′ ) is of Weil type, and the Hodge conjecture holds for A.
Proof. 
Write M = K L with L = M + totally real of degree six. The Galois closure L ˜ of L is totally real and does not contain K, so the Galois group of the Galois closure of M is Z / 2 × Γ , with Γ = Gal ( L ˜ / Q ) a transitive subgroup of S 6 , acting on the twelve embeddings X = { ± } × { 1 , ⋯ , 6 } of M; complex conjugation is the nontrivial element of Z / 2 . A CM type is Φ I = { ( + , i ) : i ∈ I } ∪ { ( − , j ) : j ∉ I } for a subset I ⊆ { 1 , ⋯ , 6 } , and ( A , K ) is of Weil type exactly when | I | = 3 . The space H 2 p ( A , C ) has the basis e S , S ⊆ X , | S | = 2 p , of wedge products of eigenvectors of M, and by Pohlmann’s theorem [100] the Hodge classes are spanned by the e S with | g S ∩ Φ I | = p for every g ∈ Z / 2 × Γ . The imaginary quadratic subfields of M are K and the fields Q ( − d e ′ ) for the quadratic subfields Q ( e ′ ) of L, which correspond to the Γ -stable partitions of { 1 , ⋯ , 6 } into two triples; their Weil classes are spanned by the e S with S a section of X → { 1 , ⋯ , 6 } constant on { 1 , ⋯ , 6 } , or constant on each triple of the partition with opposite signs on the two triples. Relabelling { 1 , ⋯ , 6 } replaces ( Γ , I ) by a conjugate pair, so it suffices to take I = { 1 , ⋯ , k } , 0 ≤ k ≤ 6 , and Γ among all 279 transitive subgroups of S 6 . For each of these 1953 pairs the program cmsixfolds.jl computes the Hodge classes and finds that those in H 4 , H 8 and H 10 are products of divisor classes, and those in H 6 are spanned by products of divisor classes and the Weil classes of the imaginary quadratic subfields (Appendix C); the Hodge classes in H 2 are divisor classes by the Lefschetz theorem. Divisor classes are algebraic, and the Weil classes are algebraic by Corollary 15, applied to each K ′ .    □
For Γ = S 6 or A 6 the Hodge numbers are dim Hdg p ( A ) = 6 , 15 , 20 , 15 , 6 for p = 1 , ⋯ , 5 when | I | ∉ { 0 , 3 , 6 } , 6 , 15 , 22 , 15 , 6 when | I | = 3 , the two extra classes in degree six being W ( A , K ) , and 36 , 225 , 400 , 225 , 36 when I is empty or everything, A being then isogenous to E 6 . For M = Q ( ζ 21 ) , Q ( ζ 28 ) , Q ( ζ 36 ) , of the 64 CM types 40 have Hodge rings generated by divisors and 24, two Galois orbits of twelve, need exactly the two Weil classes of one imaginary quadratic subfield. Theorem 29 is classical for nondegenerate types; for the types of Weil type its decisive input is Markman’s theorem, through Corollary 15.
Remark  11
(McMullen’s curve). Mostaed [82] considers the intersection of McMullen’s Kobayashi-geodesic curve in the Hilbert modular sixfold of L = Q ( cos π 21 ) with the Weil loci W K , proves that it is finite and that its points, if any, are points of complex multiplication by the degree-twelve field M = K L , with Galois group Z / 2 × Z / 2 × Z / 3 , and asserts that the Weil classes there are inaccessible to the known algebraicity theorems. By Corollary 15 and Theorem 29 the Weil classes at every such point are algebraic and the Hodge conjecture holds for the sixfold; whether the intersection is nonempty is not touched.

6.10. Seven Consecutive Shifts at E 0 8

At n = 4 the same route was taken in Section A.4 on E 0 8 , E 0 = C / Z [ i ] , with the 128 line bundles L ζ , ζ ∈ μ 4 4 , ∏ ζ = ± 1 , whose signed Chern characters add up to a multiple of the Weil class, and three multiples M 1 , M 2 , M 3 of the polarisation; the criterion asks for dim Ext 2 ( E , E ) = 104 . In the setting of Lemma A11 the pieces of one parity, ∏ ζ = ϵ , have shifts of one parity and the pieces of the other parity have shifts of the other, and Corollary A3 excludes every convolution whose 128 shifts lie in six consecutive integers. We extend this to seven. The groups used below are the groups H k ( W , V ) between two pieces in the harmonic model of the convolution Lemma A5, as in Theorem A13 and Lemma A12: for pieces W, V that differ in exactly three coordinates and have d W = d V + 1 , the groups H 3 ( W , V ) , of total dimension 640 over the 28 ratios W / V of this kind, are reached by no coboundary, and a term of d E on them passes through a piece at the shift d W + 1 , d W + 2 , d V − 1 or d V − 2 .
Proposition  35
(Two separated adjacent pairs). Let E be a convolution as in Lemma A11, let g ≥ 4 , and suppose that the shifts of the pieces L ζ take exactly the four values s, s − 1 , s − g and s − g − 1 . Then for every order, every choice of the shifts of the M i and every placement of the t i , dim Ext 2 ( E , E ) ≥ 640 , and E does not meet the polarised criterion.
Proof. 
Let P, Q, R, T be the sets of pieces at the shifts s, s − 1 , s − g , s − g − 1 ; all four are nonempty. Two shifts that differ by one belong to pieces of opposite parity, so P and Q have opposite parities, and so do R and T; P and R have the same parity when g is even and opposite parities when g is odd. Let W and V differ in exactly three coordinates with d W = d V + 1 . Then W ∈ P and V ∈ Q , or W ∈ R and V ∈ T , since these are the only pairs of occupied shifts at distance one. A term of d E on H 3 ( W , V ) passes through a piece at one of the shifts s + 1 , s + 2 , s − 2 , s − 3 when W ∈ P , and s − g + 1 , s − g + 2 , s − g − 2 , s − g − 3 when W ∈ R . Since g ≥ 4 none of these is among s, s − 1 , s − g , s − g − 1 , so no term acts and the groups H 3 ( W , V ) inject into Ext 2 ( E , E ) .
Let Γ be the group of ratios of two pieces, of order 128, acting simply transitively on the pieces, and let S 3 ⊂ Γ be the 28 ratios that move exactly three coordinates, with the weight given by the dimensions of the groups H 3 ; every ratio in S 3 changes the parity. Put X = P ∪ T if g is even and X = P ∪ R if g is odd. In both cases X is a union of one set of each parity, and its complement is the union of the other two. The edges of the Cayley graph of ( Γ , S 3 ) join pieces of opposite parity, so the edges between X and its complement join P to Q and R to T, and these are exactly the pairs ( W , V ) above. Hence the dimensions of the surviving groups add up to the weight of the cut of X. The set X is neither empty nor everything, and the proof of Proposition A11 shows, through Lemma A13, that every such cut weighs at least 640.    □
For g = 4 this is Proposition A11. Together with the results of Section A.4 it settles most arrangements in seven consecutive integers.
Corollary  16
(Seven consecutive shifts). Let E be a convolution as in Lemma A11 whose 128 pieces L ζ have their shifts in { s , s − 1 , ⋯ , s − 6 } . Let U ⊆ { s , s − 2 , s − 4 , s − 6 } and U ′ ⊆ { s − 1 , s − 3 , s − 5 } be the values taken by the two parities. Then dim Ext 2 ( E , E ) ≥ 640 for every order, every choice of the shifts of the M i and every placement of the t i , except possibly when U is one of
{ s , s − 2 , s − 6 } , { s , s − 4 , s − 6 } , { s , s − 2 , s − 4 , s − 6 } ,
U ′ is one of
{ s − 1 } , { s − 3 } , { s − 5 } , { s − 1 , s − 5 } ,
and ( U , U ′ ) is neither ( { s , s − 2 , s − 6 } , { s − 5 } ) nor ( { s , s − 4 , s − 6 } , { s − 1 } ) .
Proof. 
If the shifts do not take both values s and s − 6 they lie in six consecutive integers and Corollary A3 applies; so s , s − 6 ∈ U . If U ′ contains two values two apart it is { s − 1 , s − 3 } , { s − 3 , s − 5 } or { s − 1 , s − 3 , s − 5 } , and Theorem A15 applies to the parity at U ′ for every U. Otherwise U ′ is { s − 1 } , { s − 3 } , { s − 5 } or { s − 1 , s − 5 } . Let U = { s , s − 6 } . For U ′ = { s − 3 } the two parities sit at s − 3 and at s − 3 ± 3 , and Theorem A16 gives dim Ext 2 ≥ 67584 . For U ′ = { s − 1 } , Proposition A9 with y = s − 1 applies, since the pieces at y + 1 = s exist and no piece has the shift s − 2 or s + 2 ; for U ′ = { s − 5 } its second form with y = s − 5 applies, since the pieces at s − 6 exist and no piece has the shift s − 4 or s − 8 . For U ′ = { s − 1 , s − 5 } the shifts take exactly the values s, s − 1 , s − 5 , s − 6 , and Proposition 35 applies with g = 5 . The same two applications of Proposition A9 settle U = { s , s − 4 , s − 6 } with U ′ = { s − 1 } , where no piece has the shift s − 2 or s + 2 , and U = { s , s − 2 , s − 6 } with U ′ = { s − 5 } , where no piece has the shift s − 4 or s − 8 .    □
The ten pairs ( U , U ′ ) left open in Corollary 16 are excluded, together with every other arrangement, order, placement of the t i and group action, by Theorem A9, which gives dim Ext 2 ( E , E ) ≥ 105 for every convolution of these pieces; Proposition 35, Corollary 16 and the results of Section A.4 give the larger bound 640 in the arrangements they cover.
So the route from the split eightfold E 0 8 to all sixfolds for Q ( i ) through convolutions of these line bundles is closed, with or without a group action: in every such convolution the obstruction ev E ( v 0 ) along one Weil deformation survives on the diagonal. What remains for the Weil classes of sixfolds is one of three things: a decomposition of a multiple of a Weil class into pieces in which every piece whose class is not a multiple of θ lies on a cycle of excess zero (Theorem A10 and Remark A3 for line bundles, Proposition 28 for threefold pieces, Theorem A17 for pieces of any kind), as the configuration of Proposition A12 is at the level of graphs, with the cases of Section A.5 left open; an object on a definite secant plane of an abelian fourfold to which Perry’s G-semiregularity [98], for an object that is not a simple exterior product (Theorems 33 and 34), or an affirmative answer to Markman’s Question 11.4 [79] applies, the latter holding at first order (Remark A1), since the mechanism of [79, Lemma 11.3] fails on eightfolds and the indefinite secant planes miss the polarised family (Section 6.11); or a construction at the very general non-split sixfold, which Markman’s construction does not produce (Corollary 14) and which Proposition 34 and Theorem 28 approach through dense unions of six-dimensional loci without reaching it.

6.11. Markman’s Route on Eightfolds

By Proposition 33 the Weil classes of every abelian sixfold of Weil type for K follow from those of the polarised abelian eightfolds of split Weil type for K, and Markman’s construction, applied to an abelian fourfold X, produces abelian eightfolds X × X ∨ of split Weil type (Corollary 14). At n = 3 Markman proves semiregularity through the criterion of [79, Lemma 11.3]: for the object E = Φ ( F ¯ 2 ⊠ F ¯ 1 ) built from two secant objects, the characteristic map ev E : H T 2 → Ext 2 ( E , E ) is surjective with kernel the annihilator of ch ( E ) , so that σ E is injective by (A4). We show that at n = 4 this mechanism fails on every secant plane, and that on the secant planes of an indefinite class no mechanism reaches the very general polarised Weil eightfold.
Let X be a complex torus of dimension n, V = H 1 ( X , Q ) ⊕ H 1 ( X , Q ) ∨ with the pairing ( w + ξ , w ′ + ξ ′ ) V = ξ ( w ′ ) + ξ ′ ( w ) , and S = H * ( X , Q ) , on which V acts by w + ξ ↦ w ∧ ( · ) + ξ ⌟ ( · ) , so that v v ′ + v ′ v = ( v , v ′ ) V in the Clifford algebra. A pure spinor λ ∈ S C determines the maximal isotropic subspace W λ = { v ∈ V C : v λ = 0 } , and for complementary maximal isotropic subspaces W λ , W ′ the map ⋀ * W ′ → S C , y ↦ y λ , is an isomorphism [78, Sect. 2]. The Mukai pairing is 〈 x , y 〉 = ∫ X τ ( x ) y with τ = ( − 1 ) j on H 2 j , so that χ ( F , F ′ ) = 〈 ch F , ch F ′ 〉 by Riemann–Roch, the Todd class being 1. For h ∈ H 2 ( X , Q ) write h [ k ] = h k / k ! and define Pf ( h ) by ∫ X h [ n ] = Pf ( h ) . For q ∈ Q > 0 and Pf ( h ) ≠ 0 put
λ = exp ( − q h ) , λ 1 = Re λ = ∑ k ( − q ) k h [ 2 k ] , λ 2 = Im λ q = ∑ k ( − q ) k h [ 2 k + 1 ] ,
and call B Q ( h , q ) = Q λ 1 ⊕ Q λ 2 the secant plane; an object F with ch ( F ) ∈ B Q ( h , q ) ∖ 0 is a secant object. Every c = r λ 1 + s λ 2 is a λ + a ¯ λ ¯ with a = r 2 − s − q 2 q , and a ≠ 0 for c ≠ 0 . Since 〈 λ , λ ¯ 〉 ≠ 0 (Lemma 31 for n = 4 ), W λ ∩ W λ ¯ = 0 , and V C = W λ ⊕ W λ ¯ is the eigenspace decomposition of Markman’s action of K = Q ( − q ) on X × X ∨ , whose summands V σ = W λ and V σ ¯ = W λ ¯ are sub-Hodge structures with Hodge numbers ( n , n )  [78, Sect. 2]. The subspace V 0 , 1 = H 0 , 1 ( X ) ⊕ H 0 ( X , T X ) of V C is maximal isotropic, and under the Hochschild–Kostant–Rosenberg isomorphism the contraction of H T k ( X ) = ⋀ k V 0 , 1 into H * ( X ) is Clifford multiplication. Put P = V σ ∩ V 0 , 1 and Q = V σ ¯ ∩ V 0 , 1 ; both have dimension n, and V 0 , 1 = P ⊕ Q .
Lemma  30
(Contraction ranks on a secant plane). Let n ≥ 2 and c = a λ + b λ ¯ with a , b ≠ 0 . The contraction H T k ( X ) → H * ( X , C ) , y ↦ y · c , has rank 1 for k = 0 , 2 n k for 1 ≤ k ≤ n − 1 , 1 for k = n , and 0 for k > n . In particular Ann 1 ( c ) = 0 , and for n ≥ 3 , Ann 2 ( c ) = P ∧ Q ≅ P ⊗ Q , of dimension n 2 .
Proof. 
V 0 , 1 is isotropic, so its elements anticommute in the Clifford algebra, and ⋀ k V 0 , 1 = ⨁ i + j = k ⋀ i P · ⋀ j Q , the factors commuting up to sign. Since P ⊆ W λ and Q ⊆ W λ ¯ , a product p I q J with p I ∈ ⋀ i P , q J ∈ ⋀ j Q kills λ when i ≥ 1 and λ ¯ when j ≥ 1 ; so for k ≥ 1 the image of H T k is a ⋀ k Q · λ + b ⋀ k P · λ ¯ . Applied to the complementary pair W λ , W ′ = V σ ¯ ⊇ Q , the isomorphism ⋀ * W ′ ≅ S C shows that dim ⋀ k Q · λ = n k , and symmetrically dim ⋀ k P · λ ¯ = n k . Grade S C by S m = ⋀ m V σ ¯ · λ . For w ∈ V σ ¯ , w ( y λ ) = ( w ∧ y ) λ , so the vectors killed by V σ ¯ = W λ ¯ form the line S 2 n , which contains λ ¯ ; the elements of V σ kill λ and act on S • by contraction through ( , ) V , lowering the degree by one. Hence ⋀ k Q · λ ⊆ S k and ⋀ k P · λ ¯ ⊆ S 2 n − k , and for k ≠ n the sum is direct, of dimension 2 n k . For k = n let q top and p top span ⋀ n Q and ⋀ n P . Then q top λ is killed by Q, since ⋀ n + 1 Q = 0 , and by P, which kills λ and anticommutes with Q; so it is killed by the maximal isotropic space V 0 , 1 , and so is p top λ ¯ . Both are nonzero and lie on the pure-spinor line of V 0 , 1 , so they are proportional and the rank is 1. The cases k = 0 and k > n are clear. In degree one the rank is 2 n = dim V 0 , 1 , so Ann 1 = 0 . For n ≥ 3 , dim Ann 2 = 2 n 2 − n ( n − 1 ) = n 2 , and P ∧ Q kills λ and λ ¯ , so Ann 2 = P ∧ Q .    □
Lemma  31
(The Euler form on a secant plane). Let n = 4 . For c = r λ 1 + s λ 2 = a λ + a ¯ λ ¯ ∈ B Q ( h , q ) ,
〈 c , c 〉 = 2 | a | 2 〈 λ , λ ¯ 〉 = 8 q ( q r 2 + s 2 ) Pf ( h ) , 〈 λ , λ ¯ 〉 = 16 q 2 Pf ( h ) .
So the Euler form is definite on B Q ( h , q ) , with the sign of Pf ( h ) .
Proof. 
For n = 4 the Mukai pairing is symmetric. Since h is real, τ ( λ ) = exp ( − − q h ) = λ ¯ , so 〈 λ , λ 〉 = ∫ λ ¯ λ = ∫ 1 = 0 , likewise 〈 λ ¯ , λ ¯ 〉 = 0 , and 〈 λ , λ ¯ 〉 = ∫ exp ( − 2 − q h ) = ( 2 − q ) 4 ∫ h [ 4 ] = 16 q 2 Pf ( h ) . Hence 〈 c , c 〉 = 2 a a ¯ 〈 λ , λ ¯ 〉 with | a | 2 = r 2 4 + s 2 4 q .    □
For the ideal sheaf of a point, ch = 1 − [ pt ] lies on the split plane spanned by 1 and [ pt ] , and χ = − 2 ; the class λ 2 = h − h 3 / 24 for q = 1 4 and Pf ( h ) = − 1 has 〈 λ 2 , λ 2 〉 = − 2 as well.
Theorem  30
(The Lemma 11.3 mechanism at n = 4 ). For i = 1 , 2 let X i be an abelian fourfold, G i ⊆ X i a finite subgroup acting by translations, π i : X i → Y i = X i / G i the quotient, and F ¯ i ∈ D b ( Y i ) an object with Ext < 0 ( F ¯ i , F ¯ i ) = 0 and π i * ch ( F ¯ i ) ∈ B Q ( h i , q i ) ∖ 0 . Let E = F ¯ 2 ⊠ F ¯ 1 on Y 2 × Y 1 , or its image under a derived equivalence, and suppose that
(i)
ker ev E : H T 2 → Ext 2 ( E , E ) = Ann 2 ( ch E ) , and
(ii)
ev E is surjective onto Ext 2 ( E , E ) .
Then dim Ext k ( F ¯ i , F ¯ i ) = 1 , 8 , 12 , 8 , 1 for k = 0 , ⋯ , 4 and χ ( F ¯ i , F ¯ i ) = − 2 , for i = 1 , 2 . Consequently 〈 λ i , λ ¯ i 〉 < 0 , that is Pf ( h i ) < 0 , for both i.
Proof. 
Conditions (i) and (ii) are invariant under derived equivalences, which identify Hochschild cohomology, Ext groups, characteristic maps and the action on cohomology compatibly [78, Sect. 8.3]; so we may take E = F ¯ 2 ⊠ F ¯ 1 . The isogeny π i induces isomorphisms on rational cohomology and on H T * , compatible with contraction, so Lemma 30 gives the ranks rk 0 , rk 1 , rk 2 = 1 , 8 , 12 of the contraction with ch ( F ¯ i ) on H T 0 , H T 1 , H T 2 of Y i . Write e a = dim Ext a ( F ¯ 2 , F ¯ 2 ) and e b ′ = dim Ext b ( F ¯ 1 , F ¯ 1 ) . By the Künneth formula and Ext < 0 = 0 , Ext 2 ( E , E ) = ⨁ a + b = 2 Ext a ( F ¯ 2 , F ¯ 2 ) ⊗ Ext b ( F ¯ 1 , F ¯ 1 ) with a , b ∈ { 0 , 1 , 2 } , and H T 2 ( Y 2 × Y 1 ) = ⨁ a + b = 2 H T a ⊗ H T b with ev E = ⨁ ev F ¯ 2 a ⊗ ev F ¯ 1 b  [78, Lemma 8.4.1], [55]. The contraction of H T a ⊗ H T b into ch F ¯ 2 ⊗ ch F ¯ 1 is, up to sign, the tensor product of the contractions; the contraction of H j ( ⋀ i T ) into a class of type ( p , p ) lands in type ( p − i , p + j ) , which shifts the difference of the two Hodge indices by i + j = a , so the images for different ( a , b ) are independent and the contraction on H T 2 of the product has rank ∑ a + b = 2 rk a rk b ′ . On each factor ker ev ⊆ Ann by (A4), so e a ≥ dim im ev a ≥ rk a and e b ′ ≥ rk b ′ . By (i) dim im ev E = ∑ rk a rk b ′ , and by (ii) it equals dim Ext 2 ( E , E ) = ∑ e a e b ′ . Termwise e a e b ′ ≥ rk a rk b ′ > 0 , so every term is an equality, and since all ranks are positive, e a = rk a and e b ′ = rk b ′ for a , b ≤ 2 . Serre duality on Y i , whose canonical bundle is trivial, gives dim Ext k = 1 , 8 , 12 , 8 , 1 and χ = − 2 . Finally π * O X = ⨁ ψ ∈ G ^ L ψ with torsion line bundles L ψ of Chern character 1, so χ ( π * F ¯ , π * F ¯ ) = ∑ ψ χ ( F ¯ , F ¯ ⊗ L ψ ) = | G | χ ( F ¯ , F ¯ ) , and χ ( π i * F ¯ i , π i * F ¯ i ) has the sign of Pf ( h i ) by Lemma 31.    □
The forced dimensions are 1 , 2 n , n ( n − 1 ) in degrees 0 , 1 , 2 for every n. Serre duality on a threefold asks e 2 = e 1 , that is n ( n − 1 ) = 2 n , which holds exactly at n = 3 , where Markman’s mechanism works; at n = 4 the condition is the sign of χ . Theorem 30 excludes every secant plane of an ample class, every secant plane on a very general principally polarised abelian fourfold, where h ∈ Q Θ , Markman’s genus-four examples [78, Ex. 8.2.3, 8.2.4], whose Chern characters are polynomials in Θ , and the object [ O X → ν * O Z ˜ ] ( 3 Θ ) of [79, Sect. 12] on a principally polarised fourfold, whose Chern character lies in the span of exp ( ± − d Θ ) (Lemma 32) and which has Ext < 0 = 0 : in total degree − 1 the hypercohomology spectral sequence of a two-term complex [ A → B ] receives only Hom ( B , A ) = Hom ( ν * O Z ˜ , O X ) = 0 , the first sheaf being torsion. For these objects only the weaker criterion of [79, Question 11.4] remains. The kernel half of the criterion is a statement about products.
Proposition  36
(The kernel condition). In the setting of Theorem 30, condition (i) holds if and only if, for i = 1 , 2 ,
ev 1 ( p ) · ev 1 ( q ) = 0 in Ext 2 ( F ¯ i , F ¯ i ) for all p ∈ P i , q ∈ Q i ,
the product being the Yoneda product.
Proof. 
By the proof of Theorem 30, Ann 2 of the product is ⨁ a + b = 2 ker ( c a ⊗ c b ) , with ker ( c a ⊗ c b ) = ker c a ⊗ H T b + H T a ⊗ ker c b for the contractions c a , c b of the factors. By Lemma 30, ker c 0 = ker c 1 = 0 and ker c 2 = P ∧ Q , so Ann 2 = ( P 2 ∧ Q 2 ) ⊗ 1 ⊕ 1 ⊗ ( P 1 ∧ Q 1 ) . The summands ev a ⊗ ev b land in independent Künneth summands, so ker ev E = ⨁ ker ( ev a ⊗ ev b ) ; here ker ( ev 1 ⊗ ev 1 ) = 0 since ker ev 1 ⊆ Ann 1 = 0 , and ev 0 is injective, 1 ↦ id , so ker ( ev 0 ⊗ ev 2 ) = 1 ⊗ ker ev F ¯ 1 2 and symmetrically, with ker ev F ¯ i 2 ⊆ P i ∧ Q i . So (i) holds if and only if ev F ¯ i 2 ( P i ∧ Q i ) = 0 for both i. The Hochschild–Kostant–Rosenberg isomorphism is an isomorphism of algebras H T * ≅ H H * when the Todd class is 1 [31], and the characteristic map is a map of algebras, so ev 2 ( p ∧ q ) = ev 1 ( p ) ev 1 ( q ) .    □
Markman’s displayed object of [79, Sect. 12] carries the twist O X ( Θ ) ; the twist that puts its Chern character on a secant plane is O X ( 3 Θ ) , and it is the only one.
Lemma  32
(The twist of Markman’s object). Let ( X , Θ ) be a principally polarised abelian fourfold, θ = c 1 ( O X ( Θ ) ) , d ≥ 3 odd and n = ( d + 9 ) / 2 . Let D 1 , ⋯ , D n be translates of Θ, indexed by Z / n , such that ⋂ u ∈ U D u is smooth of codimension | U | for 2 ≤ | U | ≤ 4 and empty for | U | ≥ 5 , let Z = ⋃ a D a ∩ D a + 1 , let N be a set of isolated double points of Z and ν : Z ˜ → Z the partial normalisation along N, and put F N ( k ) = [ O X → ν * O Z ˜ ] ⊗ O X ( k Θ ) . Then ch F N ( k ) lies in the span of exp ( ± − q θ ) for some q ∈ Q > 0 if and only if k = 3 , q = d and | N | = ( d + 9 ) ( 2 d − 1 ) . In that case
ch F N ( 3 ) = 1 + 3 θ − d 2 θ 2 − d 2 θ 3 + d 2 24 θ 4 = α e s θ + β e − s θ ,
with s = − d , α = 1 2 ( 1 + 3 s ) , β = 1 2 ( 1 − 3 s ) and α β = 1 4 ( 1 + 9 d ) ≠ 0 .
Proof. 
Write t = 1 − e − θ and D U = ⋂ u ∈ U D u . At a point x at most four divisors pass, their local equations f u are part of a coordinate system, and the ideal of Z a = D a ∩ D a + 1 is ( f a , f a + 1 ) or the unit ideal. In each multidegree of the polynomial ring in the f u the augmented Mayer–Vietoris complex 0 → O / ⋂ a I a → ⨁ a O / I a → ⨁ a < b O / ( I a + I b ) → ⋯ is the augmented cochain complex of a full simplex, or zero, hence exact, and the analytic local ring is flat over the polynomial ring. So [ O Z ] = ∑ J ( − 1 ) | J | − 1 [ O D U ( J ) ] , over nonempty sets J of edges of the n-cycle with U ( J ) the set of their vertices, the terms with | U ( J ) | ≥ 5 vanishing, and ch O D U = t | U | by the Koszul resolution. A single edge gives n terms + t 2 , two adjacent edges n terms − t 3 , the sets { e a , e a + 2 } and { e a , e a + 1 , e a + 2 } cancel, and two disjoint non-adjacent edges give n ( n − 5 ) / 2 terms − t 4 :
ch O Z = n t 2 − n t 3 − n ( n − 5 ) 2 t 4 .
At p ∈ N , Z is the transversal union of two smooth surfaces S 1 , S 2 with I S 1 + I S 2 = I p , so 0 → O S 1 ∪ S 2 → O S 1 ⊕ O S 2 → O p → 0 is exact and ch ν * O Z ˜ = ch O Z + | N | θ 4 / 24 . With t 2 = θ 2 − θ 3 + 7 12 θ 4 , t 3 = θ 3 − 3 2 θ 4 and t 4 = θ 4 ,
ch F N ( k ) = e k θ 1 − n θ 2 + 2 n θ 3 + g θ 4 , g = − 25 n 12 + n ( n − 5 ) 2 − | N | 24 ,
with coefficients c 0 = 1 , c 1 = k , c 2 = k 2 2 − n , c 3 = k 3 6 − k n + 2 n . Membership in the span of e ± s θ with s 2 = − q means c 2 = − q 2 c 0 , c 3 = − q 6 c 1 and c 4 = q 2 24 c 0 . The first gives q = 2 n − k 2 , and then the second reads 2 n ( 1 − k 3 ) = 0 , so k = 3 and q = 2 n − 9 = d . For k = 3 , c 4 = 27 8 − 7 n 12 + n ( n − 5 ) 2 − | N | 24 , and c 4 = d 2 24 gives | N | = 12 n 2 − 74 n + 81 − d 2 = ( 2 d − 1 ) ( d + 9 ) . Finally α + β = 1 and s ( α − β ) = 3 .    □
The number ( d + 9 ) ( 2 d − 1 ) is the one in [79, Sect. 12]; with the displayed twist k = 1 the Chern character lies on no secant plane. The isolated double points of Z number 24 · n ( n − 5 ) 2 = ( d + 9 ) ( 3 d − 3 ) , so admissible sets N exist. Markman’s question, whether F N satisfies the weaker criterion of [79, Question 11.4], does not depend on the twist (Proposition 37), so everything below applies to the displayed object as well.
Proposition  37
(Markman’s weaker criterion). Let F be an object on an abelian fourfold. Then ker ev F 2 = Ann 2 ( ch F ) , which says that σ F is injective on ev F ( H T 2 ) , the criterion of [79, Question 11.4], if and only if σ ( Ext 2 ( F , F ) ) = H T 2 ⌟ ch F . If ch F = a λ + a ¯ λ ¯ ∈ B Q ( h , q ) with a ≠ 0 and this holds, then also σ ( Ext 1 ( F , F ) ) = H T 1 ⌟ ch F , of dimension 8. Both conditions are invariant under F ↦ F ⊗ M for a line bundle M.
Proof. 
By Step 1 of the proof of Theorem 33, tr ( ev ( v ) ∘ f ) = 〈 v , σ ( f ) 〉 ϵ , where 〈 , 〉 ϵ is the Serre pairing of H T k with ⨁ p H 4 − k + p ( Ω p ) with a sign on each block; it is perfect. For v ∈ Ann 2 and w ∈ H T 2 , multiplicativity of ev and tr = σ 0 give 〈 v , σ ( ev w ) 〉 ϵ = tr ev ( v ∧ w ) = ( ( v ∧ w ) ⌟ ch F ) H 4 ( O ) = 0 , since contraction is an action of the algebra H T * . So H T 2 ⌟ ch F = σ ( ev H T 2 ) ⊆ ( Ann 2 ) ⊥ , and the two have the same dimension 28 − dim Ann 2 ; hence they are equal. If ev ( Ann 2 ) = 0 , then σ ( Ext 2 ) ⊆ ( Ann 2 ) ⊥ = H T 2 ⌟ ch F ⊆ σ ( Ext 2 ) . Conversely, if σ ( Ext 2 ) ⊆ ( Ann 2 ) ⊥ , then tr ( ev ( v ) ∘ f ) = 0 for every f ∈ Ext 2 ( F , F ) , so ev ( v ) = 0 by Serre duality, and ker ev ⊆ Ann holds always by (A4). For the second statement let I 3 = H T 1 ∧ ( P ∧ Q ) = ⋀ 2 P ⊗ Q ⊕ P ⊗ ⋀ 2 Q , of dimension 48 in H T 3 of dimension 56. Since ev ( I 3 ) = 0 , σ ( Ext 1 ) ⊆ I 3 ⊥ , of dimension 8, and it contains σ ( ev H T 1 ) = H T 1 ⌟ ch F , of dimension 8 by Lemma 30. For the twist, f ↦ f ⊗ id M identifies the Ext groups, σ F ⊗ M ( f ⊗ id ) = ch ( M ) σ F ( f ) because At ( F ⊗ M ) = At ( F ) ⊗ 1 + 1 ⊗ At ( M ) with commuting summands, and H T k ⌟ ( e c x ) = e c ( H T k ⌟ x ) for c = c 1 ( M ) , since e − c ( ∂ ⌟ ) e c = ( ∂ + ι ∂ c ) ⌟ extends to an automorphism of the algebra H T * .    □
Theorem  31
(A necessary condition for Markman’s object). In the setting of Lemma 32 let | N | = ( d + 9 ) ( 2 d − 1 ) . For p ∈ N and a divisor D i through p let π p ( i ) be the projection of T 0 X onto the line ⋂ k T p D k , over the other three divisors through p, along T p D i . If F N ( 3 ) , equivalently F N ( 1 ) , satisfies the criterion of [79, Question 11.4], then for every i
∑ p ∈ N ∩ D i π p ( i ) = ( 4 d − 2 ) id , in particular | N ∩ D i | = 16 d − 8 .
Proof. 
We use the standard properties of the Atiyah class of a perfect complex F on X × Δ , Δ a disc [60, Ch. IV–V], [11]: ch ( F ) = tr exp ( − At ( F ) ) in Hodge cohomology, additive on K 0 , multiplicative and natural, and the restriction of At ( F ) to the fibre X 0 , followed by Ω X × Δ | X 0 → Ω X , is At ( F | X 0 ) . Write F = F | X 0 , κ ∈ Ext 1 ( F , F ) for the d t -component of At ( F ) | X 0 , and δ ( F ) for the d t -component of ch ( F ) | X 0 .
Step 1: variation. With At ( F ) | X 0 = A + κ d t , A = At ( F ) , and ( κ d t ) 2 = 0 , graded cyclicity of the trace for these even elements gives δ p ( F ) = ( − 1 ) p + 1 σ p ( κ ) in H p + 1 ( Ω p ) . Moreover δ is additive, δ ( A ⊗ B ) = δ ( A ) ch ( B 0 ) + ch ( A 0 ) δ ( B ) , and if G is the family of translates of an object G by c ( t ) , c ( 0 ) = 0 , then δ ( G ) = ε ι c ′ ( 0 ) ch G for a universal sign ε , by naturality applied to ( x , t ) ↦ ( x + c ( t ) , t ) and ι being a derivation.
Step 2: moving one divisor. Fix i and a ∈ T 0 X , and move D i to D i + t a , keeping the other divisors fixed. General position is open, so it persists for small t, and the Mayer–Vietoris and Koszul expressions of the proof of Lemma 32 hold in K 0 ( X × Δ ) . A point p ∈ N ∩ D i moves holomorphically with velocity p ˙ = π p ( i ) ( a ) : differentiating the equations gives n k ( p ˙ ) = 0 for the normals n k of the other three divisors and n i ( p ˙ − a ) = 0 . Points of N off D i stay. Partial normalisation along the moving points gives a flat family F with fibre F N ( 3 ) . Put m = ε ι a θ and M i = ∑ p ∈ N ∩ D i π p ( i ) . By Step 1, δ ( O ( − D W ) ) = − [ i ∈ W ] m e − | W | θ , and ∑ i ∈ W ⊆ U ( − 1 ) | W | e − | W | θ = − e − θ t | U | − 1 , so δ ( O D U ) = m e − θ t | U | − 1 for i ∈ U . Counting the sets J with i ∈ U ( J ) as in Lemma 32, two edges, three adjacent pairs, cancelling consecutive quadruples and 2 ( n − 5 ) disjoint non-adjacent pairs, gives δ ( O Z ) = m e − θ ( 2 t − 3 t 2 + ( 1 − d ) t 3 ) , while a moving point contributes ε ι p ˙ ( θ 4 / 24 ) = 1 6 θ 3 ∧ ε ι π p ( i ) a θ . Since θ 4 ∧ H 0 , 1 = 0 ,
δ ( F ) = − e 3 θ m e − θ 2 t − 3 t 2 + ( 1 − d ) t 3 + 1 6 θ 3 ∧ ε ι M i a θ = − 2 m θ + d − 1 3 m − 1 6 ε ι M i a θ θ 3 .
Step 3: the test. For y ∈ H 1 ( O ) , y ⌟ ch = y ∧ ch , and for ∂ ∈ H 0 ( T ) , ∂ ⌟ ch = ι ∂ θ ∧ ch ′ with ch ′ = s ( α e s θ − β e − s θ ) ; ι ∂ θ runs over H 0 , 1 and α β ≠ 0 , so H T 1 ⌟ ch F N ( 3 ) = { x ′ e s θ + y ′ e − s θ : x ′ , y ′ ∈ H 0 , 1 } . By hard Lefschetz θ k ∧ is injective on H 0 , 1 for k ≤ 3 , so ∑ k ≤ 3 θ k u k , u k ∈ H 0 , 1 , lies in this space if and only if u 2 = − d 2 u 0 and u 3 = − d 6 u 1 . The signs ( − 1 ) p + 1 of Step 1 agree on p and p + 2 , so the test applies to δ ( F ) as to σ ( κ ) .
Step 4. If the criterion holds, Proposition 37 gives σ ( κ ) ∈ H T 1 ⌟ ch F N ( 3 ) . In Step 2, u 0 = u 2 = 0 , and u 3 = − d 6 u 1 reads ( d − 1 3 ) m − 1 6 ε ι M i a θ = d 3 m , that is ι ( M i − ( 4 d − 2 ) ) a θ = 0 . As θ is nondegenerate and a is arbitrary, M i = ( 4 d − 2 ) id , and taking traces of the rank-one projections gives | N ∩ D i | = 4 ( 4 d − 2 ) .    □
The average of | N ∩ D i | over i is 4 | N | / n = 16 d − 8 , so Theorem 31 is consistent with the count, and it excludes every N that is not equidistributed among the divisors. The same computation for the 24 points of a complete quadruple D U , | U | = 4 , gives ∑ p ∈ D U π p ( i ) = 6 id for each i ∈ U ; so a union of complete quadruples meets the condition of the theorem exactly when every divisor lies in ( 2 d − 1 ) / 3 of them, which with 24 | Q | = ( d + 9 ) ( 2 d − 1 ) forces d ≡ 23 mod 24 . If F N ( 3 ) fails the criterion, every object E = F N ( 3 ) ⊠ F ′ , or its image under a derived equivalence, in the setting of Theorem 30 fails condition (i) by Proposition 36, and ev E ( w ) , for w ∈ Ann 2 ( ch E ) with ev E ( w ) ≠ 0 , is a nonzero class in ker σ E invariant under every group fixing ev E ( H T 2 ) , so E is not G-semiregular for such G. Whether the equidistributed configurations satisfy the criterion for d ≥ 5 , and the weak form of G-semiregularity for these objects, remain open.
Condition (i) alone makes σ E injective on the image of ev E , and its proof imposes no dimension condition; so for the definite planes Markman’s route needs secant objects with ev ( P i ) · ev ( Q i ) = 0 together with an affirmative answer to [79], or Perry’s G-semiregularity [98], which asks for injectivity of σ on Ext 2 ( E , E ) G ; for simple exterior products, with any equivariant structure, it too is decided by the Euler form (Theorems 33 and 34). The planes with Pf ( h ) < 0 , allowed by Theorem 30, are excluded by integrality, and their eightfolds by the sign of Markman’s invariant form.
Theorem  32
(Integrality on secant planes). Let X be a complex torus of dimension four, h ∈ H 2 ( X , Q ) with Pf ( h ) ≠ 0 and q ∈ Q > 0 . Every c ∈ B Q ( h , q ) ∩ H ev ( X , Z ) satisfies 〈 c , c 〉 ∈ 8 Z , and 〈 c , c 〉 ≠ 0 if c ≠ 0 .
Proof. 
Write h = t h 0 with h 0 primitive integral and t ∈ Q × ; replacing ( h , q ) by ( h 0 , q t 2 ) does not change λ , so let h be primitive integral. By Frobenius’ normal form [19, Sect. 3.1] there are a Z -basis x 1 , y 1 , ⋯ , x 4 , y 4 of H 1 ( X , Z ) and integers 1 = d 1 ∣ d 2 ∣ d 3 ∣ d 4 with h = ∑ i d i x i y i . Then h [ k ] = ∑ | I | = k d I ( x y ) I with d I = ∏ i ∈ I d i and ( x y ) I = ∏ i ∈ I x i y i distinct members of a Z -basis of H 2 k ( X , Z ) , so h [ k ] = D k u k with u k primitive and D k = gcd I d I = d 1 ⋯ d k , and Pf ( h ) = ± D 4 . The class λ 1 = 1 − q h [ 2 ] + q 2 h [ 4 ] lives in degrees 0 , 4 , 8 and λ 2 = h − q h [ 3 ] in degrees 2 , 6 , so c = r λ 1 + s λ 2 is integral if and only if r , r q D 2 , r q 2 D 4 , s , s q D 3 ∈ Z . The two classes are orthogonal for the Mukai pairing, which pairs degree 2 j with degree 8 − 2 j , and by Lemma 31, with D 4 = D 3 d 4 ,
〈 s λ 2 , s λ 2 〉 = 8 q s 2 Pf ( h ) = ± 8 d 4 s ( s q D 3 ) , 〈 r λ 1 , r λ 1 〉 = 8 q 2 r 2 Pf ( h ) = ± 8 r ( r q 2 D 4 ) ,
both in 8 Z . The form is definite by Lemma 31, so 〈 c , c 〉 ≠ 0 for c ≠ 0 .    □
For X = E 4 , h = e 1 + e 2 + e 3 − e 4 , with Pf ( h ) = − 1 , and q = 1 4 the integral lattice is 16 Z λ 1 ⊕ 4 Z λ 2 , with 〈 c , c 〉 = − 128 and − 32 on the generators; in particular λ 2 = h − h 3 / 24 is not integral.
Corollary  17
(Untwisted objects). Let F ¯ be an object on an abelian fourfold Y with ch ( F ¯ ) ∈ B Q ( h , q ) for some h and q. Then χ ( F ¯ , F ¯ ) ∈ 8 Z . In particular, in Theorem 30 conditions (i) and (ii) never hold together, for either sign of Pf ( h i ) .
Proof. 
For a complex torus T, ch ( K top * ( T ) ) = H * ( T , Z ) , by Atiyah’s Künneth theorem [8] for K * ( ( S 1 ) 2 g ) and ch : K * ( S 1 ) ≅ H * ( S 1 , Z ) ; so ch ( F ¯ ) is integral and Theorem 32 applies. In Theorem 30, Y i is an abelian fourfold and π i * is a ring isomorphism on rational cohomology, so ch ( F ¯ i ) ∈ B Q ( ( π i * ) − 1 h i , q i ) ∩ H ev ( Y i , Z ) , and χ ( F ¯ i , F ¯ i ) ∈ 8 Z contradicts χ = − 2 .    □
Perry’s criterion asks only for injectivity of σ on invariants, with no dimension condition, and still forces the Euler characteristic when the equivariant structure is a product.
Theorem  33
(Perry’s criterion on products of secant objects). For i = 1 , 2 let X i be an abelian fourfold, F i ∈ D b ( X i ) with ch ( F i ) ∈ B Q ( h i , q i ) ∖ 0 , and H i ⊆ X i × X i ∨ a finite subgroup acting genuinely on D b ( X i ) by g = t a * ( − ) ⊗ P ξ , with an H i -equivariant structure on F i . Let E = F 2 ⊠ F 1 , or its image under a derived equivalence with the group transported, carry the restriction of the product structure to a subgroup G ⊆ H 2 × H 1 , and assume Ext < 0 ( E , E ) G = 0 . If E is G-semiregular, then for i = 1 , 2 :
(i)
dim Ext k ( F i , F i ) H i = 1 , 8 , 12 , 8 , 1 for k = 0 , ⋯ , 4 , and 0 otherwise;
(ii)
χ ( F i , F i ) = − 2 | H i | , so 4 divides | H i | and Pf ( h i ) < 0 ; in particular ± h i is not ample;
(iii)
H i is not a group of translations. In particular no exterior product of two objects whose Chern characters lie on secant planes of abelian fourfolds, and no image of one under a derived equivalence, is semiregular with vanishing negative self-extensions.
Proof. 
Fix an abelian fourfold X, an object F with c = ch F ∈ B Q ( h , q ) ∖ 0 and a group H as in the statement, and write σ p ( f ) = 1 p ! tr ( ( f ⊗ id ) ∘ At p ) and ev ( v ) = 1 p ! ( id ⊗ v ) ∘ At p for v ∈ H j ( ⋀ p T X ) , so that σ ∘ ev is the contraction with c (A4), [12,31], and ev is a map of algebras [31,55].
Step 1: adjunction. For v ∈ H j ( ⋀ p T X ) and f ∈ Ext 4 − p − j ( F , F ) one has tr ( ev ( v ) ∘ f ) = ± 〈 v , σ p ( f ) 〉 , the right side being the Serre pairing of H j ( ⋀ p T X ) with H 4 − j ( Ω X p ) : the trace commutes with v, which acts on the coefficient Ω p only, and is graded cyclic. Since both pairings are perfect, ker ( σ : Ext 4 − k ( F , F ) → H * ( X ) ) is the orthogonal of ev ( H T k ) for the Serre pairing ( e , f ) ↦ tr ( e ∘ f ) .
Step 2: invariants.H acts on Ext * ( F , F ) through the equivariant structure, and ev ( H T * ) is invariant, as in the proof of Corollary 11. The Serre pairing is invariant, since translations act trivially on H 4 ( X , O X ) and P ξ does not change traces, so it restricts to a perfect pairing between Ext k ( F , F ) H and Ext 4 − k ( F , F ) H ; with Step 1, σ is injective on Ext 4 − k ( F , F ) H if and only if ev maps H T k onto Ext k ( F , F ) H .
Step 3: Lefschetz. By character orthogonality ∑ k ( − 1 ) k dim Ext k ( F , F ) H = | H | − 1 ∑ g ∈ H str ( g | Ext * ( F , F ) ) . For g ≠ 1 the supertrace is a pairing of classes in the Hochschild homology of D b ( X ) with coefficients in the kernel of g [99, Thm. 3.1], that is in R Γ ( X × X , O Γ ⊗ P ⊗ L O Δ ) with Γ the graph of t a . For a ≠ 0 , Γ ∩ Δ = ∅ ; for a = 0 the complex has cohomology sheaves Δ * ( P ± ξ ⊗ Ω X i ) with P ± ξ ∈ Pic 0 nontrivial, and H * ( X , P ± ξ ) = 0  [88, §8]. So the supertrace vanishes and χ H ( F , F ) = χ ( F , F ) / | H | .
Step 4: the factors. Passing from G to H = H 2 × H 1 keeps G-semiregularity, since Ext 2 ( E , E ) H ⊆ Ext 2 ( E , E ) G , and derived equivalences preserve it by the functoriality of the semiregularity map [98, Rem. 2.3]; so let E = F 2 ⊠ F 1 . By Künneth, Ext m ( E , E ) H = ⨁ a + b = m Ext a ( F 2 , F 2 ) H 2 ⊗ Ext b ( F 1 , F 1 ) H 1 , and Ext < 0 ( E , E ) H = 0 gives Ext < 0 ( F i , F i ) H i = 0 , since id is invariant. On each summand σ E is ± σ ⊗ σ followed by the Künneth isomorphism, and the summands land in independent Künneth pieces, as in the proof of Theorem 30. The invariants in degrees 0 , 1 , 2 are nonzero: they contain id , ev ( H T 1 ) ≠ 0 because Ann 1 ( c ) = 0 , and ev ( H T 2 ) . So σ is injective on Ext a ( F i , F i ) H i for a = 0 , 1 , 2 .
Step 5: dimensions. Write e k = dim Ext k ( F , F ) H for F = F i . By Step 2, ev maps H T k onto the invariants for k = 2 , 3 , 4 . As σ is injective on ev ( H T 2 ) , ker ev 2 = Ann 2 ( c ) = P ∧ Q by Lemma 30, so e 2 = 28 − 16 = 12 . Since ev is multiplicative it kills the ideal generated by P ∧ Q , whose quotient in degrees 3 and 4 is ⋀ 3 P ⊕ ⋀ 3 Q and ⋀ 4 P ⊕ ⋀ 4 Q ; so e 3 ≤ 8 and e 4 ≤ 2 . Also ev 1 is injective with invariant image, so e 1 ≥ 8 , and duality gives e 1 = e 3 = 8 , e 0 = e 4 ∈ { 1 , 2 } , and e k = 0 for k > 4 .
Step 6: the Euler characteristic. χ H ( F , F ) = 2 e 4 − 16 + 12 ∈ { − 2 , 0 } . By Step 3 and Lemma 31, | H | χ H ( F , F ) = χ ( F , F ) ≠ 0 , so e 4 = 1 , which completes (i), and χ ( F , F ) = − 2 | H | , of the sign of Pf ( h ) . Corollary 17 gives χ ( F , F ) ∈ 8 Z , so 4 divides | H | ; an ample ± h would have Pf ( h ) > 0 . This is (ii). If H consists of translations, D b ( X ) H ≃ D b ( X / H )  [98, Rem. 6.6], F = π * F ¯ with Ext * ( F ¯ , F ¯ ) = Ext * ( F , F ) H , and ch F ¯ lies on a secant plane of the abelian fourfold X / H , so χ ( F ¯ , F ¯ ) = − 2 ∉ 8 Z contradicts Corollary 17; H = 1 is the last statement of (iii).    □
For an exterior product of simple objects the product hypothesis on the group can be dropped. For an abelian variety Y and g = ( a , ξ ) ∈ Y × Y ∨ write Φ g = t a * ( − ) ⊗ P ξ and Stab ( F ) = { g : Φ g F ≅ F } ; for simple F and g ∈ Stab ( F ) , conjugation c g ( e ) = ψ − 1 Φ g ( e ) ψ by an isomorphism ψ : F → Φ g F does not depend on ψ . Every equivariant structure on F for a group acting through such functors, genuine or projective, acts on Ext * ( F , F ) by these conjugations.
Lemma  33
(Structures on simple exterior products). Let Y 1 , Y 2 be abelian varieties, F i ∈ D b ( Y i ) , and E = F 2 ⊠ F 1 with Hom ( E , E ) = C . Let H ⊆ Stab ( E ) be a finite subgroup and H i its projection to Y i × Y i ∨ . Then F 1 and F 2 are simple, H i ⊆ Stab ( F i ) , ρ i ( g i ) = c g i is a representation of H i on Ext * ( F i , F i ) , and c ( g 2 , g 1 ) = ρ 2 ( g 2 ) ⊗ ρ 1 ( g 1 ) under the Künneth isomorphism. In particular Ext * ( E , E ) H ⊇ Ext * ( F 2 , F 2 ) H 2 ⊗ Ext * ( F 1 , F 1 ) H 1 .
Proof. 
By Künneth, Hom * ( G 2 ⊠ G 1 , G 2 ′ ⊠ G 1 ′ ) = Hom * ( G 2 , G 2 ′ ) ⊗ Hom * ( G 1 , G 1 ′ ) compatibly with composition up to Koszul signs. Since Hom ( F 2 , F 2 ) ⊗ Hom ( F 1 , F 1 ) ⊆ Hom ( E , E ) = C , both factors are simple. If E ′ = F 2 ′ ⊠ F 1 ′ ≅ E , then 1 = ∑ k dim Ext k ( F 2 , F 2 ′ ) dim Ext − k ( F 1 , F 1 ′ ) , so exactly one k 0 contributes, with lines spanned by u and v, and every isomorphism E → E ′ is a multiple of u ⊠ v ; comparing with the inverse, whose composite with u ⊠ v is the identity in Künneth bidegree ( 0 , 0 ) , shows that u : F 2 → F 2 ′ [ k 0 ] and v : F 1 → F 1 ′ [ − k 0 ] are isomorphisms. For g = ( g 2 , g 1 ) ∈ H , P ( ξ 2 , ξ 1 ) ≅ P ξ 2 ⊠ P ξ 1 and t ( a 2 , a 1 ) = t a 2 × t a 1 give Φ g E ≅ Φ g 2 F 2 ⊠ Φ g 1 F 1 , so Φ g 2 F 2 ≅ F 2 [ − k 0 ] ; iterating N = ord ( g ) times gives F 2 ≅ F 2 [ − N k 0 ] , so k 0 = 0 and g i ∈ Stab ( F i ) . The isomorphism E → Φ g E is then a multiple of ψ 2 ⊠ ψ 1 with ψ i of degree 0, so c g = c g 2 ⊗ c g 1 . Since c is multiplicative on Stab ( E ) and c g i depends only on g i , choosing lifts of g 2 , g 2 ′ ∈ H 2 to H gives c g 2 + g 2 ′ = c g 2 c g 2 ′ ; so ρ i is a representation, and the invariants of ρ 2 ⊠ ρ 1 under H 2 × H 1 ⊇ H are the tensor products of the invariants.    □
Theorem  34
(Perry’s criterion on simple products). For i = 1 , 2 let X i be an abelian fourfold, h i , b i ∈ NS ( X i ) Q with Pf ( h i ) ≠ 0 , q i ∈ Q > 0 , and F i ∈ D b ( X i ) with ch ( F i ) ∈ e − b i B Q ( h i , q i ) ∖ 0 . Let E = F 2 ⊠ F 1 be simple with Ext < 0 ( E , E ) = 0 , and H ⊆ Stab ( E ) a finite subgroup with projections H i . If σ E is injective on Ext 2 ( E , E ) H , then for i = 1 , 2 :
(i)
dim Ext k ( F i , F i ) H i = 1 , 8 , 12 , 8 , 1 for k = 0 , ⋯ , 4 , and 0 otherwise;
(ii)
χ ( F i , F i ) = − 2 | H i | , so Pf ( h i ) < 0 , and ± h i is not ample;
(iii)
if b i = 0 , then 4 divides | H i | , so H i ≠ 0 .
The same conclusions hold if Ψ ( E ) , for an equivalence Ψ : D b ( X 2 × X 1 ) → D b ( A ) onto an abelian variety, carries any equivariant structure for a finite group G acting on D b ( A ) through functors Φ g , g ∈ A × A ∨ , as in [98, Thm. 1.2], and is G-semiregular; then H is the image of G under Orlov’s isomorphism A × A ∨ ≅ ( X 2 × X 1 ) × ( X 2 × X 1 ) ∨  [97, Cor. 2.13].
Proof. 
For Ψ ( E ) : Ψ − 1 Φ g Ψ ≅ Φ f ( g ) for Orlov’s isomorphism f, so the structure transports to isomorphisms E ≅ Φ f ( g ) E , whose conjugations are c f ( g ) because E is simple, Ψ identifies Ext 2 ( E , E ) H with Ext 2 ( Ψ E , Ψ E ) G , and σ is compatible with Ψ  [98, Rem. 2.3]. So let E = F 2 ⊠ F 1 . The B-field b ∈ NS ( X ) Q acts on H * ( X ) by e b , and in the Clifford algebra of H 1 ⊕ H 1 , e b ξ e − b = ξ − ξ ⌟ b for ξ ∈ H 1 ; for ξ ∈ H 0 ( T X ) and b of type ( 1 , 1 ) , ξ ⌟ b ∈ H 0 , 1 . So β = Ad ( e b ) is an automorphism of V 0 , 1 and of the algebra H T * = ⋀ V 0 , 1 , and y ⌟ ( e − b c 0 ) = e − b ( β ( y ) ⌟ c 0 ) . Hence for c = e − b c 0 with c 0 ∈ B Q ( h , q ) ∖ 0 the contraction ranks are 1 , 8 , 12 , 8 , 1 , Ann 1 ( c ) = 0 , Ann 2 ( c ) = β − 1 ( P ) ∧ β − 1 ( Q ) , the quotient of H T * by the ideal it generates has dimensions 1 , 8 , 12 , 8 , 2 , and 〈 c , c 〉 = 〈 c 0 , c 0 〉 has the sign of Pf ( h ) , because 〈 e − b x , e − b y 〉 = 〈 x , y 〉 (Lemmas 30 and 31). By Lemma 33, Ext * ( F i , F i ) H i for the conjugation representations ρ i enter Ext 2 ( E , E ) H as in Step 4 of the proof of Theorem 33, and Steps 1–3 hold for ρ i : ev is natural for isomorphisms and At ( Φ g F ) = t a * At ( F ) , so ev ( H T * ) is invariant; the Serre pairing is invariant; and the Lefschetz formula [99, Thm. 3.1] needs one functor at a time, not a group action, so ∑ k ( − 1 ) k dim Ext k ( F i , F i ) H i = χ ( F i , F i ) / | H i | . Steps 5 and 6 then give (i) and χ ( F i , F i ) = − 2 | H i | , which has the sign of Pf ( h i ) ; for b i = 0 , Corollary 17 gives χ ∈ 8 Z .    □
So for simple exterior products of secant objects, on definite planes and on B-shifted planes, Perry’s criterion fails for every finite group acting through translations and twists and every equivariant structure, the projections H i being allowed to be of Heisenberg type, and on definite planes the exclusion uses the sign of χ alone. For objects that are not exterior products the same mechanism gives nothing.
Proposition  38
(Non-product objects). Let A = X 2 × X 1 with X i abelian fourfolds, c i ∈ e − b i B Q ( h i , q i ) ∖ 0 with ± h i ample, E ∈ D b ( A ) simple with Ext < 0 ( E , E ) = 0 and ch E = c 2 ⊗ c 1 , G ⊆ Stab ( E ) finite, and e k = dim Ext k ( E , E ) G . Let R = ( 1 , 16 , 88 , 208 , 274 , 208 , 88 , 16 , 1 ) and D 6 = 112 be the coefficients of p ( t ) 2 and the t 6 coefficient of p ˜ ( t ) 2 , p = 1 + 8 t + 12 t 2 + 8 t 3 + t 4 , p ˜ = p + t 4 . If E is G-semiregular, then e k = e 8 − k , e 0 = 1 , e k ≥ R k , 88 ≤ e 2 = e 6 ≤ 112 and ∑ k ( − 1 ) k e k = χ ( E , E ) / | G | > 0 ; and for every integer N some integer vector satisfies all these constraints except the last with Euler number N. If σ E is injective on all of Ext * ( E , E ) G , then e = R and χ ( E , E ) = 4 | G | .
Proof. 
Steps 1–3 of the proof of Theorem 33 hold on the eightfold A with 4 replaced by 8. The contraction with c 2 ⊗ c 1 on H T k ( A ) = ⨁ a + b = k H T a ⊗ H T b has rank ∑ a + b = k p a p b = R k , the summands landing in independent Künneth pieces, and ker ev ⊆ Ann gives e k ≥ R k . G-semiregularity gives ker ev 2 = Ann 2 = P 2 ∧ Q 2 ⊕ P 1 ∧ Q 1 and, ev being multiplicative and onto the invariants in degree 6, e 6 ≤ D 6 . χ ( E , E ) = 〈 c 2 , c 2 〉 〈 c 1 , c 1 〉 > 0 by Lemma 31. The vector R satisfies the constraints with Euler number p ( − 1 ) 2 = 4 ; raising e 4 , or e 3 = e 5 , moves the Euler number to any integer. If σ is injective on all invariants, ev is onto them in every degree with kernel Ann , so e = R .    □
For E = F 2 ⊠ F 1 as in Theorem 34 with H = H 2 × H 1 the invariant vector is also p · p = R , with Euler number ( − 2 ) ( − 2 ) = 4 : the contradiction comes from the sign of each factor, χ ( F i , F i ) < 0 < 〈 c i , c i 〉 , and at the level of the eightfold the two signs cancel. An extension to objects that are not exterior products needs an input that sees the factors separately.
So Perry’s criterion fails on every definite secant plane for every product structure and for every structure on a simple exterior product, and on an indefinite plane Theorem 35 below keeps its eightfolds away from the polarised Weil locus. What it leaves are non-simple factors with structures not restricted from a product, Brauer-twisted factors, objects that are not exterior products, on which the Euler mechanism is silent (Proposition 38), and the weak form of G-semiregularity that Perry’s theorem uses.
Theorem 32 says nothing about objects twisted by a Brauer class, about descents along a Heisenberg group G ⊆ X × X ∨ , or about planes e − b B Q ( h , q ) with a B-field; on definite planes Theorem 34 uses only the sign of χ and covers the last two for simple exterior products. For b ∈ Q h a search over a box of integral vectors finds no class with 〈 c , c 〉 = − 2 . The second invariant is Markman’s form. Let A = X × X ∨ with Markman’s action η of K = Q ( − q ) attached to the plane P = B Q ( h , q ) , f = η ( − q ) , and let Ξ P ( x , y ) = ( f x , y ) V be the 2-form Ξ t of Corollary 14 for t = − q , so that H = q ( , ) V + − q Ξ P is the hermitian form of signature ( 4 , 4 ) ; Ξ P is rational, nondegenerate and of type ( 1 , 1 )  [78, Sect. 2.4]. For a complex structure J for which Ξ P is of type ( 1 , 1 ) , g J ( x , y ) = Ξ P ( x , J y ) is symmetric and nondegenerate, and ± Ξ P is a Kähler class exactly when g J is definite.
Lemma  34
(The sign of the invariant form). Let n = 4 , b , h ∈ NS ( X ) Q , λ = exp ( − b + − q h ) , W 1 , 0 = W λ ∩ V 1 , 0 , and let ( p + , p − ) be the signature of the hermitian form u ↦ ( u , u ¯ ) V on W 1 , 0 .
(a)
dim W 1 , 0 = 4 , and ( p + , p − ) is the signature of the hermitian form of h, up to exchanging the two entries.
(b)
g has signature ± 2 ( 4 − p + + p − ) , 2 ( 4 + p + − p − ) , so Ξ P is ample or antiample if and only if { p + , p − } = { 0 , 4 } , that is, if and only if ± h is ample.
(c)
sign 〈 λ , λ ¯ 〉 = sign Pf ( h ) = ( − 1 ) p − . So if ± Ξ P is ample then 〈 λ , λ ¯ 〉 > 0 , and if Pf ( h ) < 0 then g is indefinite.
Proof. 
e − b acts by a rational isometry fixing the Hodge structure, so we may take b = 0 . Put t = − q . For w + ξ ∈ V C , ( w + ξ ) exp ( t h ) = ( w + t ξ ⌟ h ) ∧ exp ( t h ) , so W λ = { − t ξ ⌟ h + ξ } is the graph of ξ ↦ − t ξ ⌟ h . Here V 1 , 0 = H 1 , 0 ( X ) ⊕ H 0 , 1 ( X ) ∨ , and for ξ vanishing on H 1 , 0 ( X ) the class ξ ⌟ h lies in H 1 , 0 ( X ) because h is of type ( 1 , 1 ) ; so W 1 , 0 is the graph over H 0 , 1 ( X ) ∨ , of dimension 4. For u = − t ξ ⌟ h + ξ we get ( u , u ¯ ) V = 2 t h ( ξ ¯ , ξ ) , and ξ ↦ − 1 h ( ξ ¯ , ξ ) is the hermitian form of h up to sign; this gives (a). For (b), [78, Sect. 2.4] gives g ( v , v ) = 2 q − ( v 1 , 0 , v 1 , 0 ¯ ) V + ( v 0 , 1 , v 0 , 1 ¯ ) V with independent components v 1 , 0 ∈ W 1 , 0 and v 0 , 1 ∈ W 0 , 1 = W λ ∩ V 0 , 1 . Since W λ is isotropic, W 0 , 1 is the conjugate of the hermitian orthogonal of W 1 , 0 in V 1 , 0 , on which the form has signature ( 4 − p + , 4 − p − ) , the form on V 1 , 0 having signature ( 4 , 4 ) ; doubling real dimensions gives (b). (c) is Lemma 31 together with sign ∫ h 4 = ( − 1 ) p − .    □
So the polarised case, ± h ample, is exactly the case excluded from the Lemma 11.3 mechanism by the sign of χ in Theorem 30, and the indefinite case does not reach the polarised family.
Theorem  35
(Indefinite secant planes miss the polarised Weil locus). Let h ∈ NS ( X ) Q be indefinite with Pf ( h ) ≠ 0 , P = B Q ( h , q ) , and A = X × X ∨ with η and Ξ P as above.
(i)
Up to a scalar, Ξ P is the only 2-form invariant under the stabiliser Spin ( V ) P , and it is neither ample nor antiample on any member of Markman’s period domain D P = Spin ( V R ) P · I A . The very general member of D P has NS Q = Q Ξ P and is not projective.
(ii)
Let c 1 , c 2 ∈ P ∖ 0 , B ∈ H 2 ( A , Q ) , and w = e B φ ( c 1 ⊗ τ c 2 ) , where φ is the action on cohomology of Orlov’s equivalence Φ : D ( X × X ) → D ( X × X ∨ )  [96], so that w = e B ch Φ ( F 1 ⊠ F 2 ∨ ) when c i = ch F i . Let A → S be a connected analytic family of complex tori with fibre A over s 0 along which the parallel transport of w stays of Hodge type. Then no fibre is a polarised abelian eightfold A ′ with Hdg 2 ( A ′ ) = Q θ ′ and Hdg 4 ( A ′ ) = Q θ ′ 2 .
The very general member of every polarised family of abelian eightfolds of Weil type, for every imaginary quadratic field and every discriminant, has these Hodge groups (Theorem A1); so this holds whatever object realises w, equivariant or not, twisted or not, and whatever semiregularity statement is applied at A.
Proof. 
Step 1. Spin ( V ) P is of finite index in SU ( V , H )  [78, Sect. 3.1]. Over C , SU ( H ) ≅ SL ( W ) with V C = W ⊕ W * , and the SL ( W ) -invariants of ⋀ 2 V C = ⋀ 2 W ⊕ ⋀ 2 W * ⊕ W ⊗ W * form a line, spanned by Ξ P .
Step 2. By Lemma 34(a),(b), g is indefinite at A, so ± Ξ P is not ample there.
Step 3. In a connected family of complex structures J along which Ξ P stays of type ( 1 , 1 ) , g J is symmetric, nondegenerate and continuous in J, so its signature is constant.
Step 4. The Spin ( V ) P -invariant classes stay of Hodge type on D P  [78, Sect. 4], so by Step 3, Ξ P is indefinite on all of D P . The rational 2-forms of type ( 1 , 1 ) on all of D P are the invariants of the group generated by the circles { cos ϑ + sin ϑ J } , J ∈ D P , a nontrivial connected normal subgroup of the simple group SU ( H ) ( R ) , hence all of it; so they form Q Ξ P , and since H 2 ( A , Q ) is countable, NS Q = Q Ξ P at a very general member, which is not projective. This proves (i).
Step 5. Let Ω 0 = c 1 ( P ) / 2 for the Poincaré bundle P and φ ˜ ( x ⊗ y ) = e − Ω 0 φ ( x ⊗ τ y ) ; φ ˜ is equivariant from m ⊗ m † to ρ , where m is the spin action, m g † = τ m g τ and ρ is the action on H * ( A ) by algebra automorphisms [78, Sects. 1.3, 6.1]. From Φ P ( k 0 ^ ) = O , Φ P ( O ) = k 0 [ − 4 ] and Φ P ( P | x × X ∨ ) = k − x [ − 4 ]  [84] one gets φ ˜ ( 1 ⊗ 1 ) = [ pt X ∨ ] , φ ˜ ( 1 ⊗ pt ) = e − Ω 0 , φ ˜ ( pt ⊗ 1 ) = e Ω 0 and φ ˜ ( pt ⊗ pt ) = [ pt X ] . The group Spin ( V C ) acts transitively on ordered pairs of transversal even pure-spinor lines, so there is g with m g ( 1 ) = c λ and m g ( pt ) = c ′ λ ¯ ; since τ λ = λ ¯ , m g † ( 1 ) = c λ ¯ and m g † ( pt ) = c ′ λ . Equivariance gives φ ˜ ( λ ⊗ λ ) = ( c c ′ ) − 1 e − Ω and φ ˜ ( λ ¯ ⊗ λ ¯ ) = ( c c ′ ) − 1 e Ω with Ω = ρ g ( Ω 0 ) , while φ ˜ ( λ ⊗ λ ¯ ) and φ ˜ ( λ ¯ ⊗ λ ) lie in ρ g ( H 8 ) = H 8 . An element of Spin ( V C ) P fixing λ and λ ¯ is conjugated by g into the stabiliser of the lines of 1 and pt , which preserves H 1 ( X ) ⊕ H 1 ( X ) ∨ and Ω 0 ; so Ω is Spin ( V C ) P -invariant, Ω = ϑ Ξ P with ϑ ∈ C × by Step 1, and Ω 2 ≠ 0 because Ω 0 is nondegenerate.
Step 6. Write c i = A i λ + A ¯ i λ ¯ with A i ≠ 0 , P 0 = A 1 A 2 / ( c c ′ ) , Q 0 = A ¯ 1 A ¯ 2 / ( c c ′ ) , so P 0 Q 0 ≠ 0 , and B 1 = B + Ω 0 , which is rational. Then w = e B 1 P 0 e − Ω + Q 0 e Ω + ( degree ≥ 8 ) . With r = P 0 + Q 0 , the rank of w,
w 2 = r B 1 + ( Q 0 − P 0 ) Ω , w 4 = r 2 B 1 2 + ( Q 0 − P 0 ) B 1 Ω + r 2 Ω 2 .
If r ≠ 0 , w 4 − w 2 2 / ( 2 r ) = ( 2 P 0 Q 0 / r ) Ω 2 = m Ξ P 2 , and m ∈ Q × because the left side is rational and nonzero. If r = 0 , w 2 = − 2 P 0 Ω = m ′ Ξ P with m ′ ∈ Q × .
Step 7. On a real vector space of dimension 16, a nondegenerate 2-form Ξ satisfies ι v ι u Ξ 2 = 2 Ξ ( u , v ) Ξ − ι u Ξ ∧ ι v Ξ , which has rank at most 2 if and only if Ξ ( u , v ) = 0 . So Ξ ′ 2 = s Ξ 2 with s ≠ 0 forces Ξ ′ ( u , v ) = 0 ⇔ Ξ ( u , v ) = 0 ; then Ξ ′ ( u , · ) and Ξ ( u , · ) have the same kernel for every u, Ξ ′ = κ Ξ with κ constant by linearity, and κ = ± s .
Step 8. Along S, Ξ P stays of type ( 1 , 1 ) : if r = 0 it is a multiple of the Hodge class w 2 ; if r ≠ 0 , Ξ P 2 stays Hodge, so for u in the circle of the Hodge structure ( u · Ξ P ) 2 = u · Ξ P 2 = Ξ P 2 , Step 7 gives u · Ξ P = ± Ξ P , and continuity in u gives +. By Steps 2 and 3, g has the same indefinite signature on every fibre. At a fibre A ′ as in (ii), Ξ P 2 ∈ Hdg 4 ( A ′ ) = Q θ ′ 2 if r ≠ 0 , so Ξ P = ± s θ ′ by Step 7, and Ξ P ∈ Q θ ′ if r = 0 . Either way g is definite at A ′ , a contradiction.    □
Remark  12
(A split polarisation on the same eightfold). For X = E 1 × ⋯ × E 4 and h = ∑ i ϵ i e i with signs ϵ i not all equal, A = ∏ i S i with S i = E i × E i ∨ , and η acts factorwise by ϵ i η 0 . Each S i with ± η 0 is a surface of Weil type, and Markman’s ample invariant class ω S i for n = 1 is compatible with η 0 and with − η 0 , the compatibility condition being quadratic in η. So ω ′ = ∑ i ω S i is an η-compatible polarisation, of discriminant ( − 1 ) 4 = 1  [78, Sect. 3.1]: A is a polarised abelian eightfold of split Weil type. The Spin ( V ) P -invariant class is the different, indefinite class ∑ i ϵ i ω S i . So A lies both on the split polarised family of ω ′ and on D P , and by Theorem 35 the secant classes stay of Hodge type only along the second; on A itself the Weil classes are products of classes of degree two on the S i , as in the proof of Proposition 33.
For X = E 4 , h = e 1 + e 2 + e 3 − e 4 and q = 1 the degree-two part of e − c 1 ( P ) / 2 ch ( E ) is a multiple of ω = − 4 ( h ⊞ h ∨ ) , h ∨ = ∑ ϵ i e ^ i , with coefficients of signs ( − , − , − , + , − , − , − , + ) on the eight elliptic factors, and for six objects of nonzero rank the class κ ( E ) = e − c 1 ( E ) / r ch ( E ) has κ 1 = 0 and κ 2 = 2 P 0 Q 0 r ω 2 up to the normalisation of ω , with coefficients 1 16 , 1 16 , 1 8 , 5 8 , 5 8 , 145 8 . These values, the ranks of Lemma 30 for n = 3 , 4 and both signs of q, the Euler form of Lemma 31, the rank 48 = ∑ a + b = 2 rk a rk b of the Künneth contraction at n = 3 , the divisors D k = d 1 ⋯ d k of forty integral forms and the lattices of Theorem 32 are checked in julia/secantranks.jl, julia/kunnethrank.jl, julia/kappa.jl and julia/lattice.jl; the forced dimensions, the identity n ( n − 1 ) = 2 n ⇔ n = 3 and the congruences of Theorem 32 are checked in lean/HodgeClosure.lean.
So Markman’s route from fourfolds to split eightfolds is closed for every untwisted object (Corollary 17) and, through any mechanism, for every indefinite secant plane (Theorem 35). What remains of it is a definite secant plane, ± h ample, on which the Lemma 11.3 criterion fails by Theorem 30 and Perry’s G-semiregularity fails for every product structure and every structure on a simple exterior product by Theorems 33 and 34, with G-semiregularity for an object that is not a simple exterior product [98], on which the Euler mechanism is silent (Proposition 38), or an affirmative answer to [79, Question 11.4], in its place; by Proposition 36 the kernel half of that criterion is the vanishing of the products ev ( p ) ev ( q ) , p ∈ P i , q ∈ Q i , and for Markman’s object of [79, Sect. 12] it forces the equidistribution of Theorem 31.

6.12. Weil Sixfolds from Cyclic Triple Covers of Curves of Genus Four

Abelian covers of curves give explicit Weil families. For an étale cyclic cover of degree ℓ ∈ { 3 , 4 , 6 } of a curve B of genus four, defined by a line bundle η of order ℓ, the pairs ( B , η ) have nine moduli, as many as the dimension of the domain D F , F = Q ( ζ ℓ ) , n = 3 , that receives the period map, and the image of this map is open near every point at which the multiplication map
μ η : H 0 ( K B ⊗ η ) ⊗ H 0 ( K B ⊗ η − 1 ) ⟶ H 0 ( K B ⊗ 2 )
is surjective (proof of Proposition 39(iv)). For ℓ = 3 the Prym map of these covers is generically finite, by Faber [38], and surjectivity of μ η at a general point follows. We exhibit an explicit point at which μ η is an isomorphism.
Proposition  39
(An explicit cover). For λ ∈ C ∖ { 0 , 1 } let B λ ⊂ P 1 × P 1 be the curve x 3 + y 3 + 1 + λ x 3 y 3 = 0 , in affine coordinates x, y, and η = O ( 1 , − 1 ) | B λ .
(i)
B λ is a smooth curve of genus four with K B λ = O ( 1 , 1 ) | B λ .
(ii)
η has order exactly three.
(iii)
H 0 ( K ⊗ η ) and H 0 ( K ⊗ η − 1 ) are spanned by 1 , x , x 2 and by 1 , y , y 2 , and μ η is an isomorphism.
(iv)
The period map of the étale cyclic triple covers of curves of genus four, with values in the domain D F of ( Q ( − 3 ) , 3 ) -Weil type that contains the χ-part of H 1 of the cover C λ → B λ defined by η, is a local isomorphism at ( B λ , η ) . Its image contains an open subset of D F , and the Weil classes are algebraic on every member of D F as soon as they are algebraic on the abelian sixfolds H χ ( C t ) for the covers C t of a non-meagre subset of a neighbourhood of ( B λ , η ) .
Proof. 
(i) In bihomogeneous coordinates the curve is F = x 1 3 y 0 3 + x 0 3 y 1 3 + x 0 3 y 0 3 + λ x 1 3 y 1 3 , of bidegree ( 3 , 3 ) . In the chart x 0 = y 0 = 1 the partial derivatives are 3 x 2 ( 1 + λ y 3 ) and 3 y 2 ( 1 + λ x 3 ) . At a singular point with x = 0 the equation gives y 3 = − 1 and then ∂ y F = 3 y 2 ≠ 0 ; similarly y ≠ 0 . So x 3 = y 3 = − 1 / λ , and the equation becomes 1 − 1 / λ = 0 , excluded. The equation is symmetric in x and y; the substitution ( x , y ) = ( 1 / x ′ , λ − 1 / 3 y ′ ) , after multiplication by x ′ 3 , carries it to the same form in x ′ , y ′ with λ replaced by 1 / λ ; and the substitution ( x , y ) = ( λ − 1 / 3 / x ′ , λ − 1 / 3 / y ′ ) , after multiplication by λ x ′ 3 y ′ 3 , carries it to itself. So the other three charts reduce to the first. By adjunction K B = ( K P 1 × P 1 + B ) | B = O ( 1 , 1 ) | B , of degree ( 1 , 1 ) · ( 3 , 3 ) = 6 = 2 g − 2 , so g = 4 .
(ii) Let D x and D y be the intersections of B with the lines x = ∞ and y = ∞ . At x = ∞ the equation reads 1 + λ y 3 = 0 , so D x consists of three points ( ∞ , c ) with c 3 = − 1 / λ , and at each of them u = 1 / x is a local parameter, since the derivative of 1 + u 3 y 3 + u 3 + λ y 3 in y is 3 λ c 2 ≠ 0 ; along B, 1 + λ y 3 = − u 3 ( 1 + y 3 ) , so φ = y 3 + 1 / λ vanishes to order three there. Its other zeros would have y 3 = − 1 / λ and x ≠ ∞ , hence x 3 ( 1 + λ y 3 ) = 0 = 1 + y 3 , which forces λ = 1 . Its poles are at D y , three points at each of which 1 / y is a local parameter by the same computation, of order three. So div φ = 3 D x − 3 D y , and since D x ∈ | O ( 1 , 0 ) | B | and D y ∈ | O ( 0 , 1 ) | B | , η ⊗ 3 ≅ O B . If η were trivial, D y would lie in | O B ( 1 , 0 ) | . From 0 → O ( − 2 , − 3 ) → O ( 1 , 0 ) → O B ( 1 , 0 ) → 0 and H 0 ( O ( − 2 , − 3 ) ) = H 1 ( O ( − 2 , − 3 ) ) = 0 by the Künneth formula, every section of O B ( 1 , 0 ) is a + b x , so every divisor in that system is a fibre x = c of B. The three points of D y have x 3 = − 1 / λ , three distinct values of x, so D y is not such a fibre.
(iii) K ⊗ η = O ( 2 , 0 ) | B and K ⊗ η − 1 = O ( 0 , 2 ) | B . The kernels O ( − 1 , − 3 ) and O ( − 3 , − 1 ) of the restrictions have no cohomology in degrees 0 and 1, again by the Künneth formula, so restriction gives H 0 ( K ⊗ η ) = H 0 ( P 1 × P 1 , O ( 2 , 0 ) ) , spanned by 1 , x , x 2 , and H 0 ( K ⊗ η − 1 ) = H 0 ( O ( 0 , 2 ) ) , spanned by 1 , y , y 2 . The product of the two is H 0 ( O ( 2 , 2 ) ) , of dimension nine, and μ η is its restriction to B, with values in H 0 ( O B ( 2 , 2 ) ) = H 0 ( K ⊗ 2 ) , of dimension 3 g − 3 = 9 . The kernel of the restriction is H 0 ( O ( − 1 , − 1 ) ) = 0 , so μ η is injective between spaces of dimension nine, hence an isomorphism.
(iv) Write π : C λ → B = B λ for the cover defined by η , so that π * O C λ = ⨁ i = 0 2 η i and μ 3 acts on η i through the character ζ ↦ ζ i . The eigenspace of H 1 ( C λ , C ) for the character χ of exponent one is H 1 ( B , L ) for the unitary local system L of rank one underlying η , with Hodge pieces H 0 ( K ⊗ η ) of type ( 1 , 0 ) and H 1 ( η ) ≅ H 0 ( K ⊗ η − 1 ) ∨ of type ( 0 , 1 ) . Since deg η = 0 and η ≠ O B , h 0 ( η ) = h 0 ( η − 1 ) = 0 , and Riemann–Roch gives h 0 ( K ⊗ η ) = h 0 ( K ⊗ η − 1 ) = g − 1 = 3 . The sum H χ of the χ - and χ ¯ -eigenspaces is defined over Q , carries the polarisation induced by the intersection form and an action of F = Q ( − 3 ) , and the Hermitian form on the χ -eigenspace has signature ( 3 , 3 ) ; so H χ is of ( F , 3 ) -Weil type and the abelian sixfold H χ ( C λ ) lies in a domain D F ≅ U ( 3 , 3 ) / ( U ( 3 ) × U ( 3 ) ) of dimension nine. A line bundle of finite order deforms uniquely with the curve, so the pairs ( B , η ) are locally parametrised by H 1 ( T B ) , of dimension 3 g − 3 = 9 . By Griffiths’ description of the differential of the period map [123, Ch. 10], the differential at ( B , η ) is the map H 1 ( T B ) → Hom ( H 0 ( K ⊗ η ) , H 1 ( η ) ) given by cup product, and after Serre duality, H 1 ( T B ) ∨ = H 0 ( K ⊗ 2 ) and H 1 ( η ) ∨ = H 0 ( K ⊗ η − 1 ) , its transpose is μ η . By (iii) the differential is an isomorphism and the period map is a local isomorphism at ( B λ , η ) ; its image contains an open set Ω ⊆ D F . A local biholomorphism sends a non-meagre subset of a neighbourhood to a non-meagre subset of Ω , so under the hypothesis of the last assertion the algebraic locus of the Weil classes in D F is not meagre, and by Lemma 2 it is all of D F .    □
The curve B λ is a cyclic cover of P 1 × P 1 in disguise: it is invariant under μ 3 2 acting on x and y, and its quotient is the line X + Y + 1 + λ X Y = 0 in the coordinates X = x 3 , Y = y 3 . For ℓ = 4 and ℓ = 6 the same argument needs a curve of genus four with a line bundle of order four or six for which μ η is surjective, and we have not found one.
All these étale families are split, by the next proposition, and cyclic covers of curves reach a non-split form only in one way, and never on an open set.
Proposition  40
(Which cyclic covers are non-split). Let C → B = C / 〈 σ 〉 be a cyclic cover of order ℓ ∈ { 3 , 4 , 6 } , B of genus h, with r branch points of local monodromies c j ∈ ( Z / ℓ ) ∖ { 0 } , let K = Q ( ζ ℓ ) and let H χ = ker Φ ℓ ( σ ) ⊆ H 1 ( C , Q ) , with the K-hermitian form h χ determined by E ( k x , y ) = Tr K / Q ( k δ − 1 h χ ( x , y ) ) , where E is the intersection form and δ = − 3 or 2 i generates the different of O K . Suppose dim K H χ = 6 and H χ is of Weil type.
(i)
For ℓ ∈ { 3 , 4 } , H χ is split.
(ii)
For ℓ = 6 , det h χ ≡ − 2 m 3 modulo N ( K × ) with m 3 = 2 h − 2 + r ′ , r ′ = # { j : 3 ∤ c j } ; so H χ is non-split if and only if r ′ is odd. In particular every étale cover gives a split form.
(iii)
The covers of a fixed topological type form a family of dimension 3 h − 3 + r ≤ 9 , with equality only for étale covers of curves of genus four, which are split. The non-split families have dimension at most seven. So no family of cyclic covers has a period image containing an open subset of a non-split domain, and the family of Proposition 39 lies in a split domain, where [78, Thm. 1.5.1] applies.
Proof. 
Let Λ = H 1 ( C , Z ) , unimodular for E, and Λ χ = Λ ∩ H χ , a free O K -module. Since σ preserves E, E ( f ( σ ) x , y ) = E ( x , f ( σ − 1 ) y ) ; as Φ ℓ is palindromic and coprime to R = ( x ℓ − 1 ) / Φ ℓ , H χ is orthogonal to ker R ( σ ) and Λ Q = H χ ⊕ ker R ( σ ) . Since Λ χ is an O K -module, x lies in its E-dual exactly when δ − 1 h χ ( x , Λ χ ) lies in the inverse different δ − 1 O K , so the h χ -dual and the E-dual of Λ χ coincide, so | det E | Λ χ | = | N K / Q ( det G ) | = ( det G ) 2 for a Gram matrix G of h χ , and det G ∈ Q . By the Riemann bilinear relations h χ has signature ( p , q ) = ( dim H χ , ζ 1 , 0 , dim H χ , ζ 0 , 1 ) , so det G has sign ( − 1 ) q ; Weil type means p = q = 3 . Changing Λ χ to a sublattice of finite index changes det G by a norm, and replacing the K-structure by its conjugate replaces h χ by − h ¯ χ .
(i) Let G ′ = 〈 σ 〉 for ℓ = 3 and G ′ = 〈 σ 2 〉 for ℓ = 4 , of order N = 3 or 2; then ker R ( σ ) ∩ Λ = Λ G ′ . For a saturated sublattice L of the unimodular Λ with L Q ⊕ L Q ⊥ = Λ Q , every functional on L is E ( x , · ) for some x ∈ Λ , which gives | det E | L | = [ Λ : L ⊕ L ⊥ ] = | det E | L ⊥ | . For the quotient map π : C → C / G ′ , π * π ! = N , π ! π * = ∑ g ∈ G ′ g and E ( π ! x , π ! y ) = N E ′ ( x , y ) , so π ! Λ ′ ⊆ Λ G ′ ⊆ N − 1 π ! Λ ′ and | det E | Λ G ′ | is a power of N. Hence ( det G ) 2 is a power of N and det G = ± N j . Since 3 = N ( − 3 ) and 2 = N ( 1 + i ) , det h χ ≡ ( − 1 ) q ; this argument uses neither the dimension nor Weil type. For q = 3 it is the split class ( − 1 ) 3 of Theorem 27.
(ii) Here Φ 3 ( x 2 ) = Φ 3 ( x ) Φ 6 ( x ) , so ker Φ 3 ( σ 2 ) = H 3 ⊕ H χ with H 3 = ker Φ 3 ( σ ) , of K-dimension m 3 , and the argument of (i) applied to σ 2 gives det h 3 ′ det h χ ′ ≡ ( − 1 ) q ′ for the forms and the number q ′ of negative entries defined by the σ 2 -structure. On H χ the two K-structures agree, so h χ ′ = h χ ; on H 3 , σ 2 = σ − 1 , so h 3 ′ = − h ¯ 3 , with det h 3 ′ = ( − 1 ) m 3 det h 3 , and q ′ = p 3 + 3 for the signature ( p 3 , q 3 ) of h 3 . The automorphism σ 3 is trivial on H 3 , so for ρ : C → C / 〈 σ 3 〉 the map ρ ! identifies the χ -part of the quotient, a cyclic triple cover of B, with H 3 , multiplying E by 2; by the argument of (i) for that cover, det h 3 ≡ 2 m 3 ( − 1 ) q 3 . Combining, det h χ ≡ ( − 1 ) q 3 + p 3 + m 3 + 3 2 m 3 = − 2 m 3 . The number 2 is not a norm from Q ( − 3 ) , since it is inert there, so det h χ ≡ − 1 exactly when m 3 is even. The cyclic triple cover C / 〈 σ 3 〉 → B is branched exactly at the points with 3 ∤ c j , and the Chevalley–Weil formula for the dimension H 1 ( B , j * L ) of the eigenspace of a faithful character gives m 3 = 2 h − 2 + r ′ .
(iii) By the same formula dim K H χ = 2 h − 2 + r , so 2 h − 2 + r = 6 and the family, with moduli ( B , branch points ) , has dimension 3 h − 3 + r = 5 + h ≤ 9 , with equality only for h = 4 , r = 0 , where (i) and (ii) with r ′ = 0 give split forms. A non-split family needs ℓ = 6 and r ′ odd; h = 3 forces r = 2 and c 1 + c 2 ≡ 0 mod 6 , so 3 divides both or neither and r ′ is even; hence h ≤ 2 and the dimension is at most seven. The domain of SU ( 3 , 3 ) has dimension nine.    □
For example the covers with ℓ = 6 , h = 2 and monodromies ( 3 , 1 , 4 , 4 ) have r ′ = 3 and give a seven-dimensional family of non-split Weil sixfolds; whether its Weil classes are algebraic is open, and even if they are, Lemma 2 does not carry them to the very general member, since seven-dimensional loci are meagre in the nine-dimensional domain.
Hyper-Kähler varieties cannot carry the Weil classes of a sixfold either, for a reason that disappears for fourfolds. For a rational Hodge structure M let hg ( M ) be the Lie algebra of its Hodge group and X M ∈ hg ( M ) C the derivative of h | U ( 1 ) , acting by p − q on M p , q . Call Mbad if hg ( M ) C has a simple ideal s ≅ sl 2 n with n ≥ 3 on which the component of X M is conjugate under Aut ( s ) to diag ( 1 n , ( − 1 ) n ) , and good otherwise.
Theorem  36
(No transport from good Hodge structures). Let A be an abelian 2 n -fold of Weil type with n ≥ 3 and Hodge group SU ( V , H ) , as for the very general member of every Weil domain, split or not.
(i)
H 1 ( A ) is bad. Good Hodge structures are closed under direct sums, tensor products, duals, subquotients and Tate twists; every weight-two Hodge structure with h 2 , 0 ≤ 2 is good, and so are the Kuga–Satake abelian varieties of such structures and the cohomology of every compact hyper-Kähler manifold.
(ii)
If Y is smooth projective with H * ( Y ) good, then Hg ( Y × A ) = Hg ( Y ) × Hg ( A ) , the Hodge classes of Y × A are Hdg ( Y ) ⊗ Hdg ( A ) , and every morphism of Hodge structures from a subquotient of H * ( Y ) to H * ( A m ) has image consisting of Hodge classes.
(iii)
For n = 3 , every morphism f : A → Y to a projective variety with h 2 , 0 ( Y ) ≤ 5 is constant or finite onto its image, and if Y is hyper-Kähler, f ( A ) is isotropic for the symplectic form and dim Y ≥ 12 .
Proof. 
(i) su ( V , H ) C = sl ( V ι ) for one embedding ι of K, and Weil type means that X acts on V ι with eigenvalues 1 and − 1 , each n times. For the closure properties: the Hodge group of a tensor construction or subquotient of M is the image of Hg ( M ) , being a Q -group containing the image of h ( U ( 1 ) ) ; a simple ideal of a quotient of a reductive Lie algebra is the image of a simple ideal, with the image of its component of X; and Hg ( M 1 ⊕ M 2 ) ⊆ Hg ( M 1 ) × Hg ( M 2 ) maps onto both factors, so a simple ideal s of hg ( M 1 ⊕ M 2 ) C projects isomorphically onto an ideal of some hg ( M k ) C , which is a direct summand because it commutes with the projection of the complementary ideal, and the component of X M k there is the projection of that of X. For weight two, a bad M has an irreducible summand W s ⊗ W ′ of M C with W s a nontrivial sl 2 n -module on which X s has eigenvalue range at most 4; with 〈 ω i , X s 〉 = min ( i , 2 n − i ) the range of V ( λ ) is 2 ∑ a i min ( i , 2 n − i ) , so W s is one of V ι , ⋀ 2 V ι , S 2 V ι , their duals, or the adjoint representation, whose extreme eigenspaces have dimensions n, n 2 , n + 1 2 and n 2 , all at least 3 for n ≥ 3 . The extreme eigenspaces of X on W s ⊗ W ′ are products of those of the factors, the eigenvalues of X lie in { − 2 , 0 , 2 } and the range on this summand is at least 2, so h 2 , 0 ≥ 3 or h 0 , 2 ≥ 3 , and h 2 , 0 = h 0 , 2 . For a Kuga–Satake abelian variety the complex structure is given by a lift h ˜ of h to the spin group, with Ad ∘ h ˜ = h on U ( 1 ) , so ( hg , X ) of its H 1 is isomorphic to that of the weight-two structure. On the cohomology of a compact hyper-Kähler manifold M the Weil operator acts through the rational Lie algebra of Looijenga–Lunts and Verbitsky as the image of the Weil operator of H 2 ( M )  [67], so ( hg ( H * ( M ) ) , X ) is a quotient of ( hg ( H 2 ( M ) ) , X ) , which is good since h 2 , 0 ( M ) = 1 .
(ii) Let g = hg ( H * ( Y ) ⊕ H 1 ( A ) ) ⊆ g Y ⊕ g A . The intersection g ∩ ( 0 ⊕ g A ) is an ideal of g A = su ( V , H ) , which is simple. If it were zero, g would project isomorphically onto g Y and g A would be a quotient of g Y compatible with X, so H * ( Y ) would be bad. Hence g ⊇ 0 ⊕ g A and g = g Y ⊕ g A ; the Hodge groups are connected, so Hg ( Y × A ) = Hg ( Y ) × Hg ( A ) , and the invariants of a product group on H * ( Y ) ⊗ H * ( A ) are the tensor products of the invariants. The image of a morphism from a subquotient of H * ( Y ) to H * ( A m ) is good by (i), and its Hodge algebra is a quotient of the simple su ( V , H ) , so it is zero.
(iii) Over C , H 2 ( A ) = ⋀ 2 ( V ι ⊕ V ι ¯ ) decomposes under sl ( V ι ) as C θ ⊕ sl ( V ι ) ⊕ ( ⋀ 2 V ι ⊕ ⋀ 2 V ι ∨ ) , which over Q gives sub-Hodge structures Q θ ⊕ U 35 ⊕ U 30 ; U 35 and U 30 are irreducible, the second because complex conjugation exchanges ⋀ 2 V ι and ⋀ 2 V ι ¯ , and h 2 , 0 ( U 35 ) = 9 , h 2 , 0 ( U 30 ) = 6 from the eigenvalues of X. The image of f * : H 2 ( Y ) → H 2 ( A ) is a sub-Hodge structure with h 2 , 0 ≤ 5 , so it lies in Q θ . For an ample L on Y, f * c 1 ( L ) = c θ with c ≥ 0 ; if c = 0 every curve of A is contracted and f is constant, and if c > 0 , f * L is ample and f is finite onto its image. If Y is hyper-Kähler with symplectic form σ , then f * σ ∈ H 2 , 0 ( A ) ∩ C θ = 0 , so f ( A ) , of dimension six, is isotropic, and an isotropic subvariety has dimension at most 1 2 dim Y .    □
So the Weil classes of A are not the image, under any algebraic correspondence, of a sub-Hodge structure of the cohomology of a hyper-Kähler or Kuga–Satake variety, or of a product of such varieties: a cycle on Y × A has class in Hdg ( Y ) ⊗ Hdg ( A ) and gives a Weil class of A only if it already has one as a Künneth component. For n = 2 the obstruction is absent: ⋀ 2 V ι of sl 4 has extreme eigenspaces of dimension one, a Hodge structure of K3 type, and this is the piece through which the fourfold arguments of [76], [39] and [119] pass.
Remark  13
(Degenerations of non-split sixfolds). For a degeneration of non-split Weil sixfolds, the logarithm N of the monodromy is K-linear and H-skew with N 2 = 0  [23], so I = Im N is isotropic, since H ( N x , N y ) = − H ( x , N 2 y ) = 0 , and ker N = I ⊥ . By Theorem 27 r = dim K I ≤ 2 , and V ≅ r H ⊥ I ⊥ / I makes I ⊥ / I a non-split hermitian space of K-dimension 6 − 2 r . If, as the limit mixed Hodge structure suggests, Gr 1 W is then a non-split Weil fourfold or a Weil surface, its Weil classes are algebraic, by [78, Cor. 1.6.1] or by the Lefschetz theorem on ( 1 , 1 ) -classes; carrying algebraicity from the boundary into the interior is the variational step, and it is open.

6.13. Integral Characters at Sextic CM Fields

For a sextic CM field F = F 0 ( − q ) , with F 0 totally real cubic and q ∈ F 0 totally positive, Proposition A17 shows that a perfect complex F with ch ( F ) = v in the flat secant space S ( 0 , q ) and the least possible Ext 0 , Ext 1 , Ext 2 exists only if χ ( v , v ) lies below an explicit threshold, and Proposition A18 shows that the condition of the count holds whenever χ ( v , v ) ≤ − 26 , computes χ ( v , v ) = − 8 Nm ( q ) c 0 2 + Tr ( Nm ( q ) q − 1 f 2 ) + Tr ( q g 2 ) + c 3 2 on the rational points v ( c 0 , f , g , c 3 ) , and exhibits integral points with χ = − 32 for q = 1 .
Remark  14
(No exclusion over twelve cubic fields). For the twelve totally real cubic fields of discriminant 49, 81, 148, 169, 229, 257, 316, 321, 361, 404, 469, 473, whose rings of integers are generated by one element, and for q = 1 and q = k + α with α the generator and k the least integer making q totally positive, the lattice of integral points of S ( 0 , q ) was computed exactly, in the model of Proposition A18, with the Gram matrix of − χ / 8 , which is integral. No nonzero integral point has − χ ≤ 24 , and since χ ∈ 8 Z on the lattice, every nonzero integral point has χ ≤ − 32 and meets the condition of Proposition A17(iii). For q = 1 the least value − χ = 32 is attained by four points of shape N w = 2 , r 2 = 30 , r 3 = 40 ; for q = k + α it is 32 for the discriminants 49, 81 and 288, 384, 912, 352, 480, 600, 960, 544, 2400, 1208 for the other ten. So over these fields the count of Proposition A17 excludes no integral character of a minimal object, and the route through it is closed there. This is a computation in the repository [9], and nothing in the paper depends on it.

6.14. The Exceptional Classes of a Mumford Fourfold

Keep Setting 1 and write ψ for the polarisation form on V, T = D 0 , q λ ( x , y ) = Tr F / Q ( λ trd ( x y ) ) for λ ∈ F , and m a for multiplication by a ∈ F on T. The copy T X ( − 1 ) ⊆ V ⊗ V , x ↦ ( v , w ) ↦ ψ ( x v , w ) , is symmetric and carries the form induced by ψ ⊗ ψ , equal to − 4 trd ( x y ) summed over the three real places, that is q 1 up to the scalar − 4 . The exceptional classes of X × X are the elements m a , a ∈ F ∖ Q , of End Hdg T X ( − 1 ) = F Proposition A21, and in Theorem A23 their algebraicity is derived from the Kuga–Satake class of a K3 surface S λ with T ( S λ ) Q ≅ ( T , q λ ) . We replace S λ by an abelian variety.
Setting  6
(The sixfold Z K ). Let K = F ( − δ ) ⊆ D be a CM quadratic extension of F inside D, δ ∈ F totally positive; every maximal commutative CM subfield of D is of this form since D is definite at two real places. Let D carry the left action of G = D 1 and the right action of the torus K 1 = ker ( N K / F ) . Over Q ¯ , D = D 1 ⊕ D 2 ⊕ D 3 with D i = V i ⊗ W i , V i the standard representation of the i-th factor of G and W i = χ i ⊕ χ i − 1 that of the i-th factor of K 1 . Fix a point t of the Shimura curve C of X, with Hodge cocharacter μ X nontrivial only at the split place σ 3 , and let μ Z = μ ′ μ ′ ′ , with μ ′ the action of μ X by left multiplication and μ ′ ′ the cocharacter of K 1 of weight ± 1 2 on χ 1 ± 1 , χ 2 ± 1 and 0 on χ 3 ± 1 . Then D C has type ( 1 , 0 ) on V 1 χ 1 , V 2 χ 2 , V 3 1 , 0 W 3 and ( 0 , 1 ) on the complements, so D carries a rational Hodge structure of weight one with h 1 , 0 = 6 and Hodge group inside G · K 1 . For c ∈ K with c ¯ = − c put ψ c ( x , y ) = Tr F / Q trd ( x ¯ y c ) . The abelian sixfold Z K = Z K , t is the one with H 1 ( Z K ) = D and polarisation ψ c for c in the cone where ψ c is positive; over a finite cover of C these form an abelian scheme [110], and we use only the fibre at t.
Lemma  35.
ψ c is alternating, invariant under left multiplication by D 1 and right multiplication by K 1 , and the ψ c , c ∈ K with c ¯ = − c , fill the three-dimensional space of all G × K 1 -invariant alternating forms on D. Hence NS ( Z K ) Q has dimension at least three at every point t of C, and exactly three at every t outside a countable subset of C.
Proof. 
The reduced trace is invariant under conjugation and cyclic, so ψ c ( y , x ) = Tr trd ( y ¯ x c ) = Tr trd ( c ¯ x ¯ y ) = Tr trd ( x ¯ y c ¯ ) , and ψ c is alternating exactly when c ¯ = − c . For g with nrd ( g ) = 1 , trd ( g x ¯ g y c ) = trd ( x ¯ g ¯ g y c ) = trd ( x ¯ y c ) . For k ∈ K 1 , trd ( x k ¯ y k c ) = trd ( k ¯ x ¯ y k c ) = trd ( x ¯ y k c k ¯ ) = trd ( x ¯ y c ) because K is commutative and k k ¯ = 1 . Over Q ¯ the invariant alternating forms live in the three summands ⋀ 2 ( V i ⊗ W i ) G × K 1 = ⋀ 2 V i ⊗ ( χ i χ i − 1 ) , one dimension each, and the cross terms D i ⊗ D j carry the nontrivial characters χ i ± 1 χ j ± 1 ; so the space has dimension 3 = dim Q K − , which the ψ c fill since c ↦ ψ c is injective. The Hodge group H t of Z K , t lies in G · K 1 , so these three forms are Hodge classes and dim NS ( Z K , t ) Q ≥ 3 ; a smaller group can have more invariants, so this is only a lower bound.12 For the upper bound at a very general t, recall that the Hodge group at a very general point of a family is the generic Hodge group H, and that by Deligne and André [4] the connected monodromy group is a normal subgroup of its derived group. The monodromy of the family over C acts through left multiplication by an arithmetic subgroup of G, Zariski dense in G because G is Q -simple and isotropic at the split place, so G ⊆ H ⊆ G · K 1 . The image of H in the torus G · K 1 / G , a quotient of the three-dimensional torus Res F / Q K 1 by a finite group, is the smallest Q -subtorus through which the image of μ ′ ′ factors. The cocharacter group of Res F / Q K 1 , tensored with Q , is Q 3 , one coordinate for each real place of F, and the Galois group permutes the coordinates transitively up to sign; the orbit of μ ′ ′ = ( 1 2 , 1 2 , 0 ) contains its three cyclic shifts, which are linearly independent. So H = G · K 1 at a very general t, and its invariant alternating forms are exactly the three forms ψ c , all of type ( 1 , 1 ) .    □
Lemma  36
(Where T sits in Z K ).
(i)
The K 1 -invariants of ⋀ 2 D i form a four-dimensional space: S 2 V i ⊗ ⋀ 2 W i , the adjoint representation of sl ( V i ) , and ⋀ 2 V i ⊗ 1 . The sum T Z of the three adjoint pieces is defined over Q , is a sub-Hodge structure of H 2 ( Z K ) , and is isomorphic, as a Hodge structure with G-action, to T X ( − 1 ) .
(ii)
The form induced on ⋀ 2 D by ψ c restricts on T Z to q c 2 , up to a rational scalar independent of the place; here c 2 = − N K / F ( c ) = − f 2 δ ∈ F for c = f − δ .
Proof. 
(i) ⋀ 2 ( V i ⊗ W i ) = S 2 V i ⊗ ⋀ 2 W i ⊕ ⋀ 2 V i ⊗ S 2 W i ; the torus acts trivially on ⋀ 2 W i = χ i χ i − 1 and on one summand of S 2 W i = χ i 2 ⊕ 1 ⊕ χ i − 2 . The cross terms carry nontrivial characters. The Galois group permutes the places, so T Z is defined over Q ; as a G-representation it is the adjoint representation T. The cocharacter μ ′ ′ acts trivially on K 1 -invariants and the centre acts on ⋀ 2 D with weight two, so the Hodge structure on T Z is the one induced by the adjoint action of μ ′ , twisted to weight two; this is the recipe that gives T X ( − 1 ) at the same point t.
(ii) By Schur’s lemma the induced form is, on each irreducible T i , a multiple of trd ( x y ) . In the split model at the place i the element c acts on D i by σ i ( c ) diag ( 1 , − 1 ) , whose square is σ i ( c 2 ) ; the multiple is σ i ( c ) 2 times a rational constant that is the same at every place, since the computation in the weight basis is identical at the three places. The sum over the places is q c 2 up to that constant.    □
Figure 13. The sixfold link. Two copies of the Mumford fourfold X t and the sixfold Z K , t sit over the same point t of the Shimura curve C; the class ξ t in the mixed Künneth component identifies the two copies of T and is flat along C. The stars mark the dense set Σ K of points of complex multiplication, at which ξ t is algebraic by Theorem 41.
Figure 13. The sixfold link. Two copies of the Mumford fourfold X t and the sixfold Z K , t sit over the same point t of the Shimura curve C; the class ξ t in the mixed Künneth component identifies the two copies of T and is flat along C. The stars mark the dense set Σ K of points of complex multiplication, at which ξ t is algebraic by Theorem 41.
Preprints 244139 g013
Theorem  37
(The sixfold link). Let t ∈ C be a point at which X = X t has Hodge group G, and let ξ be a nonzero element of Hom Hdg ( T X ( − 1 ) , T Z ) ⊆ H 4 ( X × X × Z K ) .
(i)
Hom Hdg ( T X ( − 1 ) , T Z ) = F · ι for the identification ι of Lemma 36, so ξ = m e ∘ ι for a unique e ∈ F × . The G · K 1 -invariant classes of degree four on X × X × Z K have dimension 26; they are Hodge classes at every t, at a very general t they are all the Hodge classes of degree four, and there ξ and ψ ⊗ NS ( Z K ) are the only ones in the Künneth component H 1 ( X ) ⊗ H 1 ( X ) ⊗ H 2 ( Z K ) .
(ii)
If ξ is algebraic, then ξ † ξ , the composite with the adjoint for the Lefschetz pairing on X × X and the intersection pairing on Z K , is an algebraic endomorphism of T X ( − 1 ) equal to m a ( e ) with a ( e ) = κ e 2 c 2 ∈ F × , κ ∈ Q × .
(iii)
For every b ∈ F ⊆ K ⊆ End 0 ( Z K ) the pull-back b * acts on T Z by m b 2 ; so the real multiplication of T Z by F is algebraic, for every K.
(iv)
If ξ is algebraic, the algebraic operators ξ † ∘ b * ∘ ξ , b ∈ F , span { m x : x ∈ F } on T X ( − 1 ) ; the exceptional classes of X × X are algebraic at t, and so is every Hodge class on every power of X. This holds for every CM extension K ⊆ D of F.
(v)
Conversely, if the exceptional classes are algebraic and some algebraic cycle on X r × Z K acts nonzero on T Z , then every ξ is algebraic.
Proof. 
(i) The Hodge group of X × X × Z K lies in G · K 1 and maps onto Hg ( X ) = G ; it acts on T X ( − 1 ) and on T Z through that image, since K 1 acts trivially on both. Both are the irreducible G-representation T with End G T = F Proposition A21, so Hom Hdg = F ι . The count is the dimension of the kernel of the six raising and lowering operators of sl 2 3 on the torus-weight-zero part of ⋀ 4 ( V ⊕ V ⊕ D ) , computed by Künneth type: ( 1 , 1 , 2 ) gives 3 + 3 , the line F ι and ψ ⊗ NS ( Z K ) ; ( 2 , 2 , 0 ) gives 4 = Q ⊕ F , the Hodge classes of X × X in H 2 ⊗ H 2 ; ( 0 , 0 , 4 ) gives 6 = S 2 NS ( Z K ) ; ( 2 , 0 , 2 ) and ( 0 , 2 , 2 ) give 3 each; and ( 4 , 0 , 0 ) , ( 0 , 4 , 0 ) , ( 3 , 1 , 0 ) , ( 1 , 3 , 0 ) one each, products of divisor classes. The sum is 26. The Hodge group of the product lies in G · K 1 , so these 26 classes are Hodge classes at every t, and 26 is a lower bound for the Hodge classes of degree four. At a very general t the argument of Lemma 35, applied to the family X × X × Z K over C, whose monodromy is again Zariski dense in G, gives Hodge group exactly G · K 1 , and then the 26 classes are all of them; the description of the Künneth component ( 1 , 1 , 2 ) follows.
(ii) The adjoint of a correspondence for the Lefschetz pairing on H 2 ( X × X ) and the polarised intersection pairing on Z K is the transposed cycle composed with cup products with powers of the polarisations and the inverse Lefschetz isomorphism, which is algebraic on an abelian variety [63, §2]; so ξ † ξ is algebraic. By Lemma 36, T X ( − 1 ) carries − 4 q 1 and T Z carries a rational multiple of q c 2 , and for ξ = m e , q c 2 ( m e x , m e y ) = Tr ( c 2 e 2 trd ( x y ) ) = q 1 ( c 2 e 2 x , y ) , so ξ † ξ = κ m e 2 c 2 .
(iii) Right multiplication by K on D = H 1 ( Z K ) commutes with the Hodge structure (Setting 6), so F ⊆ K ⊆ End 0 ( Z K ) . An element b ∈ F is central in D, so it acts on D i = D ⊗ F , σ i Q ¯ by the scalar σ i ( b ) , and on T Z , i = S 2 V i ⊗ ⋀ 2 W i ⊆ ⋀ 2 D i by σ i ( b ) 2 ; that is, b * | T Z = m b 2 . If n b is an endomorphism for an integer n > 0 , then b * = n − 2 ( n b ) * is a rational multiple of the action of the graph, hence algebraic.
(iv) By (ii) and (iii), ξ † ∘ b * ∘ ξ is algebraic, and since m b 2 commutes with ξ = m e ∘ ι it equals m a ( e ) b 2 on T X ( − 1 ) , and zero on the complement of T X ( − 1 ) in H 2 ( X × X ) , because ξ lies in T X ( − 1 ) ∨ ⊗ T Z . The squares span F over Q , as 2 b = ( b + 1 ) 2 − b 2 − 1 , so these operators span { m x : x ∈ F } . Let x ∈ F ∖ Q . Compose m x with the product map μ : H 2 ( X × X ) ⊇ T X ( − 1 ) → H 2 ( X ) and its adjoint, both built from ψ , the diagonal and θ and hence algebraic Lemma A27: the composite μ ∘ ( m x ⊗ 1 ) ∘ μ † acts on the primitive part U of H 2 ( X ) by the element ν ( Tr x − x ) / 2 of F, irrational when x is. With the identity and the projector onto Q θ this generates Q × F = End Hdg H 2 ( X ) , so by Proposition A21 the exceptional classes of X t × X t are algebraic, and then every Hodge class on every power of X t is algebraic by Theorem A24. The classes are flat along C, but by Lemma 2 algebraicity at t gives algebraicity on the stratum through t and, through the Hecke correspondences of C, on a countable dense set; whether some stratum is all of C is not decided by this argument, so the conclusion is stated at t.
(v) By Theorem A24 the Hodge conjecture holds for all powers of X once the exceptional classes are algebraic. A cycle on X r × Z K acting nonzero on the irreducible T Z carries it isomorphically onto a summand of H 2 k ( X r ) ( k − 1 ) ; projecting onto that summand and composing with an isomorphism to T X ( − 1 ) ⊆ H 2 ( X × X ) , a Hodge class on X r + 2 and algebraic, gives an algebraic nonzero element of Hom Hdg ( T Z , T X ( − 1 ) ) = F ; composing with the algebraic m a , a ∈ F , gives every element of that line, and transposing gives every ξ .    □
So the real multiplication of T is algebraic on the side of Z K for every ( F , D , K ) , and the open problem is only to transport it to X. The hypothesis on the exceptional classes in (v) is removed by the next theorem, which uses the motivic Galois group. For an abelian variety Y the Lefschetz standard conjecture holds and homological and numerical equivalence agree [72], so the homological motives generated by Y form a semisimple [61] neutral Tannakian category with the Betti fibre functor, after the usual change of the commutativity constraint by the algebraic Künneth projectors [63]. Its group G mot ( Y ) ⊆ GL ( H 1 ( Y ) ) is reductive, contains the Mumford–Tate group MT ( Y ) , commutes with End 0 ( Y ) , and for all n , p the algebraic classes in H 2 p ( Y n ) ( p ) are its invariants; hence G mot ( Y ) = MT ( Y ) if and only if the Hodge conjecture holds for all powers of Y. We write A * for algebraic classes and B * for Hodge classes.
Proposition  41
(Powers of Z K ). At a point t at which Z K , t has Hodge group G · K 1 , in particular at a very general point, the ring of Hodge classes of Z K , t b is generated by divisor classes for every b ≥ 1 , and the Hodge conjecture holds for all powers of Z K , t . At every t ∈ C , G mot ( Z K , t ) lies in the image G U of G m × G × K 1 .
Proof. 
Over Q ¯ the group G · K 1 is ∏ i GL ( P i ) with P i = V i χ i two-dimensional, and D i = P i ⊕ P i ∨ since V i is self-dual; so H 1 ( Z K b ) ⊗ Q ¯ = ⨁ i ( P i ⊕ P i ∨ ) ⊕ b . By the first fundamental theorem for GL ( P i )  [41, App. F] the invariants in a mixed tensor power of the P i and P i ∨ are spanned by complete contractions. The exterior algebra is a quotient of the tensor algebra and taking invariants of a reductive group is exact, so the invariants in ⋀ * H 1 ( Z K b ) are spanned by products of invariants of degree two. These are the Hodge classes of degree two, which are divisor classes, so the Hodge ring is generated by divisors, and the Hodge conjecture follows by the Lefschetz theorem. At every t, K ⊆ End 0 ( Z K , t ) and the three forms ψ c of Lemma 35 are algebraic, so G mot ( Z K , t ) lies in the group of K-linear g with ψ c ( g x , g y ) = ν ( g ) ψ c ( x , y ) for all c; over Q ¯ this is { ( A i , ν A i − T ) i } ≅ GL 2 3 × G m restricted to a common ν , connected of dimension 13, and it contains the connected image of G m × G × K 1 , of dimension 13; so the two coincide.    □
At a very general point dim Hdg p ( Z K ) = 1 , 3 , 6 , 7 , 6 , 3 , 1 for p = 0 , ⋯ , 6 , and dim Hdg p ( Z K 2 ) = 1 , 12 , 78 , 316 , 879 for p = 0 , ⋯ , 4 (Appendix C).
Theorem  38
(The link is all or nothing). Let t ∈ C with Hg ( X t ) = G , X = X t , Z = Z K , t for a CM extension K ⊆ D of F, and let M ⊆ GL ( V ) × GL ( D ) be the image of G m × G × K 1 . The following are equivalent.
(i)
Some nonzero link class ξ t is algebraic.
(ii)
Every link class ξ t is algebraic.
(iii)
For some r ≥ 1 some algebraic cycle on X r × Z acts nonzero on T Z , that is, T Z is isomorphic, as a homological motive, to a direct summand of some h m ( X r ) ( j ) .
(iv)
Every M -invariant class on every X a × Z b is algebraic.
Each of them implies that the exceptional classes of X × X are algebraic. If moreover t is very general, they are equivalent to
(v)
for some a , b there is an algebraic class on X a × Z b outside A * ( X a ) ⊗ B * ( Z b ) ,
and exactly one of the following holds:(I)every Hodge class on every X a × Z b is algebraic;(II)for all a , b the algebraic classes on X a × Z b are exactly A * ( X a ) ⊗ B * ( Z b ) . Case(I)holds if and only if one ξ t is algebraic.
At a very general t the M -invariant classes are all the Hodge classes, so (iv) is the Hodge conjecture for all X t a × Z K , t b .
Proof. 
Put N = G mot ( X ) and H = G mot ( X × Z ) , a subgroup of the fibre product N × G m G mot ( Z ) over the Tate character that surjects onto both factors and contains MT ( X × Z ) ; ρ : G → Sp ( V ) is the representation.
Step 1. (i)⇒(iii): ξ is such a cycle, with r = 2 . (iv)⇒(ii)⇒(i): the ξ are M -invariant.
Step 2: (iii) implies N = MT ( X ) = G m · ρ ( G ) . Let Γ : T Z → h m ( X r ) ( j ) be the nonzero morphism of motives. T Z is a simple Hodge structure, T being irreducible for G, so Γ is an isomorphism onto a submotive T ′ , and Γ ∘ b * ∘ Γ − 1 , b ∈ F , is an algebraic action of F on T ′ by Theorem 37(iii). So T ′ is an N-stable subspace on which N commutes with F and ρ ( G ) ⊆ N acts by the adjoint representation. Let n = Lie ( N ∩ Sp ( V ) ) ⊇ g = Lie ρ ( G ) . As V is irreducible for g , the centre of the reductive algebra n acts by scalars in sp ( V ) , hence is zero, and n is semisimple. Over Q ¯ , T ′ = ⨁ i T i ′ with T i ′ the eigenspaces of F, each n -stable of dimension three, with sl 2 ( i ) acting on T i ′ as so ( T i ′ ) and trivially on T j ′ , j ≠ i . The image of n in gl ( T i ′ ) is semisimple and contains so 3 , a maximal subalgebra of sl 3 , so it is so 3 or sl 3 , the same for all i since the Galois group permutes the i transitively and n is defined over Q . If it were sl 3 , the image m of n in sl 3 3 would surject onto each factor and contain so 3 3 ; then m meets the i-th factor in an ideal containing so 3 , so m = sl 3 3 , of rank six, impossible for a quotient of a subalgebra of sp 8 , of rank four. So the image is so 3 3 , the image of g , and n = k × k ′ with k the kernel and k ′ ≅ sl 2 3 . The irreducible faithful n -module V is A ⊠ B with B a faithful irreducible sl 2 3 -module, so dim B ≥ 8 , dim A = 1 and k = 0 . Hence n = g and N 0 = G m · ρ ( G ) . An element n ∈ N normalises g and preserves each T j ′ , hence each ideal sl 2 ( i ) = { x ∈ g : x T j ′ = 0 for j ≠ i } ; automorphisms of sl 2 are inner, so Ad ( n ) | g = Ad ( ρ ( g ) ) for some g, and n ρ ( g ) − 1 centralises g , hence is a scalar. Thus N = MT ( X ) , the Hodge conjecture holds for all powers of X, and the exceptional classes are algebraic.
Step 3. (iii)⇒(ii): by Step 2 the exceptional classes are algebraic, and Theorem 37(v) gives every ξ .
Step 4. (i)⇒(iv). By Step 2 and Proposition 41, H ⊆ MT ( X ) × G m G U , and H fixes ξ up to the Tate character. Over Q ¯ an element of MT ( X ) × G m G U is ( z ρ ( g ) , z ′ g ′ k ) with g , g ′ ∈ G , k ∈ K 1 and z 2 = z ′ 2 ; it acts on T X ( − 1 ) by z 2 Ad ( g ) and on T Z by z ′ 2 Ad ( g ′ ) , and ξ = m e ∘ ι with m e commuting with Ad . So it fixes ξ if and only if Ad ( g ) = Ad ( g ′ ) , that is g ′ = g ε with ε in the centre μ 2 3 of G. On D a central element ε is right multiplication by the element ( ε 1 , ε 2 , ε 3 ) of K 1 ( Q ¯ ) , and − 1 on D is the element ( − 1 , − 1 , − 1 ) of K 1 ; so ( z ρ ( g ) , z ′ g ′ k ) = ( z ρ ( g ) , z g k ′ ′ ) with k ′ ′ ∈ K 1 , an element of M . Hence H ⊆ M , and the algebraic classes, the H-invariants, contain the M -invariants.
Step 5: t very general. Then G mot ( Z ) = MT ( Z ) by Proposition 41, Hg ( Z ) = G · K 1 , and Hg ( X × Z ) is the image of G × K 1 . Let H Z = { z : ( 1 , z ) ∈ H } , a normal subgroup of MT ( Z ) . If H Z ⊇ Hg ( Z ) then H = N × G m MT ( Z ) , and the algebraic classes in H i ( X a ) ⊗ H j ( Z b ) are the invariants of ( N ∩ Sp ) × Hg ( Z ) , that is A i ( X a ) ⊗ B j ( Z b ) : this is (II), which contradicts (v). If H Z ⊇ G , then from ( ρ ( g ) , g k ) ∈ H and ( 1 , g − 1 ) ∈ H we get ( ρ ( g ) , k ) ∈ H for all g , k , hence ( 1 , k ) ∈ H and H Z ⊇ Hg ( Z ) . Otherwise H Z ∩ G is a proper normal subgroup of the Q -simple group G, hence finite, H Z 0 is a central torus, and for n ∈ H Z the map g ↦ g n g − 1 n − 1 is a homomorphism from the connected group G · K 1 to a finite group, hence trivial; so H Z lies in the centre, which acts on T Z trivially. Then the representation of H on T Z factors through H / H Z = N , that is, the motive T Z lies in the Tannakian subcategory generated by h ( X ) and is a direct summand of some h m ( X r ) ( j ) : this is (iii), and (I) holds by Steps 2 to 4. So (v) implies (iii), and if (iii) fails then (II) holds. The cases (I) and (II) exclude each other, and (i) implies (v), because ξ ∉ A 2 ( X 2 ) ⊗ B 2 ( Z ) : its component on the side of Z lies in T Z , which contains no Hodge class.    □
At a very general point the counts of Hodge classes on X a × Z b , against those of B * ( X a ) ⊗ B * ( Z b ) , are 10 / 10 , 20 / 17 , 29 / 23 , 35 / 25 in degrees 4 , 6 , 8 , 10 on X × Z , 26 / 23 , 98 / 65 , 293 / 151 in degrees 4 , 6 , 8 on X 2 × Z , and 134 / 122 , 995 / 662 in degrees 4 , 6 on X 2 × Z 2 (Appendix C); by Theorem 38 all the mixed classes are algebraic as soon as one ξ is, and in case (II) none of them is. The first mixed classes on X × Z K itself live in degree six, through the primitive part U of H 2 ( X ) rather than through T.
Theorem  39
(The tenfold link). Let U ⊆ H 2 ( X ) be the primitive part, over Q ¯ the sum of the U k = S 2 V i ⊗ S 2 V j , { i , j , k } = { 1 , 2 , 3 } , and let F act on U by m x ′ , with m x ′ = σ k ( x ) on U k .
(i)
The K 1 -invariants of ⋀ 4 D contain each U k exactly once, inside ⋀ 2 D i ⊗ ⋀ 2 D j , and their sum U Z ⊆ H 4 ( Z K ) is a rational sub-Hodge structure isomorphic to U ( − 1 ) . At t with Hg ( X t ) = G , Hom Hdg ( U X ( − 1 ) , U Z ) = F · ι ′ ; through the polarisation of H 2 ( X ) its elements are Hodge classes η ∈ H 2 ( X ) ⊗ H 4 ( Z K ) ⊆ H 6 ( X × Z K ) . At a very general t the Hodge classes in this Künneth component are B 2 ( X ) ⊗ B 4 ( Z K ) ⊕ F ι ′ , of dimension 9 = 6 + 3 , and the Hodge classes of X × Z K of degree at most four are sums of products of classes of the two factors.
(ii)
b ∈ F ⊆ End 0 ( Z K ) acts on U Z by m N ( b ) 2 b − 2 ′ , and these elements span F over Q .
(iii)
If one nonzero η is algebraic at a point t with Hg ( X t ) = G , then F acts algebraically on U ⊆ H 2 ( X t ) , the exceptional classes of X t × X t are algebraic, and the Hodge conjecture holds for all powers of X t .
(iv)
At a very general t, η t is algebraic if and only if ξ t is, if and only if the Hodge conjecture holds for all X t a × Z K , t b .
(v)
For K = F · k and t ∈ Σ K every η t is algebraic.
Proof. 
(i) The G-type U k needs even degree in V i and in V j and none in V k , so it can only occur in ⋀ 2 D i ⊗ ⋀ 2 D j , whose K 1 -invariants are ( S 2 V i ⊕ ⋀ 2 V i ) ⊗ ( S 2 V j ⊕ ⋀ 2 V j ) by Lemma 36(i); this contains S 2 V i ⊗ S 2 V j once, and it is the image of T Z , i ⊗ T Z , j under the cup product. The Galois group permutes the places, so U Z is defined over Q ; K 1 acts trivially, and the Hodge structure is induced by μ ′ as in Lemma 36, which gives U ( − 1 ) by weight. The U k are pairwise non-isomorphic irreducible G-modules, so End G U = Q ¯ 3 over Q ¯ and the Hom space is an F-line. The dimension counts are computed in Appendix C.
(ii) b * acts on ⋀ 2 D i ⊗ ⋀ 2 D j by σ i ( b ) 2 σ j ( b ) 2 = N ( b ) 2 / σ k ( b ) 2 . The Q -span of the multiplicative set { N ( b ) 2 b − 2 } is a subalgebra, hence a subfield of F; it is not Q , since N ( b ) 2 b − 2 ∈ Q forces b 2 ∈ Q and then b ∈ Q , a cubic field having no quadratic subfield. So it is F.
(iii) The adjoint η † for the polarisations is algebraic on abelian varieties, η † η = m a ′ ′ ∈ End Hdg ( U X ) = F is nonzero because η is injective and the polarisation forms are definite on each Hodge type, and η † b * η = m a ′ N ( b ) 2 b − 2 ′ ; by (ii) these span { m x ′ : x ∈ F } . With the identity and the projector onto Q θ this is End Hdg H 2 ( X ) = Q × F ; conclude by Proposition A21 and Theorem A24.
(iv) If ξ t is algebraic, Theorem 38 gives every Hodge class on X × Z K . Conversely η is an algebraic class on X × Z K outside A * ( X ) ⊗ B * ( Z K ) , since U Z contains no Hodge class, so (v) of Theorem 38 holds.
(v) Theorem 41(iii).    □
Proposition  42
(Hyper-Kähler carriers give ± 1 ). Let Y 1 , ⋯ , Y m be projective hyper-Kähler varieties of K 3 [ n j ] type whose transcendental spaces are Hodge isometric to ( T , r j q λ j ) , r j ∈ Q × , λ j ∈ F totally positive. Every composite of rational Hodge isometries between them that returns to Y j acts on T by ± 1 , and every composite of Hodge similarities by an element of Q .
Proof. 
A Hodge isometry ( T , r j q λ j ) → ( T , r k q λ k ) is a Hodge endomorphism of T, hence m a for some a ∈ F , and it is an isometry iff r j λ j = r k λ k a 2 . A composite returning to j is m a with a 2 = 1 , and a composite of similarities is m a with a 2 ∈ Q , which in a cubic field forces a ∈ Q . The isometries are algebraic by [26,77], which is why the statement matters: they realise m b between S λ and S λ b − 2 and between their moduli spaces of sheaves, but every return is trivial.    □

6.15. The Link at the Points of Complex Multiplication

The class ξ is flat along C, so the natural first test of Theorem 37 is the countable set of special points. When K contains an imaginary quadratic field the test succeeds on a dense set, with published inputs only.
Setting  7
(CM conventions). Let K = F · k with k = Q ( − d ) imaginary quadratic. Fix ι : k → C ; the embeddings of K are τ i , restricting to σ i on F and to ι on k, and their conjugates τ ¯ i , i = 1 , 2 , 3 . Let E be an elliptic curve with complex multiplication by k, of type { ι } , and A an abelian threefold with complex multiplication by K of type Φ A = { τ ¯ 1 , τ ¯ 2 , τ 3 } . The type Φ A is not lifted from k, so A is simple; E and A are determined up to isogeny by their types. Let Σ K ⊆ C be the set of points whose CM field, the F-subalgebra of D generated by the stabiliser torus, is F-isomorphic to K.
Theorem  40
(Hodge classes on E a × A b ). Let k act on B = A × E through k ⊆ K on A and through the complex multiplication on E.
(i)
( B , k ) is an abelian fourfold of Weil type: the ι-eigenspace of k on H 1 ( B ) ⊗ Q ¯ is spanned by x τ 1 , x τ 2 of type ( 0 , 1 ) and x τ 3 , y ι of type ( 1 , 0 ) , and its Weil plane W B ⊆ H 4 ( B , Q ) satisfies W B ⊗ Q ¯ = Q ¯ w ⊕ Q ¯ w ¯ with w = x τ 1 ∧ x τ 2 ∧ x τ 3 ∧ y ι .
(ii)
For all a , b ≥ 0 the ring of Hodge classes of E a × A b is generated by divisor classes and the classes φ * w ′ , φ ∈ Hom ( E a × A b , B ) ⊗ Q , w ′ ∈ W B .
(iii)
B carries a k-compatible polarisation of discriminant one. Hence W B is algebraic [76, Thm. 1.5], and the Hodge conjecture holds for every E a × A b .
Proof. 
(i) A vector x τ of the τ -eigenline of H 1 ⊗ Q ¯ of a CM abelian variety has type ( 1 , 0 ) exactly when τ lies in the CM type; read this off Φ A and { ι } . The Rosati involution of any polarisation of a CM abelian variety is complex conjugation on its CM field, so the product polarisation is k-compatible.
(ii) H 1 ( E a × A b ) ⊗ Q ¯ has the basis x τ ( c ) ( c ≤ b , τ an embedding of K) and y ϵ ( c ) ( c ≤ a , ϵ ∈ { ι , ι ¯ } ), eigenvectors of the torus ( K × ) b × ( k × ) a with pairwise distinct characters on wedge monomials. The Hodge classes are stable under this torus, hence spanned by the monomials they contain, and a monomial is a Hodge class iff all its Galois conjugates have type ( p , p )  [100]. For a monomial with n i , ± factors x τ i resp. x τ ¯ i and m ± factors y ι resp. y ι ¯ put d i = n i , + − n i , − and d E = m + − m − . The conjugates of Φ A are { τ ¯ i , τ ¯ j , τ l } , { i , j , l } = { 1 , 2 , 3 } , and their complements; the type condition for a conjugate and its complement gives by subtraction d l − d i − d j + d E = 0 for l = 1 , 2 , 3 , hence d 1 = d 2 = d 3 = d E . So a Hodge monomial is, up to sign, a product of conjugate pairs x τ ( c ) ∧ x τ ¯ ( c ′ ) , y ι ( c ) ∧ y ι ¯ ( c ′ ) and of quadruples q = x τ 1 ( c 1 ) x τ 2 ( c 2 ) x τ 3 ( c 3 ) y ι ( c 4 ) or q ¯ . For fixed c , c ′ the span of the pairs is Galois stable and of type ( 1 , 1 ) , hence spanned by divisor classes. For α ∈ K b , β ∈ k a the homomorphism φ α , β ( a 1 , ⋯ , e a ) = ( ∑ α i a i , ∑ β j e j ) has φ * x τ = ∑ i τ ( α i ) x τ ( i ) and φ * y ι = ∑ j ι ( β j ) y ι ( j ) , so the coefficient of the quadruple with indices ( c 1 , c 2 , c 3 , c 4 ) in φ * w is τ 1 ( α c 1 ) τ 2 ( α c 2 ) τ 3 ( α c 3 ) ι ( β c 4 ) . These are distinct monomials in the coordinates of ( K b × k a ) ⊗ Q ¯ , linearly independent as functions on its Zariski dense set of rational points; so the φ * w , φ * w ¯ span all quadruples.
(iii) NS ( A ) Q = { Tr K / Q ( a − d x y ¯ ) : a ∈ F } and NS ( E ) Q = Q Tr k / Q ( − d x y ¯ ) , all k-compatible; positivity is a sign condition on a at the three real places. The hermitian form of h a , b = ( a , b ) on the k-space H 1 ( A × E , Q ) = K ⊕ k is, up to a rational multiple, Tr K / k ( a x y ¯ ) ⊕ b x y ¯ ; on the k-basis of K given by a Q -basis f i of F the first block is the rational matrix Tr F / Q ( a f i f j ) , of determinant N F / Q ( a ) d F up to a square. So the discriminant is b N F / Q ( a ) d F ∈ Q × / N ( k × ) , and b = N F / Q ( a ) d F , which has the sign required for positivity, makes it trivial. The hypotheses of [76, Thm. 1.5] are signature ( 2 , 2 ) , compatibility and discriminant one; they hold. Pull-backs and products of algebraic classes are algebraic, so (ii) gives the rest.    □
Theorem  41
(The link on a dense set of special points). Let K = F · k ⊆ D as in Setting 7.
(i)
Σ K is countable and dense in C ( C ) .
(ii)
For t ∈ Σ K , X t ∼ E × A and Z K , t ∼ E 3 × A . In particular Hom ( X t , Z K , t ) ≠ 0 .
(iii)
For t ∈ Σ K and all a , b ≥ 0 the Hodge conjecture holds for X t a × Z K , t b . Every link class ξ t and every exceptional class of X t × X t is algebraic.
Proof. 
(ii) First let t = t 0 be a fixed point of K 1 . Write D = K ⊕ K j with j x = x ¯ j . Left multiplication by x ∈ K 1 and right multiplication by y ∈ K act on K by x y and on K j by x y ¯ , so both summands are sub-Hodge structures stable under the right action of K, and Z K , t 0 ∼ B 1 × B 2 with B 1 , B 2 CM threefolds for K. In the notation of Setting 6 the right τ i -eigenline of K is λ i χ i and that of K j is λ i − 1 χ i , where λ i is the character τ i of the left torus. Type ( 1 , 0 ) sits on V 1 χ 1 , V 2 χ 2 and V 3 1 , 0 W 3 , with V 3 1 , 0 the λ 3 η -line, η = ± 1 according to the fixed point. So the CM types are { τ 1 , τ 2 , τ 3 η } and { τ 1 , τ 2 , τ 3 − η } ( τ + = τ , τ − = τ ¯ ). One is the lift of { ι } , giving E ⊗ k K ∼ E 3 ; the other is the conjugate of Φ A , giving A after twisting the action by complex conjugation. For X t 0 : the torus S = Res F / Q K 1 acts on V = Cor F / Q D through G with the eight distinct characters λ s = ∏ λ i s i , s ∈ { ± 1 } 3 , so End S V is a commutative étale algebra of degree eight commuting with the Hodge structure. The Galois orbits of the sign vectors are { ± ( 1 , 1 , 1 ) } , with stabiliser Gal ( K ˜ / k ) , and the other six, with the stabiliser of ( − 1 , − 1 , 1 ) fixing the index 3; so End S V = k × K and V = V E ⊕ V A . The character λ s has type ( 1 , 0 ) iff s 3 = η ; for η = 1 the type of V E is { ι } and that of V A is { τ 3 , τ ¯ 2 , τ ¯ 1 } = Φ A , and η = − 1 gives the conjugate types and the same isogeny classes. For general t ∈ Σ K the CM field is g K g − 1 , g ∈ D × , by Skolem–Noether; left multiplication by g − 1 commutes with the right K-action and carries h t to h t 0 for a fixed point t 0 , and the representation of G on V extends to D × , so both H 1 ( Z K , t ) and H 1 ( X t ) are isomorphic to their values at t 0 .
(i) By Skolem–Noether every F-embedding of K in D is conjugate to the given one, and K 1 has exactly one fixed point t 0 in the upper half plane, since K ⊗ σ 3 R = C makes K 1 a compact torus in GL 2 ( R ) . So Σ K consists of images of fixed points of the countably many tori g K 1 g − 1 , g ∈ D × ( F ) , and is countable; it contains the images of the points g · t 0 with g ∈ D × ( F ) and σ 3 ( nrd g ) > 0 , the fixed points of g K 1 g − 1 in the upper half plane. D × ( F ) is the set of F-points of a Zariski open subset of affine four-space and Q ⊆ F is dense in R , so these g are dense in GL 2 + ( R ) = ( D ⊗ σ 3 R ) × , + , which acts transitively and continuously on the upper half plane; the orbit of t 0 is dense, and so is its image in C.
(iii) By (ii), X t a × Z K , t b ∼ E a + 3 b × A a + b ; apply Theorem 40. The classes ξ t and the exceptional classes are Hodge classes at every t (Theorem 37(i)).    □
Table 7. The link class ξ t along the Shimura curve C. “Algebraic” means proved here with published inputs; the alternative of Lemma 2 says the algebraic locus of ξ is C or countable.
Table 7. The link class ξ t along the Shimura curve C. “Algebraic” means proved here with published inputs; the alternative of Lemma 2 says the algebraic locus of ξ is C or countable.
Points t ∈ C Decomposition of X t , Z K , t Status of ξ t
t ∈ Σ K , K = F · k E × A , E 3 × A algebraic (Thm. 41)
CM, field F · k ′ , k ′ ≠ k X t ∼ E ′ × A ′ powers of X t algebraic; ξ t 1 open
CM, field without imag. quadr. subfield simple, octic End 0 X t ξ t 1 open (degree-12 Weil class)
not CM (Prop. 44) simple, End 0 X = Q open; equivalent to HC for all X a × Z K b (Thm. 38); implied by one algebraic γ t (Thm. 44); not reached by divisors, Weil classes and homomorphisms (Prop. 45)
Neither link is reached at a very general point by the published algebraicity theorems, which produce Weil classes and pull-backs of them, or work through hyper-Kähler varieties.
Proposition  43
(No Weil class and no hyper-Kähler carrier). Let t be very general.
(i)
Let E ′ be a number field, A ′ an abelian variety with E ′ ⊆ End 0 ( A ′ ) and H 1 ( A ′ ) free of rank r over E ′ , w ∈ ⋀ E ′ r H 1 ( A ′ ) , and f : X a × Z K b → A ′ a homomorphism. Then f * w = 0 , or f * w lies in a sum of Künneth components H i ( X a ) ⊗ H j ( Z K b ) with i ∈ 8 Z . Consequently ξ is not in the algebra generated by divisor classes and all such pull-backs, and neither is η.
(ii)
( Z K , K ) is not of Weil type, its signatures at the three places being ( 2 , 0 ) , ( 2 , 0 ) , ( 1 , 1 ) , and ( Z K , k ) for an imaginary quadratic k ⊆ K has signature ( 5 , 1 ) .
(iii)
No projective hyper-Kähler variety of generalised Kummer type or of OG6 type has a transcendental Hodge structure containing T.
Proof. 
(i) X is simple with End 0 X = Q  [87], and End 0 Z K is the commutant of G · K 1 in End Q D , which is the field K, so Z K is simple; Hom ( X , Z K ) = 0 . E ′ preserves the isotypic decomposition A ′ ∼ A X ′ × A Z ′ × A o ′ into powers of X, powers of Z K and the rest, so H 1 ( A ′ ) = H X ⊕ H Z ⊕ H o as E ′ -modules of ranks r X + r Z + r o = r , and ⋀ E ′ r H 1 ( A ′ ) = ⋀ E ′ r X H X ⊗ E ′ ⋀ E ′ r Z H Z ⊗ E ′ ⋀ E ′ r o H o . The homomorphism f has zero component into A o ′ , so f * w = 0 unless r o = 0 . If A X ′ ∼ X m with m > 0 , then E ′ ⊆ End 0 ( X m ) = M m ( Q ) , so Q m is an E ′ -vector space, m = m ′ [ E ′ : Q ] , and r X = 8 m / [ E ′ : Q ] = 8 m ′ . So the degree of f * w in the variables of X lies in 8 Z . In degree four the generators are divisor classes, which are sums of classes pulled back from the two factors since Hom ( X , Z K ) = 0 , and pull-backs f * w of degree four, which have X-degree zero; so the degree-four part of the algebra lies in B 4 ( X a ) ⊕ B 2 ( X a ) ⊗ B 2 ( Z K b ) ⊕ H 4 ( Z K b ) , while ξ has its component in T X ( − 1 ) ⊗ T Z and T X ( − 1 ) ∩ B 2 ( X 2 ) = 0 . In degree six on X × Z K a product of generators has X-degree two only if exactly one divisor class of X occurs among its factors, so the component in H 2 ( X ) ⊗ H 4 ( Z K ) lies in Q θ ⊗ H 4 ( Z K ) , while η ∈ U X ⊗ U Z and U ∩ Q θ = 0 .
(ii) The χ i -eigenspace of H 1 ( Z K ) is V i χ i , of type ( 1 , 0 ) for i = 1 , 2 and of type ( 1 , 0 ) + ( 0 , 1 ) for i = 3 (Setting 6); for k add over the three places.
(iii) A variety of generalised Kummer type has b 2 = 7 and one of OG6 type b 2 = 8 , and a projective variety has Picard number at least one, so the transcendental rank is at most six or seven, while T has rank nine.    □
By (iii), the constructions of Floccari and Fu through varieties of OG6 type [39] and of van Geemen and Rapagnetta through hyper-Kähler sixfolds of K3[3] type [119], which start from an abelian fourfold of Weil type of discriminant one, do not apply at a very general point, where End 0 X t = Q ; at the points of complex multiplication of C, where X t is of Weil type (Remark 15), they give a second proof of the Hodge conjecture for the powers of X t when the discriminant is one. For F = Q ( ζ 7 ) + and Q ( ζ 9 ) + and the quaternion algebras of the triangle groups Δ ( 2 , 3 , 7 ) and Δ ( 2 , 3 , 9 ) , ramified at two real places, Bouchet, Hanselman, Pieper and Schiavone write down two explicit families of curves of genus four over Q ( t ) whose Jacobians are Mumford fourfolds [16]; so for these ( F , D ) the fourfolds X t of this section are Jacobians of explicit curves, and Theorem 14 settles (F3′) for the corresponding K3 surfaces S λ , while the exceptional classes of X t × X t remain open.
Remark  15
(Every special point, and the residual class). At any CM point t with CM field L = F ( ℓ ) ⊆ D , T = F ℓ ⊕ L j ′ as Hodge structures, F ℓ ( − 1 ) consists of Hodge classes, and ξ t = ξ t 0 + ξ t 1 with ξ t 0 a sum of products of divisor classes and ξ t 1 ∈ Hom Hdg ( ( L j ′ ) X , ( L j ′ ) Z ) ; so ξ t is algebraic iff ξ t 1 is. The fourfold X t is always of Weil type: End 0 X t contains k * = Q ( − N F / Q δ ) when L = F ( − δ ) , with signature ( 2 , 2 ) , and the computation of Theorem 40(ii) shows that the Hodge ring of every power of X t is generated by divisors and pull-backs of its Weil plane. The Hodge conjecture for all powers of X t then follows from [76, Thm. 1.5] whenever X t carries a k * -compatible polarisation of discriminant one, which julia/mumfordresidual.jl exhibits for F = Q ( ζ 7 ) + , Q ( ζ 9 ) + and the cubic field of discriminant 148, and at every point of complex multiplication from Markman’s theorem for all abelian fourfolds of Weil type [78,79]. When L has no imaginary quadratic subfield, ξ t 1 is equivalent to a single Hodge line over the degree-twelve CM field M = L · k * , the pull-back of a Weil class for M on an abelian variety isogenous to X t 3 × B 1 2 × B 2 2 with dim M H 1 = 4 ; no algebraicity theorem covers Weil classes for CM fields of degree larger than two, and [80] sets up this case without closing it.

6.16. Non-Special Points, a Squaring Lemma and a Third Link

The statements made above at a very general point hold at every point of C that is not a point of complex multiplication, and the middle cohomology of Z K carries a canonical tensor decomposition that gives a third link class, with an explicit geometric prediction in the families of [16]. We call t ∈ C a CM point if MT ( X t ) is a torus.
Proposition  44
(Every non-CM point is generic). Let t ∈ C not be a CM point and K ⊆ D a CM extension of F. Then MT ( X t ) = G m · G , MT ( Z K , t ) = G m · G · K 1 and MT ( X t × Z K , t ) = M . Hence Lemma 35 with dim NS ( Z K , t ) Q = 3 , Proposition 41, the counts of Theorem 37(i), the statements of Theorem 38 for a very general point, Theorem 39(i) and Proposition 43 hold at t.
Proof. 
MT ( X t ) is a connected reductive subgroup of G m · G containing the weight, and its derived group H is a connected semisimple subgroup of G. Over Q ¯ , G = SL 2 3 , and a connected semisimple subgroup of SL 2 3 with nontrivial projections is a product of copies of SL 2 , each embedded diagonally, through isomorphisms, into the factors of one block of a partition of { 1 , 2 , 3 } ; the projections of H are all trivial or all nontrivial because Gal ( Q ¯ / Q ) permutes the three factors transitively, and the partition is Galois stable, hence { 1 } , { 2 } , { 3 } or { 1 , 2 , 3 } , as a transitive subgroup of S 3 contains a three-cycle. In the diagonal case the components of the Hodge cocharacter in G ad at the three places are identified by isomorphisms, but μ X is nontrivial only at σ 3 . So H = 1 , and t is a CM point, or H = G . For Z = Z K , t : MT ( Z ) ⊆ G m · G · K 1 (Setting 6), and its image in G ad contains the image of μ X , whose Mumford–Tate group is G ad ; so MT ( Z ) der ⊆ G maps onto G ad and equals G. Its image in the torus ( G m · G · K 1 ) / G contains the images of the weight and of μ ′ ′ , which generate that torus by the proof of Lemma 35. So MT ( Z ) = G m · G · K 1 . Finally MT ( X t × Z ) ⊆ M maps onto MT ( Z ) , and the kernel of M → GL ( D ) is finite (an element acting trivially on D acts on V by a sign, as in Step 4 of the proof of Theorem 38); both groups are connected of the same dimension, so they are equal. The cited statements use only these groups.    □
Since the families descend to a number field, the algebraic locus of a flat class such as ξ is C or a countable union of points defined over Q ¯ (Lemma 2); so algebraicity of ξ at one point outside C ( Q ¯ ) is equivalent to algebraicity on all of C.
Over Q ¯ write D τ for the eigenspace of the right action of K on D for an embedding τ of K, so D τ i = V i χ i and D τ ¯ i = V i χ i − 1 with τ i , τ ¯ i the embeddings above σ i . A CM type Φ of K is a choice of one embedding above each σ i ; put N Φ = ∏ τ ∈ Φ τ and D Φ = ⨂ τ ∈ Φ D τ ⊆ ⋀ 3 D .
Theorem  42
(Tensor decomposition of H 3 ( Z K ) ). Let t ∈ C be any point, X = X t and Z = Z K , t .
(i)
The V-isotypic part Π of H 3 ( Z , Q ) = ⋀ 3 D for the action of G is a sub-Hodge structure of dimension 64, and Π ⊗ Q ¯ = ⨁ Φ D Φ over the eight CM types.
(ii)
M = ( H 1 ( X ) ∨ ⊗ Π ) G is a sub-Hodge structure of H 1 ( X ) ∨ ⊗ H 3 ( Z ) of dimension 8, of CM type, of weight two, with Hodge numbers ( 2 , 4 , 2 ) , and evaluation H 1 ( X ) ⊗ M → Π is an isomorphism of Hodge structures. Over Q ¯ , M = ⨁ Φ m Φ with K acting on the line m Φ by N Φ , of type ( a , 2 − a ) for a = | Φ ∩ { τ 1 , τ 2 } | .
(iii)
The Hodge structure M = M t does not depend on t: it is the space of flat sections of H 1 ( X ) ∨ ⊗ H 3 ( Z ) over a finite cover of C.
(iv)
If K = F · k as in Setting 7 and Φ 0 = { τ 1 , τ 2 , τ 3 } , then M L = m Φ 0 ⊕ m Φ ¯ 0 is defined over Q , of type ( 2 , 0 ) + ( 0 , 2 ) , and M L ≅ T ( E × E ) , the transcendental part of H 2 ( E × E ) , for every elliptic curve E with complex multiplication by k. Hence Π L = ev ( H 1 ( X ) ⊗ M L ) ≅ H 1 ( X ) ⊗ T ( E × E ) , and H 1 ( X t ) is a direct summand of H 1 ( Z K , t ) ⊗ 3 ⊗ H 1 ( E ) ⊗ 2 ( 2 ) at every t ∈ C .
Proof. 
(i) Over Q ¯ , ⋀ 3 ⨁ τ D τ = ⨁ ⨂ τ ⋀ n τ D τ with ∑ n τ = 3 . As a G-module D τ is the standard representation V i of the factor above σ i , so ⋀ 0 D τ and ⋀ 2 D τ are trivial; a summand is isomorphic to V = V 1 ⊗ V 2 ⊗ V 3 exactly when n τ = 1 for one τ above each σ i and n τ = 0 otherwise, that is, when it is D Φ . V is defined over Q as a G-module, so Π is defined over Q ; it is stable under G m , G and the commuting torus K 1 , hence under MT ( Z ) , so it is a sub-Hodge structure. (ii) The Hodge structure of X × Z is given by a cocharacter of M , in which G is normal, so the G-invariants of H 1 ( X ) ∨ ⊗ H 3 ( Z ) form a sub-Hodge structure on which M acts through the torus M / G : it is of CM type. V is absolutely irreducible with End G V = Q , so dim M = dim Hom G ( V , Π ) = 8 and evaluation is an isomorphism of G-modules and of Hodge structures. Right multiplication by K acts by endomorphisms of Z, hence on M by Hodge endomorphisms, with the eight distinct characters N Φ . The factor of D Φ above σ i , i = 1 , 2 , has type ( 1 , 0 ) if τ i ∈ Φ and ( 0 , 1 ) otherwise, and the factor above σ 3 carries the Hodge structure of V 3 ; so D Φ ≅ V ⊗ C ( a , 2 − a ) and m Φ has type ( a , 2 − a ) . Two types have a = 2 , two have a = 0 and four have a = 1 . (iii) Over a finite cover of C carrying both families the monodromy is an arithmetic subgroup of G ( Q ) , acting through ρ on H 1 ( X ) and by left multiplication on D, Zariski dense in G; so the flat global sections are the G-invariants M t , and by the theorem of the fixed part [4,33] they form a sub-Hodge structure independent of t. (iv) N Φ 0 ( x ) = τ 1 ( x ) τ 2 ( x ) τ 3 ( x ) = ι ( N K / k ( x ) ) . The Galois group permutes the τ i when it fixes k and maps Φ 0 to Φ ¯ 0 otherwise, so { N Φ 0 , N Φ ¯ 0 } is a Galois orbit and M L is defined over Q ; by (ii) it has type ( 2 , 0 ) on m Φ 0 and ( 0 , 2 ) on m Φ ¯ 0 . K × acts through N K / k ( K × ) , which spans k, so M L is a one-dimensional k-vector space with k acting on its ( 2 , 0 ) -line by ι . The same holds for T ( E × E ) ⊆ H 1 ( E ) ⊗ H 1 ( E ) , spanned over Q ¯ by y ι ⊗ y ι ′ and y ι ¯ ⊗ y ι ¯ ′ , with k acting on the first factor; a weight-two Hodge structure on a one-dimensional k-vector space with h 2 , 0 = h 0 , 2 = 1 is determined by the embedding acting on H 2 , 0 , and replacing that embedding by its conjugate does not change the abstract Hodge structure. So M L ≅ T ( E × E ) , and evaluation gives Π L ≅ H 1 ( X ) ⊗ T ( E × E ) . Hence H 1 ( X ) ≅ Hom ( T ( E × E ) , Π L ) ⊆ Π L ⊗ T ( E × E ) ( 2 ) , using the polarisation. ⋀ 3 is a direct summand of ⊗ 3 , T ( E × E ) ⊆ H 1 ( E ) ⊗ 2 , and polarisable Hodge structures are semisimple, so H 1 ( X ) is a direct summand.    □
At a point that is not a CM point there is no Hodge class in H 1 ( X ) ⊗ H 3 ( Z K ) itself (Appendix C.5): the twist by E × E is needed.
Theorem  43
(Squaring lemma). Let X be a Mumford fourfold with Hg ( X ) = G , Y an abelian variety such that every Hodge class on every power of Y is algebraic, and B an abelian variety of CM type. Suppose that an algebraic class γ on X × Y × B acts nonzero on H 1 ( X ) , that is γ * : H 1 ( X ) → H * ( Y × B ) is nonzero.
(a)
Every Hodge endomorphism of H 1 ( X ) ⊗ H 1 ( X ) is algebraic as a class on X 4 ; the exceptional classes of X × X are algebraic, and the Hodge conjecture holds for every power of X.
(b)
If X = X t and Y = Z K , t at a point t that is not a CM point, then the link classes ξ t are algebraic, and the Hodge conjecture holds for every X t a × Z K , t b .
Proof. 
The Künneth projectors are algebraic, so we may take a component U ⊗ N = H i ( Y ) ⊗ H j ( B ) on which γ * is nonzero, a Hodge class in Hom ( V , U ⊗ N ) up to a Tate twist. Let H = MT ( X × Y × B ) and H ′ its derived group. Then H ′ maps onto Hg ( X ) der = G , acts trivially on N because its image in the torus MT ( B ) is trivial, and fixes γ * . Put R = Hom H ′ ( V , U ) , a sub-Hodge structure as H ′ is normal in H , and let ev : V ⊗ R → U be the inclusion of the V-isotypic part (V is absolutely irreducible for H ′ ). As N is H ′ -trivial, γ * ∈ Hom H ′ ( V , U ⊗ N ) = R ⊗ N ; write γ * = ∑ ℓ ρ ℓ ⊗ n ℓ . A polarisation Q N of N is algebraic on B × B , the Lefschetz decomposition and pairing being algebraic on abelian varieties. A polarisation Q U of U is H ′ -invariant, the V-isotypic part is orthogonal to the others, and there Q U = ψ ⊗ Q R , since the invariant bilinear forms on V are Q ψ ; the Weil operators satisfy C U = C V ⊗ C R , so ± Q R polarises R.
Let c = ( 1 ⊗ Q N ⊗ 1 ) ∘ ( γ * ⊗ γ * ) : V ⊗ V → U ⊗ U . It is the action of the class obtained from γ × γ on X 2 × ( Y × B ) 2 by intersecting with Q N and pushing forward, so it is algebraic and a morphism of Hodge structures, and c ( v ⊗ w ) = ( ev ⊗ ev ) ( v ⊗ w ⊗ s ) with s = ∑ ℓ , ℓ ′ Q N ( n ℓ , n ℓ ′ ) ρ ℓ ⊗ ρ ℓ ′ ∈ R ⊗ R . Contracting s with Q R gives ( Q R ⊗ Q N ) ( γ * , γ * ) , nonzero by the Hodge–Riemann relations, since γ * is a nonzero rational class of type ( p , p ) in the polarised Hodge structure ( R ⊗ N , ± Q R ⊗ Q N ) . So s ≠ 0 , and c is injective.
(a) The adjoint c † for the polarisations of H 1 ( X ) ⊗ 2 and U ⊗ 2 is algebraic, by the Lefschetz theory on X 2  [63,72] and by the hypothesis on Y. The Hodge endomorphism c † c of V ⊗ V is positive definite for Q ( x , C x ¯ ) , hence invertible, and its inverse is a polynomial in it. For ϕ ∈ End Hdg ( V ⊗ V ) the class c ϕ ( c † c ) − 1 c † is a Hodge class on Y 4 , hence algebraic, and ϕ = ( c † c ) − 1 c † c ϕ ( c † c ) − 1 c † c . In particular the operators m x , x ∈ F , on T X ( − 1 ) , extended by zero on the other isotypic parts, are algebraic: these are the exceptional classes, and Theorem A24 gives every power of X. (b) Now U ⊗ U ⊆ H * ( Z 2 ) , and c ( T X ( − 1 ) ) ≅ T X ( − 1 ) ≅ T Z (Lemma 36). By semisimplicity a nonzero Hodge morphism c ( T X ( − 1 ) ) → T Z extends to a Hodge morphism ϕ : U ⊗ U → H 2 ( Z ) , a Hodge class on Z 3 , algebraic by Propositions 44 and 41. The algebraic morphism ϕ ∘ c from H 1 ( X ) ⊗ 2 to T Z is G-equivariant, so it vanishes on the isotypic parts of V ⊗ V other than T, which occurs once; it is a nonzero element of F · ι (Theorem 37(i)), an algebraic ξ t . Theorem 38 gives the rest.    □
The hypothesis that B is of CM type is what makes H ′ act trivially on N; for B = X and γ the diagonal one has R = 0 .
Theorem  44
(The third link). Let K = F · k and E be as in Setting 7, and for t ∈ C let Λ t be the space of flat Hodge classes in H 1 ( X t ) ∨ ⊗ H 3 ( Z K , t ) ⊗ H 2 ( E × E ) , suitably twisted; write γ t for a nonzero element.
(a)
Through M L ∨ ≅ T ( E × E ) ( 2 ) ⊆ H 2 ( E × E ) ( 2 ) , the maps v ↦ ∑ i ev ( v ⊗ x m i ) ⊗ m i ∨ , x ∈ k , with ( m i ) a basis of M L and ( m i ∨ ) the dual basis, lie in Λ t and form a two-dimensional space of flat sections, of Hodge type at every t; at a point that is not a CM point it is the space of all Hodge classes in that Künneth component. Every γ t is injective on H 1 ( X t ) .
(b)
For t ∈ Σ K every γ t is algebraic.
(c)
At a point t that is not a CM point, if one γ t is algebraic then every link class ξ t is algebraic, the exceptional classes of X t × X t are algebraic, and the Hodge conjecture holds for every X t a × Z K , t b .
(d)
At a point t that is not a CM point, γ t is algebraic if and only if every Hodge class on every X t a × Z K , t b × E c is algebraic. In particular, if γ t is algebraic then so are the Weil classes of the abelian tenfold ( Z K , t × E 4 , k ) , with k acting through K on Z K , t and by the conjugate of the complex multiplication on E, of signature ( 5 , 5 ) .
(e)
The algebraic locus of γ is C or a countable dense subset containing Σ K .
Proof. 
(a) By Theorem 42(iii), (iv) these maps correspond to the elements x of End k ( M L ) = k through evaluation, so they are flat, and they are Hodge classes because M L is a sub-Hodge structure on which k acts by Hodge endomorphisms. The count at a point that is not a CM point is in Appendix C.5. There H 1 ( X t ) is simple, so γ t ≠ 0 is injective; at the other points γ t is the flat transport of an injective map. (b) By Theorem 41(ii), X t × Z K , t × E 2 ∼ E 6 × A 2 , and Theorem 40(iii) applies. (c) Theorem 43(b) with Y = Z K , t and B = E × E . (d) If γ t is algebraic, then by (c) and Theorem 38 G mot ( X × Z ) = M , and G mot ( E ) = MT ( E ) = T k = Res k / Q G m ; so G mot ( X × Z × E ) lies in the fibre product P = M × G m T k over the weight characters. Over Q ¯ , Λ ⊗ Q ¯ = Q ¯ γ + ⊕ Q ¯ γ − with γ + in the component of ( Φ 0 , ι ) and γ − in that of ( Φ ¯ 0 , ι ¯ ) , and P acts on them by characters λ + and λ − = λ + − 1 , with λ + ( z ρ ( g ) , z g u ; e ) = N Φ 0 ( u ) ι ¯ ( e ) / ι ( e ) for u ∈ K 1 . A nonzero rational γ has both components nonzero, being fixed by complex conjugation, so G mot ( X × Z × E ) ⊆ H ′ = ker λ + . Now MT ( X × Z × E ) ⊆ H ′ and maps onto MT ( X × Z ) = M (Proposition 44); H ′ → M is surjective with kernel { e : ι ( e ) ι ¯ ( e ) = 1 , ι ( e ) = ι ¯ ( e ) } = { ± 1 } , so dim H ′ = dim M ≤ dim MT ( X × Z × E ) . H ′ is connected: the loop θ ↦ u ( θ ) ∈ K 1 ( C ) with χ 1 = e i θ , χ 2 = χ 3 = 1 is closed in M , and its lift to H ′ through e ( θ ) with ι ( e ( θ ) ) = e i θ / 2 , ι ¯ ( e ( θ ) ) = e − i θ / 2 runs from the identity to the element − 1 of the kernel. Hence H ′ = MT ( X × Z × E ) = G mot ( X × Z × E ) , and every Hodge class on every X a × Z b × E c is algebraic. The converse is trivial. ( Z K , k ) has signature ( 5 , 1 ) (Proposition 43(ii)) and ( E , k ) with the conjugate action has signature ( 0 , 1 ) , so ( Z K × E 4 , k ) is of Weil type of signature ( 5 , 5 ) , and its Weil classes are Hodge classes. (e) Lemma 2, (b) and Theorem 41(i).    □
The Weil classes of (d) are the first Hodge classes on a product Z K , t × E c that are not generated by divisors: there is none for c ≤ 3 and a plane in degree ten for c = 4 (Appendix C.5). The converse of (d) does not follow from the classes known so far: the element g 0 = ( 1 , e 0 ) ∈ P with ι ( e 0 ) = i , ι ¯ ( e 0 ) = − i fixes every Hodge class on every X a × Z b , lies in MT ( E ) , fixes the Weil classes of (d), on which P acts through λ + ± 2 , and has λ + ( g 0 ) = − 1 .
For F = Q ( ζ 7 ) + and Q ( ζ 9 ) + , Bouchet, Hanselman, Pieper and Schiavone consider the genus-four curves C n , t ( n = 7 , 9 ) whose Jacobians are the Mumford fourfolds over the triangle curves, and the hypergeometric curves X 7 , t : y 84 = x 13 ( x − 1 ) 43 ( x − t ) 29 and X 9 , t : y 36 = x 5 ( x − 1 ) 19 ( x − t ) 13 , whose Pryms are isogenous to fourth powers of abelian sixfolds A n , t with an action of K 7 = Q ( ζ 7 ) + ( − 21 ) , resp. K 9 = Q ( ζ 9 ) + ( i ) ; they conjecture that H 1 ( X n , t , Q ) ⊗ 3 ⊗ H 1 ( E n , Q ) ⊗ 2 ( 2 ) has a summand isomorphic to H 1 ( C n , t , Q ) , with E 7 , E 9 elliptic curves with complex multiplication by Q ( − 21 ) , Q ( i )  [16, Conj. 5.1, 5.2].
Corollary  18
(The conjectures of Bouchet–Hanselman–Pieper–Schiavone, conditionally). Let n ∈ { 7 , 9 } and assume
(I)
for every t there is a point τ ( t ) of the Shimura curve of ( F , D ∞ ) with H 1 ( C n , t , Q ) ≅ H 1 ( X τ ( t ) , Q ) and H 1 ( A n , t , Q ) ≅ H 1 ( Z K n , τ ( t ) , Q ) , possibly after replacing the action of K n by its conjugate,
which is what the descriptions of the two period maps in [16, Thm. 1.2, Thm. 4.15, §5.3], together with the results on the hypergeometric families quoted there, give. Then H 1 ( C n , t , Q ) is a direct summand of H 1 ( X n , t , Q ) ⊗ 3 ⊗ H 1 ( E n , Q ) ⊗ 2 ( 2 ) for every t, which is [16, Conj. 5.1, 5.2]. If at one fibre t that is not a CM point some algebraic cycle on C n , t × E n m × A n , t r acts nonzero on H 1 ( C n , t ) , for instance a correspondence on C n , t × E n 2 × X n , t 3 inducing the summand composed with the projection to the Prym part, then the exceptional classes of Jac ( C n , t ) 2 are algebraic and the Hodge conjecture holds for every power of Jac ( C n , t ) .
Proof. 
K n = F · k with k = Q ( − 21 ) , resp. Q ( i ) , and K n embeds in D ∞ , a CM extension of F embedding in every quaternion algebra over F ramified only at real places. Theorem 42(iv) at τ ( t ) gives the summand of H 1 ( Z K n ) ⊗ 3 ⊗ H 1 ( E n ) ⊗ 2 ( 2 ) , and H 1 ( A n , t ) is a direct summand of H 1 ( X n , t ) , the image of an idempotent of Q [ μ N ] on the Prym part; conjugating the action of K n does not change T ( E n × E n ) . The second statement is Theorem 43(a) with X = Jac ( C n , t ) , Y = A n , t ≅ Z K n , τ ( t ) and B = E n m , the hypothesis on Y holding by Propositions 44 and 41.    □
Neither ξ , η nor γ is reached by the operations that produce the known algebraic classes.
Proposition  45
(No-go for divisors, Weil classes and homomorphisms). Let X be a Mumford fourfold with Hg ( X ) = G , and let W be the smallest system of subspaces W ( X a × Y ) ⊆ H * ( X a × Y , Q ) , for all a ≥ 0 and all abelian varieties Y with Hom ( X , Y ) = 0 , that contains
(1)
the divisor classes,
(2)
all classes pulled back from Y, algebraic or not,
(3)
all f * w with f : X a × Y → A ′ a homomorphism to an abelian variety A ′ with an action of a number field E ′ for which H 1 ( A ′ ) is free over E ′ , and w in the Weil space ⋀ E ′ r H 1 ( A ′ ) ,
and is closed under cup products and under pull-back and push-forward along homomorphisms X a × Y → X b × Y ′ . Then W ( X a × Y ) ⊆ H * ( X a ) Sp ( V ) ⊗ H * ( Y ) , where Sp ( V ) acts diagonally on H 1 ( X a ) = V ⊗ Q a , and H * ( X a ) Sp ( V ) is the subalgebra generated by divisor classes. In particular the exceptional classes of X × X , the link classes ξ, η and γ are not in W .
Proof. 
Hom ( X , Y ) = 0 implies Hom ( Y , X ) = 0 by Poincaré reducibility, and End ( X ) = Z , so a homomorphism X a × Y → X b × Y ′ is a pair ( M , ϕ ) with M an integral b × a matrix and ϕ : Y → Y ′ ; M * and M * commute with the diagonal Sp ( V ) . (1) NS ( X a × Y ) Q = NS ( X a ) Q ⊕ NS ( Y ) Q , since Hom ( X , Y ^ ) = 0 , and NS ( X a ) Q = Q ψ ⊗ Sym a ( Q ) is Sp ( V ) -invariant. (2) is clear. (3) A ′ ∼ X m × A ′ ′ with Hom ( X , A ′ ′ ) = 0 , and E ′ preserves this decomposition, so H 1 ( A ′ ) = ( V ⊗ Q m ) ⊕ H 1 ( A ′ ′ ) as E ′ -modules, with E ′ acting on Q m through E ′ ⊆ M m ( Q ) = End 0 ( X m ) . For an embedding σ of E ′ the σ -eigenspace of V ⊗ Q m ⊗ Q ¯ is V ⊗ U σ , and Sp ( V ) acts on the line ⋀ top ( V ⊗ U σ ) by det dim U σ = 1 . The Weil space is spanned over Q ¯ by the lines ⋀ top ( V ⊗ U σ ) ⊗ ⋀ top H σ ′ ′ , so it lies in H * ( X m ) Sp ( V ) ⊗ H * ( A ′ ′ ) , and f = ( M , ϕ ) pulls it back into H * ( X a ) Sp ( V ) ⊗ H * ( Y ) . Cup products and ( M , ϕ ) * , ( M , ϕ ) * preserve this subspace. By the first fundamental theorem for Sp ( V )  [41, App. F] the invariants in ⋀ * ( V ⊗ Q a ) are generated by those of degree two, Q ψ ⊗ Sym a ( Q ) , which are divisor classes. The exceptional classes and ξ have components in T X ( − 1 ) ⊆ V ⊗ V , whose Sp ( V ) -invariants are zero; η has its component in U ⊆ H 2 ( X ) , and U ∩ Q θ = 0 ; and γ has degree one in the variables of X, where V Sp ( V ) = 0 .    □
Proposition 45 contains Proposition 43(i): even the algebraicity of all Weil classes of all number fields in all dimensions would not give ξ , η or γ by these operations. A proof has to produce a cycle whose class breaks the Sp ( V ) -symmetry, that is, a cycle tied to the geometry of X t , such as the correspondence of Corollary 18, or a deformation argument from Σ K .

6.17. Where the Link Stops

Theorem 37 replaces the Kuga–Satake variety of Theorem A23, of dimension 2 7 , by an abelian sixfold over the same curve, and it replaces the irrationality that had to be extracted from anticommuting Clifford blocks by the real multiplication of Z K , which is algebraic for every K (Theorem 37(iii)): for every K every link closes the target. By Theorem 38 the sixfold link is all or nothing: one algebraic cycle on some X r × Z K acting nonzero on T Z gives the Hodge conjecture for every X a × Z K b , and without it the algebraic classes on these products are exactly A * ( X a ) ⊗ B * ( Z K b ) . It is an equivalent form of the target, not a reduction to something smaller.
The missing step is the algebraicity of one class ξ ∈ H 4 ( X × X × Z K ) at one point of C that is not a CM point; by Proposition 44 every such point behaves like a very general one, and by Theorem 44 one algebraic γ t would suffice. It is exceptional: its component in V ⊗ V is the nine-dimensional T, not a multiple of ψ , so it is not invariant under Sp ( V ) and not in the Lefschetz algebra of X × X × Z K (Lemma 1); the theorems that prove the conjecture when the Hodge ring is generated by divisors do not apply, and by Propositions 43 and 45 neither divisor classes, Weil classes for any field in any dimension and homomorphisms, nor hyper-Kähler varieties of Kummer or OG6 type can be fed in: what is needed is a cycle that sees the geometry of X t , of which the correspondence predicted by Bouchet, Hanselman, Pieper and Schiavone (Corollary 18) is the most concrete candidate; the tenfold link η of Theorem 39 is equivalent to ξ and meets the same obstruction. When K = F · k , Theorem 41 gives ξ t on the dense set Σ K , so by Lemma 2 the algebraic locus of ξ is either countable and dense or all of C; deciding between the two is the variational Hodge conjecture for one flat section over a curve, and it is open. When K contains no imaginary quadratic field, the smallest open instance is the residual class ξ 1 of Remark 15 at the fixed point of K 1 : one Hodge line over a CM field of degree twelve on a tenfold. This is the price of staying inside abelian varieties, and it is the exact statement that (F2) needs at its smallest instance.
Table 8. Every statement of the paper with its status. “Proved” means proved in the text from the cited results; “open” means not proved here or elsewhere to our knowledge.
Table 8. Every statement of the paper with its status. “Proved” means proved in the text from the cited results; “open” means not proved here or elsewhere to our knowledge.
statement where status what it needs
HC ⇔ ( F 2 ) ∧ ( F 3 ′) Thm. 1 proved –
(F3′) for fourfolds with zero-cycles on a threefold Prop. 3 proved from [24] –
(F2) ⇔ AB Thm. 8 proved –
six minimal sufficient sets Prop. 4 proved, checked in Lean –
dichotomy of the algebraic locus Lem. 2 proved –
(F2) as propagation from CM points Thm. 9 proved –
Weil loci dense with CM base point Prop. 6 proved –
chains make e with e 2 ∈ Q algebraic Thm. 10 proved a chain of the right multiplier
symplectic chains reach exactly 2 , 3 , 6 Prop. 20 proved –
6 algebraic on two-dimensional families Thm. 11 proved –
− 5 , − 7 algebraic on one-dimensional families Thm. 13 proved –
HC on all powers and Hilbert schemes of the Fricke surfaces Thm. 13 proved –
similarities act by multipliers d 1 d 2 Prop. 9 proved –
finite-group correspondences give abelian E Prop. 10 proved –
HC for X × X , X in the dihedral families Cor. 2 proved –
(F3′) for S λ , F = Q ( ζ 7 ) + or Q ( ζ 9 ) + , D = D ∞ Thm. 14 proved –
where the dihedral mechanism stops Prop. 11 proved –
(F3′) for S λ , other cubic F or other D Sect. 5.10 open a non-similarity correspondence
every Q ( d ) in transcendental rank six, HC on all powers Thm. 16 proved from [42] –
which T ( S ) Q embed in some V m Lem. 11 proved –
ρ ≥ 17 : Kuga–Satake, HC , (F3′) on S n , S [ n ] Cor. 3 proved from [43] –
ρ = 16 , Witt index two: the same Cor. 4 proved from [43] –
Witt index of the trace form of degree-two real multiplication Thm. 17 proved –
families of Witt index two, resp. one, for every real quadratic field Cors. 5, 6 proved –
dense curves of Witt index two in every family with real multiplication of degree two Thm. 18 proved from [43] –
the Hodge locus of the class on S 1 × S 2 is the Hecke graph Thm. 19 proved from [26] –
algebraic points dense in the Hecke family; Shioda–Inose representative Cors. 7, 8 proved from [26,81] –
kernels of equivalences, Euler forms, Kummer graphs do not attain the bound Cor. 9, Props. 18, 19 proved –
the semiregularity count on S 1 × S 2 Prop. 16 proved given (HH), (HH′) –
an object at a point of the Hecke graph with dim Ext 2 = R ( v ) Sect. 5.10 open such an object
Q ( 2 ) in rank ≤ 10 ; Q ( 3 ) , Q ( 6 ) in rank six: HC , (F3′) on S × S , S [ 2 ] Thm. 12, Cor. 1 proved from [115] –
reach of symplectic quotients in rank ≥ 8 Prop. 8 proved –
ρ = 16 , Witt index one, e ∉ { 2 , 3 , 6 } , e.g. e = 5 , 7 in Table 4 Cor. 6 open a correspondence not from [43]
Hodge classes on all powers of S: E and ⋀ E m T Prop. 13 proved from [116] –
K3 surfaces of Kummer type over Q : (F3′) and HC on all powers Prop. 14 proved –
Fricke Kuga–Satake fourfold: B ∼ E K × B 3 Prop. 15 proved –
⋀ E m T for the other surfaces of Sect. 5 Rem. 2 open a cycle on S m
(F3′) in general – open –
product object of polarised shape Thm. 21 proved –
direct sums never semiregular Prop. 22 proved –
annihilator of the secant pieces is the product locus Thm. 22 proved –
secant and line-bundle pieces orthogonal Prop. 24 false for every realisation –
(P2′) at a product point Sect. 6.5 open Ext * of secant sheaves
the Ext groups at the CM point E 0 6 Lem. 15 proved –
no two-step cancellation of diagonal classes Prop. 25 proved –
three line bundles at E 0 6 Prop. 26 proved, fail –
isolated Ext blocks at E 0 6 Prop. 27 proved, fail –
a threefold piece outside every closed chain Prop. 28 proved, fail –
further pieces within one shift of the thirty-two Cor. 12 proved, fail –
twists along θ h ; discriminant of D h Rems. 7, 8 proved –
at most five line bundles, any twists, every polarisation Thm. 23 proved, fail –
six or seven line bundles; eight outside two arrangements Lem. 23, Rem. 9 enumeration, fail –
the rest of the support cuts a threefold in an ample divisor Thm. 24 proved –
two threefolds meeting along a surface carry nothing Thm. 25 proved –
objects on the thirty-two threefolds of one family Cor. 13 proved, fail –
objects on the sixty-four threefolds of both families Thm. 26 proved, fail –
a smooth complete intersection cannot carry the rest Prop. 31 proved, fail –
an ample divisor in the support empties the localised obstruction Prop. 32 proved –
(P2′) at E 0 6 : two arrangements of eight, or ten or more line bundles Rem. 9 open the chain products
E 0 8 : shifts in seven consecutive integers Cor. 16, Thm. A9 proved, fail (all) –
E 0 6 , E 0 8 : every line-bundle convolution, every placement and group Thm. A9 proved, fail –
line-bundle convolutions: pieces off Q θ lie on cycles of excess zero Thm. A10 proved –
pieces of any kind: convex sets; exclusions (Cor. A4); transverse threefolds with at most 2 n − 3 multiples of θ Thms. A17, A18 proved, fail –
260 line bundles at E 0 6 passing every graph test Prop. A12 proved, enumeration –
the 260 pieces: Koszul cubes, whole-cube sets (Prop. A13), one generic cube Thms. A19, A20 proved, fail; enumeration –
the 260 pieces: the remaining twisting cochains Sect. A.5 open Ext 2 as a function of δ
equivariant criterion; two-shift bounds at E 0 8 Cor. 11, Thm. A21 proved from [98]; enumeration –
an étale Prym map of triple covers at genus four Prop. 39 proved –
sextic CM fields: the count excludes nothing Rem. 14 enumeration –
split ( 2 n + 2 ) -folds give all 2 n -folds; odd pieces Props. 33, 34 proved from [78] –
Weil classes of sixfolds of CM type Cor. 15 proved from [78] –
HC for sixfolds with CM by a degree-twelve field K L Thm. 29 proved from [78], enumeration –
Markman’s construction gives split forms only; trace form Thm. 27, Cor. 14 proved –
Weil classes on the quaternionic loci of every non-split family Thm. 28 proved –
the Lemma 11.3 mechanism on eightfolds; integrality on secant planes Thm. 30, Cor. 17 proved, fail –
indefinite secant planes miss the polarised Weil eightfolds Thm. 35 proved –
Weil classes of a very general non-split sixfold Sects. 6.9, 6.11 open a definite secant plane with [98] or [79, Qu. 11.4], or a decomposition passing Thm. A17
real multiplication of T Z algebraic; one link ⇒ exceptional classes, every K Thm. 37 proved –
the link is all or nothing Thm. 38 proved –
the tenfold link η is equivalent to ξ Thm. 39 proved –
link ξ t on the dense set Σ K , K = F · k Thm. 41 proved from [76] –
no Weil class for any field and no Kummer or OG6 carrier reaches ξ , η Prop. 43 proved –
non-CM points are generic; H 3 ( Z K ) as a tensor product Prop. 44, Thm. 42 proved –
squaring lemma; one algebraic γ t suffices Thms. 43, 44 proved –
the summands conjectured in [16] Cor. 18 proved given (I) –
divisors, Weil classes and homomorphisms do not reach ξ Prop. 45 proved –
the link ξ at a very general point Sect. 6.17 open a cycle that sees the geometry of X t
dominant maps, blow-ups, derived equivalences preserve the defect Prop. 47 proved –
residues of complete intersections Prop. 48 proved rational residues
HC for α as a vanishing mass gap Prop. 49 proved –
exceptional classes of X × X – open –
(F2) in general – open –
the Hodge conjecture – open (F2) and (F3′)

7. What Is Proved and What Is Open

Table 8 lists the statements. Three remarks on what the table says.
The two open statements are not independent problems that happen to add up to the conjecture; they are two halves of one phenomenon. Both say that the algebraic locus of a flat Hodge class, a countable union of closed strata through a dense set of points of complex multiplication, is not meagre. For (F2) the dense set is supplied by explicit cycles at CM points and the first instance is a single class on an abelian fourteenfold or a single Ext group on an eightfold; for (F3′) the dense set is supplied by the K3 surfaces with complex multiplication and by the chains of Section 5, and the first instance is a cubic real multiplication on one K3 surface of Picard number 13, which Theorem 14 supplies for the fields Q ( ζ 7 ) + and Q ( ζ 9 ) + with the algebra ramified at two real places, and which stays open for every other pair of a cubic field and an algebra.
Each construction of this paper ends at a question with a definite answer. Whether some convolution of the pieces of Theorem 21 has dim Ext 2 = 57 is a finite computation in the Ext algebra of two sheaves on a fourfold; at the point E 0 6 of complex multiplication every group entering it is written down in Lemma 15, Proposition 26 answers it in the negative for every convolution with three powers of L, Proposition 27 answers it in the negative for two threefolds at one shift and for every placement of the thirty-five pieces, Proposition 28 with Corollary 12 for every further piece of positive twist within one shift of the thirty-two, Theorem 23 for at most five line bundles at any twists and every polarisation, the enumeration of Remark 9, with Lemma 23, for six and seven and for eight outside two arrangements, Corollary 13 and Theorem 26 for every convolution of threefold pieces, of one family or of both, at any twists and shifts, and for the configurations left, the two arrangements of eight and ten or more line bundles, with or without threefolds, it is the rank of the products along the chains of Proposition 25 and along the cycles through the line bundles, the support of the object being constrained by Theorem 24 and Propositions 30 and 31. Whether the link ξ is algebraic is a question about one Hodge class on one abelian variety, equivalent by Theorem 38 to the existence of one algebraic cycle on some X r × Z K acting nonzero on T Z ; Theorem 41 answers it on a dense set of special points when K contains an imaginary quadratic field, so that what remains is whether the algebraic locus of a flat class over a curve, known to be dense, is the whole curve. Whether a correspondence acting on T ( S λ ) otherwise than by a similarity exists is a question about one K3 surface; for F = Q ( ζ 7 ) + or Q ( ζ 9 ) + and D = D ∞ Theorem 14 answers it, and for every other pair ( F , D ) it is open, with Proposition 11 marking the three places where the present mechanism fails. For the Weil classes of sixfolds the line-bundle convolutions at E 0 6 and E 0 8 are excluded for every arrangement and every group action (Theorem A9), and so is the semiregularity mechanism of Markman’s Lemma 11.3 on eightfolds (Theorem 30, Corollary 17), while the secant planes of indefinite classes miss the polarised family (Theorem 35); what remains is an object on a definite secant plane with Perry’s G-semiregularity for an equivariant structure that is not a product (Theorem 33) or an affirmative answer to Markman’s Question 11.4, a decomposition satisfying Theorem A17, which the 260 pieces of Proposition A12 do for every twisting cochain at the level of graphs and which Theorems A19 and A20 decide only for the cochains they cover, or the very general non-split sixfold. For real multiplication of degree two in transcendental rank at least eight, outside Q ( 2 ) in ranks eight and ten and the lattice conditions of Proposition 8, it is an object at one point of the Hecke graph with dim Ext 2 equal to the bound R ( v ) of Proposition 16, which suffices if the Hochschild inputs (HH) and (HH′) hold (Section 5.10). For the link of a Mumford fourfold it is a cycle that sees the geometry of X t : divisor classes, Weil classes and homomorphisms cannot produce it (Proposition 45). A negative answer to the first closes that route, as the analogous question on E 0 6 closed; a positive answer to any of them proves the corresponding statement at its first instance.

7.1. Specialisation, Degeneration and Blow-Ups

Three devices are often proposed for passing from the members where a class is known to be algebraic to the others, and we record exactly what each one proves.
Proposition  46.
Let π : X → B be a smooth projective family over an irreducible base and ξ a flat Hodge class on its fibres.
(i)
If ξ t is algebraic at the generic point of a curve C ⊆ B then it is algebraic at every point of C. The converse implication, from one point of C to its generic point, holds when ξ t 0 is the Chern character of a perfect complex whose semiregularity map is injective, and in no greater generality is it known.
(ii)
Let X ˜ be the blow-up of a smooth projective variety X along a smooth closed subvariety Z of codimension c. Then Hdg k ( X ˜ ) / Alg k ( X ˜ ) ≅ Hdg k ( X ) / Alg k ( X ) ⊕ ⨁ i = 1 c − 1 Hdg k − i ( Z ) / Alg k − i ( Z ) . In particular the Hodge conjecture for X ˜ is equivalent to the Hodge conjecture for X and for Z, and a blow-up never reduces the defect of a variety.
(iii)
For the Weil classes of D n , n and the exceptional class of a Mumford fourfold, every degeneration to a fibre with a semistable model on which the limit class is algebraic has its lift to the generic fibre as its missing step, and for the Mumford fourfold no degeneration keeping the class of Hodge type exists, since the Hodge locus of the class is a compact Shimura curve.
Proof. 
(i) Specialisation of cycles along a curve [44, Ch. 20] sends an algebraic cycle on the generic fibre to one on each special fibre with the same flat cohomology class; this is the direction used by Schoen’s degeneration and by Proposition 6. In the other direction the algebraic locus is a countable union of closed sets by Lemma 2, and a point in one stratum says nothing about the stratum being the whole base; the one proved lift is the semiregularity theorem of Bloch, Buchweitz and Flenner [11,20], which is the content of Theorem 20. (ii) Manin’s decomposition of the motive of a blow-up [75], h ( X ˜ ) = h ( X ) ⊕ ⨁ i = 1 c − 1 h ( Z ) ( − i ) , is realised by the algebraic correspondences π * and j * ∘ p * ∘ ( powers of the relative hyperplane class ) , which are compatible with Hodge classes and with algebraic classes; so both the Hodge classes and the algebraic classes of X ˜ decompose accordingly, and the quotients add. Since every summand is non-negative, the defect of X ˜ is at least that of X, and a blow-down along a smooth centre Z lowers the defect by exactly that of Z in the shifted degrees. (iii) A flat class has a limit in the limit mixed Hodge structure of a semistable fibre, and on a totally degenerate fibre the graded pieces are combinatorial and every class is algebraic on the components; the return to the generic fibre is the lift of (i), with the same unproved step. For a Mumford fourfold the Hodge locus of the exceptional class is the compact Shimura curve of Proposition A22, which has no boundary, so no fibre at which the class is still of Hodge type degenerates.    □

7.2. Dominating Maps, Birational Models and Derived Equivalences

Write δ p ( X ) = dim Hdg p ( X ) − dim Alg p ( X ) for the defect of the conjecture in degree 2 p , δ ab p ( X ) = dim Hdg p ( X ) − dim Ab p ( X ) for the defect of (F3′), and δ ( X ) = ∑ p δ p ( X ) . The conjecture for X is δ ( X ) = 0 . The following proposition records how the surgeries of birational geometry move these numbers; Figure 14 draws it.
Proposition  47
(Defects under surgery). Let X, Y be smooth complex projective varieties.
(i)
If f : Y → X is surjective and generically finite, then δ p ( Y ) ≥ δ p ( X ) and δ ab p ( Y ) ≥ δ ab p ( X ) for every p. In particular no blow-up along a singular centre, weighted blow-up, resolution of a modification or finite cover of X has a smaller defect than X.
(ii)
If X and Y are birational, then δ 2 ( X ) = δ 2 ( Y ) and δ d − 1 ( X ) = δ d − 1 ( Y ) , d = dim X ; for 3 ≤ p ≤ d − 2 the difference δ p ( X ) − δ p ( Y ) is a signed sum of defects δ p − j ( Z ) , j ≥ 1 , of smooth projective varieties Z of dimension less than d.
(iii)
If Γ ∈ CH * ( X × Y ) Q induces an isomorphism from H * ( X , Q ) onto H * ( Y , Q ) whose inverse is induced by a cycle, for instance the kernel of an equivalence between the derived categories of X and Y, then δ ( X ) = δ ( Y ) and δ ab ( X ) = δ ab ( Y ) .
(iv)
Let X → Δ be a projective family, smooth over the punctured disc, with a flat Hodge class ξ on the smooth fibres that is algebraic on a crepant resolution X ^ of the special fibre X 0 , or on the components of a semistable model of X 0 . Then ξ is algebraic on X t for t ≠ 0 if and only if the specialisation H 2 p ( X 0 ) → H 2 p ( X t ) sends an algebraic class to an algebraic class, which is the statement (V) for the family; the monodromy plays no role, since it fixes ξ.
Proof. 
(i) f * f * is multiplication by the degree, so f * is an injective morphism of Hodge structures and f * Hdg p ( X ) ⊆ Hdg p ( Y ) . If f * α is algebraic then so is deg ( f ) α = f * f * α ; hence Alg p ( Y ) ∩ f * Hdg p ( X ) = f * Alg p ( X ) and f * embeds Hdg p ( X ) / Alg p ( X ) into Hdg p ( Y ) / Alg p ( Y ) . For Ab the same argument applies because f * of a class Γ * γ reached from an abelian variety is ( Γ f ∘ Γ ) * γ , reached from the same one. The coarse space of a weighted blow-up has quotient singularities and f * f * = id with rational coefficients, so the argument is unchanged there.
(ii) By weak factorisation [3] a birational map between smooth projective varieties is a composite of blow-ups and blow-downs along smooth centres with smooth projective intermediate varieties, and Proposition 46(ii) gives the change of δ p at each step as ± ∑ j = 1 c − 1 δ p − j ( Z ) . For p = 2 the terms are δ 1 ( Z ) and δ 0 ( Z ) , which vanish by the Lefschetz ( 1 , 1 ) theorem; for p = d − 1 they are δ d − 1 − j ( Z ) with dim Z ≤ d − 2 , hence δ q ( Z ) with q ≥ dim Z − 1 after j ≥ 1 , and these vanish by hard Lefschetz and Lefschetz ( 1 , 1 ) on Z.
(iii) By the projection formula Γ * sends ⨁ p Hdg p ( X ) into ⨁ p Hdg p ( Y ) , since [ Γ ] is a sum of classes of type ( i , i ) and cup product and push-forward are morphisms of Hodge structures up to Tate twists; it sends algebraic classes to algebraic classes because [ Γ ] ∪ pr X * [ Z ] is the class of a cycle and the push-forward of a cycle is a cycle; and it sends Ab to Ab by composing correspondences. The inverse does the same in the other direction, so the two total quotients are isomorphic. The kernel of a Fourier–Mukai equivalence and the kernel of its inverse are cycles by Mukai’s and Orlov’s description of the inverse functor [96], which is the equivalence with kernel the dual of the original one shifted by the dimension.
(iv) ξ is monodromy invariant, so by the local invariant cycle theorem it is the specialisation of a class in H 2 p ( X 0 ) , which pulls back to X ^ or to the components of the semistable model; the hypothesis says that the pulled back class is algebraic there, and push-forward keeps it algebraic on X 0 . Cycles specialise from the generic fibre to the special fibre [44, Ch. 20] and not the other way, so the remaining step is exactly the one in the statement. If the limit cycle on X 0 lifts to the family, its restriction to X t has class ξ t ; conversely if ξ t is algebraic the specialisation of a representing cycle is a cycle on X 0 with class the limit of ξ , and the specialisation map sends that class to ξ t . The vanishing cycles, on which the monodromy acts, are orthogonal to the invariant part, and ξ is invariant, so no condition on them enters.    □
Items (i) to (iii) say that every surgery available in birational and derived geometry carries the defect of X into its output as long as the open class survives there, and loses the class only in outputs of lower dimension, the centres, where it has become a defect of those varieties. Both open statements of this paper sit in degree four, (F2) for the Weil classes and the exceptional classes, (F3′) for the classes of H 4 ( S × S ) on a K3 surface, and in degree four the defect is a strict birational invariant by (ii) and a derived invariant by (iii); so the open classes are present, and not known to be algebraic, on every birational model and every Fourier–Mukai partner of the varieties that carry them. Item (iv) is the reason that degenerations to products of elliptic curves, where every Hodge class is algebraic [91], do not close (F2): the class returns to the smooth fibre only through (V). Flows of Kähler metrics change no complex structure between their surgeries, and the surgeries themselves are the contractions and flips of the minimal model programme, covered by (i) and (ii); the surgery of Perelman has no counterpart for Hodge structures, since it cuts a manifold along spheres and a Hodge structure has no local pieces to cut.
The devices therefore relocate the two open statements without weakening them. The one true reduction in shape is Theorem 9: the whole of (F2) is equivalent to a lift from CM points along polarised abelian schemes, and the lift along each scheme is a question about one semiregular object at one point.

7.3. Residues and Mass-Minimising Currents

Two analytic descriptions of algebraic classes are exact, and we record what each reduces the conjecture to. The first is Griffiths’ residue calculus on hypersurfaces [48]. For a smooth hypersurface X = { F = 0 } ⊂ P n + 1 of even dimension n = 2 m and degree d, with Jacobian ring R ( F ) and Ω = ∑ i ( − 1 ) i x i d x 0 ∧ ⋯ d x i ^ ⋯ ∧ d x n + 1 , the residues of the forms P Ω / F q + 1 identify H prim n − q , q ( X ) with R ( F ) ( q + 1 ) d − n − 2 .
Proposition  48
(Residues of complete intersections). Let X = { F = 0 } be as above, and suppose that F = ∑ i = 1 m + 1 f i g i with Z = { f 1 = ⋯ = f m + 1 = 0 } a complete intersection of codimension m + 1 in P n + 1 whose ideal is generated by the f i .
(i)
Let θ be the hyperplane class and J H the Jacobian determinant of ( f 1 , g 1 , ⋯ , f m + 1 , g m + 1 ) , a polynomial of degree ( m + 1 ) ( d − 2 ) . Then
[ Z ] = deg Z d θ m − m ! d res J H Ω F m + 1 m , m
[122, Thm. 1.1].
(ii)
Let J ( F ) ⊆ R ( F ) ( m + 1 ) ( d − 2 ) be the span of the classes of the determinants J H over all such decompositions of F. If the rational span of the residues of J ( F ) and θ m contains Hdg m ( X ) , the Hodge conjecture holds for X.
Proof. 
(i) is the cited theorem; the degree of the determinant is ∑ i ( deg f i + deg g i ) − ( n + 2 ) = ( m + 1 ) d − 2 m − 2 . (ii) By (i) every residue of an element of J ( F ) is a rational combination of θ m and of classes of complete intersections in X, so it is algebraic. The Hodge conjecture for a hypersurface holds outside the middle degree by the Lefschetz hyperplane theorem, and in the middle degree it is the statement that Hdg m ( X ) consists of algebraic classes.    □
The proposition replaces a cycle by a polynomial, and the open step becomes the statement that the polynomials of J ( F ) have enough rational residues. Deciding whether a given element of R ( F ) has a rational residue requires its periods, so the residue calculus moves the conjecture from cycles to transcendental numbers without closing it. On a Fermat hypersurface ∑ i x i d = 0 the linear subspaces cut out by x 2 i = ζ i x 2 i + 1 , ζ i d = − 1 , are of this form, since x 2 i d + x 2 i + 1 d factors into linear forms; whether their residues and those of the other decompositions exhaust Hdg m ( X ) is again a question about periods.
The second description uses currents and is the analogue for cycles of the flows of geometric analysis. Let X be a smooth complex projective variety of dimension N, h an ample class with Kähler form ω , and 0 < p < N . For β ∈ H 2 ( N − p ) ( X , Z ) let M ( β ) be the infimum of the masses of the integral currents in β , put vol ( β ) = ∫ β ω N − p / ( N − p ) ! , and call g ( β ) = M ( β ) − vol ( β ) the gap of β . We identify H 2 p ( X , Z ) with H 2 ( N − p ) ( X , Z ) by Poincaré duality.
Proposition  49
(Mass-minimising currents).
(i)
Every class β contains an integral current of mass M ( β ) , and g ( β ) ≥ 0 , with equality if and only if β is the class of an effective algebraic cycle with integer coefficients.
(ii)
Let α ∈ H 2 p ( X , Z ) . The image of α in H 2 p ( X , Q ) is algebraic if and only if g ( e α + k h p ) = 0 for some integers e ≥ 1 and k ≥ 0 .
(iii)
There are X and α with g ( α + k h p ) > 0 for every k ≥ 0 although g ( e α ) = 0 for some e ≥ 1 : for a very general hypersurface X ⊂ P 4 of degree D divisible by ℓ 3 for an integer ℓ > 1 prime to 6, p = 2 , and the generator α of H 4 ( X , Z ) .
Proof. 
(i) By the compactness theorem of Federer and Fleming [40] the integral currents in β of mass at most M ( β ) + 1 form a compact set for the flat topology, and mass is lower semicontinuous, so the infimum is attained. For an integral current T in β , Wirtinger’s inequality, that ω N − p / ( N − p ) ! has comass one with equality exactly on the complex ( N − p ) -planes with their canonical orientation, gives M ( T ) ≥ T ω N − p / ( N − p ) ! = vol ( β ) , the last value depending only on β because ω is closed [53]. If equality holds, T is calibrated by ω N − p / ( N − p ) ! , hence a closed positive current of bidimension ( N − p , N − p )  [53], and by King’s theorem [62] a closed positive integral current of that bidimension is ∑ i n i [ V i ] with n i positive integers and V i irreducible analytic subvarieties, which are algebraic by Chow’s theorem. Conversely, for an effective cycle Z = ∑ n i V i the tangent planes of the smooth locus of each V i are complex, so the current [ Z ] has mass vol ( [ Z ] ) and g ( [ Z ] ) = 0 .
(ii) If g ( e α + k h p ) = 0 , then e α + k h p is the class of an effective cycle by (i), and α = [ Z ] − k h p / e in H 2 p ( X , Q ) . Conversely let α be rationally algebraic, so that e 0 α − [ Z + ] + [ Z − ] is a torsion class for an integer e 0 ≥ 1 and effective cycles Z ± ; multiplying by its order, e α = [ Z + ] − [ Z − ] in H 2 p ( X , Z ) . Write Z − = ∑ i c i V i . For r large the ideal sheaf of each V i twisted by O ( r h ) is globally generated, so p general members of | r h | containing V i meet in a complete intersection Y i of codimension p containing V i , and [ Y i ] = r p h p ; the cycle Z i ′ of Y i minus V i is effective. Then e α + ∑ i c i r p h p = [ Z + + ∑ i c i Z i ′ ] , effective, and (i) gives g = 0 with k = ∑ i c i r p .
(iii) By the Lefschetz hyperplane theorem H 4 ( X , Z ) ≅ Z α with ∫ X α · h = 1 , and h 2 = D α , the class of a complete intersection, so g ( D α ) = 0 . By Kollár’s theorem [65], in the form of [114, Thm. 2], every curve on X has degree divisible by ℓ. The class α + k h 2 has degree 1 + k D , prime to ℓ, so it is not the class of an effective cycle, and g ( α + k h 2 ) > 0 by (i).    □
The analytic part of the variational approach is therefore complete. A minimiser exists in every class without any flow, so a mass-decreasing flow of a cycle, with whatever surgeries along its singularities, can at best reach it; what such a flow cannot provide is the equality case, and the equality case is the whole conjecture: by (ii) the Hodge conjecture for a class α is the vanishing of the gap of e α + k h p for one pair ( e , k ) . Part (iii) shows that the multiplier e cannot be omitted, since integral minimisers of a class that is rationally algebraic need not be calibrated for any k, and so no argument by integral currents alone, including one modelled on Ricci flow with surgery, proves the conjecture without passing to rational coefficients. For the Weil classes of D n , n and the exceptional classes of a Mumford fourfold the gap of e α + k h p is not known to vanish for any ( e , k ) .
Apart from Proposition 16(b),(c), Corollary 18 and Remark A4, which name their hypotheses, nothing in the paper assumes an open statement, and the statements that rest on an enumeration are marked as such. The machine checks of Appendix D recompute the finite quantities; beyond those enumerations they do not carry any proof.

Funding

No funding of any kind, internal or external, was received for this work. The authors met its costs themselves.

Institutional Review Board Statement

The work raises no ethical question; it involves no human participants, no animals and no personal data.

Data Availability Statement

No data were generated. The Lean 4 file and the Julia checks of Appendix D are in the repository github.com/creelie/fermat-hodge- [9]. Nothing in the paper rests on them: every statement proved in the text has its proof there, and the statements obtained by enumeration are marked as such.

Conflicts of Interest

None of the authors has a competing interest, and neither do the institutions with which they are or were affiliated, which had no part in the work.

Appendix A. Weil Classes, the Polarised Criterion and Mumford Fourfolds: Statements and Proofs

This appendix states and proves the results on Weil classes, on the polarised semiregularity criterion, on convolutions of line bundles, on the closure graph, on Mumford fourfolds and on K3 surfaces with real multiplication that the body of the paper uses. It is self-contained apart from the published literature it cites. The numbering is that of the appendix; the notation is that of the paper, with the additions fixed in each subsection.

Appendix A.1. Abelian Varieties of Weil Type for a CM Field

Setting  8
(Weil type for a CM field). Let F be a CM field of degree 2 m with maximal totally real subfield F 0 , and write α ↦ α ¯ for the nontrivial automorphism of F / F 0 , which is complex conjugation under every embedding. The real places of F 0 are τ 1 , ⋯ , τ m , and each extends to a conjugate pair σ τ , σ ¯ τ of embeddings of F into C . Let V be an F-vector space of dimension 2 n , let H : V × V → F be a nondegenerate hermitian form, F-linear in the first variable, and fix a totally imaginary ξ ∈ F × , that is ξ ¯ = − ξ . Put
E ( x , y ) = Tr F / Q ξ H ( x , y ) ,
a nondegenerate alternating Q -bilinear form with E ( α x , y ) = E ( x , α ¯ y ) . For each τ let V τ = V ⊗ F 0 , τ R , a vector space of dimension 2 n over F ⊗ F 0 , τ R ≅ C , and let ( p τ , q τ ) be the signature of the hermitian form H τ induced by H on it. We say that ( V , H ) is of ( F , n ) -Weil typeif p τ = q τ = n for every τ, and we call the class of ( − 1 ) n det H in F 0 × / N F / F 0 ( F × ) its discriminant δ; it is totally positive. A polarised abelian variety ( A , ι , E ) of dimension 2 m n with ι : F → End 0 ( A ) and H 1 ( A , Q ) ≅ V as F-modules with polarisation E is of ( F , n , δ ) -Weil type when ( V , H ) is. We work throughout with V = H 1 ; a complex structure on V R is the same datum as one on the tangent space. The Weil classes of A are
W F ( A ) = ⋀ F 2 n V ⊆ ⋀ Q 2 n V = H 2 n ( A , Q ) ,
the Q -subspace whose complexification is ⨁ σ ⋀ 2 n V σ , where V σ = V ⊗ F , σ C runs over the 2 m eigenspaces of F on V C ; it is a one-dimensional F-vector space, of dimension 2 m over Q . Write
W ( F , n , δ ) : every abelian variety of ( F , n , δ ) -Weil type has algebraic Weil classes.
The period domain D F is the set of complex structures J on V R commuting with F and polarised by E, and the group G F = Res F 0 / Q SU ( V , H ) acts on it. We fix an arithmetic subgroup Γ ⊂ G F ( Q ) and write S F = Γ ∖ D F for the quotient, p : D F → S F for the projection, π : A → S F for the universal family and Σ F ⊆ D F for the set of s at which W F ( A s ) is algebraic.
Definition  A1
(The imaginary quadratic case). For m = 1 the field F = K = Q ( − d ) is imaginary quadratic and F 0 = Q . We then write V C = V + ⊕ V − for the two eigenspaces of K, W ( A ) = W K ( A ) for the two-dimensional rational space of Weil classes, θ ∈ ⋀ 2 V for the class of the polarisation E, and D n , n = D K . We fix generators α ± of the lines ⋀ 2 n V ± , so that W ( A ) ⊗ C = C α + ⊕ C α − ; a real class of W ( A ) ⊗ R is u α + + u ¯ α − once α − = α + ¯ . Since E ( α x , y ) = E ( x , α ¯ y ) , the form E vanishes on V + × V + and on V − × V − , so
θ ∈ V + ⊗ V − ⊆ ⋀ 2 V C ,
and H identifies V − with the dual of V + as a representation of SU ( V , H ) ⊗ C = SL ( V + ) .
Theorem  A1
(The Hodge ring in the imaginary quadratic case). Let A be an abelian 2 n -fold of Weil type for the imaginary quadratic field K with Hg ( A ) = SU ( V , H ) . Then
Hdg * ( A ) = Q [ θ ] ⊕ W ( A ) , W ( A ) · θ = 0 ,
that is, the space of Hodge classes of degree 2 k is Q θ k for k ≠ n and Q θ n ⊕ W ( A ) for k = n . The equality W ( A ) · θ = 0 holds for every abelian variety of Weil type for K.
Proof. 
The Hodge classes are the invariants of Hg ( A ) in ⋀ * V , and taking invariants commutes with extension of scalars, so we compute the invariants of SL ( V + ) ≅ SL 2 n ( C ) in
⋀ k V C = ⨁ r + s = k ⋀ r V + ⊗ ⋀ s V + ∨ .
An invariant is a vector of weight zero. Fix a basis of V + , the diagonal torus, and for T , U ⊆ { 1 , ⋯ , 2 n } the monomials v T ⊗ v U ∨ in the basis and its dual. The weight of v T ⊗ v U ∨ is 1 T − 1 U modulo ( 1 , ⋯ , 1 ) , with entries in { − 1 , 0 , 1 } , so it is zero only for T = U , for ( T , U ) = ( { 1 , ⋯ , 2 n } , ∅ ) and for ( T , U ) = ( ∅ , { 1 , ⋯ , 2 n } ) . The last two occur for ( r , s ) = ( 2 n , 0 ) and ( 0 , 2 n ) and give the invariant lines ⋀ 2 n V + and ⋀ 2 n V − , which span W ( A ) ⊗ C . For r = s the invariants in ⋀ r V + ⊗ ⋀ r V + ∨ ≅ End ( ⋀ r V + ) form the line of scalars, by Schur’s lemma, since ⋀ r V + is irreducible. By (A1) the class θ r lies in this summand; it is invariant, and nonzero for r ≤ 2 n because θ is a nondegenerate form on a space of dimension 4 n . This gives the list. Finally α ± θ ∈ ⋀ 2 n + 1 V ± ⊗ V ∓ = 0 by (A1), with no hypothesis on the Hodge group.    □
Proposition  A1
(The Weil classes of a CM field). Keep Setting 8, with ( V , H ) of ( F , n ) -Weil type.
(i)
A polarised complex structure J is the same as an H τ -orthogonal decomposition V τ = V τ + ⊕ V τ − for every τ, with H τ definite of the sign of Im σ τ ( ξ ) on V τ + and of the opposite sign on V τ − ; then dim V σ τ 1 , 0 = dim V τ + = n and dim V σ ¯ τ 1 , 0 = dim V τ − = n , so every class in W F ( A J ) is of type ( n , n ) . The spaces W F ( A J ) form a flat subsystem W F of R 2 n π * Q of rank 2 m on which G F acts trivially, and D F ≅ ∏ τ D n , n is connected of dimension m n 2 .
(ii)
If one nonzero element of W F ( A ) is algebraic, every element is.
(iii)
If Hg ( A s ) = G F , the Hodge ring of A s is generated by NS ( A s ) Q and by W F , so the Hodge conjecture for A s in every degree is equivalent to the algebraicity of W F . For n ≥ 2 the space NS ( A s ) Q is the m-dimensional space of classes η t = [ E ( t · , · ) ] , t ∈ F 0 .
(iv)
Two members of ( F , n , δ ) -Weil type have isomorphic ( V , H ) , so every one of them is a fibre of π.
Proof. 
(i) A complex structure commuting with F preserves each V τ , since F 0 ⊗ R acts diagonally, and is C -linear on it for the structure F ⊗ F 0 , τ R ≅ C , so it is ± − 1 on two complementary C -subspaces V τ ± . Write σ τ ( ξ ) = − 1 r τ with r τ ∈ R × . For x ∈ V τ ϵ and y ∈ V τ ϵ ′ one has H τ ( J x , J y ) = ϵ ϵ ′ H τ ( x , y ) , so E ( J x , J y ) = E ( x , y ) forces E ( x , y ) = 0 for ϵ ≠ ϵ ′ , and replacing x by − 1 x then forces H τ ( x , y ) = 0 : the decomposition is orthogonal. For x ∈ V τ ϵ , E ( x , J x ) = 2 Re − 1 r τ ϵ − 1 ¯ H τ ( x , x ) = 2 ϵ r τ H τ ( x , x ) , so positivity is the sign condition stated. Conversely such a decomposition defines a polarised J. Under V τ ⊗ R C = V σ τ ⊕ V σ ¯ τ the space V 1 , 0 , the − 1 -eigenspace of J on V C , meets V σ τ in the image of V τ + and V σ ¯ τ in the conjugate of the image of V τ − , which gives the two equalities of dimensions. Since H τ has signature ( n , n ) and is definite, of opposite signs, on the complementary orthogonal subspaces V τ ± , both have dimension n. So each V σ is the sum of its intersections with V 1 , 0 and V 0 , 1 , both of dimension n, and a generator α σ of ⋀ 2 n V σ is of type ( n , n ) . The subspace ⋀ F 2 n V of H 2 n ( A s , Q ) is defined by the action of F on V alone and does not depend on s, so it is flat; SU ( V , H ) acts on ⋀ F 2 n V by the F-determinant, which is 1. Finally, the choice of V τ + is a point of the bounded symmetric domain of SU ( V τ , H τ ) ≅ SU ( n , n ) , of dimension n 2 , and D F is the product over τ .
(ii) For α ∈ O F the endomorphism ι ( α ) acts on ⋀ F 2 n V by α 2 n , through an algebraic correspondence, so the algebraic classes in W F form a Q -subspace stable under multiplication by every α 2 n . By the polarisation identity, in characteristic zero every product x 1 ⋯ x 2 n of elements of a field is a rational combination of 2 n -th powers of sums of the x i ; taking x 2 = ⋯ = x 2 n = 1 shows that the 2 n -th powers span F over Q , and since every element of F is a rational multiple of an element of O F , so do the α 2 n with α ∈ O F . Hence the Q -span of { α 2 n w } is F w = W F for every w ≠ 0 .
(iii) Hodge classes are the invariants of the Hodge group, which here is G F . Over C it is ∏ τ SL ( U τ ) with U τ = V σ τ , and H identifies V σ ¯ τ with U τ ∨ , so ⋀ * V C = ⨂ τ ⋀ * ( U τ ⊕ U τ ∨ ) with the τ -th factor of the group acting on the τ -th tensor factor alone. The invariants of a product group in a tensor product of representations are the tensor product of the invariants, and the invariants of SL ( U ) in ⋀ * ( U ⊕ U ∨ ) are generated by the pairing ψ ∈ U ⊗ U ∨ and the two determinants, by the weight computation in the proof of Theorem A1. The classes η t are Hodge classes: J commutes with t and preserves E, hence preserves the F 0 -twisted form E t = E ( t · , · ) = Tr ( ξ t H ) , which for totally positive t is again a polarisation of A s . The complexified span of the η t is ⨁ τ C ψ τ , and the complexification of W F contains all the determinants. So the Q -algebra generated by NS Q and W F has complexification equal to the ring of invariants, hence equals Hdg * ( A s ) . Since divisor classes are algebraic, the conjecture reduces to W F , and by (ii) to one class. For the last sentence, the invariants in ⋀ 2 V C are ⨁ τ C ψ τ when n ≥ 2 : a summand ⋀ 2 U τ or ⋀ 2 U τ ∨ has an SL ( U τ ) -invariant only if dim U τ = 2 n equals 2, and a summand mixing two different τ has none. So NS ( A s ) Q has dimension m, and it contains the classes η t , which form an m-dimensional space because t ↦ η t is injective on F 0 , E being nondegenerate.
(iv) Two hermitian forms over F / F 0 of the same rank, the same discriminant and the same signature at every real place of F 0 are isomorphic [106, Ch. 10, §6]. So the datum ( F , n , δ ) determines ( V , H ) up to isomorphism, and the Hodge structure of any member is, by (i), a polarised complex structure on that V R , that is a point of D F .    □
Lemma  A1
(The Hodge group of a very general member). Keep Setting 8. For every s ∈ D F one has Hg ( A s ) ⊆ G F , and there is a countable union Z of proper closed analytic subsets of D F such that Hg ( A s ) = G F for every s ∉ Z .
Proof. 
The Hodge group of A s commutes with End 0 ( A s ) ⊇ F and preserves E, so it lies in Res F 0 / Q U ( V , H ) : an F-linear map preserving E = Tr F / Q ( ξ H ) preserves H, because H is recovered from E and the action of F: the trace form of F is nondegenerate, so H ( x , y ) is determined by the values E ( α x , y ) = Tr F / Q ( ξ α H ( x , y ) ) , α ∈ F . It fixes every Hodge class, in particular the Weil classes, which are Hodge classes by Proposition A1(i); an element g of Res F 0 / Q U ( V , H ) acts on ⋀ F 2 n V by det F ( g ) , so det F ( g ) = 1 and g ∈ G F .
For the second statement let T be the set of rational tensors t ∈ ⨁ a , b V ⊗ a ⊗ ( V ∨ ) ⊗ b that are not fixed by G F , a countable set, and for t ∈ T let Z t be the set of s at which t is a Hodge class, of type ( p , p ) with 2 p = a − b , for the Hodge structure induced by A s . The condition is that t lies in the p-th step of the Hodge filtration, which varies holomorphically with s, so Z t is a closed analytic subset of D F . It is proper. Otherwise t is fixed by h s ( U ( 1 ) ) for every s ∈ D F , where h s is the Hodge representation, which acts on the ( p , q ) part through z q − p , and hence by the subgroup N of G F ( R ) that these tori generate. Since D F is one orbit of G F ( R ) + and h g s = g h s g − 1 , the group N is a connected normal subgroup of G F ( R ) + = ∏ τ SU ( V τ , H τ ) , so its Lie algebra is an ideal of ⨁ τ su ( V τ , H τ ) , a sum of some of these simple summands. It contains every summand, because h s ( − 1 ) acts on V τ as the complex structure J, which is − 1 on V τ + and − − 1 on V τ − by Proposition A1(i), and is therefore not central in the factor SU ( V τ , H τ ) . So N = G F ( R ) + , which is Zariski dense in the connected group G F , and t would be fixed by G F , contrary to t ∈ T .
Put Z = ⋃ t ∈ T Z t . The Hodge group of A s is the subgroup of GL ( V ) fixing every rational Hodge class in the spaces V ⊗ a ⊗ ( V ∨ ) ⊗ b  [35, §3]. For s ∉ Z each such class is fixed by G F , so G F ⊆ Hg ( A s ) , and equality follows from the first statement.    □
Theorem  A2
(A base point in every family, for every CM field). Let ( V , H ) be of ( F , n ) -Weil type with discriminant δ. Then D F contains a point s 0 with the following properties.
(i)
A s 0 is isogenous to B Φ n × B Φ ¯ n , where B Φ is an abelian m-fold with complex multiplication by F of type Φ = { σ : Im σ ( ξ ) > 0 } and B Φ ¯ its conjugate; there is an H-orthogonal basis e 1 , ⋯ , e 2 n of V with H ( e i , e i ) totally positive for i ≤ n and totally negative for i > n , and F e i = H 1 ( B Φ , Q ) for i ≤ n , F e i = H 1 ( B Φ ¯ , Q ) for i > n .
(ii)
For i ≤ n put i ′ = i + n and let D i i ′ ⊆ H 2 ( A s 0 , Q ) be the F-balanced part of F e i ∧ F e i ′ , spanned by the classes
δ i ( f ) = ∑ k ( ω k f e i ) ∧ ( ω k * e i ′ ) , f ∈ F ,
for dual Q -bases ( ω k ) , ( ω k * ) of F with respect to the trace form. Every δ i ( f ) is a Hodge class of type ( 1 , 1 ) , hence a divisor class, and D i i ′ is a one-dimensional F-vector space with α · δ i ( f ) = δ i ( α f ) .
(iii)
The Weil classes are the products
w ( f ) = ∑ k 1 , ⋯ , k n − 1 δ 1 ( ω k 1 f ) · δ 2 ( ω k 1 * ω k 2 ) ⋯ δ n − 1 ( ω k n − 2 * ω k n − 1 ) · δ n ( ω k n − 1 * ) , f ∈ F ,
which lie in W F ( A s 0 ) , span it as f runs over F, and are polynomials with rational coefficients in divisor classes. In particular W F ( A s 0 ) is algebraic, and Σ F ≠ ∅ for every F, every n ≥ 1 and every δ.
Proof. 
(i) Since H τ has signature ( n , n ) at every τ , the determinant det H has sign ( − 1 ) n at every real place, so b = ( − 1 ) n det H , a representative of δ , is totally positive, and the diagonal form H ′ = diag ( b , 1 , ⋯ , 1 , − 1 , ⋯ , − 1 ) with n positive and n negative entries has the same dimension, the same discriminant and the same signatures as H. Hermitian forms over F / F 0 are classified by these invariants [106, Ch. 10, §6], so ( V , H ) ≅ ( F 2 n , H ′ ) and the standard basis is the required orthogonal basis e i , with a i = H ( e i , e i ) . Write σ τ ( ξ ) = − 1 r τ and define J on V τ = ⨁ i C e i to be ϵ i , τ − 1 on C e i , with ϵ i , τ = sign ( r τ ) sign ( τ ( a i ) ) . The decomposition V τ + = ⨁ ϵ i , τ = + C e i , V τ − = ⨁ ϵ i , τ = − C e i is H τ -orthogonal, with H τ of sign sign ( r τ ) on V τ + and of the opposite sign on V τ − , so Proposition A1(i) makes J a point s 0 of D F . On F e i ⊗ R = ∏ τ C e i the structure J is multiplication by the element ( ϵ i , τ − 1 ) τ of F ⊗ Q R , so each F e i is a rational sub-Hodge structure of weight one with F-action, polarised by the restriction of E; that is, it is the H 1 of an abelian m-fold with complex multiplication by F, of CM type Φ i = { σ : ( F e i ) σ ⊆ V 1 , 0 } , which by the computation in Proposition A1(i) is { σ τ ϵ i , τ } , where σ + = σ and σ − = σ ¯ . For i ≤ n the a i are totally positive, so ϵ i , τ = sign ( r τ ) and Φ i = { σ : Im σ ( ξ ) > 0 } = Φ ; for i > n the signs are reversed and Φ i = Φ ¯ . So V = ⨁ i F e i splits A s 0 up to isogeny as stated.
(ii) Write L i , σ = ( F e i ) σ , a line, so that V C = ⨁ i , σ L i , σ and L i , σ has type ( 1 , 0 ) exactly when σ ∈ Φ i . For dual bases, ∑ k σ ( ω k ) σ ′ ( ω k * ) = 1 if σ = σ ′ and 0 otherwise, because the matrices ( σ ( ω k ) ) and ( σ ( ω k * ) ) are inverse transposes by the definition of the trace form. Let ℓ i , σ be the component of e i in L i , σ , a generator of that line, so that ω e i = ∑ σ σ ( ω ) ℓ i , σ for ω ∈ F . Hence in F e i ⊗ F e i ′ ⊗ C = ⨁ σ , σ ′ L i , σ ⊗ L i ′ , σ ′ the class δ i ( f ) has component σ ( f ) ℓ i , σ ∧ ℓ i ′ , σ on the diagonal σ = σ ′ and zero off it. In particular it is F-balanced: α ∈ F acting through the first factor and through the second agree on it, both multiplying the component at σ by σ ( α ) , and this is the identity α · δ i ( f ) = δ i ( α f ) . Since Φ i ′ = Φ ¯ i , exactly one of L i , σ , L i ′ , σ has type ( 1 , 0 ) , so every component, hence δ i ( f ) , has type ( 1 , 1 ) ; and δ i ( f ) is rational, being a sum of wedges of rational vectors. By the Lefschetz theorem on ( 1 , 1 ) classes it is a divisor class. As f varies the components σ ( f ) vary independently, so D i i ′ ⊗ C = ⨁ σ L i , σ ⊗ L i ′ , σ and D i i ′ ≅ F .
(iii) The complexification of W F is spanned by α σ = ± ∏ i ≤ n ℓ i , σ ∧ ℓ i ′ , σ , a generator of ⋀ 2 n V σ = ⋀ 2 n ⨁ i L i , σ . In the product w ( f ) , the component indexed by ( σ 1 , ⋯ , σ n ) , with σ i the embedding chosen in the i-th factor, carries the coefficient σ 1 ( f ) ∏ j = 1 n − 1 ∑ k j σ j ( ω k j ) σ j + 1 ( ω k j * ) , which vanishes unless σ 1 = ⋯ = σ n and then equals σ 1 ( f ) . So w ( f ) = ∑ σ σ ( f ) α σ up to the sign of the reordering, an element of W F ⊗ C which is rational because it is a rational polynomial in rational classes; as f runs over F the vectors ( σ ( f ) ) σ span C 2 m , so the w ( f ) span W F . Each w ( f ) is a rational combination of n-fold products of the divisor classes δ i ( · ) , and a product of divisor classes is algebraic.    □

Appendix A.2. The Algebraic Locus of the Weil Classes

Theorem  A3
(Density and strata of the algebraic locus). Fix ( F , n , δ ) with n ≥ 2 and keep Setting 8.
(i)
W ( F , n , δ ) holds if and only if Σ F = D F , and then the Hodge conjecture holds in every degree for every member of the family with Hg ( A s ) = G F .
(ii)
Σ F is stable under G F ( Q ) , every G F ( Q ) -orbit in D F is dense, and Σ F is dense.
(iii)
Σ F is the preimage under p of a countable union of Zariski closed subsets of S F ; either Σ F = D F , or Σ F is meagre.
(iv)
The following are equivalent: Σ F is closed; Σ F has an interior point; Σ F = D F ; the Hodge conjecture holds in every degree for A s for all s outside a countable union of proper closed analytic subsets of D F .
Proof. 
(i) By Proposition A1(i) and (iv), every abelian variety of ( F , n , δ ) -Weil type has its Hodge structure at some point of D F , so W ( F , n , δ ) is the algebraicity of W F at every point, and by Proposition A1(iii) it gives the conjecture at the members with full Hodge group.
(ii) An element g ∈ G F ( Q ) carries J to g J g − 1 and induces, after clearing denominators, an isogeny A J → A g J ; pull-back and push-forward along an isogeny carry algebraic classes to algebraic classes, and g fixes W F pointwise by Proposition A1(i). So Σ F is stable.
For the density we show that the closure L of G F ( Q ) in G F ( R ) = ∏ τ SU ( V τ , H τ ) contains the identity component, which acts transitively on D F . First, the form H is isotropic over F 0 . It is indefinite at every real place, its signature being ( n , n ) . At every finite place v its trace form is a quadratic form in 4 n ≥ 8 variables over the local field F 0 , v , hence isotropic, and a hermitian form is isotropic as soon as its trace form is [106, Ch. 10, §1]. Since hermitian forms over the quadratic extension F / F 0 satisfy the Hasse principle [106, Ch. 10, §6], H is isotropic. Let X = { v ∈ V : H ( v , v ) = 0 } be the cone of isotropic vectors, a quadric cone in the 4 n -dimensional F 0 -space V with a nonzero F 0 -point. Projection from that point shows that the projective quadric of its lines is F 0 -rational, so it satisfies weak approximation at the archimedean places, F 0 being dense in F 0 ⊗ R = ∏ τ R . Hence the F 0 -points of X are dense in its points over F 0 ⊗ R , which are the products ∏ τ X τ of the isotropic cones of the H τ .
For an isotropic vector v ∈ V let U v ⊆ SU ( V , H ) be the unipotent radical of the stabiliser of the line F v , a unipotent F 0 -group. The exponential identifies U v ( F 0 ) with an F 0 -vector space and U v ( F 0 ⊗ R ) = ∏ τ U v , τ ( R ) with the corresponding real space, so U v ( F 0 ) is dense in ∏ τ U v , τ ( R ) . Hence L contains ∏ τ U v , τ ( R ) for every rational isotropic v, and its Lie algebra l contains u v , τ for every τ and every rational isotropic v. The map v ↦ u v , τ is continuous on X τ ∖ 0 and rational v are dense in ∏ τ X τ , so the closed subspace l ∩ su ( V τ , H τ ) contains u v , τ for every isotropic v ∈ V τ . The span of these is stable under the adjoint action of SU ( V τ , H τ ) , which permutes the isotropic lines, so it is a nonzero ideal of the simple Lie algebra su ( n , n ) , hence everything. Thus l ⊇ ⨁ τ su ( V τ , H τ ) and L ⊇ G F ( R ) + , so every G F ( Q ) -orbit in D F is dense. Since Σ F is nonempty by Theorem A2 and G F ( Q ) -stable, it contains a dense orbit.
(iii) The quotient S F is a quasi-projective variety by the theorem of Baily and Borel [10]; after passing to a neat subgroup Γ of finite index it is smooth and carries the polarised abelian scheme π : A → S F . By Proposition A1(i) the group Γ ⊆ G F ( Q ) fixes W F pointwise, so every w ∈ W F descends to a flat section of R 2 n π * Q over S F that is a Hodge class at every point. Fix a Q -basis w 1 , ⋯ , w 2 m of W F . By Lemma 2(i) the algebraic locus Σ ( w j ) ⊆ S F of each w j is a countable union of Zariski closed subsets, and so is the finite intersection ⋂ j Σ ( w j ) , whose preimage is Σ F ; by Proposition A1(ii) it equals each Σ ( w j ) . The alternative is Lemma 2(ii), since p is a local homeomorphism and the preimage of a nowhere dense closed set is nowhere dense and closed.
(iv) If Σ F is closed it is closed and dense by (ii), hence everything. If it has an interior point it is not meagre, by Baire’s theorem, and (iii) gives Σ F = D F . If Σ F = D F , the Hodge conjecture holds for every member with Hg ( A s ) = G F by (i), and these are the members outside the set Z of Lemma A1. Finally, if the Hodge conjecture holds outside a countable union Z ′ of proper closed analytic subsets, then Σ F contains the complement of Z ′ , since the Weil classes are Hodge classes at every point; this complement is not meagre, and (iii) gives Σ F = D F .    □
Corollary  A1
(The algebraic locus, for every CM field). For every CM field F of degree 2 m , every n ≥ 2 and every discriminant δ, the polarised period domain D F of the family of ( F , n , δ ) -Weil type is connected of dimension m n 2 , its very general member has Hodge group G F = Res F 0 / Q SU ( V , H ) , and the algebraic locus Σ F of the Weil classes
(i)
contains a point s 0 at which A s 0 is isogenous to a product of abelian varieties with complex multiplication by F and the Weil classes are polynomials with rational coefficients in divisor classes;
(ii)
is dense in D F ;
(iii)
is the preimage of a countable union of Zariski closed subsets of S F , and is either D F or meagre.
Proof. 
Proposition A1(i), Lemma A1, Theorem A2 and Theorem A3(ii), (iii).    □
Proposition  A2
(The K-valued Weil class is multiplicative). Let K = Q ( − d ) , write δ K = − d ∈ C , and let A be of Weil type for K with H 1 ( A , Q ) = V and a K-basis u 1 , ⋯ , u 2 n of V. Put x j = u j , y j = − d u j ∈ V and
ω ( A ) = d − n ∏ j = 1 2 n d x j + δ K y j = ω 1 ( A ) + δ K ω 2 ( A ) , ω 1 ( A ) , ω 2 ( A ) ∈ ⋀ 2 n V .
Then ω ( A ) spans ⋀ 2 n V − , its conjugate spans ⋀ 2 n V + , and ω 1 ( A ) , ω 2 ( A ) form a rational basis of W ( A ) . If A 1 , A 2 are of Weil type for K, of dimensions 2 n 1 , 2 n 2 , then A 1 × A 2 with the diagonal action and the product polarisation is of Weil type, and for the basis obtained by concatenation
ω 1 ( A 1 × A 2 ) = ω 1 ( A 1 ) ω 1 ( A 2 ) − d ω 2 ( A 1 ) ω 2 ( A 2 ) , ω 2 ( A 1 × A 2 ) = ω 1 ( A 1 ) ω 2 ( A 2 ) + ω 2 ( A 1 ) ω 1 ( A 2 ) ,
that is, ω ( A 1 × A 2 ) = ω ( A 1 ) ω ( A 2 ) , the products being cup products after pull-back along the two projections.
Proof. 
The vector d x j + δ K y j = d u j + δ K − d u j ∈ V C satisfies − d · ( d x j + δ K y j ) = d y j − d δ K x j = − δ K ( d x j + δ K y j ) , so it lies in the eigenspace on which − d acts by − δ K ; call it V − . These 2 n vectors are the images of the K-basis under the projection V → V − , up to the factor 2 d , so they form a basis of V − and their product spans ⋀ 2 n V − . The conjugate spans ⋀ 2 n V + , so ω 1 = 1 2 ( ω + ω ¯ ) and ω 2 = ( 2 δ K ) − 1 ( ω − ω ¯ ) are rational, lie in W ( A ) ⊗ C and span it. The product is of Weil type because H 1 ( A 1 × A 2 ) = V 1 ⊕ V 2 as K-modules with polarisation E 1 ⊕ E 2 , so the hermitian form is H 1 ⊕ H 2 , of signature ( n 1 + n 2 , n 1 + n 2 ) . In (A2) for the concatenated basis the product splits as
ω ( A 1 × A 2 ) = d − n 1 ∏ j ≤ 2 n 1 ( d x j + δ K y j ) d − n 2 ∏ j > 2 n 1 ( d x j + δ K y j ) = ω ( A 1 ) ω ( A 2 ) ,
with no sign, the factors being taken in the order of the basis. Expanding ( ω 1 ( A 1 ) + δ K ω 2 ( A 1 ) ) ( ω 1 ( A 2 ) + δ K ω 2 ( A 2 ) ) with δ K 2 = − d and comparing rational parts and δ K -parts, which is legitimate because the classes ω j ( A i ) are rational and 1 , δ K are linearly independent over Q , gives (A3).    □

Appendix A.3. The Polarised Semiregularity Criterion

Setting  9
(Polyvector fields, evaluation and semiregularity). Let A be an abelian variety of dimension g. Write
H H * ( A ) = ⨁ p , q H p A , ⋀ q T A = ⋀ * H 0 , 1 ( A ) ⊕ H 0 ( T A ) ,
graded by p + q ; by the Hochschild–Kostant–Rosenberg isomorphism this is the Hochschild cohomology of A, as a graded ring because the Todd class of A is trivial [28,31]. Its part of degree one, H H 1 ( A ) = H 0 , 1 ( A ) ⊕ H 0 ( T A ) , has dimension 2 g , and its part of degree two is
H T 2 ( A ) = H 0 , 2 ( A ) ⊕ H 1 ( T A ) ⊕ H 0 ( ⋀ 2 T A ) , dim H T 2 ( A ) = 2 g 2 .
The algebra H H * ( A ) acts on H * ( A , C ) = ⋀ * ( H 1 , 0 ⊕ H 0 , 1 ) : a class of H 0 , 1 acts by wedge product and a class of H 0 ( T A ) = ( H 1 , 0 ) ∨ by contraction; the two families of operators involve disjoint generators, so they anticommute and the action is an action of the exterior algebra. We write ξ ⌟ γ for the action of ξ on γ, and Ann H H k ( γ ) , Ann H T 2 ( γ ) for the classes ξ of the given degree with ξ ⌟ γ = 0 . On H 1 ( T A ) = Hom ( H 1 , 0 , H 0 , 1 ) the action is the derivation that replaces one factor of type ( 1 , 0 ) by its image, so Ann H T 2 ( γ ) ∩ H 1 ( T A ) is the space of first order deformations of A along which γ stays of Hodge type to first order [48].
Let E be a nonzero perfect complex on A with Atiyah class At ( E ) ∈ Ext 1 ( E , E ⊗ Ω A 1 ) , normalised so that ch ( E ) = tr exp At ( E ) . The evaluation map is
ev E : H T 2 ( A ) → Ext 2 ( E , E ) , ev E ( z , v , π ) = z · id E + v ⌟ At ( E ) + 1 2 π ⌟ At ( E ) 2 ,
the contraction with the exponential of the Atiyah class [55, Theorem A], [12], and the semiregularity map is
σ E : Ext 2 ( E , E ) ⟶ ⨁ q ≥ 0 H q + 2 ( A , Ω A q ) , σ E ( ζ ) = ∑ q 1 q ! tr ζ ∘ At ( E ) q .
They satisfy
σ E ev E ( ξ ) = ξ ⌟ ch ( E ) ( ξ ∈ H T 2 ( A ) )
by [12, Proposition 6.2.1 and Corollary 6.3.2]. This form of the identity uses that the Todd class of A is 1; on a general variety the comparison is twisted by td 1 / 2 (Remark 5). More generally ev E extends to a homomorphism of graded rings H H * ( A ) → Ext * ( E , E ) and σ E to all degrees, with σ E ∘ ev E = ( · ) ⌟ ch ( E )  [12, Section 6], [28,31]; the form in degrees other than two is used only in Proposition A17(iii). Buchweitz and Flenner use the opposite sign for the Atiyah class, which changes only the sign of the middle term of ev E and does not affect any statement below.
Theorem  A4
(Buchweitz and Flenner [11]; Pridham [102]). Let X be a smooth projective complex variety, E a perfect complex on X, and X ↪ X → Spec R a deformation of X over an Artin local C -algebra along which ch ( E ) stays of Hodge type. If the semiregularity map
σ E : Ext X 2 ( E , E ) ⟶ ⨁ q ≥ 0 H q + 2 ( X , Ω X q ) , σ E ( ζ ) = ∑ q 1 q ! tr ζ ∘ At ( E ) q ,
is injective, then E extends to a perfect complex on X .
Lemma  A2
(Injective semiregularity deforms the complex). Let ( S , s 0 ) be a germ of a complex manifold, T → S a smooth proper family of complex tori with fibre T 0 over s 0 , and E • a bounded complex of holomorphic vector bundles on T 0 whose Chern character extends to a flat section of Hodge classes over S. If σ E is injective, there is a neighbourhood U of s 0 such that for every s ∈ U the fibre T s carries a bounded complex of holomorphic vector bundles whose Chern character is the parallel transport of ch ( E ) .
Proof. 
By the theorem of Buchweitz and Flenner in Pridham’s form (Theorem A4, [11,102]; the argument applies to compact complex manifolds), every obstruction to extending E along an Artinian deformation of T 0 inside S vanishes. The deformations of the complex structure inside S together with those of the holomorphic structures and differentials of E • are governed by a differential graded Lie algebra of Dolbeault type with finite dimensional cohomology, and Kuranishi’s method [66], in the form of [46, Sections 2–3], gives an analytic germ ( M , m 0 ) with a morphism π : ( M , m 0 ) → ( S , s 0 ) and a complex of holomorphic vector bundles on the pulled back family of tori deforming E • , whose completion is a hull of the formal deformation functor. The vanishing of the obstructions says that π is formally smooth, hence smooth, so its image contains a neighbourhood U of s 0 . Chern characters are locally constant in families, which gives the last statement. When T → S is the analytic germ of a polarised abelian scheme and Ext < 0 ( E , E ) = 0 , the same conclusion follows from Lieblich’s algebraic stack of complexes [71], which is locally of finite presentation over the base with tangent space Ext 1 ( E , E ) and obstruction space Ext 2 ( E , E ) , and then U is the image of a smooth morphism, hence Zariski open.    □
Remark  A1
(Twisted and first order deformations). Perry proves that a semiregular perfect complex E 0 with Ext < 0 ( E 0 , E 0 ) = 0 on a fibre of a smooth proper family deforms as a complex twisted by a Brauer class that vanishes on that fibre, over an étale neighbourhood of the locus where exp ( B 0 ) ch ( E 0 ) stays of Hodge type, and that this class then stays algebraic [98, Thm. 1.1]; this answers [78, Conj. 7.3.9]. It does not decide [79, Question 11.4], which asks whether injectivity of σ E on the image of ev E suffices. At first order it does: the obstruction to deforming E along v ∈ H 1 ( A , T A ) is ev E ( v ) = v ⌟ At ( E )  [57], and σ E ( ev E ( v ) ) = v ⌟ ch ( E ) by (A4), which vanishes when v is tangent to the Hodge locus of ch ( E ) ; so if σ E is injective on ev E ( H T 2 ) , then E deforms to first order along every first order deformation in that locus. The obstructions of higher order need not lie in the image of ev E , and the question remains open there.
Theorem  A5
(The annihilator is the mixed part of a splitting). Let A be an abelian 2 n -fold of Weil type for K with n ≥ 2 , write V ± 1 , 0 , V ± 0 , 1 for the Hodge components of V ± , each of dimension n, and T ± = ( V ± 1 , 0 ) ∨ ⊂ H 0 ( T A ) , and let ω = a α + + b α − ∈ W ( A ) ⊗ C with a b ≠ 0 . Put
P = Ann H H 1 ( α + ) = V + 0 , 1 ⊕ T − , Q = Ann H H 1 ( α − ) = V − 0 , 1 ⊕ T + .
(i)
dim P = dim Q = 2 n and H H 1 ( A ) = P ⊕ Q ; in particular Ann H H 1 ( ω ) = 0 .
(ii)
Ann H H 2 ( ω ) = P ∧ Q , of dimension 4 n 2 ; it is V + 0 , 1 ⊗ V − 0 , 1 ⊕ Hom K H 1 , 0 , H 0 , 1 ⊕ T + ⊗ T − , and it equals Ann H H 2 ( α + ) ∩ Ann H H 2 ( α − ) .
(iii)
Contraction with α − is injective on ⋀ * P , and contraction with α + is injective on ⋀ * Q ; the two images meet in exactly one line, in degree 2 n .
(iv)
For every k ≥ 1 ,
Ann H H k ( ω ) = ⨁ i + j = k i , j ≥ 1 ⋀ i P ∧ ⋀ j Q ⊕ ϵ k ,
where ϵ k = 0 for k ≠ 2 n and ϵ 2 n is the line of(iii). So dim Ann H H k ( ω ) = 4 n k − 2 2 n k + δ k , 2 n for k ≥ 1 , while Ann H H 0 ( ω ) = 0 , and the annihilator is the ideal generated by its degree-two part together with that one line.
Proof. 
Choose bases x 1 , ⋯ , x n of V + 1 , 0 and y 1 , ⋯ , y n of V − 1 , 0 , and bases x ¯ i of V + 0 , 1 and y ¯ j of V − 0 , 1 (the bar is notation only: complex conjugation carries V + 1 , 0 to V − 0 , 1 ), so that α + = x 1 ⋯ x n x ¯ 1 ⋯ x ¯ n and α − = y 1 ⋯ y n y ¯ 1 ⋯ y ¯ n , and write ∂ x i , ∂ y j for the dual basis of H 0 ( T A ) = T + ⊕ T − .
(i) Of the 4 n basis operators, x ¯ i ∧ kills α + because x ¯ i already occurs in it, and ∂ y j ⌟ kills α + because no y occurs in it; so P ⊇ V + 0 , 1 ⊕ T − . Conversely, the classes y ¯ j ∧ α + are linearly independent, as are the classes ∂ x i ⌟ α + , and the two families lie in different summands of H * , so no nonzero combination of the operators y ¯ j ∧ and ∂ x i ⌟ kills α + : P = V + 0 , 1 ⊕ T − , of dimension 2 n . The computation for Q is the same with the roles of the two eigenspaces exchanged, and P ∩ Q = 0 with dim P + dim Q = 4 n , so H H 1 = P ⊕ Q . A class of degree one that kills ω kills α + and α − separately, the two images lying in different summands and a b ≠ 0 , so it lies in P ∩ Q = 0 .
(iii) A monomial x ¯ I ∂ y J ∈ ⋀ * P sends α − to ± x ¯ I y [ n ] ∖ J y ¯ [ n ] , and these are linearly independent for distinct pairs ( I , J ) , which is the injectivity; the statement for Q is symmetric. For the intersection, the image of ⋀ * P is spanned by monomials with no factor x and all n factors y ¯ , and the image of ⋀ * Q , by the same computation, by monomials with no factor y and all n factors x ¯ . Both images are spanned by monomials, so their intersection is spanned by the monomials common to both, and the only such monomial is x ¯ [ n ] y ¯ [ n ] . It is the image of π P = x ¯ [ n ] ∂ y [ n ] and of π Q = y ¯ [ n ] ∂ x [ n ] , both of degree 2 n . So the intersection is the single line spanned by that class.
(ii) and (iv) Decompose ⋀ k H H 1 = ⨁ i + j = k ⋀ i P ∧ ⋀ j Q . A summand with i , j ≥ 1 annihilates α + , because it has a factor in P and the action is an action of the algebra, and annihilates α − for the same reason with Q; so it lies in the annihilator of ω . The sum of these summands over all k is the ideal generated by P ∧ Q , since each of their monomials has, after reordering, a factor p ∧ q with p ∈ P and q ∈ Q . On the remaining two summands, ξ P ∈ ⋀ k P has ξ P ⌟ ω = b ξ P ⌟ α − and ξ Q ∈ ⋀ k Q has ξ Q ⌟ ω = a ξ Q ⌟ α + , so a combination ξ P + ξ Q annihilates ω if and only if b ξ P α − = − a ξ Q α + . By (iii) both contractions are injective and their images meet only in degree 2 n , in one line; so this happens only for k = 2 n , and then only on the line spanned by b − 1 π P − a − 1 π Q , with π P and π Q scaled so that π P ⌟ α − = π Q ⌟ α + . In degree zero the annihilator is zero because ω ≠ 0 . Taking k = 2 and n ≥ 2 , so that 2 ≠ 2 n , gives (ii), the dimension being dim P · dim Q = 4 n 2 ; the general count is 4 n k − 2 n k − 2 n k plus the exceptional line.    □
Setting  10
(Polarised characters). Let A be an abelian 2 n -fold of Weil type for the imaginary quadratic field K, n ≥ 2 , with polarisation class θ, in the notation of Definition A1, and let ω = u α + + u ¯ α − , u ≠ 0 , be a nonzero real class of W ( A ) ⊗ R . For complex numbers c 0 , ⋯ , c 2 n put
p ( θ ) = ∑ k = 0 2 n c k θ k , γ = ω + p ( θ ) , μ k = k ! c k ,
and let ρ ∈ { 0 , 1 , 2 , 3 } be the rank of the ( 2 n − 1 ) × 3 Hankel matrix H = [ μ j + i ] , 0 ≤ j ≤ 2 n − 2 , 0 ≤ i ≤ 2 , theHankel rankof γ. A class of polarised shape N ω + ∑ c k θ k with N ≠ 0 is of this form after replacing ω by N ω . Write r ( γ ) for the rank of the contraction H T 2 ( A ) → H * ( A , C ) , ξ ↦ ξ ⌟ γ .
Choose a basis e 0 , ⋯ , e 2 n − 1 of V + with e j ∈ V + 1 , 0 for j < n and e j ∈ V + 0 , 1 for j ≥ n . By (A1) and since θ is nondegenerate of type ( 1 , 1 ) , there is a unique basis f 0 , ⋯ , f 2 n − 1 of V − with θ = ∑ j t j , t j = e j ∧ f j , and then f j ∈ V − 0 , 1 for j < n and f j ∈ V − 1 , 0 for j ≥ n . Write w ( x ) for the wedge product with a class x of type ( 0 , 1 ) and c ( x ) for the contraction with the tangent vector dual to a basis vector x of type ( 1 , 0 ) . For a < n ≤ b put
z a b = w ( e b ) w ( f a ) , v a b 1 = w ( e b ) c ( e a ) , v a b 2 = w ( f a ) c ( f b ) , π a b = c ( e a ) c ( f b ) ,
in V + 0 , 1 ⊗ V − 0 , 1 ⊆ H 0 , 2 , in Hom ( V + 1 , 0 , V + 0 , 1 ) and Hom ( V − 1 , 0 , V − 0 , 1 ) ⊆ H 1 ( T A ) , and in T + ⊗ T − ⊆ H 0 ( ⋀ 2 T A ) . These 4 n 2 classes form a basis of the space P ∧ Q of Theorem A5.
Theorem  A6
(The polarised criterion). In Setting 10:
(i)
For n ≥ 3 ,
Ann H T 2 ( γ ) = ⨁ a < n ≤ b x z a b + y 1 v a b 1 + y 2 v a b 2 + w π a b : ( x , y 1 + y 2 , w ) ∈ ker H ,
so that dim Ann H T 2 ( γ ) = n 2 ( 4 − ρ ) and r ( γ ) = 4 n 2 − n 2 ( 4 − ρ ) = ( 4 + ρ ) n 2 − 2 n , whatever u ≠ 0 is. For n = 2 and p = 0 the same holds, with ρ = 0 and r ( γ ) = 2 n ( 2 n − 1 ) .
(ii)
The n 2 classes v a b 1 − v a b 2 span the tangent space of the polarised family D n , n at A, and they lie in Ann H T 2 ( γ ) . For n ≥ 3 , if c k ≠ 0 for some 1 ≤ k ≤ 2 n − 1 , then Ann H T 2 ( γ ) ∩ H 1 ( T A ) , the first order Hodge locus of γ, is exactly this tangent space; and if ρ = 3 the whole of Ann H T 2 ( γ ) is this tangent space, with no component in H 0 , 2 or in H 0 ( ⋀ 2 T A ) .
(iii)
Let E be a perfect complex on A with ch ( E ) = γ . Then dim Ext 2 ( E , E ) ≥ r ( γ ) , with no hypothesis on σ E . If equality holds, ev E is onto with kernel Ann H T 2 ( γ ) and σ E is injective.
(iv)
Let E be a perfect complex on A with ch ( E ) = γ and σ E injective. Then E deforms with A over an open subset of D n , n , the Weil classes are algebraic on every member of the family D n , n , and the Hodge conjecture holds in every degree for every member with Hodge group SU ( V , H ) .
Proof. 
(i) Let the torus ( C × ) 2 n act on V C by e j ↦ τ j e j , f j ↦ τ j − 1 f j . It preserves the Hodge decomposition and θ , and scales α ± by ( ∏ j τ j ) ± 1 ; the action of H H * ( A ) is equivariant. The basis of H H 1 ( A ) formed by the operators w ( e b ) , w ( f a ) , c ( e a ) , c ( f b ) consists of weight vectors of weights ± ϵ j , so a weight of H T 2 ( A ) has at most two nonzero coordinates, and for n ≥ 3 two such weights never differ by a nonzero multiple of 1 = ( 1 , ⋯ , 1 ) ∈ Z 2 n . Decompose ξ ∈ H T 2 ( A ) into weight components ξ λ . Then ξ λ ⌟ p ( θ ) , ξ λ ⌟ α + and ξ λ ⌟ α − have weights λ , λ + 1 and λ − 1 , and all the weights so obtained from the different λ are distinct. Hence
Ann H T 2 ( γ ) = Ann H T 2 ( p ( θ ) ) ∩ Ann H T 2 ( α + ) ∩ Ann H T 2 ( α − ) ,
and the last two meet in P ∧ Q by Theorem A5(ii) and its proof. For n = 2 and p = 0 the equality Ann H T 2 ( ω ) = P ∧ Q is Theorem A5(ii) itself. The space P ∧ Q is the sum over a < n ≤ b of the four-dimensional weight spaces of weight ϵ b − ϵ a spanned by z a b , v a b 1 , v a b 2 , π a b , and Ann ( p ( θ ) ) is stable under the torus, so the computation splits pair by pair. Fix a < n ≤ b , drop the indices, and put m = e b ∧ f a and θ ′ = θ − t a − t b . The t j are even with t j 2 = 0 , so θ k = θ ′ k + k θ ′ k − 1 ( t a + t b ) + k ( k − 1 ) θ ′ k − 2 t a t b , and since c ( e a ) t a = f a and c ( f b ) t b = − e b , direct evaluation gives
z ⌟ θ k = m θ ′ k , v 1 ⌟ θ k = k m θ ′ k − 1 , v 2 ⌟ θ k = k m θ ′ k − 1 , π ⌟ θ k = k ( k − 1 ) m θ ′ k − 2 .
So for ξ = x z + y 1 v 1 + y 2 v 2 + w π and λ = y 1 + y 2 ,
ξ ⌟ p ( θ ) = m ∧ ∑ j = 0 2 n − 2 θ ′ j j ! x μ j + λ μ j + 1 + w μ j + 2 .
The classes m ∧ θ ′ j , 0 ≤ j ≤ 2 n − 2 , are nonzero, because θ ′ is a nondegenerate form on the span of the 4 n − 4 basis vectors other than e a , f a , e b , f b , and they have distinct degrees; so ξ kills p ( θ ) exactly when ( x , λ , w ) ∈ ker H . The map ( x , y 1 , y 2 , w ) ↦ ( x , λ , w ) is onto with kernel spanned by v 1 − v 2 , which gives 1 + ( 3 − ρ ) dimensions for each of the n 2 pairs.
(ii) A first order deformation of A in H 1 ( T A ) = Hom ( H 1 , 0 , H 0 , 1 ) carries the action of K exactly when it is K-linear, that is, lies in Hom ( V + 1 , 0 , V + 0 , 1 ) ⊕ Hom ( V − 1 , 0 , V − 0 , 1 ) , with basis the v a b 1 and v a b 2 ; and it keeps the polarisation exactly when it kills θ . By the formulas of (i) with k = 1 , a combination ∑ y a b 1 v a b 1 + y a b 2 v a b 2 kills θ exactly when y a b 1 + y a b 2 = 0 for all pairs. So the tangent space of D n , n , which has dimension n 2 by Proposition A1(i), is spanned by the classes v a b 1 − v a b 2 . These kill every θ k by the same formulas, and they kill α ± because they lie in P ∧ Q ; so they lie in Ann H T 2 ( γ ) for every n ≥ 2 . For n ≥ 3 , by (i) the intersection of Ann H T 2 ( γ ) with H 1 ( T A ) consists of the combinations with ( 0 , y 1 + y 2 , 0 ) ∈ ker H in each pair, that is, with ( y 1 + y 2 ) μ j + 1 = 0 for 0 ≤ j ≤ 2 n − 2 ; if some μ k with 1 ≤ k ≤ 2 n − 1 is nonzero this forces y 1 + y 2 = 0 . If ρ = 3 then ker H = 0 .
(iii) By (A4), ker ev E ⊆ Ann H T 2 ( γ ) , so dim Ext 2 ( E , E ) ≥ rank ev E ≥ dim H T 2 ( A ) − dim Ann H T 2 ( γ ) = r ( γ ) . If equality holds, ev E is onto and its kernel has the dimension of Ann H T 2 ( γ ) , hence equals it. Every class of Ext 2 ( E , E ) is then ev E ( ξ ) for some ξ , and σ E ( ev E ( ξ ) ) = ξ ⌟ γ vanishes only for ξ ∈ Ann H T 2 ( γ ) = ker ev E ; so σ E is injective.
(iv) Since ch ( E ) = γ is rational, ω is a nonzero rational Weil class and the c k are rational. The classes θ and ω are flat sections of Hodge classes over D n , n (Proposition A1(i)), so γ is. Since A is projective, E is isomorphic to a bounded complex of vector bundles, and Lemma A2 gives an open neighbourhood U of the point of A in D n , n over which E deforms: every A s , s ∈ U , carries a bounded complex of holomorphic vector bundles E s with ch ( E s ) = γ s . The abelian variety A s is projective, so E s is algebraic and its Chern character is the class of an algebraic cycle with rational coefficients. Its component of degree 2 n is ω s + c n θ n , so ω s is algebraic, and then every class of W ( A s ) is algebraic by Proposition A1(ii). So the algebraic locus Σ K contains the open set U, and Σ K = D n , n by Theorem A3(iv). The last statement is Theorem A1.    □
Remark  A2
(The range of n and the Hankel rank). The formula r ( γ ) = ( 4 + ρ ) n 2 − 2 n is proved for n ≥ 3 , which is the range in which Theorem 20 is applied in the paper; the values used there are r = 57 for n = 3 and r = 104 for n = 4 , both with ρ = 3 . For n = 2 the two weights ϵ 2 + ϵ 3 and − ϵ 0 − ϵ 1 of H T 2 differ by 1 , the three annihilators in the proof of (i) are coupled in one block, and dim Ann H T 2 ( γ ) can exceed 4 ( 4 − ρ ) by at most two; parts (ii) (first sentence), (iii) and (iv) hold for every n ≥ 2 with r ( γ ) defined as the rank of the contraction. The matrix H has three columns, so ρ ≤ 3 ; ρ = 3 for a general polynomial p, and ρ ≤ 2 exactly when μ 0 , ⋯ , μ 2 n satisfy a linear recurrence x μ j + λ μ j + 1 + w μ j + 2 = 0 , 0 ≤ j ≤ 2 n − 2 , with ( x , λ , w ) ≠ 0 , that is, when p is one of a e s θ + b e t θ , ( a + b θ ) e t θ , a e t θ + b θ 2 n , a θ 2 n − 1 + b θ 2 n , according to the roots of w T 2 + λ T + x .
The criterion cannot be met by a complex whose Chern character is a multiple of a Weil class alone. This is proved on the larger family of Weil tori.
Setting  11
(Weil tori). Let A be an abelian 2 n -fold of Weil type for K, V = H 1 ( A , Q ) with its action of K, and V ⊗ C = V + ⊕ V − the two eigenspaces, each of dimension 2 n . For complementary n-dimensional subspaces X , Y ⊂ V + the subspace X ⊕ Y ¯ of V ⊗ C is the space of holomorphic 1-forms of a complex structure on the real torus underlying A; call the resulting complex torus T X , Y . On T X , Y the field K acts with signature ( n , n ) , so T X , Y is a Weil torus in the sense of [124], and every complex structure on that real torus commuting with K and of signature ( n , n ) arises in this way. The pairs ( X , Y ) form a connected complex manifold S of dimension 2 n 2 , containing the point of A; the polarised Weil family is the submanifold of dimension n 2 on which the polarisation of A is of type ( 1 , 1 ) , and the tangent space of S is the space Hom K ( H 1 , 0 , H 0 , 1 ) of K-linear first order deformations in Theorem A5(ii). Since α + spans ⋀ 2 n V + = ⋀ n X ⊗ ⋀ n Y and X is holomorphic while Y is antiholomorphic, the Weil classes are of type ( n , n ) on every T X , Y . We say that a property holds for a very general member of S if it holds outside a countable union of proper analytic subsets.
Lemma  A3
(The Hodge classes of a very general Weil torus). For a very general member T of S, Hdg k ( T ) = 0 for 1 ≤ k ≤ 2 n − 1 , k ≠ n , and Hdg n ( T ) is the rational Weil plane.
Proof. 
There are countably many rational classes, and the Hodge locus of each is an analytic subset of the connected manifold S; so a rational class that is of Hodge type at a very general member is of Hodge type on all of S. Let G = Res K / Q GL 2 n be the group of K-linear automorphisms of V. Its real points act transitively on S, since GL ( V + ) acts transitively on pairs of complementary subspaces. The Mumford–Tate group M of a very general member lies in G, since K acts by endomorphisms of every member, and it contains the Mumford–Tate group of every member, whose Hodge classes include those of the very general one. So the real Lie algebra of M contains the complex structure J s of every member, which is the derivative of its Hodge cocharacter, and therefore contains the Ad G ( R ) -span of one J s . This span is an ideal of g R = gl 2 n ( C ) containing the non-central element J s , which acts on V + by i on X and by − i on Y; hence it contains sl 2 n ( C ) , and M contains G der = Res K / Q SL 2 n . Over C this is SL ( V + ) × SL ( V − ) acting on ⋀ * ( V + ⊕ V − ) = ⋀ * V + ⊗ ⋀ * V − , whose invariants are spanned by 1, α + , α − and α + α − . Hodge classes are invariant under M, which gives the statement.    □
Lemma  A4
(A very general Weil torus has no subvarieties). A very general member T of S has no analytic subsets other than finite sets and T itself.
Proof. 
Let Z ⊂ T be an irreducible analytic subset of codimension c with 0 < c < 2 n . For every Kähler class κ the class [ Z ] satisfies ∫ T [ Z ] κ 2 n − c = ∫ Z κ 2 n − c > 0 , so [ Z ] is a nonzero Hodge class of degree 2 c . By Lemma A3 this forces c = n and [ Z ] into the Weil plane. Take for κ 0 the class of a positive definite hermitian form that is block diagonal for the splitting H 1 , 0 = X ⊕ Y ¯ . Its blocks lie in X ⊗ X ¯ and Y ¯ ⊗ Y , both inside V + ⊗ V − , so every power of κ 0 has a factor from V + and a factor from V − in each term, and α ± ∧ κ 0 n = 0 because α ± spans the top exterior power of V ± . So ∫ T [ Z ] κ 0 n = 0 , which is impossible.    □
Proposition  A3
(Coherent sheaves on a very general Weil torus). Let n ≥ 2 and let T be a very general member of S. Every coherent analytic sheaf F on T has ch k ( F ) = 0 in H 2 k ( T , Q ) for 1 ≤ k ≤ 2 n − 1 . For n = 2 this is Voisin’s theorem that every coherent sheaf on a general Weil torus of dimension four has c 2 ( F ) = 0 [124, Theorem 10], whose proof we follow.
Proof. 
Chern classes of coherent sheaves on T exist in rational Deligne cohomology, extend those of bundles and are additive [47], so they are Hodge classes. By Lemma A4 the torsion of F, the locus where the torsion-free quotient is not locally free, and the cokernel of the map from that quotient to its reflexive hull are all supported on finite sets, and a sheaf supported on a finite set has Chern character in the top degree only. So it suffices to treat a reflexive sheaf G. Every coherent sheaf on T has c 1 = 0 , since Hdg 1 ( T ) = 0 (Lemma A3); so every torsion-free sheaf has slope zero for every Kähler class, and a torsion-free sheaf is stable for one Kähler class exactly when it has no saturated subsheaf of intermediate rank, which does not depend on the class. A Jordan–Hölder filtration of G therefore has stable torsion-free factors Q j , stable for every Kähler class, and ch ( G ) = ∑ j ch ( Q j ) . By the theorem of Bando and Siu [25] the stable reflexive sheaf Q j * * carries an admissible Hermite–Einstein metric for every Kähler class κ and satisfies ∫ T 2 r c 2 − ( r − 1 ) c 1 2 ( Q j * * ) κ 2 n − 2 ≥ 0 , with equality exactly when Q j * * is a projectively flat vector bundle. Here c 1 = 0 and c 2 ( Q j * * ) ∈ Hdg 2 ( T ) . For n ≥ 3 that group is zero; for n = 2 it is the Weil plane, and then ∫ T c 2 ( Q j * * ) κ 0 2 = 0 for the class κ 0 of the proof of Lemma A4. In either case equality holds for some κ , so Q j * * is a projectively flat bundle with c 1 = 0 , hence flat, and its rational Chern classes of positive degree vanish. Since Q j and Q j * * differ only on a finite set, ch k ( Q j ) = 0 for 1 ≤ k ≤ 2 n − 1 , and the same holds for G and for F.    □
Theorem  A7
(The pure form of the criterion is never met). Let n ≥ 2 , let A be an abelian 2 n -fold of Weil type for K, ω a nonzero rational Weil class, and E a perfect complex on A with ch ( E ) − N ω ∈ H 0 ( A ) ⊕ H 4 n ( A ) , N ≠ 0 . Then σ E is not injective. In particular, if ch ( E ) = N ω then dim Ext 2 ( E , E ) > 2 n ( 2 n − 1 ) .
Proof. 
Suppose that σ E is injective. Since A is projective, E is isomorphic to a bounded complex of holomorphic vector bundles. The class ch ( E ) is of Hodge type on every member of the family S of Setting 11, so by Lemma A2 every member T s with s in a neighbourhood U of the point of A in S carries a bounded complex of holomorphic vector bundles E s • whose Chern character is the parallel transport of ch ( E ) . Choose s ∈ U very general. By Proposition A3, ch n ( E s ) = ∑ i ( − 1 ) i ch n ( E s i ) = 0 , while it is the transport of N ω , which is not zero. This contradiction proves that σ E is not injective. If ch ( E ) = N ω , then dim Ext 2 ( E , E ) ≥ 2 n ( 2 n − 1 ) and equality would force σ E to be injective, by Theorem A6(i) for p = 0 and (iii).    □

Appendix A.4. Convolutions of Line Bundles on E 0 6 and E 0 8

This subsection treats the convolutions, the iterated cones of shifted line bundles, at the member E 0 2 n , E 0 = C / Z [ i ] , of the family of Weil type for K = Q ( i ) . Their Chern character is the same signed sum of exponentials as for a direct sum, but their Ext groups are cut down by the maps between the pieces. Proposition A4 writes a signed sum of 2 · 4 n − 1 line bundles that is exactly a real Weil class; for every convolution of those pieces with three multiples of the polarisation, in any order and with any shifts, twists and placement of the multiples, Theorem A9 gives dim Ext 2 ≥ r ( γ ) + 1 , also for the invariant part under any finite group acting by translations and twists. The larger bounds dim Ext 2 ≥ 69 > 57 at n = 3 (Theorem A12) and dim Ext 2 ≥ 640 > 104 at n = 4 for the arrangements of Theorem A13 to Corollary A3 measure how far those convolutions are from the criterion. A perfect complex E with ch ( E ) = γ of polarised shape is said to meet the polarised criterion if dim Ext 2 ( E , E ) = r ( γ ) , the equality case of Theorem A6(iii).
Setting  12
(Convolutions). Let A be an abelian variety of dimension g with a flat Kähler metric, and let L 1 , ⋯ , L m be line bundles on A with hermitian metrics of constant curvature, hermitian forms H 1 , ⋯ , H m and pairwise distinct first Chern classes. For integers d 1 , ⋯ , d m put P a = L a [ − d a ] . Aconvolutionof P 1 , ⋯ , P m is an object E given by a one-sided twisted complex13
⨁ a A 0 , * ( L a ) [ − d a ] , ∂ ¯ + δ , δ = ( δ a b ) a < b , δ a b ∈ A 0 , 1 + d a − d b L a − 1 ⊗ L b ,
with ∂ ¯ δ + δ ∘ δ = 0 , so that ch ( E ) = ∑ a ( − 1 ) d a e c 1 ( L a ) . Write A ( a , b ) = A 0 , * ( L a − 1 ⊗ L b ) and H ( a , b ) for its harmonic forms, so that H q ( a , b ) ≅ H q ( A , L a − 1 ⊗ L b ) , and write π, G and h = ∂ ¯ * G for the harmonic projection, the Green operator and the homotopy, with 1 − π = ∂ ¯ h + h ∂ ¯ and π h = h π = h 2 = 0 . A form of degree q in A ( a , b ) is a morphism P a → P b of degree w = q − d a + d b ; we call q its Ext degree and, when w = 1 , call q − 1 itsexcess. When H b − H a is nondegenerate, H ( a , b ) is concentrated in the single degree n ( H b − H a ) , the number of negative eigenvalues [19, Chapter 3].
Lemma  A5
(The harmonic model). In Setting 12:
(i)
E is isomorphic to a convolution of the same pieces with h δ = 0 , and then δ = δ H − h ( δ ∘ δ ) with δ H = π δ harmonic. Substituting repeatedly, every δ a b is a finite sum of planar binary trees whose leaves are components δ c d H with c < d , whose internal vertices compose, and whose internal edges carry − h .
(ii)
Ext * ( E , E ) is the cohomology of H E = ⨁ a , b H ( a , b ) , graded by w, with the differential d E = ∑ k ≥ 0 π Q ( − h Q ) k , where Q = [ δ , − ] .
(iii)
For x ∈ H ( a , b ) , the component of d E ( x ) in H ( c , e ) is a sum of terms ± π T , one for each planar binary tree T with k ≥ 2 leaves, internal vertices composing, − h on the internal edges other than the root, whose leaves read in order are the components along a path c = a 0 → a 1 → ⋯ → a k = e in which one step is x : a → b and the others are components δ a j a j + 1 H with a j < a j + 1 . The Ext degree of such a term is the sum of the Ext degrees of its leaves minus k − 2 .
The terms with k leaves in (iii) are the Massey products of length k of the pieces;14 (iii) is all that is used of them below.
Proof. 
(i) Suppose h δ c d = 0 for d − c < r . Conjugating by 1 + g , with g of degree 0 supported at distance r, replaces δ a b for b − a = r by X a b + ∂ ¯ g a b , where X a b depends only on δ , and leaves the components at smaller distance unchanged. With g a b = − h X a b the new component is X a b − ∂ ¯ h X a b , and h ∂ ¯ h = h ( 1 − π − h ∂ ¯ ) = h gives h ( X a b − ∂ ¯ h X a b ) = 0 . Induction on r gives h δ = 0 , and then δ = π δ + ∂ ¯ h δ + h ∂ ¯ δ = δ H − h ( δ ∘ δ ) by the Maurer–Cartan equation. The component ( δ ∘ δ ) a b involves only components at distance less than b − a , so the substitution ends. (ii) The endomorphism complex of E is ⨁ a , b A ( a , b ) with the differential ∂ ¯ + Q . Blockwise, the inclusion of harmonic forms, π and − h form a strong deformation retraction of ( ⨁ A ( a , b ) , ∂ ¯ ) onto ( H E , 0 ) , and Q raises b − a , so h Q is nilpotent. The basic perturbation lemma gives a homotopy equivalence of ( ⨁ A ( a , b ) , ∂ ¯ + Q ) with ( H E , d E ) , with the chain maps ι E = ∑ k ( − h Q ) k on H E and π E = π ∑ k ( − Q h ) k . (iii) Each Q in d E composes with one component of δ on the left or on the right; writing that component as in (i) gives the trees, the edges from (i) and those from d E being exactly the internal edges other than the root. Composition adds degrees, h lowers them by one, and a binary tree with k leaves has k − 2 internal edges.    □
Proposition  A4
(A pure Weil class from line bundles). Let E 0 = C / Z [ i ] , A = E 0 2 n with coordinates z 1 , ⋯ , z 2 n and n ≥ 2 , let K = Q ( i ) act by i on z 1 , ⋯ , z n and by − i on z n + 1 , ⋯ , z 2 n , and let θ be the product principal polarisation; then A is an abelian 2 n -fold of Weil type for Q ( i ) , with complex multiplication. Write B j for the surface with coordinates ( z j , z n + j ) , so that A = B 1 × ⋯ × B n . For p ∈ Z and ζ ∈ μ 4 n let L ζ be a line bundle whose hermitian form is p ζ j ζ ¯ j p on B j and zero between different surfaces, and put α = ∏ j i 2 d z j ∧ d z ¯ n + j .
(i)
α and α ¯ span the Weil classes of A, and
∑ ∏ ζ = 1 e c 1 ( L ζ ) − ∑ ∏ ζ = − 1 e c 1 ( L ζ ) = 2 · 4 n − 1 ( α + α ¯ ) .
(ii)
For distinct integers t 1 , t 2 , t 3 and signs σ i , the character γ = 2 · 4 n − 1 ( α + α ¯ ) + ∑ i σ i e t i θ has Hankel rank 3, so r ( γ ) = 7 n 2 − 2 n for n ≥ 3 . A convolution with this character of the 2 · 4 n − 1 bundles L ζ , with ( − 1 ) d = ∏ ζ , and of three line bundles M i with c 1 ( M i ) = t i θ has 2 · 4 n − 1 + 3 pieces and ( 2 · 4 n − 1 + 3 ) n ( 2 n − 1 ) diagonal classes of degree two: 525 against r = 57 at n = 3 , and 3668 against 104 at n = 4 .
Proof. 
The hermitian forms have imaginary parts integral on Z [ i ] 2 n , so the line bundles exist. On B j write c 1 of the restriction as p ( a j + b j ) + ζ j c j + ζ ¯ j c j ′ , where a j = i 2 d z j ∧ d z ¯ j , b j = i 2 d z n + j ∧ d z ¯ n + j , c j = i 2 d z j ∧ d z ¯ n + j and c j ′ = i 2 d z n + j ∧ d z ¯ j . Since c j 2 = c j ′ 2 = 0 ,
e c 1 ( L ζ ) = ∏ j f j ( ζ j ) , f j ( z ) = e p ( a j + b j ) 1 + z c j + z ¯ c j ′ + c j c j ′ .
For u ∈ μ 4 , 1 2 ( u + u ¯ ) is u when u = ± 1 and 0 when u = ± i , so the left side of (i) is 1 2 ∏ j ∑ z z f j ( z ) + ∏ j ∑ z z ¯ f j ( z ) . Now ∑ z ∈ μ 4 z = ∑ z 2 = 0 and ∑ | z | 2 = 4 , so ∑ z z f j ( z ) = 4 e p ( a j + b j ) c j ′ = 4 c j ′ , because a j c j ′ and b j c j ′ repeat a differential, and likewise ∑ z z ¯ f j ( z ) = 4 c j . This gives 2 · 4 n − 1 ( ∏ c j ′ + ∏ c j ) = 2 · 4 n − 1 ( α ¯ + α ) . The differentials d z j , j ≤ n , and d z ¯ n + j span the i-eigenspace of K on H 1 ( A , C ) , and α spans its top exterior power, which is a Weil class, and α ¯ the conjugate one. (ii) The moments μ m = ∑ i σ i t i m make the Hankel matrix of Setting 10 a product of Vandermonde matrices of rank 3, and Theorem A6(i) gives r for n ≥ 3 . The counts are immediate.    □
The pieces L ζ and L ζ ′ have a degenerate difference as soon as ζ and ζ ′ share a coordinate. The rest of this subsection computes what the Massey products can do to the diagonal classes of such a convolution. The pieces are external products over the surfaces B j , and the first tool is a bound on how far a product can lower the degree on each surface.
Lemma  A6
(Degree drop on a product). Let A = B 1 × ⋯ × B n with a product metric, and let every L a be an external product with a product metric. Decompose the leaves of a term of Lemma A5(iii) with k leaves into tensor products of harmonic forms on the factors. Then the output has degree s j − d j on B j , where s j is the sum of the degrees of the leaves on B j and d j ≥ 0 , ∑ j d j = k − 2 , and
d j ≤ max ( 0 , u j − 2 ) ,
where u j is the number of leaves whose component on B j is not a multiple of the unit section of O B j (such a unit occurs only where source and target agree on B j ).
Proof. 
The Laplacian is Δ = ∑ j Δ j , so h = ∑ j h j with h j = ∂ ¯ j * G = G ∂ ¯ j * , which lowers the degree on B j by one and keeps the others; expand every internal edge accordingly, and let d j be the number of edges carrying h j . Let V j , U j and W j be the forms whose component on B j is Δ j -harmonic, a multiple of the unit, or in the image of ∂ ¯ j * . The operators G and h i , i ≠ j , preserve all three, h j kills V j and W j , π kills W j , and multiplication by an element of U j preserves V j , U j and W j . For an edge e carrying h j let N ( e ) be the set of leaves below e whose component on B j is not a unit. If | N ( e ) | ≤ 1 the value below e lies in V j and the term vanishes. If e lies below the next such edge e ′ with N ( e ) = N ( e ′ ) , everything composed in between lies in U j , the value at e ′ lies in W j , and the term vanishes; if N ( e ) is the set of all u j such leaves for the highest such edge, the value at the root lies in W j and π kills it. So in a nonzero term the sets N ( e ) are distinct, of size at least two, different from the whole set, and pairwise nested or disjoint. Together with the whole set they form a family of that kind, which has at most u j − 1 members, since its members are the internal vertices of a forest with u j leaves in which every internal vertex has at least two children. Hence d j ≤ u j − 2 .    □
Proposition  A5
(Nothing enters the diagonal). Keep Proposition A4 with n ≥ 3 , and let t 1 , t 2 , t 3 be distinct with | t i − p | ≥ 2 . For every convolution E of the pieces L ζ and M i , in any order, with any shifts of the given parities and with the pieces twisted by any line bundles of degree zero, no element of degree one of H E has a differential with a nonzero diagonal component. Consequently
dim Ext 2 ( E , E ) ≥ dim ker d E | D 2 , D 2 = ⨁ a H 2 ( A , O A ) .
Proof. 
A term of Lemma A5(iii) that reaches a diagonal class in H 2 ( c , c ) = H 2 ( A , O A ) has k leaves of degree one running around a closed path c → ⋯ → c ; the pieces before the step x increase from c and those after it increase to c, so the path passes through k distinct pieces; and the output has Ext degree ∑ j q j − ( k − 2 ) = 2 , where q j are the Ext degrees of the leaves, so the total excess ∑ j ( q j − 1 ) is 0. The step from L ζ to L ζ ′ has Ext degree 1 on each surface where ζ j ≠ ζ j ′ , the difference there having eigenvalues ± | ζ j ′ − ζ j | , and at least 0 elsewhere. If ∏ ζ = ∏ ζ ′ the two differ on at least two surfaces; otherwise the step has even Ext degree, the shifts having opposite parities. Either way its excess is at least 1. The difference between L ζ and M i has eigenvalues t i − p ± 1 on every surface, so it is definite: the step has Ext degree 0 and excess − 1 if it goes up the value ϕ , where ϕ ( L ζ ) = p and ϕ ( M i ) = t i , and Ext degree 2 n and excess 2 n − 1 if it goes down; the same holds between the M i . A maximal run of the cycle through the M i starts and ends at the value p, or is the whole cycle, so it has at least one step down and at most three steps up; its excess is at least 2 n − 1 − 3 ≥ 2 . So every cycle has excess at least 1. Twisting by line bundles of degree zero leaves the forms unchanged and can only make some H ( a , b ) vanish. For the inequality, the diagonal components of cocycles of degree two vanish on coboundaries, so dim Ext 2 ( E , E ) is at least the dimension of their image, which contains the kernel of d E on D 2 .    □
At n = 2 the bound on a run is 0 and a cycle through the three multiples can reach the diagonal; this is why the analysis below starts at n = 3 .
The hypothesis | t i − p | ≥ 2 can be removed, and then the diagonal of the obstruction decides the question for every convolution of these pieces at once. The first input is the range of degrees of the cohomology of a line bundle on an abelian surface, including the degenerate case.
Lemma  A7
(Cohomology of a line bundle on an abelian surface). Let N be a line bundle on an abelian surface B whose hermitian form has i + positive and i − negative eigenvalues. Then H q ( B , N ) = 0 unless i − ≤ q ≤ 2 − i + . Consequently, if A = B 1 × ⋯ × B n and N = N 1 ⊠ ⋯ ⊠ N n , then H q ( A , N ) = 0 unless i − ≤ q ≤ 2 n − i + , where i ± now count the eigenvalues of the hermitian form of N, the orthogonal sum of those of the N j .
Proof. 
If the form is nondegenerate, H * ( B , N ) is concentrated in degree i −  [88, §16], [19, Chapter 3]. If it is zero, N ∈ Pic 0 ( B ) and 0 ≤ q ≤ 2 . Otherwise exactly one eigenvalue λ is nonzero. The kernel W of the form is the kernel of its imaginary part, which is integral on the lattice Λ , so W ∩ Λ is a lattice in W and E W = W / ( W ∩ Λ ) is an elliptic curve in B; let f : B → C = B / E W be the quotient, an elliptic curve. The form vanishes on W × C 2 , so ϕ N ( x ) = t x * N ⊗ N − 1 restricts trivially to E W for every x ∈ B , and N has isomorphic restrictions to all fibres of f. If N | E W is nontrivial, of degree zero, then f * N = R 1 f * N = 0 and H * ( B , N ) = 0 . If it is trivial, N ≅ f * N ¯ by the seesaw principle, and deg N ¯ has the sign of λ . Then R f * N ≅ N ¯ ⊗ R f * O B , and R f * O B ≅ O C ⊕ O C [ − 1 ] : indeed R 1 f * O B ≅ O C , the family being isotrivial with translation-invariant H 0 , 1 of the fibre, and the extension splits because Ext 2 vanishes on a curve. So H q ( B , N ) = H q ( C , N ¯ ) ⊕ H q − 1 ( C , N ¯ ) , which is nonzero only for q ∈ { 0 , 1 } if λ > 0 and only for q ∈ { 1 , 2 } if λ < 0 . In all cases i − ≤ q ≤ 2 − i + . The second statement follows from the Künneth formula.    □
Theorem  A8
(Nothing enters the diagonal, for every placement). The conclusion of Proposition A5 holds for all pairwise distinct integers t 1 , t 2 , t 3 , without the hypothesis | t i − p | ≥ 2 : for n ≥ 3 and every convolution E of the pieces L ζ and M i , in any order, with any shifts of the given parities and any twists of degree zero, every closed path of distinct pieces whose steps have degree one and nonzero harmonic groups has total excess at least 1, so no element of degree one of H E has a differential with a nonzero diagonal component.
Proof. 
As in the proof of Proposition A5, it suffices to show that every closed path a 0 → a 1 → ⋯ → a k = a 0 through k ≥ 2 distinct pieces, with steps of degree one and Ext degrees q j , has ∑ j ( q j − 1 ) ≥ 1 ; twisting by bundles of degree zero leaves the hermitian forms unchanged and can only make harmonic groups vanish. The steps between two bundles L ζ have excess at least 1, by the argument of that proof. Put d i = t i − p and let the value be ϕ ( L ζ ) = p , ϕ ( M i ) = t i . On every surface B j the form of M i ⊗ L ζ − 1 has the eigenvalues d i ± 1 . By Lemma A7, a step L ζ → M i has Ext degree q ≥ 0 if d i ≥ 1 , q = n if d i = 0 , q ≥ n if d i = − 1 and q = 2 n if d i ≤ − 2 ; for a step M i → L ζ the same holds with d i replaced by − d i . The form of M k ⊗ M i − 1 is ( t k − t i ) times the identity, so a step between multiples has q = 0 if it raises ϕ and q = 2 n if it lowers it. Hence a step that strictly raises ϕ has excess at least − 1 ; every other step touching a multiple has excess at least n − 1 , and exactly 2 n − 1 if it lowers ϕ between two multiples or between a multiple and a bundle L ζ at value distance at least 2.
A closed path through bundles L ζ only has excess at least k ≥ 2 . A closed path through multiples only has at most three steps, at least one of which lowers ϕ ; its excess is at least 2 n − 1 − 2 = 2 n − 3 . Otherwise cut the path into its steps between bundles L ζ , each of excess at least 1, and its maximal runs L ζ → M a 1 → ⋯ → M a r → L ζ ′ through 1 ≤ r ≤ 3 distinct multiples. A run starts and ends at the value p, so at least one of its r + 1 steps does not raise ϕ . If two or more do not, its excess is at least 2 ( n − 1 ) − ( r − 1 ) ≥ 2 n − 4 . If exactly one step s does not raise ϕ , three cases occur. If s lies between multiples, it lowers ϕ , and the run has excess at least 2 n − 1 − r ≥ 2 n − 4 . If s is the first step, the values increase strictly from t a 1 to p along the run, so t a 1 ≤ p − r ; for r ≥ 2 the step s has excess 2 n − 1 and the run at least 2 n − 4 , and for r = 1 the run has excess at least ( n − 1 ) − 1 = n − 2 . If s is the last step, the same argument applies with t a r ≥ p + r . Since 2 n − 4 ≥ n − 2 , every run has excess at least n − 2 ≥ 1 , and so has every closed path.    □
The diagonal components of a class of degree two are therefore well defined. The next lemma computes them on the image of the characteristic map, in the generality of Setting 12. Call a cycle of excess zero through a piece P a a closed path a = a 0 → a 1 → ⋯ → a k = a through k ≥ 2 distinct pieces with nonzero harmonic groups H q j ( a j , a j + 1 ) , each step of degree q j − d a j + d a j + 1 = 1 , and ∑ j ( q j − 1 ) = 0 .
Lemma  A8
(The diagonal of the characteristic map). In Setting 12, let P a be a piece through which no cycle of excess zero passes. Then the component in H 2 ( a , a ) = H 2 ( A , O A ) of a cocycle of degree two of ( H E , d E ) depends only on its class, and defines a linear map Δ a : Ext 2 ( E , E ) → H 2 ( A , O A ) with
Δ a ev E ( ξ ) = ev L a ( ξ ) ( ξ ∈ H T 2 ( A ) ) , in particular Δ a ev E ( v ) = v ⌟ c 1 ( L a )
for v ∈ H 1 ( A , T A ) , where ev E ( v ) = v ⌟ At ( E ) is the obstruction to deforming E along v [11,57].
Proof. 
By Lemma A5(iii), a term of d E ( x ) , x of degree one, with a nonzero component in H 2 ( a , a ) runs along a closed path from a to a whose steps have degree one; as in the proof of Proposition A5, its pieces are distinct and its total excess is 0, so it is a cycle of excess zero through P a . There is none, so coboundaries of degree two have zero component in H 2 ( a , a ) , and Δ a is well defined. As in the proof of Proposition 23(i), the Atiyah class of E is represented in the Dolbeault model by [ D ′ , ∂ ¯ + δ ] , where D ′ is the sum of the ( 1 , 0 ) -parts of the Chern connections; this is ⨁ a F a + [ D ′ , δ ] , with F a the constant curvature form of L a , normalised to represent c 1 ( L a ) , and [ D ′ , δ ] strictly upper triangular because δ is. So ξ ⌟ exp At is block upper triangular, with diagonal blocks ξ ⌟ exp F a , which are forms with constant coefficients. The quasi-isomorphism π E = π ∑ k ( − Q h ) k of Lemma A5(ii) has Q = [ δ , − ] , which maps upper triangular elements to strictly upper triangular ones, while h preserves blocks; so the diagonal block of π E ( z ) is π ( z a a ) for upper triangular z. Forms with constant coefficients are harmonic for the flat metric, so the component of ev E ( ξ ) in H 2 ( a , a ) is the part of type ( 0 , 2 ) of ξ ⌟ exp c 1 ( L a ) , which is ev L a ( ξ ) , and v ⌟ c 1 ( L a ) for ξ = v .    □
Lemma  A9
(A Weil deformation moving every piece). In Proposition A4, let v 0 ∈ H 1 ( A , T A ) = Hom ( H 1 , 0 , H 0 , 1 ) send d z 1 ↦ d z ¯ n + 2 and d z n + 2 ↦ d z ¯ 1 and the other d z j to 0; in the usual notation v 0 = d z ¯ n + 2 ⊗ ∂ z 1 + d z ¯ 1 ⊗ ∂ z n + 2 . Then v 0 ∈ T D n , n , v 0 ⌟ θ = 0 , v 0 ⌟ c 1 ( M i ) = 0 , and for every ζ ∈ μ 4 n
v 0 ⌟ c 1 ( L ζ ) = i 2 ζ 1 d z ¯ n + 2 ∧ d z ¯ n + 1 + ζ ¯ 2 d z ¯ 1 ∧ d z ¯ 2 ≠ 0 .
Proof. 
The i-eigenspace V + of K on H 1 ( A , C ) is spanned by d z j and d z ¯ n + j , j ≤ n , and V − by d z n + j and d z ¯ j . The map v 0 sends d z 1 ∈ V + 1 , 0 to d z ¯ n + 2 ∈ V + 0 , 1 and d z n + 2 ∈ V − 1 , 0 to d z ¯ 1 ∈ V − 0 , 1 , so it is K-linear. It acts on H 2 as the derivation that replaces a factor d z 1 by d z ¯ n + 2 and a factor d z n + 2 by d z ¯ 1 (Setting 9). In the notation of the proof of Proposition A4, the only forms of c 1 ( L ζ ) = ∑ j [ p ( a j + b j ) + ζ j c j + ζ ¯ j c j ′ ] containing d z 1 or d z n + 2 are a 1 , c 1 , b 2 and c 2 ′ , and
v 0 ⌟ a 1 = i 2 d z ¯ n + 2 ∧ d z ¯ 1 , v 0 ⌟ b 2 = i 2 d z ¯ 1 ∧ d z ¯ n + 2 , v 0 ⌟ c 1 = i 2 d z ¯ n + 2 ∧ d z ¯ n + 1 , v 0 ⌟ c 2 ′ = i 2 d z ¯ 1 ∧ d z ¯ 2 .
The first two cancel; since θ = ∑ j ( a j + b j ) this gives v 0 ⌟ θ = 0 , hence v 0 ⌟ c 1 ( M i ) = t i v 0 ⌟ θ = 0 , and the displayed formula. Its two monomials are linearly independent and | ζ 1 | = 1 , so it is not zero. A K-linear first order deformation that kills θ is tangent to D n , n , by the proof of Theorem A6(ii).    □
Theorem  A9
(No convolution of the pieces meets the criterion). Let n ≥ 3 , let t 1 , t 2 , t 3 be pairwise distinct integers, and let E be a convolution of the bundles L ζ and M 1 , M 2 , M 3 of Proposition A4 on A = E 0 2 n , in any order, with any shifts of the given parities for the L ζ , any shifts of the M i and any twists of degree zero, and let γ = ch ( E ) . Then ev E ( v 0 ) ≠ 0 and σ E ( ev E ( v 0 ) ) = 0 , so
dim Ext 2 ( E , E ) ≥ r ( γ ) + 1 = 7 n 2 − 2 n + 1 ,
which is 58 at n = 3 and 105 at n = 4 . If G ⊆ A × A ∨ is a finite group acting on D perf ( A ) as in Corollary 11 and E carries a G-equivariant structure, then also dim Ext 2 ( E , E ) G ≥ r ( γ ) + 1 . In particular no such convolution meets the polarised criterion or the equivariant criterion, at E 0 6 or at E 0 8 .
Proof. 
The class γ = 2 · 4 n − 1 ( α + α ¯ ) + ∑ i σ i e t i θ , with σ i = ± 1 given by the shifts of the M i , has Hankel rank 3, so r ( γ ) = 7 n 2 − 2 n by Proposition A4(ii). By Theorem A8 no cycle of excess zero passes through any piece, so Lemma A8 applies to every piece, and with Lemma A9 the component of ev E ( v 0 ) at each L ζ is v 0 ⌟ c 1 ( L ζ ) ≠ 0 ; so ev E ( v 0 ) ≠ 0 . Since v 0 ∈ T D n , n ⊆ Ann H T 2 ( γ ) by Theorem A6(ii), (A4) gives σ E ( ev E ( v 0 ⌟ ) ) = v 0 γ = 0 . Let U = ev E ( H T 2 ( A ) ) . By (A4), σ E ( U ) is the image of the contraction with γ , of dimension r ( γ ) , and U ∩ ker σ E contains ev E ( v 0 ) ≠ 0 ; so dim Ext 2 ( E , E ) ≥ dim U ≥ r ( γ ) + 1 . If E is G-equivariant, U ⊆ Ext 2 ( E , E ) G by the proof of Corollary 11, and the same count applies inside Ext 2 ( E , E ) G . Equality in the criterion would make σ E injective on Ext 2 ( E , E ) , respectively on Ext 2 ( E , E ) G .    □
The numerical bounds of Theorems A12 and A13 to Corollary A3, of Section 6.10 and of Theorem A21 below are larger than r ( γ ) + 1 and remain valid where they apply; Theorem A9 covers every arrangement, every placement of the t i and every group G. The same argument gives a necessary condition for an arbitrary convolution of line bundles on an arbitrary abelian variety of Weil type.
Theorem  A10
(A necessary condition for line-bundle convolutions). Let A be an abelian 2 n -fold of Weil type for K, n ≥ 2 , with polarisation θ, and let E be a convolution of line bundles as in Setting 12, possibly G-equivariant as in Theorem A9, with ch ( E ) = γ = N ω + ∑ k c k θ k of polarised shape, N ≠ 0 . Then some piece has c 1 ( L a ) ∉ Q θ , and if one such piece lies on no cycle of excess zero, then dim Ext 2 ( E , E ) G ≥ r ( γ ) + 1 , so E does not meet the criterion or its equivariant form. Equivalently, in a convolution of line bundles meeting either criterion, every piece whose class is not a multiple of θ lies on a cycle of excess zero.
Proof. 
The classes α ± lie in ⋀ 2 n V ± and θ k in ⋀ k V + ⊗ ⋀ k V − , so ω ∉ Q [ θ ] ; if every c 1 ( L a ) were in Q θ , then γ = ∑ ± e c 1 ( L a ) would lie in Q [ θ ] . Let P a be a piece with c = c 1 ( L a ) ∉ Q θ on no cycle of excess zero, so that Δ a is defined and Δ a ( ev E ( v ) ) = v ⌟ c by Lemma A8. Suppose v ⌟ c = 0 for all v ∈ T D n , n . The polarisation identifies V − with V + ∨ , V − 1 , 0 with ( V + 0 , 1 ) ∨ and V − 0 , 1 with ( V + 1 , 0 ) ∨ , and a K-linear first order deformation kills θ exactly when its two components are X and − X ∨ ; so T D n , n is the block Hom ( V + 1 , 0 , V + 0 , 1 ) of sl ( V + ) , acting on H * ( A , C ) = ⋀ * ( V + ⊕ V + ∨ ) by derivations. The class c is real and the complex conjugates of these derivations again kill θ ; they form the opposite block Hom ( V + 0 , 1 , V + 1 , 0 ) , and they kill c too. The two off-diagonal blocks of sl ( V + ) for the decomposition V + = V + 1 , 0 ⊕ V + 0 , 1 generate sl ( V + ) , because their brackets span the trace-free diagonal blocks. So c is invariant under sl ( V + ) . In ⋀ 2 ( V + ⊕ V + ∨ ) = ⋀ 2 V + ⊕ ( V + ⊗ V + ∨ ) ⊕ ⋀ 2 V + ∨ the invariants are the line of the identity, spanned by θ , since ⋀ 2 V + is irreducible and nontrivial for dim V + = 2 n ≥ 4 . So c ∈ C θ ∩ H 2 ( A , Q ) = Q θ , a contradiction. Hence some v ∈ T D n , n has Δ a ( ev E ( v ) ) ≠ 0 , and the count of the proof of Theorem A9, ordinary and equivariant, applies to v with r ( γ ) the rank of the contraction with γ .    □
Remark  A3
(Cycles of length two). Let the pieces be external products over A = B 1 × ⋯ × B n , as in Lemma A6, and let H b − H a have i + positive and i − negative eigenvalues. By Lemma A7, H ( a , b ) lives in degrees [ i − , 2 n − i + ] and H ( b , a ) in degrees [ i + , 2 n − i − ] . A cycle of excess zero through exactly the two pieces P a , P b has Ext degrees q + q ′ = 2 with q ≥ i − , q ′ ≥ i + , so H b − H a has rank at most 2. If the difference is indefinite, q = q ′ = 1 , and degree one in both directions forces d a = d b : the two pieces enter ch ( E ) with the same sign. In the decomposition of Proposition A4 the differences of rank at most two occur only between L ζ and L ζ ′ with ζ , ζ ′ differing in one coordinate; they are indefinite, with eigenvalues ± | ζ j − ζ j ′ | , and the two pieces have opposite signs. So a line-bundle decomposition of a Weil class that can satisfy Theorem A10 needs pieces of equal sign whose forms differ by an indefinite form of rank two, pieces whose forms differ by a semidefinite form of rank at most two, or cycles of excess zero through at least three pieces.
Remark  A4
(Twisted deformations). In Theorem A9, ev E ( v 0 ) ≠ β · id E for every β ∈ H 2 ( A , O A ) : the diagonal components of β · id E are all equal to β, while those of ev E ( v 0 ) vanish at the M i and vary with ζ 1 at the L ζ . If the obstruction to deforming E along v 0 as a complex twisted by a Brauer class that vanishes at A, as in [98, Thm. 1.1], is the image of ev E ( v 0 ) in Ext 2 ( E , E ) / H 2 ( A , O A ) · id E (an identification not checked here), then these convolutions do not deform to first order along v 0 even as twisted complexes. Section A.5 extends Theorem A10 to convolutions of pieces of any kind.
Proposition  A6
(Cup products on the diagonal). Keep Proposition A5 with n = 3 , so that D 2 has dimension 35 · 15 = 525 . The terms of d E on D 2 with two leaves, the cup products, vanish on the 45 diagonal classes of the M i , and on the 480 diagonal classes of the L ζ they have a kernel of dimension at least 204. So their kernel on D 2 has dimension at least 249.
Proof. 
For y = ( y a ) ∈ D 2 the cup products in H ( a , b ) are ( y a − y b ) ∧ δ a b H , since y has even degree. On a surface where the pieces differ, δ a b H has degree 1, the only degree of the cohomology there, and the product with a component of y of positive degree lands in a degree where the cohomology vanishes. The pieces M i differ from every other piece on every surface, which gives the first statement. Write y a by its type τ ∈ { 0 , 1 , 2 } 3 , | τ | = 2 , the degrees on the three surfaces: three types ( 2 , 0 , 0 ) of dimension one and three types ( 1 , 1 , 0 ) of dimension four. The degree count on each surface and the parity of Ext degrees show that L ζ and L ζ ′ can carry a nonzero cup product on the type τ only if either ζ ′ differs from ζ by a sign in one coordinate outside the support of τ , or ζ ′ is ζ multiplied by ( i , i ) or ( − i , − i ) in two coordinates and τ is 2 on the third. Let Γ τ be the graph on the 32 pieces with these edges. If each component y τ is constant on the connected components of Γ τ , every cup product vanishes. For τ = ( 2 , 0 , 0 ) the coordinate ζ 1 is constant along the edges, and the eight pieces with a given ζ 1 are connected, so Γ τ has 4 components; for τ = ( 1 , 1 , 0 ) the edges change the sign of ζ 3 only, and Γ τ has 16 components. The kernel therefore contains a space of dimension 3 · 4 · 1 + 3 · 16 · 4 = 12 + 192 = 204 .    □
Theorem  A11
(The diagonal on E 0 6 ). Let n = 3 and let E be a convolution of the 32 bundles L ζ and of M 1 , M 2 , M 3 as in Proposition A4, with p + 2 ≤ t 1 < t 2 < t 3 and signs σ i , so that ch ( E ) = 32 ( α + α ¯ ) + ∑ i σ i e t i θ and the polarised criterion asks for dim Ext 2 ( E , E ) = 57 .
(i)
The Massey products of length at least three vanish on D 2 , except those of length four along a path M 2 → M 3 → L ζ → M 1 taken in this order in the convolution, which occur only if σ 1 = σ 3 = − σ 2 and take values in H 6 ( M 2 , M 1 ) ≅ H 6 A , O A ( − ( t 2 − t 1 ) θ ) , of dimension ( t 2 − t 1 ) 6 .
(ii)
dim Ext 2 ( E , E ) ≥ 249 − min 249 , ( t 2 − t 1 ) 6 , and dim Ext 2 ( E , E ) ≥ 249 unless the order and the signs are as in (i). In particular E does not meet the polarised criterion when t 2 − t 1 ≤ 2 , since 249 − 64 = 185 > 57 , nor when the signs or the order differ from (i).
(iii)
If E meets the polarised criterion, then the length four products of (i) have rank at least 249 − 57 = 192 on the kernel of the cup products, and t 2 − t 1 ≥ 3 .
The same holds for t 1 < t 2 < t 3 ≤ p − 2 with E replaced by its dual.
Proof. 
(i) By Lemma A5(iii) a term with k ≥ 3 leaves acting on y ∈ H 2 ( O A ) at a piece c runs along a chain s = a 0 → ⋯ → a k − 1 = e of distinct pieces, increasing in the order of the convolution, whose k − 1 steps are components of degree one, with y inserted at c. Its value lies in H ( s , e ) , in Ext degree 2 + ∑ j q j − ( k − 2 ) = 3 + ε , where ε is the sum of the excesses of the steps; so − 3 ≤ ε ≤ 3 , and H ( s , e ) must be nonzero in degree 3 + ε . By the proof of Proposition A5 a step between two L ζ has excess at least 1, with equality only between pieces of opposite parities, and a step between pieces at different values of ϕ has excess − 1 and Ext degree 0 going up, excess 5 and Ext degree 6 going down. Let U and D be the numbers of steps up and down and λ the total excess of the steps between two L ζ , so that ε = 5 D − U + λ . For two distinct L ζ , L ζ ′ differing in m ≥ 1 coordinates, H ( L ζ , L ζ ′ ) vanishes above degree 6 − m ; for pieces at different values it sits in degree 0 if the target is higher and in degree 6 if it is lower.
Two observations remove chains. First, a chain of total excess − 2 has no step down, and so consists of steps between L ζ followed by at most three steps up. Second, let a → b → c be two consecutive steps of Ext degree 0, sections x a b and x b c of line bundles. The Maurer–Cartan equation π ( δ ∘ δ ) = 0 of the proof of Proposition A8(i) has, in H 0 ( a , c ) , one term for each chain from a to c of total excess − 2 . If a → b → c is the only such chain, the equation reads x a b x b c = 0 , a product of holomorphic sections being holomorphic and so harmonic; since A is connected, x a b or x b c vanishes, and every term along a chain through both steps is zero.
Here every step up ends at one of M 1 , M 2 , M 3 , so U ≤ 3 , and D ≤ 1 . If D = 0 , the chain is r ≥ 0 pieces L ζ followed by multiples in increasing order. When it ends at an L ζ , k ≥ 3 gives λ ≥ 2 , so s and e differ in one coordinate and λ = 2 ; then the two steps change the parity twice, whereas s and e have opposite parities. When it ends at a multiple, e lies above s, so ε = − 3 , which forces U = 3 , λ = 0 and the chain L ζ → M 1 → M 2 → M 3 ; by the first observation L ζ → M 1 → M 2 is the only chain of excess − 2 from L ζ to M 2 , and the second removes it. If D = 1 , then U ≥ 2 + λ , so ( U , λ ) is ( 2 , 0 ) , ( 3 , 0 ) or ( 3 , 1 ) and ε is 3, 2 or 3. For ε = 3 the piece s is a multiple and e lies below it. With ( 3 , 1 ) the three steps up would end at three multiples other than s, which do not exist. With ( 2 , 0 ) the chain has three steps, and a step down lands at an L ζ , since a fourth multiple does not exist. It is M 1 → M 2 → M 3 → L ζ , removed by the second observation at M 1 → M 2 → M 3 ; or M a → M b → L ζ → M c with t a < t b and t c < t a , that is M 2 → M 3 → L ζ → M 1 ; or M a → L ζ → M b → M c with t b < t c < t a , that is M 3 → L ζ → M 1 → M 2 , removed at L ζ → M 1 → M 2 . For ε = 2 , with ( U , λ ) = ( 3 , 0 ) , the pieces s and e are L ζ differing in one coordinate; the last step is the step down, preceded by three steps up from s, so the chain is L ζ → M 1 → M 2 → M 3 → L ζ ′ , removed at L ζ → M 1 → M 2 .
Only M 2 → M 3 → L ζ → M 1 remains. Its steps have Ext degrees 0, 6 and 0, so degree one forces d M 3 = d M 2 + 1 , d L = d M 3 − 5 and d M 1 = d L + 1 ; this gives the signs, and the value lies in H 3 + d M 2 − d M 1 ( M 2 , M 1 ) = H 6 ( M 2 , M 1 ) , whose dimension is | χ ( O A ( ( t 1 − t 2 ) θ ) ) | = ( t 2 − t 1 ) 6 . (ii) By Proposition A5 it suffices to bound the kernel of d E on D 2 . By Proposition A6 the cup products have a kernel Y of dimension at least 249, and on Y the differential is the sum of the products of (i), of rank at most ( t 2 − t 1 ) 6 . (iii) follows from (ii). For the dual, E ↦ E ∨ preserves Ext 2 and sends the configuration with p, t i to the one with − p , − t i , the set of the L ζ being stable under ζ ↦ − ζ .    □
The placements of the t i on both sides of p are not covered by Theorem A11. There two families of fourfold products occur instead of one, and no convolution carries both.
Proposition  A7
(Placements on both sides of p). Keep Theorem A11, but let t 1 ≤ p − 2 and p + 2 ≤ t 2 < t 3 .
(i)
The Massey products of length at least three vanish on D 2 unless σ 1 = σ 2 = − σ 3 . In that case the only ones that act have length four and run along one of the paths
( a ) M 2 → M 3 → M 1 → L ζ , ( b ) M 3 → M 1 → L ζ → M 2 ,
taken in this order in the convolution, with ∏ ζ = σ 3 . The products along (a) take values in H 6 ( M 2 , L ζ ) , of dimension ( t 2 − p ) 2 − 1 3 , and those along (b) in H 6 ( M 3 , M 2 ) , of dimension ( t 3 − t 2 ) 6 . A convolution carries at most one of the two kinds, since (a) puts M 2 before M 3 and (b) puts M 3 before M 2 .
(ii)
If E meets the polarised criterion, then σ 1 = σ 2 = − σ 3 , and the products along (a) or those along (b) have rank at least 192 on the kernel of the cup products. For (b) this needs t 3 − t 2 ≥ 3 . The sixteen targets of (a) have dimension 16 ( t 2 − p ) 2 − 1 3 ≥ 432 , so for (a) the count gives no condition.
For t 1 < t 2 ≤ p − 2 and t 3 ≥ p + 2 the same holds for E ∨ [ 1 ] , and the condition on the signs becomes σ 2 = σ 3 = − σ 1 .
Proof. 
(i) We use the notation and the two observations of the proof of Theorem A11(i). Now ϕ ( M 1 ) = t 1 < p < t 2 < t 3 ; a step up ends at M 2 , at M 3 , or at an L ζ coming from M 1 , so again U ≤ 3 and D ≤ 1 . If D = 0 , the values do not decrease along the chain. When e lies above s, ε = − 3 forces U = 3 , λ = 0 and the chain M 1 → L ζ → M 2 → M 3 , removed at L ζ → M 2 → M 3 , the only chain of excess − 2 from L ζ to M 3 . When e and s have the same value, they are two L ζ joined by steps between L ζ , which is excluded as before. If D = 1 , ( U , λ ) is ( 2 , 0 ) , ( 3 , 0 ) or ( 3 , 1 ) . For ε = 3 , e lies below s. With ( 3 , 1 ) the steps up end at M 2 , M 3 and an L ζ following M 1 ; then s is an L ζ , so e = M 1 , which is followed by a step: impossible. With ( 2 , 0 ) the chain has three steps and no two consecutive L ζ . As M 1 is the lowest piece, it is never the end of a step up, and a step up into an L ζ comes from M 1 ; so two consecutive steps up end at M 2 or M 3 . The chain cannot start at M 1 . If it starts at an L ζ , it ends at M 1 with the step down, after two steps up: L ζ → M 2 → M 3 → M 1 , removed at L ζ → M 2 → M 3 . If it starts at M 3 , the first step goes down and the last two go up, ending at M 2 ; the middle piece is then an L ζ entered from M 1 , and the chain is M 3 → M 1 → L ζ → M 2 , path (b). If it starts at M 2 , then e ∈ { L ζ , M 1 } , so the last two steps are not both up; the first step goes up to M 3 , the second down, and the third up into e = L ζ from M 1 : the chain is M 2 → M 3 → M 1 → L ζ , path (a). For ε = 2 , ( U , λ ) = ( 3 , 0 ) and s, e are L ζ differing in one coordinate. The last step enters e from M 1 , since a step down into e would follow three steps up from s, one more than the values allow; M 1 is entered by the step down, after two steps up from s: L ζ → M 2 → M 3 → M 1 → L ζ ′ , removed at L ζ → M 2 → M 3 . Along (a) the steps have Ext degrees 0, 6 and 0, so degree one forces d M 3 = d M 2 + 1 , d M 1 = d M 3 − 5 and d L = d M 1 + 1 . Along (b) they have Ext degrees 6, 0 and 0, which forces d M 1 = d M 3 − 5 , d L = d M 1 + 1 and d M 2 = d L + 1 . This gives the signs and puts the values in H 3 + d M 2 − d L ( M 2 , L ζ ) = H 6 ( M 2 , L ζ ) and in H 3 + d M 3 − d M 2 ( M 3 , M 2 ) = H 6 ( M 3 , M 2 ) . On B j the line bundle L ζ ⊗ M 2 − 1 has the hermitian form p − t 2 ζ j ζ ¯ j p − t 2 , with eigenvalues p − t 2 ± 1 < 0 and determinant ( t 2 − p ) 2 − 1 , which gives the first dimension. The second is that of Theorem A11(i) with t 3 , t 2 in place of t 2 , t 1 . (ii) Proposition A5 and Proposition A6 do not depend on the placement of the t i , so as in Theorem A11(ii) dim Ext 2 ( E , E ) ≥ 249 − ρ , where ρ is the rank on the kernel of the cup products of the products of (i) that E carries, and ρ = 0 unless σ 1 = σ 2 = − σ 3 . The criterion asks for 57, so ρ ≥ 192 , and ( t 3 − t 2 ) 6 ≥ 192 exactly when t 3 − t 2 ≥ 3 . For the last statement, E ↦ E ∨ [ 1 ] preserves Ext 2 , sends p and the t i to − p and the − t i , reverses the order of the pieces and changes every sign.    □
The Maurer–Cartan equation can be solved for one order of the pieces.
Proposition  A8
(An explicit convolution). Keep Theorem A11 with p = 0 and t = ( 2 , 5 , 6 ) , take the pieces to be external products L ζ = N 1 , ζ 1 ⊠ N 2 , ζ 2 ⊠ N 3 , ζ 3 and M i = M i , 1 ⊠ M i , 2 ⊠ M i , 3 of line bundles on the B j , and order them as M 2 , M 3 , the sixteen L ζ with ∏ ζ = 1 , M 1 , the sixteen L ζ with ∏ ζ = − 1 , with the shifts 0, 1, − 4 , − 3 , − 5 .
(i)
The components that the degrees allow are x 1 ∈ H 0 ( M 2 , M 3 ) , x 2 ζ ∈ H 6 ( M 3 , L ζ ) , x 3 ζ ∈ H 0 ( L ζ , M 1 ) , components in H 2 ( L ζ , L ζ ′ ) for the 48 + 96 pairs with ∏ ζ = 1 = − ∏ ζ ′ that differ in one or two coordinates, and components in H 6 ( M 2 , L ζ ′ ) , ∏ ζ ′ = − 1 , which we take to be zero. For harmonic components the Maurer–Cartan equation is exactly
x 1 x 2 ζ = 0 in H 6 ( A , M 2 − 1 ⊗ L ζ ) for every ζ , ∑ ζ x 2 ζ x 3 ζ = 0 in H 6 ( A , M 3 − 1 ⊗ M 1 ) ,
and every solution is the harmonic part of a convolution E with ch ( E ) = 32 ( α + α ¯ ) + e 5 θ − e 6 θ − e 2 θ .
(ii)
There are solutions with every x 2 ζ ≠ 0 and x 1 x 2 ζ = x 2 ζ x 3 ζ = 0 .
(iii)
For every solution dim Ext 2 ( E , E ) ≥ 2560 , so E does not meet the polarised criterion.
Proof. 
(i) A component from P a to P b has Ext degree 1 + d a − d b , and the forms give the cohomology of each pair: H ( M 2 , M 3 ) and H ( L ζ , M 1 ) sit in degree 0, H ( M 3 , L ζ ) , H ( M 2 , L ζ ′ ) and H ( M 2 , M 1 ) in degree 6, and H ( L ζ , L ζ ′ ) in degree 2 exactly for the pairs listed; no other pair has cohomology in the degree a component needs. Put δ = δ H − h ( δ ∘ δ ) recursively, as in Lemma A5(i), and Φ = ∂ ¯ δ + δ ∘ δ . Then Φ = π ( δ ∘ δ ) + h ∂ ¯ ( δ ∘ δ ) and ∂ ¯ ( δ ∘ δ ) = Φ ∘ δ − δ ∘ Φ , so by induction on b − a the twisted complex exists exactly when π ( δ ∘ δ ) = 0 . The components of π ( δ ∘ δ ) come from paths of two or more components; those that land in a nonzero group are x 1 x 2 ζ in H 6 ( M 2 , L ζ ) and ∑ ζ x 2 ζ x 3 ζ in H 6 ( M 3 , M 1 ) , while the product along M 2 → M 3 → L ζ → M 1 would lie in H 5 ( M 2 , M 1 ) = 0 and x 2 ζ composed with a component in H 2 ( L ζ , L ζ ′ ) in degree eight. The signs ( − 1 ) d are + , − , + , − , − , which gives the Chern character by Proposition A4(i). (ii) Take x 1 and the x 3 ζ to be external products of nonzero sections on the B j . On each B j the products with their factors map H 2 ( B j , M 3 − 1 L ζ ) , of dimension 35, to spaces of dimensions 24 and 16, so they have kernels of dimensions at least 11 and 19. A sum of external products whose factor on one surface lies in the first kernel and whose factor on another lies in the second is a nonzero x 2 ζ with x 1 x 2 ζ = x 2 ζ x 3 ζ = 0 , which solves (i).15 (iii) The sixteen L ζ with ∏ ζ = 1 have shift − 4 and the others − 5 , so Lemma A10 below applies to every pair of them that differ in all three coordinates: there are 112 such pairs, and their groups H 3 have dimensions adding up to 16 · 160 = 2560 .    □
The classes counted in (iii) are the ones no product can reach. Two pieces L ζ , L ζ ′ that differ in every coordinate have a difference that is indefinite on every surface, so H ( L ζ , L ζ ′ ) is ⨂ j H 1 ( B j , · ) , concentrated in degree three, of dimension D ( ζ , ζ ′ ) = ∏ j | ζ j − ζ j ′ | 2 ; no component can join them, and if their shifts differ by one their classes have degree two.
Lemma  A10
(Pieces that differ everywhere). Keep Proposition A5 with n = 3 . Suppose that for some integer r either p + 2 ≤ t 1 < t 2 < t 3 and M 1 , M 2 , M 3 have the shifts r + 1 , r + 4 , r + 5 , or t 1 ≤ p − 2 < p + 2 ≤ t 2 < t 3 and they have the shifts r − 1 , r + 3 , r + 4 or r − 1 , r + 1 , r + 4 ; these are the shifts forced by the paths of Theorem A11(i) and of Proposition A7(i), (a) and (b), with r the shift of the L ζ on the path. Let L ζ , L ζ ′ differ in all three coordinates, with d L ζ = d L ζ ′ + 1 . Then d E vanishes on H 3 ( L ζ , L ζ ′ ) , and no element of degree one has a differential with a nonzero component there. Consequently
dim Ext 2 ( E , E ) ≥ dim ker d E | D 2 + ∑ D ( ζ , ζ ′ ) ,
the sum running over the pairs of this kind.
Proof. 
Only the shifts are used, and the order of the pieces only removes components, so we ignore it. A component from P a to P b of Ext degree q has d b = d a + 1 − q . Between distinct L ζ the degree q is at least 2: there is no H 0 , and H 1 occurs only for pieces differing in one coordinate, which have opposite parities, while q = 1 would need equal shifts. So a component between two L ζ lowers the shift by at least one, and by exactly one only in degree two. For a multiple M i let μ i and ν i be the shifts of an L ζ that a component from M i reaches and from which a component reaches M i : μ i = d M i − 5 and ν i = d M i − 1 when t i > p , and μ i = d M i + 1 , ν i = d M i + 5 when t i < p . Between the multiples the degrees allow only M 2 → M 3 in the first case, M 2 → M 3 → M 1 in the second, and M 3 → M 1 in the third. Let λ i be the largest μ k and κ i the least ν k over the multiples M k reachable from M i , respectively reaching M i , through these steps, M i included. The values, in the order M 1 , M 2 , M 3 and relative to r, are
Preprints 244139 i001
and in all three cases, for all i and k,
λ i ≤ r ≤ κ k , κ i ≥ λ i + 2 , λ i ≤ μ i + 2 , ν i ≤ κ i + 2 ,
while μ i ≤ λ i and ν i ≥ κ i by definition. A run L → M k → ⋯ → M i → L ′ through multiples goes from the shift ν k ≥ κ k to μ i ≤ λ k ≤ κ k − 2 , so every chain of components from one L to another lowers the shift, by at least two if it passes through a multiple and by exactly one only if it is a single component of degree two. In particular a chain from an L of shift x into M i has x ≥ κ i , and a chain from M i to an L of shift y has y ≤ λ i .
By Lemma A5(iii), a term of d E on a class u of Ext degree q u in H ( a ′ , b ′ ) runs along a chain c → ⋯ → a ′ , the step u, and a chain b ′ → ⋯ → e , and since the degrees of the components telescope, its value lies in H ( c , e ) in Ext degree
q u + 1 + ( d c − d a ′ ) + ( d b ′ − d e ) .
Write a = L ζ and b = L ζ ′ , so that d a = d b + 1 .
Nothing reaches H 3 ( a , b ) . Let u have degree one, so q u = 1 + d a ′ − d b ′ , with chains from a to a ′ and from b ′ to b and ( a ′ , b ′ ) ≠ ( a , b ) . If a ′ and b ′ are both pieces L, then d a ′ ≤ d a and d b ′ ≥ d b , not both with equality, so q u ≤ 1 ; q u = 0 is impossible, and q u = 1 needs either a ′ = b ′ , and then a chain from a to b lowering the shift by one, a single component in H 2 ( a , b ) = 0 , or a ′ , b ′ differing in one coordinate, of opposite parities, against d a ′ = d b ′ . If a ′ = M i and b ′ is an L, then u lies in the degree of a component, so d b ′ = μ i , while d b ′ ≥ d b = d a − 1 ≥ κ i − 1 ≥ λ i + 1 > μ i . If a ′ is an L and b ′ = M k , then d a ′ = ν k and d a − d b ≥ ν k − λ k ≥ κ k − λ k ≥ 2 . If a ′ = M i and b ′ = M k , then d a − d b ≥ κ i − λ k ; for i = k this is at least 2, and for i ≠ k the class u lies in the degree of a step M i → M k , so λ k ≤ λ i and again d a − d b ≥ 2 . Each case contradicts d a − d b = 1 .
d E vanishes on H 3 ( a , b ) . Here u is the class itself and a term lands in H ( c , e ) , ( c , e ) ≠ ( a , b ) , in degree 3 + d c − d e , with chains from c to a and from b to e. If c and e are both pieces L, then d c − d e ≥ 2 ; c = e is impossible, and otherwise a class of degree at least 5 needs c, e differing in one coordinate, in degree 5, so d c − d e = 2 against their opposite parities. If c = M i and e is an L, the degree is that of a component when d e = μ i + 2 , but d e ≤ d b < d a ≤ λ i ≤ μ i + 2 . If c is an L and e = M k , it is that of a component when d c = ν k − 2 , but d c ≥ d a > d b ≥ κ k ≥ ν k − 2 . If c = M i and e = M k , then d a ≤ λ i ≤ r ≤ κ k ≤ d b , against d a > d b .
For the inequality: a coboundary has no component on D 2 (Proposition A5) and none on these groups, so the cocycles of ker ( d E | D 2 ) and of these groups span a space that meets the coboundaries only in zero. Twisting the pieces by line bundles of degree zero does not change D ( ζ , ζ ′ ) , the forms being nondegenerate on every surface, and can only remove components.    □
Theorem  A12
(No convolution of these pieces meets the criterion). Keep Proposition A5 with n = 3 : let E be any convolution of the 32 bundles L ζ and of M 1 , M 2 , M 3 , with t 1 < t 2 < t 3 and | t i − p | ≥ 2 , in any order, with any shifts of the given parities and with the pieces twisted by any line bundles of degree zero. Then dim Ext 2 ( E , E ) ≥ 69 . In particular no such convolution meets the polarised criterion, which asks for 57.
Proof. 
Replacing E by E ∨ or E ∨ [ 1 ] , which preserves Ext 2 , we may assume that the t i are as in Theorem A11 or in Proposition A7. If no fourfold product of those results acts, dim Ext 2 ( E , E ) ≥ 249 . Otherwise the path fixes the shifts of the M i as in Lemma A10, with r the shift of every L ζ on a path; let P be the sixteen pieces of the parity of r, Q the other sixteen, and m the number of pieces on paths.
The fourfold products act on D 2 only through the classes at the M i and at the m pieces on paths, so their rank is at most 45 + 15 m , and by Proposition A6 dim ker ( d E | D 2 ) ≥ 249 − 45 − 15 m = 204 − 15 m . If m ≤ 9 this is at least 69.
Let m ≥ 10 , so that at least ten pieces of P have the shift r. Every W ∈ Q has 3, 6 and 7 partners in P differing from it in one, two and three coordinates, and for the seven D = 16 six times and D = 64 once.16 If d W = r ± 1 , at most six pieces of P lack the shift r, so one of the seven has it, and Lemma A10 gives at least 16 classes. If d W ≠ r ± 1 , consider for each type τ of the form ( 1 , 1 , 0 ) the partner W τ ∈ P obtained by changing the sign of the coordinate where τ vanishes. By the proof of Proposition A6, W τ is the only piece with which W can carry a nonzero cup product on the type τ , and conversely; a cup product between them needs a component of degree two, hence shifts differing by one. If they do, the classes of type τ equal at W and W τ and zero elsewhere have vanishing cup products, and W τ , of shift d W ± 1 ≠ r , lies on no path; if they do not, the classes of type τ at W alone have vanishing cup products. Either way these 4 classes lie in ker ( d E | D 2 ) , 12 for each such W, and those of different W are independent, since W ↦ W τ is injective. Adding over Q, Lemma A10 gives dim Ext 2 ( E , E ) ≥ 16 · 12 = 192 .    □
At n = 4 there are 128 pieces L ζ , and the classes that no product reaches are again those between pieces that differ in three coordinates and have adjacent shifts. The shifts along a chain are now controlled by counting alone.
Lemma  A11
(Shifts along chains at n = 4 ). Keep Proposition A4 with n = 4 , take the L ζ to be external products N 1 , ζ 1 ⊠ ⋯ ⊠ N 4 , ζ 4 of line bundles on the B j , let t 1 , t 2 , t 3 be distinct with | t i − p | ≥ 2 , and let E be a convolution of the L ζ and of M 1 , M 2 , M 3 with shifts of the given parities. Call a component of δ H , or an element of degree one of H E in one group H ( a , b ) , a step from P a to P b ; a step of Ext degree q has d b = d a + 1 − q . Put ϕ ( L ζ ) = p and ϕ ( M i ) = t i .
(i)
A step between distinct L ζ has Ext degree at least two, so it lowers the shift by at least one, and by exactly one only in Ext degree two. A step in H 1 ( a , a ) = H 1 ( A , O A ) keeps the shift.
(ii)
A step between an L ζ and a multiple, or between two multiples, has Ext degree 0 if it goes up the value ϕ and 8 if it goes down, so it raises the shift by one or lowers it by seven.
(iii)
A sequence of steps from an L ζ to an L ζ ′ through multiples, distinct before one marked step and distinct after it, where the marked step may lie in H 1 ( a , a ) at a multiple, lowers the shift by at least four. Consequently a sequence of steps from one L ζ to another, at most one of them not a component, lowers the shift by at least one, and by exactly one only if it consists of steps between pieces L ζ , one of Ext degree two and the others in H 1 ( a , a ) .
Proof. 
(i) For ζ ≠ ζ ′ the group H ( L ζ , L ζ ′ ) is the tensor product over the surfaces of H 1 of a line bundle with a nondegenerate form, on the k ≥ 1 surfaces where ζ and ζ ′ differ, and of H * ( B j , O B j ) on the others, the factors of L ζ and L ζ ′ on those being equal. So its degrees are at least k, and degree one needs k = 1 . Then the ratio ζ j ′ / ζ j ≠ 1 in the one coordinate is the ratio of ∏ ζ ′ and ∏ ζ , which is − 1 , and the shifts have opposite parities, while a step of Ext degree one keeps the shift. (ii) The forms of M i ⊗ L ζ − 1 have the eigenvalues t i − p ± 1 on every surface, which have the sign of t i − p ; the form of M k ⊗ M i − 1 is ( t k − t i ) Id . A definite form has cohomology in degree 0 if it is positive and 2 n = 8 if it is negative [19, Chapter 3].17 (iii) Let U and D be the numbers of steps up and down outside H 1 ( a , a ) ; by (ii) the shift changes by U − 7 D . The values start and end at p and pass through the values of multiples, so D ≥ 1 . If D = 1 , the values increase strictly before the step down and strictly after it, so those before it lie above p and those after it below p, each multiple occurs once, and U ≤ 3 : the drop 7 − U is at least four.18 If D ≥ 2 , at most six multiples occur, counted with repetition, so there are at most seven steps outside H 1 ( a , a ) , U ≤ 7 − D , and the drop 7 D − U ≥ 8 D − 7 is at least nine. For the consequence, split the sequence at the pieces L ζ it visits: each part through multiples lowers the shift by at least four, and by (i) each other part lowers it by at least one, unless it lies in H 1 ( a , a ) .    □
Theorem  A13
(Two shifts on E 0 8 ). Keep Lemma A11, and let L ζ and L ζ ′ differ in exactly three coordinates and have the shifts d L ζ = d L ζ ′ + 1 . Then H ( L ζ , L ζ ′ ) has its classes of degree two in H 3 ( L ζ , L ζ ′ ) , of dimension D ( ζ , ζ ′ ) = ∏ ζ j ≠ ζ j ′ | ζ j − ζ j ′ | 2 , which is 16 or 64, and:
(i)
for every convolution, no element of degree one has a differential with a nonzero component in H 3 ( L ζ , L ζ ′ ) ;
(ii)
if the L ζ occupy two consecutive shifts, those with ∏ ζ = ϵ the shift s and the others s − 1 , for a sign ϵ and an integer s, then d E vanishes on these groups, and for every order of the pieces, every choice of the shifts of the M i and every placement of the t i ,
dim Ext 2 ( E , E ) ≥ 64 · 640 = 40960 .
In particular E does not meet the polarised criterion, which asks for dim Ext 2 ( E , E ) = 104 .
Proof. 
Write W = L ζ and V = L ζ ′ . By the proof of Lemma A11(i), H ( W , V ) has degrees 3, 4 and 5, the last two from the surface where ζ and ζ ′ agree, and its classes of degree w = 2 have Ext degree 2 + d W − d V = 3 ; the dimension of H 3 ( W , V ) is the product over the three other surfaces of the determinants | ζ j − ζ j ′ | 2 , as in Lemma A10.
(i) As in the proof of Lemma A10, a term of the differential of an element u of degree one in H ( a ′ , b ′ ) with a component in H ( W , V ) runs along a chain of components W → ⋯ → a ′ , the step u, and a chain b ′ → ⋯ → V , with ( a ′ , b ′ ) ≠ ( W , V ) . This is a sequence of steps from W to V, one of them u, lowering the shift by one. By Lemma A11(iii) it consists of steps between pieces L ζ , one of Ext degree two and at most one, u, in H 1 ( a , a ) . If u is not in H 1 ( a , a ) , it is the step of degree two and both chains are empty, against ( a ′ , b ′ ) ≠ ( W , V ) . If u is in H 1 ( a , a ) , the rest is a single component W → V of Ext degree two, in H 2 ( W , V ) = 0 .
(ii) A term of d E on x ∈ H 3 ( W , V ) runs along chains of components c → ⋯ → W and V → ⋯ → e , not both empty, and lands in H ( c , e ) in Ext degree 3 + 1 + ( d c − d W ) + ( d V − d e ) = 3 + d c − d e . No piece L ζ ′ ′ other than W lies on the first chain: by Lemma A11(iii) it would have the shift at least d W + 1 = s + 1 . Likewise only multiples follow V on the second chain, and each chain has at most three pieces besides W or V. Let U and D be the numbers of steps up and down along both chains, so that U + D ≤ 6 ; by Lemma A11(ii), d W = d c + U 1 − 7 D 1 and d e = d V + U 2 − 7 D 2 , and the Ext degree is
3 + d c − d e = 4 − U + 7 D .
If c = e , this degree is 3, so U = 7 D + 1 , and D ≥ 1 because the values return to ϕ ( c ) ≠ p ; then U ≥ 8 , which is impossible. If c ≠ e , the group H ( c , e ) sits in degree 0 if ϕ ( e ) > ϕ ( c ) and in degree 8 if ϕ ( e ) < ϕ ( c ) , by Lemma A11(ii). Degree 0 needs U = 4 + 7 D ≥ 4 with D = 0 , so the values increase strictly, the multiples are distinct and U ≤ 3 . Degree 8 needs 7 D = U + 4 , so D = 1 and U = 3 ; the values increase strictly before the step down, from ϕ ( c ) , and after it, up to ϕ ( e ) < ϕ ( c ) , so the five values along the path, with W and V counted once, are distinct, and four of them are values of multiples, of which there are three. So no term reaches a nonzero group, and d E vanishes on H 3 ( W , V ) .
By (i) and (ii) the classes of these groups are cocycles of degree two, and no nonzero combination of them is a coboundary, since every coboundary has zero components there. Each piece of the shift s has 28 partners of the other parity differing from it in exactly three coordinates,19 all of the shift s − 1 , and the 28 groups have dimensions adding up to 640; there are 64 pieces of the shift s.    □
The hypothesis of (ii) is the arrangement of the explicit convolution of Proposition A8 at n = 3 , where the sixteen pieces of each parity share a shift. The pieces must not be twisted by line bundles of degree zero here, unlike in Theorem A12: a twist that is nontrivial on the surface where ζ and ζ ′ agree removes H ( L ζ , L ζ ′ ) .
With a third shift, a component of Ext degree two can follow or precede a class of H 3 ( L ζ , L ζ ′ ) , so the differential no longer vanishes on these classes. When the three shifts are consecutive, such terms reach only the groups between the highest and the lowest shift, which are too small to remove the classes.
Theorem  A14
(Three consecutive shifts on E 0 8 ). Keep Lemma A11, and suppose that the shifts of the L ζ lie in { s , s − 1 , s − 2 } for an integer s. By their parities, for a sign ϵ, the pieces with ∏ ζ = ϵ have the shift s − 1 and the others the shift s or s − 2 ; let T and B be the sets of pieces of the shifts s and s − 2 . Then, for every order of the pieces, every choice of the shifts of the M i and every placement of the t i :
(i)
every two pieces L ζ , L ζ ′ that differ in exactly three coordinates have shifts differing by one, and the groups H 3 ( L ζ , L ζ ′ ) of Theorem A13, taken from the piece of the larger shift, have dimensions adding up to 40960;
(ii)
d E maps the sum of these groups into ⨁ c ∈ T , e ∈ B H 5 ( c , e ) , and on them it consists of cup products with a single component of Ext degree two;
(iii)
dim Ext 2 ( E , E ) ≥ 40960 − ∑ c ∈ T , e ∈ B dim H 5 ( c , e ) ≥ 40960 − 18432 = 22528 .
In particular E does not meet the polarised criterion. When T or B is empty this is Theorem A13(ii).
Proof. 
(i) The ratios ζ j ′ / ζ j in the three coordinates have product − 1 , so the two pieces have opposite parities: one has the shift s − 1 and the other s or s − 2 . Each of the 64 pieces of the shift s − 1 has 28 such partners, whose groups have dimensions adding up to 640, as in the proof of Theorem A13; this gives 64 · 640 = 40960 .
(ii) Let x ∈ H 3 ( a , b ) with d a = d b + 1 . As in the proof of Theorem A13(ii), a term of d E ( x ) runs along chains of components c → ⋯ → a and b → ⋯ → e , not both empty, and lands in H ( c , e ) in Ext degree 3 + d c − d e . The shifts of the pieces L ζ differ by at most two, so by Lemma A11(iii) no part of a chain between two pieces L ζ passes through a multiple. Hence a chain c → ⋯ → a that starts at a multiple M i is a run M i = m 0 → ⋯ → m r → L 1 of distinct multiples, r ≤ 2 , followed by pieces L ζ only; let U 1 and D 1 be the numbers of its steps up and down the value ϕ before L 1 , so that d L 1 = d M i + U 1 − 7 D 1 by Lemma A11(ii) and U 1 + D 1 = r + 1 ≤ 3 . Likewise a chain b → ⋯ → e ending at a multiple M k ends with a run from a piece L 2 with U 2 steps up and D 2 down, U 2 + D 2 ≤ 3 and d M k = d L 2 + U 2 − 7 D 2 . Put L 1 = c if c is a piece L ζ and L 2 = e if e is one, and δ = d L 1 − d L 2 ; then 1 = d a − d b ≤ δ ≤ 2 .
If c and e are both pieces L ζ , each nonempty chain lowers the shift by at least one, so 3 + d c − d e ≥ 5 , while δ ≤ 2 gives at most 5. So the degree is 5, c ∈ T , e ∈ B , and the two chains together consist of one component, of Ext degree two by Lemma A11(i): the term is a cup product of x with that component.
If c = M i and e is a piece L ζ , the degree is 3 + δ − U 1 + 7 D 1 , and H ( c , e ) sits in degree 0 or 8 by Lemma A11(ii). Degree 0 needs U 1 ≥ 3 + δ ≥ 4 , and degree 8 needs 7 D 1 − U 1 = 5 − δ ∈ { 3 , 4 } , so D 1 ≥ 1 and U 1 ≥ 3 ; both contradict U 1 + D 1 ≤ 3 . The case of a piece c and a multiple e is the same with the second run.
If c = M i and e = M k , the degree is 3 + δ − U + 7 D with U = U 1 + U 2 ≤ 6 and D = D 1 + D 2 . For i = k it is 3, so U − 7 D = δ ≥ 1 ; D = 0 is impossible, since the values would increase from t i to p along the first run and from p to t i along the second, and D ≥ 1 gives U ≥ 8 . For i ≠ k , H ( c , e ) sits in degree 0 if t k > t i and 8 if t k < t i . Degree 0 needs U = 3 + δ + 7 D , so D = 0 and U ≥ 4 ; but then the multiples of the first run have values below p and those of the second values above p, so they are distinct and U ≤ 3 . Degree 8 needs 7 D − U = 5 − δ , so D = 1 and U ≥ 3 : the runs have at least four steps, so some multiple lies on both. One run increases strictly. If it is the second, its values lie in ( p , t k ] , while those of the first lie at or above t i > t k before its step down and below p after it. If it is the first, its values lie in [ t i , p ) , while those of the second lie above p before its step down and at or below t k < t i after it. Either way no multiple lies on both runs, a contradiction.
(iii) By Theorem A13(i), which holds for every convolution, no coboundary has a component in these groups, so the kernel of d E on their sum injects into Ext 2 ( E , E ) , and by (ii) its dimension is at least 40960 minus the dimension of the targets. For pieces c, e of the same parity, H 5 ( c , e ) is nonzero only when they differ in two or three coordinates, and each piece has such partners with groups H 5 of dimensions adding up to 960.20 So the sum in (iii) is at most min ( | T | , | B | ) · 960 ≤ 32 · 960 = 30720 , which already gives dim Ext 2 ( E , E ) ≥ 10240 . The dimension of H 5 ( c , e ) depends only on the ratio e / c , so the sum is the weight of the cut ( T , B ) of a weighted Cayley graph on the group of ratios of two pieces of one parity, of order 64 and degree 960. For the vector x = ± 1 of the cut, the weight is 1 4 64 · 960 − x T A x ≤ 16 ( 960 − λ ) , with A the weighted adjacency matrix and λ its least eigenvalue. Its eigenvalues are the character sums ∑ g dim H 5 ( 1 , g ) χ ( g ) . A character is χ ( g ) = ∏ j g j k j with k ∈ ( Z / 4 ) 4 , up to adding ( 1 , 1 , 1 , 1 ) ; put ω j = i k j , W = ∑ j ω j , ϱ = ∑ j ω j 2 and Ω = ∏ j ω j . By the list in the footnote, the ratios that move the pair of coordinates { j , l } contribute 64 ω j 2 ω l 2 + 32 ℜ ( ω j ω ¯ l ) , and those that fix only the coordinate o contribute 32 ℜ Ω ω ¯ o ( W − ω o ) . Summing,
χ ∑ g dim H 5 ( 1 , g ) g = 32 ϱ 2 + 16 | W | 2 − 192 + 32 ℜ ( Ω ) | W | 2 − 4 .
If ℜ Ω = 0 , the sum of the k j is odd, so an odd number of them are odd, | ϱ | = 2 , and the value is at least − 64 . If ℜ Ω = 1 , an even number are odd: with none or four, | ϱ | = 4 and the value is at least 192; with two, ϱ = 0 , and W = 0 would force the k j to be 0 , 1 , 2 , 3 , whose sum is not divisible by 4, so | W | 2 ≥ 4 and the value is at least − 128 . If ℜ Ω = − 1 , the value is 32 ϱ 2 − 16 | W | 2 − 64 , and | W | 2 ≤ 8 + 2 ϱ 2 : when all k j have the same parity, | W | ≤ 4 = | ϱ | , and when two are odd, ϱ = 0 and W is a sum of two of ± 1 and two of ± i , so | W | 2 ≤ 8 . Hence λ ≥ − 192 ,21 and the sum is at most 16 · 1152 = 18432 .    □
The groups of Theorem A13 join two pieces whose shifts differ by one. All groups between pieces whose shifts differ by three, and thirteen kinds between pieces whose shifts differ by two, are also left alone by the differential, and with them the argument covers every arrangement of the shifts within five consecutive values. For pieces a and b call d a − d b the spread of H ( a , b ) ; its classes of degree two have Ext degree 2 + d a − d b . For ζ , ζ ′ ∈ μ 4 4 call ζ ′ / ζ = ( ζ j ′ / ζ j ) j their ratio; it moves the coordinates where it differs from 1, and it changes the parity, ∏ ζ ′ = − ∏ ζ , exactly when the product of its entries is − 1 . As before, D ( ζ , ζ ′ ) = ∏ ζ j ≠ ζ j ′ | ζ j − ζ j ′ | 2 .
Lemma  A12
(Groups of spread two and three). Keep Lemma A11, and let E be any convolution of the L ζ and the M i , in any order, with any shifts of the given parities and any placement of the t i .
(i)
For a class x of degree two in H ( L ζ , L ζ ′ ) of positive spread, every term of d E ( x ) has all its leaves between pieces L ζ ′ ′ .
(ii)
Let ∏ ζ = ∏ ζ ′ , let ζ ′ / ζ move all four coordinates, one of them by − 1 , and let d L ζ = d L ζ ′ + 2 . The classes of degree two of H ( L ζ , L ζ ′ ) form H 4 ( L ζ , L ζ ′ ) , of dimension D ( ζ , ζ ′ ) ∈ { 64 , 256 } ; no element of degree one has a differential with a nonzero component in it, and d E vanishes on it.
(iii)
Let ∏ ζ = − ∏ ζ ′ and d L ζ = d L ζ ′ + 3 . The classes of degree two of H ( L ζ , L ζ ′ ) lie in H 5 ( L ζ , L ζ ′ ) , and d E vanishes on them.
(iv)
Let ζ and ζ ′ differ in exactly three coordinates, with d L ζ = d L ζ ′ + 1 . Every term of d E on H 3 ( L ζ , L ζ ′ ) runs along chains of components c → ⋯ → L ζ and L ζ ′ → ⋯ → e between pieces L ζ ′ ′ that together lower the shift by one or two.
Proof. 
Write a = L ζ , b = L ζ ′ and δ = d a − d b . As in the proof of Theorem A13(ii), a term of d E on a class of degree two in H ( a , b ) runs along chains of components c → ⋯ → a and b → ⋯ → e , not both empty, and lands in H ( c , e ) in Ext degree 3 + d c − d e , which is 3 if c = e .
(i) Let L 1 be the first piece L ζ ′ ′ of the first chain and L 2 the last of the second, and put X = ( d L 1 − d a ) + ( d b − d L 2 ) . By Lemma A11(iii), X ≥ 0 , and X ≥ 4 if a multiple lies between L 1 and a or between b and L 2 . Before L 1 the first chain is a run of distinct multiples, with U 1 steps up and D 1 steps down the value ϕ and U 1 + D 1 ≤ 3 ; after L 2 the second chain is a run with U 2 and D 2 steps, U 2 + D 2 ≤ 3 . By Lemma A11(ii) the term has Ext degree 3 + δ + X − U + 7 D , where U = U 1 + U 2 and D = D 1 + D 2 . If c and e are pieces, this is 3 + δ + X ; a multiple on a chain makes it at least 8, but H 8 ( c , e ) ≠ 0 only for c = e , where the degree is 3. If c = M i and e is a piece, H ( c , e ) sits in degree 0 if t i < p and in degree 8 if t i > p . Degree 0 needs U 1 ≥ 3 + δ + X ≥ 4 . Degree 8 needs a step down on the run from t i to p, so D 1 ≥ 1 , and then U 1 = 7 D 1 − 5 + δ + X ≥ 3 . Both contradict U 1 + D 1 ≤ 3 , and a piece c with a multiple e is the same. Let c = M i and e = M k . For i = k the degree is 3, so U − 7 D = δ + X ≥ 1 ; D ≥ 1 would give U ≥ 8 , and D = 0 would make the values increase from t i to p and then from p to t i . For i ≠ k , H ( c , e ) sits in degree 0 if t k > t i and in degree 8 if t k < t i . Degree 0 needs U = 3 + δ + X + 7 D , so D = 0 and U ≥ 4 ; but then the first run lies below p and the second above, so the multiples are distinct and U ≤ 3 . Degree 8 needs D ≥ 1 , since with D = 0 the values would satisfy t i < p < t k ; then 7 D − U = 5 − δ − X ≤ 4 gives D = 1 and U ≥ 3 , so the runs have at least four steps and share a multiple, which the last paragraph of the proof of Theorem A14(ii) excludes.
(ii) By the proof of Lemma A11(i), H ( a , b ) is the tensor product of four groups H 1 and sits in degree 4, of dimension D ( ζ , ζ ′ ) ; the ratio is ( − 1 , − 1 , − 1 , − 1 ) , with D = 256 , or a permutation of ( i , − i , − 1 , − 1 ) , with D = 64 .22 A coboundary with a component in H ( a , b ) is a sum of terms of Lemma A5(iii) along a path a → ⋯ → b of morphisms of degree one, one of them an element u and the others components, that lowers the shift by two. By Lemma A11(iii) no multiple lies on it, and by Lemma A11(i) each step between distinct pieces lowers the shift by at least one. So the path consists of a component of Ext degree three from a to b, which lies in H 3 ( a , b ) = 0 , and a class of H 1 ( a , a ) or H 1 ( b , b ) ; or of two steps a → c → b of Ext degree two, possibly with a class of H 1 ( A , O A ) at one of the three pieces. The output has degree one on every surface. Without the class of H 1 the two steps have degrees adding up to one on every surface, so they move disjoint pairs of coordinates and have degree zero elsewhere; a step of Ext degree two changes the parity, so its two moved entries have product − 1 and are ( i , i ) or ( − i , − i ) , and ζ ′ / ζ has no entry − 1 . With the class of H 1 , Lemma A6 gives a drop of one on a surface where all three leaves have positive degree, and the output has degree at least two there.
Now let a term of d E on H 4 ( a , b ) carry the components y 1 , ⋯ , y m , m ≥ 1 , between pieces by (i), and land in H ( c , e ) ; let K be the set of coordinates fixed by the ratio of c and e, and X ≥ m the drop of the chains. The output has Ext degree 5 + X . Let n j be the number of the y i of positive degree on B j ; the class x has degree one there, so Lemma A6 bounds the drop on B j by max ( 0 , n j − 1 ) . For j ∉ K the output has degree 1 on B j , so the degrees of the y i there add up to at most max ( 0 , n j − 1 ) , which forces n j = 0 . For j ∈ K the y i undo the entry ζ j ′ / ζ j ≠ 1 , so n j ≥ 1 , and the output has degree at least 1 + n j − ( n j − 1 ) = 2 there, hence exactly 2. So 5 + X = ( 4 − | K | ) + 2 | K | , that is X = | K | − 1 . The components change the parity X times modulo two, and together their ratio is ζ / ζ ′ on K and 1 elsewhere, so
∏ j ∈ K ζ j ′ / ζ j = ( − 1 ) | K | − 1 .
For the ratio ( − 1 , − 1 , − 1 , − 1 ) the left side is ( − 1 ) | K | . For a permutation of ( i , − i , − 1 , − 1 ) the left side is real only if K contains both or neither of the coordinates moved by ± i , and then it is again ( − 1 ) | K | . So no term is nonzero.
(iii) The classes of degree two have Ext degree 5. By (i) a term of d E on them lands in Ext degree 6 + X , with X ≥ 1 the drop of the chains between pieces. For X ≥ 2 this is at least 8, which needs c = e and degree 3. For X = 1 the pieces c, e have shifts differing by four, hence the same parity, so their ratio moves at least two coordinates and H ( c , e ) has no degree above 6.
(iv) By (i) a term lands in Ext degree 4 + X , with X ≥ 1 the drop of the chains between pieces. For X ≥ 4 the degree is at least 8, which needs c = e . For X = 3 the degree is 7, which needs a ratio moving one coordinate, hence opposite parities, while d c − d e = 4 is even.    □
The groups of (ii) are reached by no coboundary and left by no term in every convolution, so they survive in Ext 2 ( E , E ) as soon as the shifts give them the spread two. That happens along every edge of a Cayley graph that crosses between two shift values of one parity, and a cut of that graph is never light.
Lemma  A13
(Cuts of a Cayley graph). Let G be a finite abelian group, S ⊂ G ∖ { 0 } a generating set and w : S → R > 0 a weight, extended by 0 to G. For every subset T ⊂ G with T ≠ ∅ , G ,
∑ c ∈ T , e ∉ T w ( e − c ) ≥ min H | H | · w ( S ∖ H ) ,
the minimum over the proper subgroups H of G. For the group G 0 of the 64 ratios of pieces of one parity, with S the thirteen ratios of Lemma A12(ii) and w = D , the minimum is 1024, attained only at H = 0 , so every such cut weighs at least 1024.
Proof. 
The left side is ∑ s ∈ S w ( s ) | ( T + s ) ∖ T | . Let H be the stabiliser of T, a proper subgroup since T ≠ ∅ , G . Then T and T + s are unions of cosets of H, and for s ∉ H the set ( T + s ) ∖ T is nonempty, since | T + s | = | T | , so it has at least | H | elements. This gives the inequality. In G 0 , written additively as the quadruples in ( Z / 4 ) 4 of sum 0, the ratios of type ( i , − i , − 1 , − 1 ) generate, since the differences of two of them give every e j − e k . The weights are 256 once and 64 otherwise, and | S ∩ H | ≤ | H | − 1 . So | H | w ( S ∖ H ) is 1024 for H = 0 , at least | H | ( 1024 − 256 − 64 ( | H | − 2 ) ) , which is 1536, 2560 and 3072, for | H | = 2 , 4 , 8 , and at least 64 | H | ≥ 1024 for | H | ≥ 16 , because some generator lies outside H.23    □
Theorem  A15
(One parity at two or three values). Keep Lemma A11, and suppose that the shifts of the pieces L ζ with ∏ ζ = ϵ take exactly two values s and s − 2 , or exactly three values s + 2 , s and s − 2 , for a sign ϵ and an integer s. Then, for every order of the pieces, all shifts of the other pieces, every choice of the shifts of the M i and every placement of the t i ,
dim Ext 2 ( E , E ) ≥ 1024 .
In particular E does not meet the polarised criterion.
Proof. 
Let T be the set of pieces of parity ϵ at the shift s; it is neither empty nor all of them. Two pieces of parity ϵ , one in T and the other not, have shifts differing by exactly two. When their ratio is one of the thirteen of Lemma A12(ii), the group H 4 from the piece of the larger shift to the other is reached by no coboundary and killed by no term, so its classes inject into Ext 2 ( E , E ) , and those of different pairs lie in different summands H ( a , b ) . Their dimensions add up to the weight of the cut ( T , ∁ T ) of the Cayley graph of Lemma A13, since the pieces of parity ϵ form a coset of G 0 ; so they add up to at least 1024.    □
Theorem  A16
(Shifts three apart). Keep Lemma A11, and suppose that the pieces L ζ with ∏ ζ = ϵ all have the shift s and the others have shifts in { s − 3 , s + 3 } . Then, for every order, every choice of the shifts of the M i and every placement of the t i ,
dim Ext 2 ( E , E ) ≥ 64 · 1072 − 128 · 8 = 67584 .
Proof. 
Every piece W of parity − ϵ and every piece V of parity ϵ have shifts differing by three. By Lemma A12(iii) the classes of degree two of the group from the higher to the lower of the two lie in H 5 , and d E vanishes on them. For each W these groups have dimensions adding up to 1072.24 A coboundary with a component in such a group runs along a path of morphisms of degree one that lowers the shift by three; no multiple lies on it, by Lemma A11(iii), and a step between distinct pieces lowers the shift by at least one, while no piece has a shift strictly between those of W and V. So the path is the component between W and V together with a class of H 1 ( A , O A ) at one of them, and the components of all coboundaries in these groups come from ⨁ a H 1 ( a , a ) over the 128 pieces, of dimension 128 · 8 = 1024 . The 64 · 1072 = 68608 classes are cocycles, and at most 1024 dimensions of them are coboundaries.    □
When one parity sits at a single shift, the groups H 3 of Theorem A13 survive next to it unless pieces of the other parity crowd it from both sides.
Proposition  A9
(A gap of four). Keep Lemma A11, let the pieces L ζ with ∏ ζ = ϵ all have the shift y, and let T be the set of pieces of parity − ϵ at the shift y + 1 . If T is not empty and no piece has the shift y − 1 or y + 3 , then for every order, every choice of the shifts of the M i and every placement of the t i , the groups H 3 ( W , V ) with W ∈ T and V differing from W in exactly three coordinates are reached by no coboundary and killed by no term, and
dim Ext 2 ( E , E ) ≥ 640 | T | ≥ 640 .
The same holds with y + 1 , y − 1 , y + 3 replaced by y − 1 , y + 1 , y − 3 and the groups H 3 ( V , W ) .
Proof. 
By Theorem A13(i), which holds for every convolution, no coboundary reaches these groups. By Lemma A12(iv) a term of d E on H 3 ( W , V ) runs along chains of components between pieces into W and out of V that lower the shift by one or two in all; the first would pass through a piece at the shift y + 2 or y + 3 , the second through one at y − 1 or y − 2 . The shifts y ± 2 have parity ϵ , whose pieces are all at y, and y − 1 , y + 3 are excluded. Each W ∈ T has 28 partners V, all at the shift y, with groups of dimensions adding up to 640. The second statement has the same proof, with the chains into V and out of W in place of those into W and out of V.    □
Corollary  A2
(Five consecutive shifts). Keep Lemma A11. If the shifts of the 128 pieces L ζ lie in five consecutive integers, then, for every order, every choice of the shifts of the M i and every placement of the t i , dim Ext 2 ( E , E ) ≥ 640 ; no such convolution meets the polarised criterion.
Proof. 
Let the values be s , ⋯ , s − 4 , with one parity at values in { s , s − 2 , s − 4 } and the other at values in { s − 1 , s − 3 } . If the second parity takes both its values, or the first takes two values two apart or all three, Theorem A15 applies. Otherwise the second parity sits at one value y and the first at one value or at { s , s − 4 } . One value each, at an odd distance at most three: Theorem A13 or Theorem A16. The first at { s , s − 4 } : for y = s − 1 the pieces at y + 1 = s are not empty and y − 1 , y + 3 are empty, and for y = s − 3 the pieces at y − 1 = s − 4 are not empty and y + 1 , y − 3 are empty, so Proposition A9 applies. The least of the bounds is 640.    □
Within six consecutive values two arrangements are left, up to exchanging the parities: each parity at one shift, five apart, and the parities at { s , s − 4 } and { s − 1 , s − 5 } . The first is settled by the diagonal, whose classes at a shift with no other piece nearby can be reached only through all three multiples, and the second by the groups of Theorem A13 once more.
Lemma  A14
(The diagonal at a lonely shift). Keep Lemma A11, and let the pieces L ζ with ∏ ζ = ϵ all have the shift s and the others the shift s − g , for a sign ϵ and an odd integer g ≥ 5 . Let y ∈ H 2 ( a , a ) be a diagonal class at a piece a, and let a term of d E ( y ) in Lemma A5(iii) be nonzero. Then a is the only piece on its path, the path passes through M 1 , M 2 and M 3 , and the shifts of a, M 1 , M 2 , M 3 are pairwise incongruent modulo 4.
Proof. 
Write the path as c → ⋯ → a → ⋯ → e , with the chain of components into a and the chain out of a not both empty, so that c ≠ e , since the components of a convolution go up the order of the pieces. Let L 1 and L 2 be the first and the last piece on it and X = d L 1 − d L 2 ; before L 1 and after L 2 the path is a run of distinct multiples, with U steps up and D steps down the value ϕ in all. As in the proof of Lemma A12(i), now with δ = 0 , the term has Ext degree 3 + X − U + 7 D by Lemma A11(ii). A sequence of components from one piece to another lowers the shift by Lemma A11(iii), and the pieces have the shifts s and s − g ; so X ∈ { 0 , g } , and X = 0 exactly when a is the only piece on the path.
If c and e are pieces, then U = D = 0 and X = g , and H 3 + g ( c , e ) = 0 since 3 + g ≥ 8 and c ≠ e . If c is a multiple and e a piece, then H ( c , e ) sits in degree 0 or 8. Degree 0 needs U = 3 + X + 7 D , so X = D = 0 and U = 3 . Degree 8 needs 7 D − U = 5 − X : for X = 0 this gives D = 1 and U = 2 ; for X = g it gives D = 0 and U = g − 5 , which is at least one because the run before L 1 is not empty and at most three, so g = 7 ; but then the path lowers the shift by seven from L 1 to e, by a single component, of Ext degree eight between distinct pieces, or by a run through multiples, and neither is possible.25 A piece c with a multiple e is the same with the runs exchanged. If c and e are multiples, both runs are nonempty and together pass through at most three multiples, so U + D ≤ 3 and U + D ≥ 2 . Degree 0 needs U = 3 + X + 7 D , so X = D = 0 and U = 3 . Degree 8 needs 7 D − U = 5 − X ; D = 0 gives U = X − 5 ≥ 2 , so X = g = 7 , which is excluded as before; D = 1 gives U = 2 + X ≤ 2 , so X = 0 and U = 2 ; D ≥ 2 gives U ≥ 9 .
So in every case X = 0 and U + D = 3 : the path consists of a and the three multiples, with three steps. A step up raises the shift by one and a step down lowers it by seven, so every step adds one modulo 4, and the four nodes of the path have the shifts d c , d c + 1 , d c + 2 , d c + 3 modulo 4 in the order of the path.26    □
Proposition  A10
(Single shifts). Keep Lemma A11, and suppose that the pieces L ζ with ∏ ζ = ϵ all have the shift s and the others all have the shift s − g , for a sign ϵ and an odd integer g. Then, for every order, every choice of the shifts of the M i and every placement of the t i ,
dim Ext 2 ( E , E ) ≥ 64 · 28 = 1792 .
Proof. 
Exchanging the parities, we may take g > 0 . For g = 1 and g = 3 this is Theorem A13 and Theorem A16. Let g ≥ 5 . If a term of d E were nonzero on a diagonal class at a piece of shift s and another on a diagonal class at a piece of shift s − g , then by Lemma A14 the shifts of the three multiples would be pairwise incongruent modulo 4 and miss both the residue of s and that of s − g , which differ, since g is odd. So at one of the two shifts d E vanishes on the 28 diagonal classes of each of the 64 pieces there, and by Proposition A5, which holds for n = 4 , these 1792 classes inject into Ext 2 ( E , E ) .27    □
Proposition  A11
(Interleaved shifts). Keep Lemma A11, and suppose that the shifts of the pieces L ζ with ∏ ζ = ϵ take exactly the two values s and s − 4 , and those of the others exactly the two values s − 1 and s − 5 . Then, for every order, every choice of the shifts of the M i and every placement of the t i , dim Ext 2 ( E , E ) ≥ 640 .
Proof. 
Let A and A ′ be the pieces of parity ϵ at the shifts s and s − 4 , and B and B ′ those of parity − ϵ at s − 1 and s − 5 ; all four are nonempty. Let W and V differ in exactly three coordinates, with d W = d V + 1 , that is W ∈ A and V ∈ B , or W ∈ A ′ and V ∈ B ′ . By Theorem A13(i) no coboundary reaches H 3 ( W , V ) , and by Lemma A12(iv) a term of d E on it passes through a piece at the shift d W + 1 or d W + 2 or at d V − 1 or d V − 2 . These are s + 1 , s + 2 , s − 2 , s − 3 for W ∈ A and s − 3 , s − 2 , s − 6 , s − 7 for W ∈ A ′ , and no piece has these shifts. So the groups inject into Ext 2 ( E , E ) .
Let Γ be the group of the ratios of two pieces, of order 128, which acts simply transitively on the pieces, and S 3 ⊂ Γ the 28 ratios that move exactly three coordinates and change the parity, with the weight D, which adds up to 640 (Theorem A13). With X = A ∪ B ′ the pairs above are exactly the edges of the Cayley graph of ( Γ , S 3 ) between X and its complement A ′ ∪ B ; so the dimensions add up to the weight of that cut, and X is neither empty nor everything. Written additively, as the quadruples in ( Z / 4 ) 4 of even sum, the differences of two ratios of type ( i , − i , − 1 , 1 ) give every e j − e k , which generate the quadruples of sum 0 modulo 4, and the elements of S 3 have sum 2 modulo 4; so S 3 generates Γ . In Lemma A13, H = 0 gives 640; for | H | = 2 , 4 , 8 , 16 at most | H | − 1 elements of S 3 lie in H, at most four of them of weight 64, so | H | w ( S 3 ∖ H ) is at least 2 · 576 , 4 · 448 , 8 · 336 and 16 · 208 ; for | H | = 32 some r ∈ S 3 lies outside H, and so does − r , which gives weight at least 32, or 64 if r = − r ; and for | H | = 64 the weight outside H is at least 16. So every such cut weighs at least 640.28    □
Corollary  A3
(Six consecutive shifts). Keep Lemma A11. If the shifts of the 128 pieces L ζ lie in six consecutive integers, then, for every order, every choice of the shifts of the M i and every placement of the t i , dim Ext 2 ( E , E ) ≥ 640 ; no such convolution meets the polarised criterion.
Proof. 
By Corollary A2 we may assume that the shifts take both values s and s − 5 of the window s , ⋯ , s − 5 ; let the parity at s have its values in { s , s − 2 , s − 4 } and the other in { s − 1 , s − 3 , s − 5 } . If one of them takes two values two apart, or three, Theorem A15 applies. Otherwise the first takes { s } or { s , s − 4 } and the second { s − 5 } or { s − 1 , s − 5 } . For { s } and { s − 5 } this is Proposition A10, and for { s , s − 4 } and { s − 1 , s − 5 } it is Proposition A11. For { s } and { s − 1 , s − 5 } , Proposition A9 applies in its second form with y = s : the shift s − 1 is taken and s + 1 , s − 3 are not. For { s , s − 4 } and { s − 5 } it applies in its first form with y = s − 5 : the shift s − 4 is taken and s − 6 , s − 2 are not. The least of the bounds is 640.    □

Appendix A.5. Pieces of Any Kind

Theorem A10 and Proposition 28 say that in a convolution meeting the criterion every piece outside every closed chain has its Chern character in Q [ θ ] ⊕ W ( A ) . The same holds for every convex set of pieces, for pieces of any kind. Throughout, A is an abelian 2 n -fold of Weil type for an imaginary quadratic field K, a member of D n , n with polarisation θ , H 1 ( A , C ) = V + ⊕ V − the eigenspace decomposition for K, V − ≅ V + ∨ through θ , α ± spanning ⋀ 2 n V ± , and T D n , n ⊆ H 1 ( A , T A ) the tangent space of the family. We say that Emeets the criterion when dim Ext 2 ( E , E ) , or dim Ext 2 ( E , E ) G for a G-equivariant E as in Corollary 11, equals the rank of the contraction H T 2 ( A ) → H * ( A ) with γ = ch ( E ) .
Setting  13
(Convolutions of perfect complexes). Let F 1 , ⋯ , F m be perfect complexes on A, each represented by a bounded complex of holomorphic vector bundles, d 1 , ⋯ , d m integers and P a = F a [ − d a ] . Let A ( a , b ) = A 0 , * ( H o m • ( F a , F b ) ) with the differential D, the sum of ∂ ¯ and the internal differentials, so that its cohomology is Ext * ( F a , F b ) . A convolution of P 1 , ⋯ , P m in this order is a one-sided twisted complex E = ( ⨁ a P a , D + δ ) with δ = ( δ a b ) a < b of total degree one and D δ + δ ∘ δ = 0 ; every object obtained by iterated cones in this order is of this form, since the Dolbeault complexes form a pretriangulated enhancement. In each block choose a subcomplex H ( a , b ) of D-closed forms mapping isomorphically to Ext * ( F a , F b ) , a projection π onto it and a homotopy h with 1 − π = D h + h D and π h = h π = h 2 = 0 , which exist for every complex by a choice of complements. The proof of Lemma A5 uses only these identities, so (i)–(iii) hold verbatim. Let Γ be the directed graph on the pieces with an edge a → b , a ≠ b , when H ( a , b ) is nonzero in total degree one; its directed cycles are the cycles of excess zero. Let Γ δ ⊆ Γ consist of the edges with δ a b H ≠ 0 , in the gauge h δ = 0 of Lemma A5(i); it has no directed cycle, δ being strictly upper triangular, and by the trees of Lemma A5(i), δ a b ≠ 0 implies a path of Γ δ from a to b. A set C of pieces is Γ -convexif every walk of Γ that starts and ends in C stays in C, and δ -convexif every walk consisting of a path of Γ δ , at most one edge of Γ and a path of Γ δ that starts and ends in C stays in C; a Γ-convex set is δ-convex for every δ. Put E C = ( ⨁ a ∈ C P a , D + δ | C ) . Fix a block diagonal ( 1 , 0 ) -connection D ′ on ⨁ a F a ; then At ( E ) is represented by [ D ′ , D + δ ] , block upper triangular with diagonal blocks At ( F a ) and off-diagonal blocks [ D ′ , δ a b ] , as in the proofs of Proposition 23 and Lemma 18.
Theorem  A17
(Convex sets of pieces). In Setting 13, let C be δ-convex. Then E C is a twisted complex and ker ev E ⊆ ker ev E C . If E meets the criterion, ordinary or equivariant, and γ = ch ( E ) has the polarised shape, then ker ev E = Ann ( γ ) ⊆ Ann ( ch E C ) and ch ( E C ) ∈ Q [ θ ] ⊕ W ( A ) .
Proof. 
Every term below is labelled by a walk: a nonzero δ x y gives a path of Γ δ from x to y, and a nonzero block ( c , d ) of ξ ⌟ exp At ( E ) needs c = d or a chain of nonzero δ -components from c to d, because exp At ( E ) is built from the diagonal blocks and the blocks [ D ′ , δ c c ′ ] . For a , b ∈ C , a nonzero term of ( δ ∘ δ ) a b = ∑ c δ a c δ c b gives a walk a ⇝ c ⇝ b of Γ δ , so c ∈ C , and the Maurer–Cartan equation of δ | C is the C × C block of that of δ ; so E C is a twisted complex. For x ∈ H E of total degree one, the component of d E ( x ) in H ( a , b ) is a sum of trees along a walk from a through components of δ , one component x c d with c = d or c → d in Γ , and components of δ to b (Lemma A5(iii)); for a , b ∈ C the walk stays in C, and π , h and composition act blockwise, so the C-block of d E ( x ) is d E C ( x C ) . Likewise, for y = ξ ⌟ exp At ( E ) , whose C-block is y C = ξ ⌟ exp At ( E C ) , the ( a , b ) block of π E ( y ) = π ∑ k ( − Q h ) k y is a sum over walks a ⇝ c , y c d , d ⇝ b , so the C-block of π E ( y ) is π E C ( y C ) . If ev E ( ξ ) = 0 , then π E ( y ) = d E ( x ) for some x, and taking C-blocks gives ev E C ( ξ ) = 0 .
If E meets the ordinary criterion, σ E is injective because σ E ∘ ev E is the contraction with γ by (A4), so ker ev E = Ann ( γ ) ; in the equivariant case ev E ( H T 2 ) ⊆ Ext 2 ( E , E ) G by the proof of Corollary 11, σ E is injective there, and the same follows. Then Ann ( γ ) = ker ev E ⊆ ker ev E C ⊆ Ann ( ch E C ) , the last inclusion by (A4) for E C . The tangent space T D n , n kills θ and W ( A ) , so it lies in Ann ( γ ) and kills ch ( E C ) ; contraction with v ∈ H 1 ( A , T A ) preserves the degree, so it kills every graded piece of ch ( E C ) , and Lemma A15 gives ch ( E C ) ∈ R [ θ ] ⊕ W R , hence ch ( E C ) ∈ Q [ θ ] ⊕ W ( A ) , the subspace being defined over Q .    □
Remark  A5
(Special cases). A single piece is Γ-convex exactly when it lies on no cycle of excess zero, so Theorem A10 and Proposition 28 are the case of one piece. A strongly connected component of Γ is Γ-convex, as is every set closed under successors or under predecessors in Γ. No assumption on the Hankel rank is used.
Lemma  A15
(Invariant classes). If c ∈ H * ( A , R ) satisfies v ⌟ c = 0 for all v ∈ T D n , n , then c ∈ R [ θ ] ⊕ W R .
Proof. 
As in the proof of Theorem A10, T D n , n is the block Hom ( V + 1 , 0 , V + 0 , 1 ) of sl ( V + ) acting by derivations on ⋀ * ( V + ⊕ V + ∨ ) , and since c is real its complex conjugates, the block Hom ( V + 0 , 1 , V + 1 , 0 ) , kill c too; the two blocks generate sl ( V + ) , so c is SL ( V + ) -invariant. With N = 2 n , ⋀ k V + ⊗ ⋀ l V + ∨ ≅ Hom ( ⋀ l V + , ⋀ k V + ) , and the ⋀ j V + are irreducible and pairwise non-isomorphic for 0 < j < N and trivial for j ∈ { 0 , N } . So the invariants are the line of the identity for each k = l , spanned by θ k , and the lines ⋀ N V + and ⋀ N V + ∨ , spanned by α + and α − .    □
For n = 3 , 4 the joint kernel of T D n , n and its conjugate on ⋀ * C 4 n is computed in julia/invcheck.jl; the kernel of T D n , n alone has dimension 344 for n = 3 and 1696 for n = 4 , so the reality of c is needed.
Lemma  A16
(Abelian subvarieties). Let n ≥ 2 and let Z ⊆ A be a translate of an abelian subvariety with 0 < dim Z < 2 n . Then [ Z ] ∉ Q [ θ ] ⊕ W ( A ) .
Proof. 
Let Z 0 be the underlying subvariety, q : A → A / Z 0 and c = 2 n − dim Z . Then [ Z ] = q * [ pt ] = e 1 ∧ ⋯ ∧ e 2 c with e i spanning q * H 1 ( A / Z 0 ) , and a constant vector u ∈ H 1 ( A , C ) satisfies u ⌟ [ Z ] = 0 exactly when u ∈ H 1 ( Z 0 , C ) ≠ 0 . We show that a nonzero element λ θ c + w of Q [ θ ] ⊕ W ( A ) in degree 2 c , with w ≠ 0 only if c = n , is killed by no u ≠ 0 . If λ ≠ 0 and w = 0 : u ⌟ θ c = c ( u ⌟ θ ) θ c − 1 , with u ⌟ θ ≠ 0 since θ is nondegenerate, and multiplication by θ c − 1 is injective on H 1 by the Lefschetz theorem, as 1 + ( c − 1 ) ≤ 2 n . If c = n , write u = u + + u − and w = a α + + b α − ; in the bigrading by ( V + , V − ) , u ⌟ α + has bidegree ( 2 n − 1 , 0 ) , u ⌟ α − has ( 0 , 2 n − 1 ) , and u ⌟ θ n has ( n − 1 , n ) and ( n , n − 1 ) , all distinct for n ≥ 2 , so the three terms vanish separately. If λ ≠ 0 the first case gives u = 0 ; if λ = 0 , w is real and nonzero, so b = a ¯ ≠ 0 , and u + ⌟ α + = 0 , u − ⌟ α − = 0 force u = 0 .    □
Corollary  A4
(Exclusions). Let E be a convolution meeting the criterion, ordinary or equivariant, with ch ( E ) of polarised shape. Then every piece O Z ⊗ F , with Z a translate of an abelian subvariety, 0 < dim Z < 2 n and F a vector bundle on Z, and every semihomogeneous vector bundle piece whose slope is not in Q θ , lies on a cycle of excess zero. Otherwise dim Ext 2 ( E , E ) , or dim Ext 2 ( E , E ) G , exceeds the rank of the contraction with ch ( E ) .
Proof. 
A piece on no cycle of excess zero is Γ -convex, so its Chern character lies in the graded subspace Q [ θ ] ⊕ W ( A ) by Theorem A17. The lowest component of ch ( O Z ⊗ F ) is rk ( F ) [ Z ] , excluded by Lemma A16. A semihomogeneous bundle has ch = r e δ , whose degree-two part r δ must lie in Q θ , since W ( A ) lives in degree 2 n ≥ 4 .    □
For semihomogeneous bundles the Ext groups lie in windows. For a hermitian form X write i ± ( X ) for the numbers of positive and negative eigenvalues.
Lemma  A17
(Cohomology windows). Let B be an abelian variety of dimension g and G a semihomogeneous vector bundle on B of slope δ, ch G = r e δ , and i ± = i ± ( δ ) .
(i)
H q ( B , G ) = 0 unless i − ≤ q ≤ g − i + .
(ii)
If G = L is a line bundle, either H * ( L ) = 0 or H q ( L ) ≠ 0 for every q ∈ [ i − , g − i + ] , and the second case holds when L restricts trivially to the connected component K ( L ) 0 of the kernel of its hermitian form; in particular when L is nondegenerate, and when L = f * M for a surjective homomorphism f with connected kernel and M nondegenerate.
For semihomogeneous bundles F, F ′ of slopes δ, δ ′ , Ext q ( F , F ′ ) = H q ( F ∨ ⊗ F ′ ) vanishes outside [ i − ( δ ′ − δ ) , g − i + ( δ ′ − δ ) ] .
Proof. 
By [83, Sects. 5–6], G ≅ ⨁ j E j ⊗ U j with E j simple semihomogeneous of slope δ and U j unipotent, an iterated extension of O , so it suffices to treat a simple E; and E ≅ f * L for an isogeny f : Y → B and a line bundle L on Y, with H q ( B , E ) = H q ( Y , L ) . Since f * f * L ≅ ⨁ x ∈ ker f t x * L , c 1 ( L ) = f * δ in NS Q ( Y ) , whose hermitian form is the pullback along the isomorphism d f and has the same inertia. Let L be a line bundle on Y and K = K ( L ) 0 . For y ∈ Y , t y * L ≅ L ⊗ P φ L ( y ) , and the character of P φ L ( y ) on the lattice of K is exp ( 2 π i Im H L ( y , · ) ) , trivial on T K ; so all fibres of π : Y → Y / K carry isomorphic restrictions L 0 ∈ Pic 0 ( K ) . If L 0 is nontrivial, H * ( K , L 0 ) = 0  [88, §8], so R π * L = 0 and H * ( L ) = 0 . If L 0 is trivial, L ≅ π * M with M nondegenerate on Y / K by the seesaw theorem, R q π * O Y ≅ O ⊗ H q ( K , O ) , and by the projection formula and the vanishing theorem [88, §16] H q ( Y , L ) ≅ H i ( M ) ( M ) ⊗ H q − i ( M ) ( K , O ) with H i ( M ) ( M ) ≠ 0 , i ( M ) = i − and dim K = g − i − − i + ; so H q ( L ) ≠ 0 exactly for q ∈ [ i − , g − i + ] . For L = f * M as in (ii), K ( L ) 0 = ker f and L is trivial on it. The last statement holds because F ∨ ⊗ F ′ is semihomogeneous with ch = r r ′ e δ ′ − δ .    □
Lemma  A18
(Subadditivity of inertia). For hermitian forms X, Y on C g , i − ( X + Y ) ≤ i − ( X ) + i − ( Y ) .
Proof. 
Let U be a subspace of dimension i − ( X + Y ) on which X + Y is negative definite, and P, Q subspaces of dimensions g − i − ( X ) , g − i − ( Y ) on which X ≥ 0 , Y ≥ 0 . On U ∩ P ∩ Q the form X + Y is both ≥ 0 and negative definite, so U ∩ P ∩ Q = 0 and i − ( X + Y ) + 2 g − i − ( X ) − i − ( Y ) ≤ 2 g .    □
Lemma  A19
(Cycles of semihomogeneous bundles). If a cycle of excess zero of length k passes through two semihomogeneous bundle pieces a, b of slopes δ a , δ b and through semihomogeneous bundles only, then rank ( δ b − δ a ) ≤ k .
Proof. 
An edge e from c to c ′ has Ext degree q e = 1 + d c − d c ′ , and ∑ e q e = k around the cycle. By Lemma A17, q e ≥ i − ( δ c ′ − δ c ) , and by Lemma A18, i − ( δ b − δ a ) is at most the sum of the q e along the arc from a to b and i + ( δ b − δ a ) = i − ( δ a − δ b ) at most the sum along the arc from b to a. Adding gives the claim.    □
Corollary  A5
(Nondegenerate slope differences). No convolution of at most 2 n − 1 semihomogeneous vector bundles whose slope differences are pairwise nondegenerate meets the criterion, ordinary or equivariant.
Proof. 
A cycle of excess zero would have length at most 2 n − 1 and pass through two pieces whose slope difference has rank 2 n , against Lemma A19. So Γ is acyclic, every piece is Γ -convex, and by Corollary A4 every slope lies in Q θ ; then ch ( E ) ∈ Q [ θ ] , which is not of polarised shape.    □
Lemma  A20
(Windows of Lagrangian pieces). Let Z, Z ′ be translates of abelian subvarieties of dimension n with T Z ∩ T Z ′ = 0 , F, F ′ vector bundles on Z, Z ′ and M a vector bundle on A. Then Ext k ( M , O Z ⊗ F ) = H k ( Z , M ∨ | Z ⊗ F ) , Ext k ( O Z ⊗ F , M ) = H k − n ( Z , F ∨ ⊗ M | Z ) , and Ext k ( O Z ⊗ F , O Z ′ ⊗ F ′ ) vanishes for k ≠ n and not for k = n .
Proof. 
The first is adjunction for the locally free M. Locally Z = { x 1 = ⋯ = x n = 0 } for a regular sequence, and the Koszul resolution shows that the local Ext sheaves of ( O Z ⊗ F , M ) vanish except in degree n, where they are F ∨ ⊗ M | Z ⊗ det N Z / A with trivial normal bundle; the local-to-global spectral sequence degenerates. For the third, Z 0 + Z 0 ′ = A , so Z ∩ Z ′ is finite and nonempty; at each point the x i restrict to a regular system of parameters on Z ′ , so the local Ext sheaves are skyscrapers in degree n, nonzero there.    □
Let F be semihomogeneous on Z with slope δ F , let λ 1 , ⋯ , λ n be the eigenvalues of the hermitian form of δ F relative to the positive form of θ | Z , and let M t be semihomogeneous on A with slope t θ . By Lemmas A17 and A20, with P = O Z ⊗ F , the Ext degrees lie in the windows [ # { λ < t } , n − # { λ > t } ] for M t → P , n + [ # { λ > t } , n − # { λ < t } ] for P → M t , { 0 } for M t → M s with s > t , { 2 n } for s < t , and { n } between two transverse pieces.
Theorem  A18
(Transverse Lagrangian pieces and multiples of θ ). Let n ≥ 3 and let E be a convolution, possibly G-equivariant, of proper pieces O Z a ⊗ F a [ − d a ] , with Z a translates of abelian subvarieties of dimension n, T Z a ∩ T Z b = 0 for a ≠ b and F a semihomogeneous on Z a , and of ℓ semihomogeneous bundles M t [ − d t ] of slopes t θ with pairwise distinct t. If ℓ ≤ 2 n − 3 , or if ℓ ≤ 2 n − 2 and no two proper pieces have | d a − d b | = n − 1 , then every proper piece lies on no cycle of excess zero, and E does not meet the criterion, ordinary or equivariant.
Proof. 
A shortest closed walk through a piece is a simple cycle, and along a cycle ∑ e ( q e − 1 ) = 0 with q e at least the lower end lo e of its window; so it suffices to show that the cost T = ∑ e ( lo e − 1 ) of every simple cycle through a proper piece is positive. From the windows, an edge P → P ′ costs n − 1 , P ( λ ) → M t costs n − 1 + # { λ > t } , M t → P ( λ ′ ) costs # { λ ′ < t } − 1 , an up step M t → M s , s > t , costs − 1 and a down step costs 2 n − 1 . Cut the cycle at its proper pieces into d direct segments P → P ′ and m segments P ( λ ) → M t 1 → ⋯ → M t r → P ( λ ′ ) , r ≥ 1 . With W down steps, ℓ u ≤ ℓ multiples of θ used, hence ℓ u − m − W up steps, and Λ ≥ 0 the sum of the counts # { λ > t first } and # { λ ′ < t last } over the segments,
T = ( m + d ) ( n − 1 ) + 2 n W − ℓ u + Λ .
(a) If m = 0 , then d ≥ 2 and T ≥ 2 ( n − 1 ) . (b) If W ≥ 1 , then T ≥ ( n − 1 ) + 2 n − ℓ . (c) If W = 0 , d ≥ 1 and m + d ≥ 2 , then T ≥ 2 ( n − 1 ) − ℓ . (d) If W = 0 , d = 0 and m ≥ 2 , each segment is one increasing run from a s to b s followed by a proper piece P s + 1 , and a s + 1 ≠ b s since the multiples are distinct; if a s + 1 < b s , every eigenvalue of P s + 1 is below b s or above a s + 1 , so P s + 1 contributes at least n to Λ ; and a s + 1 > b s for all s would give a 1 ≤ b 1 < a 2 ≤ ⋯ ≤ b m < a 1 . So Λ ≥ n and T ≥ 3 n − 2 − ℓ . (e) If W = 0 , m = 1 , d = 0 , there are one proper piece and one run t 1 < ⋯ < t r , and T = ( n − 1 ) − r + # { λ > t 1 } + # { λ < t r } ; for r ≥ 2 every λ is counted at least once, so T ≥ 2 n − 1 − ℓ , and for r = 1 , T ≥ n − 2 . For ℓ ≤ 2 n − 3 every case gives T > 0 . Under the condition on the shifts no edge P → P ′ has q = 1 + d a − d b = n , so d = 0 , cases (a) and (c) do not occur, and (b), (d), (e) give T > 0 for ℓ ≤ 2 n − 2 . The conclusion follows from Corollary A4; without proper pieces ch ( E ) ∈ Q [ θ ] .    □
The thresholds are those of the windows: an exhaustive search over window-admissible cycles, in julia/windowsearch.jl, finds cycles of cost zero at ℓ = 2 n − 2 with edges between proper pieces and at ℓ = 2 n − 1 without them, for n = 3 with up to two proper pieces and for n = 4 with one, for instance P ( 0 , 0 , 0 ) → M 0 → M 1 → M 2 → M 3 → P ( 3 , 3 , 3 ) → P ( 0 , 0 , 0 ) at n = 3 . Whether such cycles carry nonzero classes, and the cases ℓ ≥ 2 n − 2 with edges between proper pieces, ℓ ≥ 2 n − 1 without, supports that are not transverse or of mixed dimension, and equal slopes, are open.
All the tests above depend on Γ only, and Γ does not depend on δ . They do not decide every configuration. Let A = E 0 6 with the coordinates and blocks B j of Proposition A4, n = 3 . For a ∈ μ 4 let f j , a = z j + a z 3 + j : A → E 0 , a surjective homomorphism with connected kernel, and m j , a = c 1 ( f j , a * O ( 0 ) ) , whose hermitian form | z j + a z 3 + j | 2 has eigenvalues 0 and 2 on B j ; m j , a 2 = 0 , and Z ζ = ⋂ j ker f j , ζ j is the threefold of Appendix B.
Proposition  A12
(A configuration passing every graph test). For each of the 32 elements ζ ∈ μ 4 3 with ∏ ζ = ± 1 put c ( ζ ) = 0 if ∏ ζ = 1 and c ( ζ ) = 1 otherwise, and take the cube of eight line bundles
M S ( ζ ) = P ⌀ ( ζ ) ⊗ ⨂ j ∈ S f j , ζ j * O ( 0 ) ⊗ Q j ( ζ ) at the shift | S | + c ( ζ ) , S ⊆ { 1 , 2 , 3 } ,
with P ⌀ ( ζ ) ∈ Pic 0 ( A ) and Q j ( ζ ) ∈ Pic 0 ( E 0 ) , together with four line bundles A , B , C , D of classes − 2 θ , − θ , 3 θ , 4 θ at the shifts − 1 , 0 , 4 , 5 : 260 pieces.
(i)
Every convolution E of these pieces has ch ( E ) = − 32 ( α + α ¯ ) + p ( θ ) with p = − e − 2 θ + e − θ + e 3 θ − e 4 θ of Hankel rank 3, so r ( ch E ) = 57 .
(ii)
For every choice of the twists, Γ contains the 46 edges inside each cube of Lemma A21(i) and the 196 edges between pieces whose classes differ by a nondegenerate form, and these 1668 edges form a strongly connected graph on all 260 pieces.
So in every convolution of these pieces, in any order and for every twisting cochain, every piece lies on a cycle of excess zero, the only nonempty Γ-convex set is the set of all pieces, and Theorem A17 in its Γ-form, Theorem A10, Proposition 28 and Corollary A4 do not apply.
Proof. 
(i) Inside a cube, ∑ S ( − 1 ) | S | + c e ∑ j ∈ S m j = ( − 1 ) c ∏ j ( 1 − e m j ) = − ( − 1 ) c m 1 m 2 m 3 , since m j 2 = 0 and Pic 0 twists are numerically trivial, and m 1 , ζ 1 m 2 , ζ 2 m 3 , ζ 3 = [ Z ζ ] because ( f 1 , f 2 , f 3 ) : A → E 0 3 is surjective with connected kernel Z ζ . On B j , m j , z = a j + b j + z ¯ c j + z c j ′ in the notation of the proof of Proposition A4, and ( − 1 ) c ( ζ ) = ∏ ζ ; the computation there gives ∑ ∏ ζ = ± 1 ∏ ζ ∏ j m j , ζ j = 1 2 ∏ j ∑ z z m j , z + ∏ j ∑ z z ¯ m j , z = 32 ( α + α ¯ ) . The four multiples of θ give p, whose moments ∑ i σ i t i k over four distinct exponents make the 5 × 3 Hankel matrix of rank 3. (ii) The classes inside a cube differ by pullbacks along homomorphisms with connected kernel, so all window groups of Lemma A21(i) are nonzero by Lemma A17(ii), whatever the twists; a nondegenerate difference has nonzero cohomology in its index by the same lemma. Strong connectivity is a finite check.    □
The identity (i), the Hankel rank, the edge counts and the strong connectivity are checked in exact arithmetic in julia/cubeexample.jl and julia/collapse.jl. For a suitable choice of P ⌀ there is moreover a pair of edges of degree one in both directions between M { 1 , 2 } ( ζ + ) and M { 1 } ( ζ − ) with ζ 1 + = ζ 1 − , and with it the 196 nondegenerate edges and the 384 up and 384 down edges of the cubes already form a strongly connected graph with 385 cycles of length two; these cycles have semidefinite differences of rank at most two, as Remark A3 requires. For δ = 0 , Proposition 22 excludes E. The rest of this subsection decides the criterion for this configuration for the twisting cochains of two kinds. Shifts are cohomological degrees: a block Ext k ( F a , F b ) has total degree k + s b − s a .
Lemma  A21
(Windows of the configuration). In Proposition A12:
(i)
Inside a cube, Γ has an edge M S → M S ′ exactly when | S ′ ∖ S | ≤ 1 , S ≠ S ′ , which gives 46 edges; the only edges into M { 1 , 2 , 3 } come from the M j k and the only edges out of M ⌀ go to the M j , all of Ext degree 0.
(ii)
Between a cube and the multiples of θ the only groups of total degree one are, for c = 0 , A → M ⌀ and D → M ⌀ , B → M j and M { 1 , 2 , 3 } → C , and for c = 1 , B → M ⌀ , M j k → C , M { 1 , 2 , 3 } → A and M { 1 , 2 , 3 } → D .
(iii)
Among the multiples of θ the edges are A → B , C → D , of degree 0, and C → A , D → B , of degree 6.
(iv)
Between pieces of different cubes the difference is nondegenerate only for S = S ′ = { 1 , 2 , 3 } and ζ j ≠ ζ j ′ for all j, of signature ( 3 , 3 ) with the only group in degree 3 and total degree 3 + c ′ − c ≠ 1 .
Say that a cube Q satisfies ( Gen Q ) if all groups between a piece of Q and a piece of another cube with degenerate difference vanish. This holds for P ⌀ ( Q ) outside finitely many translates of proper subtori of Pic 0 ( A ) , and then Γ has no edge between Q and another cube. The twists aregenericif ( Gen Q ) holds for all 32 cubes.
Proof. 
(i) M S ′ ⊗ M S − 1 has class ∑ S ′ ∖ S m j − ∑ S ∖ S ′ m j , of inertia ( i + , i − ) = ( | S ′ ∖ S | , | S ∖ S ′ | ) . An edge needs Ext degree k = 1 + | S | − | S ′ | = 1 + | S ∖ S ′ | − | S ′ ∖ S | in the window [ | S ∖ S ′ | , 6 − | S ′ ∖ S | ] of Lemma A17, which forces | S ′ ∖ S | ≤ 1 , and these groups are nonzero by Lemma A17(ii); counting pairs gives 19 + 27 = 46 . (ii) On each block the eigenvalues of t θ − m S lie in { t , t − 2 } ; both are positive for t ∈ { 3 , 4 } and both negative for t ∈ { − 2 , − 1 } . So Ext k ( M S , L t ) = H k ( L t M S − 1 ) is concentrated in degree 0 for t ∈ { 3 , 4 } and in degree 6 for t ∈ { − 2 , − 1 } , and dually Ext k ( L t , M S ) = H k ( L − t M S ) is concentrated in degree 6 for t ∈ { 3 , 4 } and in degree 0 for t ∈ { − 2 , − 1 } ; requiring total degree one selects the listed pairs. (iii) The groups are H k ( L l ′ − l ) with k ∈ { 0 , 6 } and k = 1 + s − s ′ . (iv) On B j the difference is [ j ∈ S ′ ] m j , ζ j ′ − [ j ∈ S ] m j , ζ j , nondegenerate exactly when both terms occur with different kernels, and then of signature ( 1 , 1 ) . For ( Gen Q ) : by the proof of Lemma A17, H * ( L ′ ) = 0 unless L ′ restricts trivially to K ( L ′ ) 0 , and the P with ( L ′ ⊗ P ± 1 ) | K ( L ′ ) 0 trivial form a coset of the kernel of the surjection Pic 0 ( A ) → Pic 0 ( K ( L ′ ) 0 ) , proper since dim K ( L ′ ) 0 ≥ 1 .    □
Lemma  A22
(Koszul cubes). Let a cube Q be Γ δ -convex and suppose that, in the gauge h δ = 0 , the components of δ between pieces of Q are nonzero exactly on the 12 edges S → S ∪ { l } . Then E Q ≅ O Z ′ ⊗ N [ − ( 3 + c ) ] , where Z ′ = ⋂ j f j , ζ j − 1 ( p j ) is a translate of Z ζ , O ( p j ) ≅ O ( 0 ) ⊗ Q j , and N = M { 1 , 2 , 3 } | Z ′ ∈ Pic 0 ( Z ′ ) ; for every δ, ch ( E Q ) = − ( − 1 ) c [ Z ζ ] .
Proof. 
On an edge S → S ∪ { l } , δ = δ H − h ( δ ∘ δ ) , and a term δ S , c δ c , S ∪ l has c ∈ Q by convexity, with both factors on edges that raise | S | by one, impossible. So δ S , S ∪ l ∈ H 0 ( f l * O ( p l ) ) , which is one-dimensional because f l * O A = O E 0 : δ S , S ∪ l = c S , l s l with c S , l ∈ C × and s l the canonical section. The Q × Q block of the Maurer–Cartan equation holds by convexity (proof of Theorem A17), and D δ = 0 on holomorphic sections, so c S , j c S ∪ j , k + c S , k c S ∪ k , j = 0 on every square. The standard signs c S , l 0 = ( − 1 ) # { j ∈ S : j < l } satisfy the same equations, so c / c 0 is a C × -valued cocycle on the 2-skeleton of the cube, the boundary of a ball and simply connected; it is a coboundary λ S ∪ l / λ S , and rescaling the pieces by λ S turns δ | Q into the Koszul complex of the regular sequence ( s 1 , s 2 , s 3 ) , the f l − p l being independent linear coordinates [36, Cor. 17.5], twisted by M { 1 , 2 , 3 } . Each f l is constant on Z ′ , so f l * O ( p l ) | Z ′ is trivial. The Chern character is computed in the proof of Proposition A12(i).    □
Lemma  A23
(Regrouping). Let P be a partition of the pieces such that the quotient of Γ δ by P has no directed cycle. Then E is a convolution of the objects E P , P ∈ P , in a linear order of the blocks, and Setting 13 and Theorem A17 apply to this presentation.
Proof. 
Each edge of the support of δ expands to a path of Γ δ , so the quotient of the support of δ by P has no directed cycle either. Order the blocks by a linear extension and each block internally as before; δ stays strictly upper triangular and the twisted complex is unchanged. Grouping consecutive blocks gives a one-sided twisted complex over twisted complexes, which is E [18, Sects. 1–3], and the E P are perfect.    □
Theorem  A19
(Koszul cubes fail). In Proposition A12, suppose that the hypothesis of Lemma A22 on δ holds for all 32 cubes, and that the quotient of Γ δ by the partition into cubes and single multiples of θ has no directed cycle. Then dim Ext 2 ( E , E ) ≥ 58 , and E does not meet the criterion. The second condition holds for generic twists.
Proof. 
By Lemmas A22 and A23, E is a convolution of F ζ = O Z ζ ′ ⊗ N ζ at the shift 3 + c ( ζ ) and of A , B , C , D . By Grothendieck duality with trivial normal bundle, Ext k ( F ζ , L t ) = H k − 3 ( Z ′ , L t N − 1 ) and Ext k ( L t , F ζ ) = H k ( Z ′ , L − t N ) , and L ± t | Z ′ is definite of sign ± t ; so the first group lives in degree 3 or 6, the second in 0 or 3, and with the shifts − 1 , 0 , 4 , 5 against 3 , 4 the only groups of total degree one are D → F ζ for s F = 3 and F ζ → A for s F = 4 . For two threefolds let d = # { j : ζ j = ζ j ′ } ; Z ζ ′ ∩ Z ζ ′ ′ is empty or a union of translates of an abelian d-fold, and as in Lemma 16(i),(ii), translations and Pic 0 twists changing only global groups, which can only vanish, Ext k ≠ 0 only for k ∈ [ 3 − d , 3 + d ] . Total degree one needs k = 1 + s ζ − s ζ ′ ∈ { 0 , 1 , 2 } ; k = 0 needs d = 3 , impossible, and k = 1 needs equal shifts, hence ∏ ζ = ∏ ζ ′ , so ζ , ζ ′ differ in at least two coordinates and d ≤ 1 . So the edges of the new graph are C → A , C → D , A → B , D → B , D → F @ 3 , F @ 4 → A and F @ 4 → F @ 3 , all increasing in the order C , F @ 4 , D , A , F @ 3 , B . Every F ζ lies on no cycle, so ch ( F ζ ) = ± [ Z ζ ] ∈ Q [ θ ] ⊕ W ( A ) by Theorem A17, against Lemma A16; with Theorem 20, dim Ext 2 ( E , E ) ≥ r ( γ ) + 1 = 58 . For generic twists Γ has no edges between cubes, so a directed cycle of the quotient enters and leaves cubes through the edges of Lemma A21(ii); in each cube every entry piece is contained in every exit piece (entries in { ⌀ , { j } } and exit { 1 , 2 , 3 } for c = 0 , entry ⌀ and exits in { j k , { 1 , 2 , 3 } } for c = 1 ), and the up edges give a path of Γ δ from the entry to the exit, so a cycle of the quotient would lift to one of Γ δ .    □
After Lemma A22 the configuration is that of Setting 5 with h = ( 1 , ⋯ , 1 ) , translated threefolds, Pic 0 twists and line bundles L − 2 , L − 1 , L 3 , L 4 at the shifts − 1 , 0 , 4 , 5 , which do not form an interval; translation and Pic 0 twists can only remove edges, so Lemma 22, Lemma 21(ii) and Proposition 28 give the same conclusion. The collapsed graph is checked in julia/collapse.jl, both for generic twists and in the model in which every window group is nonzero, with 37252 edges in the uncollapsed graph; in that model a single collapsed cube is chain-free for only 4 of the 32 cubes, against 32 for generic twists.
Lemma  A24
(Components). Let K run over the components of the undirected graph underlying Γ δ . Then E = ⨁ K E K , and if E meets the criterion, ch ( E K ) ∈ Q [ θ ] ⊕ W ( A ) for every K .
Proof. 
δ has no component between different K , so the twisted complex is a direct sum. A walk consisting of a path of Γ δ , one edge of Γ and a path of Γ δ that starts in K stays in K up to the edge of Γ and then in the component of its end, so it returns to K only if it never left; K is δ -convex, and Theorem A17 applies.    □
Proposition  A13
(Convex sets of whole cubes). Let C be the union of the cubes in a set U and of some multiples of θ, and suppose C is δ-convex and E meets the criterion. Then U is empty, the set of all 32 cubes, or one of 42 sets of 16 cubes, each with eight cubes with c = 0 and eight with c = 1 ; neither set of cubes of one value of c is among them. This applies to every component of Lemma A24 made of whole cubes and multiples of θ.
Proof. 
By Lemma A22, for every δ , ch ( E C ) = − ∑ Q ∈ U ( − 1 ) c ( Q ) [ Z Q ] + ∑ ± e t θ , and by Theorem A17 its component in H 6 lies in Q θ 3 ⊕ W ( A ) , that is, in C θ 3 + C α + C α ¯ , which meets H 6 ( A , Q ) in Q θ 3 ⊕ W ( A ) . The 35 classes θ 3 , α , α ¯ , [ Z ζ ] span a space of dimension 25, the signed combinations of the [ Z ζ ] lying in C θ 3 + C α + C α ¯ form a space of dimension 10, and its vectors with coefficients in { 0 , ± 1 } of the required signs are the 44 listed sets, by julia/subsets.jl.    □
Theorem  A20
(One generic cube). Let a cube Q satisfy ( Gen Q ) and be Γ δ -convex, δ being otherwise arbitrary, and suppose that c ( Q ) = 0 and Γ δ contains an edge M j k ( Q ) → M { 1 , 2 , 3 } ( Q ) , or c ( Q ) = 1 and Γ δ contains an edge M ⌀ ( Q ) → M j ( Q ) . Then dim Ext 2 ( E , E ) ≥ 58 , and E does not meet the criterion.
Proof. 
Let c = 0 . By Lemma A23, with blocks Q and single pieces, E is a convolution of E Q and the other 252 pieces: a cycle of the quotient through Q would be a path of Γ δ leaving Q and returning. For a piece X ≠ C outside Q, Lemma A21 and ( Gen Q ) give no block Ext * ( M S ( Q ) , X ) in total degree one; Hom • ( E Q , P X ) is ⨁ S A ( M S , X ) , D + Q δ with Q δ the composition with δ | Q , and by the perturbation lemma its cohomology is that of ⨁ S H ( M S , X ) , ∑ k π Q δ ( − h Q δ ) k , which vanishes in total degree one. For C = L 3 [ − 4 ] , Ext k ( M S , L 3 ) = H k ( L 3 M S − 1 ) with 3 θ − m S positive definite, so only k = 0 occurs, in total degree 4 − | S | : nothing in degree 0, V 1 = H 0 ( L 3 M { 1 , 2 , 3 } − 1 ) in degree one and ⨁ | S | = 2 H 0 ( L 3 M S − 1 ) in degree two, so Hom 1 ( E Q , C ) = ker ( d : V 1 → degree two ) . Let Y be the set of y, | y | = 2 , with δ y , { 1 , 2 , 3 } H ≠ 0 , nonempty by hypothesis, and x its last element in the order. The x-component of d ( φ ) , φ ∈ V 1 , is a sum over chains x = c k < ⋯ < c 0 = M { 1 , 2 , 3 } with nonzero δ c i + 1 c i . For k ≥ 2 , δ c 1 , { 1 , 2 , 3 } ≠ 0 gives a path of Γ δ from c 1 to M { 1 , 2 , 3 } , whose last edge comes from some y ∈ Y with y ≥ c 1 > x by Lemma A21(i), against the choice of x. The term with k = 1 is π ( φ ∘ δ x , { 1 , 2 , 3 } ) , and in δ x , { 1 , 2 , 3 } = δ x , { 1 , 2 , 3 } H − h ( δ ∘ δ ) x , { 1 , 2 , 3 } a nonzero δ x , c δ c , { 1 , 2 , 3 } would again give some y ≥ c > x ; so δ x , { 1 , 2 , 3 } = σ is a nonzero holomorphic section of M { 1 , 2 , 3 } ⊗ M x − 1 . Then φ σ is ∂ ¯ -closed of Dolbeault degree 0, so π ( φ σ ) = φ σ , as h vanishes in degree − 1 , and d ( φ ) x = ± φ σ ≠ 0 for φ ≠ 0 , A being integral. So Hom 1 ( E Q , C ) = 0 , E Q has no edge out in the regrouped presentation, and it lies on no cycle; Theorem A17 gives ch ( E Q ) = − [ Z ζ ] ∈ Q [ θ ] ⊕ W ( A ) , against Lemma A16. For c = 1 the argument is dual: no group of total degree one enters Q except from B = L − 1 ; Ext k ( L − 1 , M S ) = H k ( L ⊗ M S ) lives in k = 0 and total degree | S | + 1 , so Hom 1 ( B , E Q ) = ker H 0 ( L P ⌀ ) → ⨁ j H 0 ( L M j ) ; with x the first M j with δ ⌀ , j H ≠ 0 , only the direct term survives and d ( ψ ) x = ± σ ψ ≠ 0 , so E Q has no edge in and lies on no cycle.    □
Theorem A20 allows arbitrary components of δ inside the cube, lateral and downward ones included; without the edge into M { 1 , 2 , 3 } , respectively out of M ⌀ , d can vanish on V 1 . Under ( Gen Q ) , Q is Γ δ -convex for c = 0 when every piece of Q entered by an edge of Γ δ from outside reaches M { 1 , 2 , 3 } along Γ δ , and for c = 1 when M ⌀ reaches every piece of Q with an edge of Γ δ leaving Q; both hold for Koszul cubes.
Lemma  A25
(An isolated top piece). Let Q have c = 0 , satisfy ( Gen Q ) , and have no edge of Γ δ into M { 1 , 2 , 3 } ( Q ) . If E meets the criterion, Γ δ has a path from M { 1 , 2 , 3 } ( Q ) to some M j k ( Q ) .
Proof. 
The edges of Γ into M { 1 , 2 , 3 } ( Q ) come only from the M j k ( Q ) (Lemma A21), and none lies in Γ δ ; so a walk of the form path, edge, path from M { 1 , 2 , 3 } back to it ends with an edge M j k → M { 1 , 2 , 3 } of Γ preceded by a path of Γ δ from M { 1 , 2 , 3 } to M j k . Without such a path { M { 1 , 2 , 3 } } is δ -convex, and Theorem A17 gives e m 1 + m 2 + m 3 ∈ Q [ θ ] ⊕ W ( A ) ; but m 1 + m 2 + m 3 has a hermitian form of rank 3, so it is not in Q θ .    □
Corollary  A6
(The shape of δ ). Suppose the twists are generic and E meets the criterion. Then
(i)
for every cube with c = 0 , Γ δ has a path from M { 1 , 2 , 3 } ( Q ) to some M S ( Q ) with | S | ≤ 2 ;
(ii)
for every cube with c = 1 , either M ⌀ ( Q ) is a sink of Γ δ , or δ B , M ⌀ ( Q ) H ≠ 0 and Γ δ has a path from some M S ( Q ) with | S | ≥ 2 through multiples of θ to M ⌀ ( Q ) .
Proof. 
(i) Without an edge of Γ δ into M { 1 , 2 , 3 } apply Lemma A25. With one, Q is not Γ δ -convex by Theorem A20; under ( Gen Q ) the only edge of Γ leaving Q is M { 1 , 2 , 3 } → C , so the escaping path starts at M { 1 , 2 , 3 } and re-enters at M ⌀ or some M j . (ii) Without an edge of Γ δ out of M ⌀ it is a sink, its only edges in Γ going to the M j . Otherwise Q is not Γ δ -convex by Theorem A20; the only edge of Γ entering Q is B → M ⌀ , and the exits are the M j k and M { 1 , 2 , 3 } .    □
So for generic twists no cube is closed and Koszul-like: every cube with c = 0 sends M { 1 , 2 , 3 } back into itself, through lateral or downward components or through the multiples of θ . Whether some twisting cochain gives dim Ext 2 ( E , E ) = 57 is open in exactly the following cases: generic twists and a cube with c = 0 without an edge of Γ δ into M { 1 , 2 , 3 } , where Lemma A25 gives a path back from M { 1 , 2 , 3 } and Hom 1 ( E Q , C ) can be nonzero; generic twists and a cube with c = 1 whose M ⌀ is a sink of Γ δ ; cubes that are not block-compatible, where a path of Γ δ leaves a cube through multiples of θ and returns, so that the convex hulls contain multiples of θ and parts of other cubes and Proposition A13 restricts only hulls made of whole cubes; and special twists, where groups between cubes with degenerate difference are nonzero and δ may use them, including Koszul cubes that are not block-compatible. Deciding them requires Ext 2 ( E , E ) as a function of δ , that is, the higher products m k , k ≥ 3 , of the theta functions along the cycles: sections composed with lateral classes of degree one and with the classes of degrees six and three on the edges through the multiples of θ . No equivariant structure has been constructed on this configuration, and whether an analogue exists at a point of complex multiplication of a non-split family for Q ( i ) is open.

Appendix A.6. Equivariant Convolutions on E 0 8

Theorem A9 excludes every equivariant convolution of the pieces of Proposition A4. This subsection determines the groups that act, describes the invariant parts of the harmonic groups, and gives larger lower bounds for dim Ext 2 ( E , E ) G in the arrangements with two shifts; they show how the bounds of Theorems A13 to Corollary A3, which rest on off-diagonal groups, change under a group action. Write h = 1 + i 2 , and par ( ζ ) ∈ Z / 2 for the parity, ∏ ζ = ( − 1 ) par ( ζ ) . Fix a piece L 1 among the L ζ . For a ∈ A let g a = t a * ( − ) ⊗ P − ϕ L 1 ( a ) , where ϕ L ( a ) = t a * L ⊗ L − 1 , and let G be the group of the g a that fix every piece up to isomorphism, a subgroup of A × A ∨ .
Proposition  A14
(The group and the invariant groups). Keep Proposition A4, with the L ζ external products of line bundles N j , ζ j on the surfaces B j , and with any twists of degree zero.
(i)
G = ∏ j G j , where G j ⊂ B j is generated by a j = ( h , h ) if all t i − p are odd, and G = 0 otherwise. The group G is isotropic for the commutator pairing of Mukai’s central extension, so it acts genuinely on D perf ( A ) , and its linearisations of a piece form a torsor under G ∨ = ∏ j G j ∨ .
(ii)
Let all t i − p be odd, and give the pieces G-linearisations; on the factor B j the linearisations of L a and L b differ by a character χ a , j χ b , j − 1 of G j . For L a = L ζ and L b = L ζ ′ differing exactly on the set S of surfaces, with D = ∏ j ∈ S | ζ j − ζ j ′ | 2 ,
dim H q ( A , L a − 1 ⊗ L b ) G = D 2 − | S | dim ⨂ j ∉ S H * ( B j , O B j ) q − | S | if χ a , j = χ b , j for all j ∉ S , 0 otherwise .
Between two multiples, and between a bundle L ζ and a multiple M i with | t i − p | ≥ 2 , the invariant part is 2 − n times the whole group, for all linearisations.
(iii)
The diagonal groups H * ( A , O A ) are fixed pointwise by G.
Proof. 
(i) The functor g a fixes L exactly when ϕ L ( a ) = ϕ L 1 ( a ) , that is, when a ∈ K ( L ⊗ L 1 − 1 ) . The forms are orthogonal sums over the surfaces, so the condition splits as a condition on each component a j ∈ B j = C 2 / Z [ i ] 2 : for a line bundle on B j with form H, K consists of the x with H ( x , λ ) ∈ Z [ i ] for all λ ∈ Z [ i ] 2 , because Im ( z μ ) ∈ Z for all μ ∈ Z [ i ] exactly when z ∈ Z [ i ] . For the ratio of two bundles L ζ with ζ j − ζ j ′ = w , the form is 0 w w ¯ 0 and the condition is x ∈ ( w − 1 Z [ i ] ) 2 ; the values | w | 2 = 2 occur, and give a j ∈ { 0 , h } 2 . Every unit multiple of h is h modulo Z [ i ] . For M i ⊗ L ζ − 1 , with d = t i − p , the conditions read d x 1 ≡ x 2 and x 1 ≡ d x 2 modulo Z [ i ] ; for x ∈ { 0 , h } 2 they give x = 0 if d is even and x ∈ { 0 , ( h , h ) } if d is odd. For M k ⊗ M i − 1 the condition is ( t k − t i ) x ∈ Z [ i ] 2 , which ( h , h ) satisfies when t k − t i is even. This gives G. The commutator pairing of two elements ( a , − ϕ L 1 ( a ) ) and ( b , − ϕ L 1 ( b ) ) is e L 1 ( a , b ) ± 1 = exp ( ± 2 π i Im H 1 ( a , b ) )  [84]; it is 1 for a , b in different factors, the form being an orthogonal sum, and for a = b = a j , since H 1 ( a j , a j ) is real. So G is isotropic, the restricted extension of G by C × is abelian and splits, and G acts genuinely; two splittings, and two linearisations of a piece, differ by a character of G.
(ii) The group G = ∏ G j acts factorwise on H * ( A , L a − 1 ⊗ L b ) = ⨂ j H * ( B j , N a , j − 1 ⊗ N b , j ) , and the invariants of an external tensor product under a product group are the tensor product of the invariants. On j ∉ S the factor is H * ( B j , O B j ) , on which translations act trivially, so G j acts by the scalar χ a , j χ b , j − 1 . On j ∈ S the form is nondegenerate, the factor is H 1 , of dimension | w | 2 , and it is the irreducible representation of weight one of the theta group of N = N a , j − 1 ⊗ N b , j  [86, Section 1]; the generator of G j lies in K ( N ) by (i) and acts through a lift a ˜ j with a ˜ j 2 = 1 . Since the commutator pairing of N is perfect on K ( N ) , some x ∈ K ( N ) has e N ( a j , x ) = − 1 , and a lift of x conjugates a ˜ j to − a ˜ j ; so the eigenvalues ± 1 of a ˜ j have equal multiplicities | w | 2 / 2 . This gives the formula. Between multiples, and between L ζ and M i with | t i − p | ≥ 2 , the forms are nondegenerate on every surface, and the same argument gives the factor 1 2 on each of the n surfaces.
(iii) Translations act trivially on H * ( A , O A ) , and the linearisation of a piece acts on its own endomorphisms by conjugation, which is trivial on scalar classes.    □
For the pieces with all t i − p odd, the groups H 3 ( L ζ , L ζ ′ ) of Theorem A13, of dimension 16 and 64, have invariant parts of dimension 2 and 8 when the characters agree on the fourth surface, and 0 otherwise; the same factor 2 − | S | applies to every off-diagonal group used in Theorems A12 and A13 to Corollary A3, whose bounds therefore do not carry over to Ext 2 ( E , E ) G . The diagonal does carry over, by (iii). The equivariant problem is an ordinary problem on an isogenous abelian variety.
Proposition  A15
(The equivariant count is an ordinary count). Let all t i − p be odd, K 0 = ⨁ j 〈 a j 〉 ⊂ A , A ′ = A / K 0 = ∏ j B j / 〈 a j 〉 and π : A → A ′ the quotient.
(i)
K 0 is stable under Z [ i ] , so A ′ is of Weil type for Q ( i ) , and π * identifies H * ( A ′ , Q ) with H * ( A , Q ) compatibly with the Hodge structures and the action of Q ( i ) ; in particular W ( A ′ ) corresponds to W ( A ) .
(ii)
Ψ = ( − ) ⊗ L 1 − 1 satisfies Ψ ∘ g a ≅ t a * ∘ Ψ for a ∈ K 0 . The G-equivariant objects of D b ( A ) correspond to objects E ′ of D b ( A ′ ) with π * E ′ ≅ Ψ ( E ) , and Ext A * ( E , E ) G = Ext A ′ * ( E ′ , E ′ ) .
(iii)
Under (ii) the linearisations of a piece correspond to the twists of its descent by the line bundles of order two in ker Pic 0 ( A ′ ) → Pic 0 ( A ) .
Proof. 
(i) i a j = i − 1 2 , i − 1 2 ≡ a j modulo Z [ i ] 2 , so 〈 a j 〉 is stable under Z [ i ] , and π is an isogeny of degree 16 commuting with Z [ i ] ; an isogeny induces an isomorphism of rational Hodge structures. (ii) t a * ( F ⊗ L 1 ) ⊗ P − ϕ L 1 ( a ) ⊗ L 1 − 1 ≅ t a * F because ϕ L 1 ( a ) = t a * L 1 ⊗ L 1 − 1 . The genuine action of G is unique up to a character (Proposition A14(i)), so after a twist it is the transport by Ψ of the action of K 0 by translations. The group K 0 acts freely, so K 0 -equivariant coherent sheaves are sheaves on A ′  [88, §7, Proposition 2], the derived category of equivariant sheaves is the equivariant derived category [37], and since | G | is invertible the morphisms in the equivariant category are the invariants of the morphisms. (iii) Two linearisations differ by a character of K 0 , and the characters of K 0 are the line bundles in the kernel of π * on Pic 0 .    □
The identification in (ii) uses the twist by L 1 , which exists at the single point A; the deformation statement along the family is the equivariant one of Corollary 11. For the counts below the lemma on unit leaves is needed; a leaf is a unit on B j if its component there is a multiple of the unit section of O B j .
Lemma  A26
(Unit leaves). In the situation of Lemma A6, let a term of Lemma A5(iii)have k ≥ 3 leaves, and suppose that one leaf ℓ is, on every surface where it is not a unit, the only leaf that is not a unit. Then the term is zero.
Proof. 
Let J be the set of surfaces where ℓ is not a unit. For j ∈ J , an edge carrying h j has at most one leaf below it that is not a unit on B j , so the term vanishes by the proof of Lemma A6 unless no edge carries h j . Then every value along the tree has a component on ∏ j ∈ J B j that is harmonic, equal to that of ℓ or to a unit, and on such forms G, π and the h i , i ∉ J , act through the remaining factors. So the term is ± ℓ J ⊗ T ′ , where T ′ is the same tree with ℓ replaced by a unit. In T ′ the unit composes with its sibling X to give X; if the parent is the root, X is not a leaf because k ≥ 3 , so X = h Y and π h Y = 0 ; otherwise the edge above the parent applies h to X, which is 0 if X is a harmonic leaf and h 2 Y = 0 if X = h Y .    □
Proposition  A16
(The diagonal at n = 4 ). Keep Lemma A11, and let E be a convolution of the 131 pieces in any order, with any shifts of the given parities and, when all t i − p are odd, any G-linearisation. Let D 2 = ⨁ a H 2 ( A , O A ) , of dimension 131 · 28 = 3668 .
(i)
dim Ext 2 ( E , E ) G ≥ dim ker ( d E | D 2 ) .
(ii)
The cup products have a kernel of dimension at least 484 = 84 + 4 · 4 · 1 + 6 · 16 · 4 on D 2 .
(iii)
Every term of d E on D 2 with k ≥ 3 leaves either runs through bundles L ζ only, along strictly decreasing shifts, or has exactly four leaves along a path of one of the following types, according to the placement: M 2 → M 3 → L → M 1 if p < t 1 < t 2 < t 3 ; M 2 → M 3 → M 1 → L or M 3 → M 1 → L → M 2 if t 1 < p < t 2 < t 3 ; L → M 3 → M 1 → M 2 or M 2 → L → M 3 → M 1 if t 1 < t 2 < p < t 3 ; M 3 → L → M 1 → M 2 if t 1 < t 2 < t 3 < p . A convolution carries at most one of these types, and all the bundles L on its paths have the same shift, so their number m is at most 64.
(iv)
If no term of the first kind occurs, dim ker ( d E | D 2 ) ≥ dim ker ( cup ) − 84 − 28 m .
(v)
Terms of the first kind vanish on the classes of the four types with a component of degree two on one surface.
Proof. 
(i) By Theorem A8 coboundaries have no diagonal components, and by Proposition A14(iii) the diagonal classes are invariant; since | G | is invertible, Ext 2 ( E , E ) G is the cohomology of the invariant part of ( H E , d E ) , and the argument of Proposition A5 applies to it. (ii) For y ∈ D 2 the cup products are ( y a − y b ) ∧ δ a b H . Decompose y a by its type τ ∈ { 0 , 1 , 2 } 4 , | τ | = 2 : four types ( 2 , 0 , 0 , 0 ) of dimension one and six types ( 1 , 1 , 0 , 0 ) of dimension four. The multiples differ from every other piece on every surface, where δ H has its only degree, so the 84 diagonal classes of the multiples lie in the kernel. Between two bundles L ζ , L ζ ′ the product with a τ -class can be nonzero only if τ vanishes on the surfaces where they differ. For fixed τ join ζ and ζ ′ when they agree on the support of τ ; the τ -classes constant on the components of this graph lie in the kernel, and the components are the sets of pieces with fixed coordinates on the support, 4 for the types ( 2 , 0 , 0 , 0 ) and 16 for the types ( 1 , 1 , 0 , 0 ) . This gives the bound, which Program cupkernel confirms. (iii) A term acting on y at a piece runs along distinct pieces s = a 0 → ⋯ → a K = e , K ≥ 2 , and its value lies in H ( s , e ) in Ext degree 3 + ϵ , ϵ = 7 D − U + λ , where U and D are the numbers of steps up and down the value ϕ and λ ≥ 0 is the total excess of the steps between bundles L ζ , each at least 1 (Lemma A11). If only bundles L ζ occur, each step lowers the shift (Lemma A11(i)). Otherwise H ( s , e ) is nonzero only in degree 0 or 8 when ϕ ( s ) ≠ ϕ ( e ) , and in degrees [ m ′ , 8 − m ′ ] when s , e are bundles differing in m ′ ≥ 1 coordinates; an exhaustive enumeration of the chains through multiples that pass this test (Program mchains; at most nine steps, since D ≤ 1 , U ≤ 5 , λ ≤ 3 ) leaves, besides the six types, chains containing two consecutive steps up a → b → c with b a multiple whose value is the only value strictly between ϕ ( a ) and ϕ ( c ) . Such a pair is the only chain of total excess − 2 from a to c, so the component of π ( δ ∘ δ ) = 0 in H 0 ( a , c ) reads x a b x b c = 0 , a product of holomorphic sections on the connected A; one factor vanishes and so does the term. The six survivors have λ = 0 and ϵ = 5 . Their steps have Ext degrees 0, 8, 0 in some order, which fixes the shift of L in terms of those of the multiples; the two types for one placement order two multiples oppositely, and the order of the pieces is fixed. (iv) The terms of the second kind depend on y only through its components at the multiples and at the m bundles on their paths. (v) Let τ 1 = 2 . On B 1 the output has degree o 1 = 2 + s ′ − d 1 ≤ 2 , where s ′ is the sum of the degrees on B 1 of the other leaves that are not units there, each at least 1; Lemma A6 gives d 1 ≤ max ( 0 , u 1 − 2 ) , so u 1 ≥ 2 would give o 1 ≥ 3 . Hence y is the only leaf that is not a unit on B 1 , it is a unit on the other surfaces, and Lemma A26 applies.    □
Theorem  A21
(Two shifts, every linearisation). Keep Proposition A16, and suppose that the bundles L ζ with ∏ ζ = ϵ have the shift s and the others the shift s − g , with g odd and g ≠ ± 3 . Then for every G-linearisation, every order, all shifts of the M i and every placement of the t i ,
dim Ext 2 ( E , E ) G ≥ 120 ( g = ± 1 ) , dim Ext 2 ( E , E ) G ≥ 1792 ( | g | ≥ 5 ) .
Proof. 
A term of the first kind in Proposition A16(iii) needs three distinct shifts, so by (iv) there dim ker ( d E | D 2 ) ≥ dim ker ( cup ) − 84 − 28 · 64 . Between two bundles L ζ a component of degree one has Ext degree at least two, so it cannot join pieces of one shift (degree one would need one differing coordinate and equal parities) and it goes from the higher shift to the lower one, in Ext degree | g | + 1 , with degree 1 on each of the | S | surfaces where the pieces differ.
If | g | ≥ 5 , a cup product of such a component with a τ -class needs τ j = 0 on S and τ j + e j ≤ 2 on the other surfaces, where e j is the degree of the component there; summing, 2 + ( | g | + 1 − | S | ) ≤ 2 ( 4 − | S | ) , impossible for | S | ≥ 1 . So the cup products vanish on D 2 , and dim Ext 2 ( E , E ) G ≥ 3668 − 84 − 1792 = 1792 .
Let g = ± 1 ; after renaming, ϵ is at s and the other parity at s − 1 . For a pair ( a , b ) , a at s and b at s − 1 , differing in exactly three coordinates and agreeing on B j , the classes of degree two lie in H 3 ( a , b ) ; by Theorem A13(i),(ii), whose proofs use only the shifts and apply to invariant δ , d E vanishes on them and no coboundary reaches them. By Proposition A14(ii) the invariant part has dimension D / 8 ∈ { 2 , 8 } if χ a , j = χ b , j and 0 otherwise. For j and z ∈ μ 4 the slice ( j , z ) is the set of the 32 pieces with ζ j = z , with the graph whose edges are the pairs of opposite parity differing in the other three coordinates, weighted by D / 8 ; each pair above lies in exactly one slice. Put f j ( x ) = χ x , j + par ( x ) in Z / 2 . Since the parities on an edge differ, the edge carries invariant classes exactly when f j takes different values at its ends, so the invariant classes in a slice have dimension the weight of the cut defined by f j . Program slices shows that the 16 slice graphs are 7-regular and connected, with weighted degree 20 and weighted minimum cut 20 [113]. So a slice on which f j is not constant contributes at least 20 invariant classes. A nonzero invariant component of δ between bundles joins pieces of opposite parities with | S | ≤ 2 and equal characters on every agreeing surface k (Proposition A14(ii)), so the slice ( k , ζ k ) of its ends is not constant. Let Σ be the set of nonconstant slices and N = | Σ | ; the cup products are then supported on pairs all of whose agreeing slices lie in Σ . Coboundaries have no components on D 2 or on the groups H 3 above, so the diagonal kernel and the invariant classes in those groups inject together into Ext 2 ( E , E ) G , and
dim Ext 2 ( E , E ) G ≥ 20 N + max 0 , κ ( Σ ) − 1876 ,
where κ ( Σ ) = 84 + ∑ τ dim ( τ ) c τ ( Σ ) bounds the kernel of the cup products from below as in Proposition A16(ii): c τ ( Σ ) is the number of components of the graph on the bundles L ζ whose edges are the pairs of opposite parity on which a component of Ext degree two can multiply a τ -class nontrivially ( τ = 0 on S, and | S | = 2 , or | S | = 1 with τ k ≤ 1 on some agreeing surface k) and all of whose agreeing slices lie in Σ . Program twolevel, over the 6885 sets Σ with N ≤ 5 , finds the minima 3668, 3668, 3632, 3545, 3428, 3311 of κ for N = 0 , ⋯ , 5 , so the bound is at least 1792, 1812, 1796, 1729, 1632, 1535; for N ≥ 6 the slices alone give 120.    □
For the linearisation χ x , j = par ( x ) every f j is constant: all 40960 classes of Theorem A13 have zero invariant part, and every component of δ between bundles L ζ vanishes, so E joins the L ζ only through the multiples and the diagonal gives dim Ext 2 ( E , E ) G ≥ 1792 . For a nontrivial group the off-diagonal bounds of the ordinary case are not available, and the two-shift bound rests on the diagonal and the cuts of the slices; Theorem A9 rests on the diagonal alone and covers every arrangement, including g = ± 3 and three or more shifts.

Appendix A.7. The Count for Sextic CM Fields

Setting  14
(The flat secant space of a sextic CM field). Let F = F 0 ( − q ) be a sextic CM field with totally real cubic subfield F 0 , real places τ 1 , τ 2 , τ 3 and q ∈ F 0 totally positive, and let X be a principally polarised abelian sixfold with real multiplication by the ring of integers of F 0 , so that H 1 ( X , C ) = U 1 ⊕ U 2 ⊕ U 3 with U j the eigenspace of τ j , of dimension four and Hodge type ( 2 , 2 ) . Let θ j ∈ H 1 , 1 ( X ) be the component of the polarisation on U j , so that θ = ∑ j θ j , θ j 3 = 0 and ∫ X θ 1 2 θ 2 2 θ 3 2 = 8 . For t ∈ F 0 and ε ∈ { ± 1 } 3 put s j = − 1 τ j ( q ) and e ε = exp ∑ j ( τ j ( t ) + ε j s j ) θ j , and let S ( t , q ) be the span of the eight e ε . The Galois group permutes the e ε , so the span is a rational subspace of the algebra generated by the θ j ; since θ j 3 = 0 , e ± s j θ j = B j ± s j θ j with B j = 1 − 1 2 τ j ( q ) θ j 2 , so the eight classes are independent and dim S ( t , q ) = 8 ; and S ( t , q ) = e ∑ j τ j ( t ) θ j S ( 0 , q ) . A real class v = ∑ ε w ε e ε has w − ε = w ¯ ε . Write N w for the number of nonzero w ε ; A j j ′ for the number of sign pairs ( σ , σ ′ ) such that w ε ≠ 0 for some ε with ε j = σ and ε j ′ = σ ′ ; ρ j for the rank of the 2 × 4 matrix ( w ε ) whose rows are indexed by ε j and whose columns by the two other signs; and, for { j , j ′ , j ′ ′ } = { 1 , 2 , 3 } and σ ′ = ± 1 , M σ ′ ( j ; j ′ ) for the 2 × 2 matrix ( w ε ) ε j , ε j ′ ′ with ε j ′ = σ ′ fixed. Let r k ( v ) be the rank of the contraction H T k ( X ) → H * ( X , C ) , ξ ↦ ξ ⌟ v , where H T k ( X ) is the part of degree k of H H * ( X ) in Setting 9, and write χ ( v , v ′ ) = ∫ X v ∨ v ′ , where v ∨ is v with the sign of its components of degree 2 mod 4 changed. For c 0 , c 3 ∈ Q and f , g ∈ F 0 let
v ( c 0 , f , g , c 3 ) = c 0 B 1 B 2 B 3 + ∑ j τ j ( f ) θ j B k B l + ∑ j τ j ( g ) B j θ k θ l + c 3 θ 1 θ 2 θ 3 ,
where { j , k , l } = { 1 , 2 , 3 } in each sum.
Proposition  A17
(The count in degree six). In Setting 14, let v ∈ S ( t , q ) be a nonzero real class.
(i)
r 1 ( v ) = 12 ,
r 2 ( v ) = 4 ( A 12 + A 13 + A 23 ) + ρ 1 + ρ 2 + ρ 3 , r 3 ( v ) = 8 N w + 2 ∑ j ≠ j ′ rk M + ( j ; j ′ ) + rk M − ( j ; j ′ ) .
So r 2 ( v ) ≤ 54 and r 3 ( v ) ≤ 112 , with equality when every w ε is nonzero and every M σ ′ ( j ; j ′ ) is invertible, which holds outside a proper Zariski closed subset of S ( t , q ) . When the support of w is a pair { ε 0 , − ε 0 } , so that N w = 2 , one has ( r 2 , r 3 ) = ( 30 , 40 ) ; this is the case of a class in the secant plane of an imaginary quadratic subfield of F.
(ii)
χ ( v , v ) = ∫ X v ∨ v = − 64 N F 0 / Q ( q ) ∑ ε | w ε | 2 < 0 . For a CM field of degree 2 m , with X of dimension 2 m and 2 m exponentials, the same computation gives ( − 4 ) m N ( q ) ∑ ε | w ε | 2 : the Euler form on the secant space is definite, of sign ( − 1 ) m .
(iii)
Let F be a perfect complex on X with ch ( F ) = v and put e k = dim Ext k ( F , F ) . Then e k ≥ r k ( v ) for every k. If e 0 = 1 , e 1 = 12 and e 2 = r 2 ( v ) , the least values, then e 3 = 2 r 2 ( v ) − 22 − χ ( v , v ) , and this is compatible with e 3 ≥ r 3 ( v ) exactly when
χ ( v , v ) ≤ − r 3 ( v ) − 2 r 2 ( v ) + 22 ,
thecondition of the count: χ ( v , v ) ≤ − 26 when ( r 2 , r 3 ) = ( 54 , 112 ) , and χ ( v , v ) ≤ − 2 when N w = 2 .
Proof. 
(i) Choose a basis x j 1 , x j 2 of U j 1 , 0 and y j 1 , y j 2 of U j 0 , 1 with θ j = x j 1 y j 1 + x j 2 y j 2 , and let ω j = y j 1 y j 2 span ⋀ 2 U j 0 , 1 . Then H * ( X , C ) and H T * ( X ) = ⋀ * ( H 0 , 1 ⊕ T ) are the tensor products over j of the corresponding algebras of the three factors, H T k ( X ) = ⨁ a 1 + a 2 + a 3 = k H T 1 a 1 ⊗ H T 2 a 2 ⊗ H T 3 a 3 , and contraction is compatible with the factorisation; v = ∑ ε w ε ⨂ j e λ j ( ε j ) θ j with λ j ( ± 1 ) = τ j ( t ) ± s j , two distinct numbers. The blocks of H T k ( X ) shift the three Hodge bidegrees of H * ( X ) by different amounts, so their images are independent and r k ( v ) is the sum of their ranks. Inside a block the bookkeeping rests on two facts about one factor: the two classes e λ j ( ± 1 ) θ j are linearly independent, and so are the subspaces U j 0 , 1 e λ j ( ± 1 ) θ j of H * ( U j ) , since u e λ θ j = u + λ u θ j for u ∈ U j 0 , 1 ; whereas ω j e λ θ j = ω j for every λ , so that the summands of a block through H T j 2 share the factor ω j and are combined, not juxtaposed. On one factor, contraction into e λ θ j is as follows. For ξ ∈ H T j 1 , ξ ⌟ e λ θ j = y λ ( ξ ) e λ θ j with y λ ( ξ ) = ξ 0 , 1 + λ ξ T ⌟ θ j ∈ U j 0 , 1 , because contraction with a tangent vector is a derivation; y λ is onto, its kernel K j λ has dimension two, and H T j 1 = K j λ j ( + 1 ) ⊕ K j λ j ( − 1 ) because λ j ( + 1 ) ≠ λ j ( − 1 ) and ξ T ⌟ θ j determines ξ T . For ξ ∈ H T j 2 , since ω j θ j = 0 , one finds ω j ⌟ e λ θ j = ω j , ( y j a ⊗ ∂ x j b ) ⌟ e λ θ j = ± λ ω j for a ≠ b and 0 for a = b , and ( ∂ x j 1 ∧ ∂ x j 2 ) ⌟ e λ θ j = ± λ 2 ω j : the image of H T j 2 is the line C ω j , ξ ⌟ e λ θ j = c ( ξ , λ ) ω j , and the pairs c ( ξ , λ j ( + 1 ) ) , c ( ξ , λ j ( − 1 ) ) fill C 2 . For k ≥ 3 , H T j k kills e λ θ j , because every element of it contains ω j or ∂ x j 1 ∧ ∂ x j 2 together with a further factor, and the further factor kills ω j .
Now count. The block H T j 1 acts on v by ξ ↦ ∑ ε w ε y λ j ( ε j ) ( ξ ) ⨂ j ′ e λ j ′ ( ε j ′ ) θ j ′ ; on the summand K j λ j ( − σ ) only the ε with ε j = σ contribute, and they map it isomorphically onto U j 0 , 1 tensored with a nonzero class as soon as one such w ε is nonzero. Both signs occur in the support of a nonzero real w, since w − ε = w ¯ ε , so the block has rank 4 and r 1 ( v ) = 12 . The block H T j 1 ⊗ H T j ′ 1 is the sum of the four summands K j λ j ( − σ ) ⊗ K j ′ λ j ′ ( − σ ′ ) , each of dimension four; the summand indexed by ( σ , σ ′ ) maps isomorphically onto U j 0 , 1 ⊗ U j ′ 0 , 1 tensored with the class ∑ σ ′ ′ w ε e λ j ′ ′ ( σ ′ ′ ) θ j ′ ′ , which is nonzero exactly when the pair ( σ , σ ′ ) occurs in the support, and to zero otherwise; so the block has rank 4 A j j ′ . The block H T j 2 maps ξ to ∑ σ c ( ξ , λ j ( σ ) ) z σ with z σ = ω j ⊗ ∑ μ w σ μ e μ , where μ runs over the four sign patterns of the other two places and the e μ are independent; as ( c ( ξ , λ j ( + 1 ) ) , c ( ξ , λ j ( − 1 ) ) ) fills C 2 , the rank is the dimension of the span of z + , z − , which is ρ j . This gives r 2 ( v ) . In degree three the block H T 1 1 ⊗ H T 2 1 ⊗ H T 3 1 is the sum of eight summands ⨂ j K j λ j ( − ε j ) , of dimension eight, on each of which exactly one term of v acts, isomorphically when w ε ≠ 0 and by zero otherwise: rank 8 N w . The block H T j 2 ⊗ H T j ′ 1 factors through C 2 ⊗ H T j ′ 1 by c ⊗ id , which is onto, and on C 2 ⊗ K j ′ λ j ′ ( − σ ′ ) the map is M σ ′ ( j ; j ′ ) ⊗ y λ j ′ ( σ ′ ) followed by tensoring with independent classes, of rank 2 rk M σ ′ ( j ; j ′ ) ; the two values of σ ′ give independent images. The blocks H T j 3 contribute nothing. This gives r 3 ( v ) . The maximal values are 48 + 6 = 54 and 64 + 6 · 8 = 112 ; the conditions for them are the nonvanishing of the w ε and of the determinants of the M σ ′ ( j ; j ′ ) , which are polynomials in the coordinates of v, nonzero on S ( t , q ) . If the support of w is a pair { ε 0 , − ε 0 } , then A j j ′ = 2 , ρ j = 2 , N w = 2 and every M σ ′ ( j ; j ′ ) has one nonzero entry, so ( r 2 , r 3 ) = ( 24 + 6 , 16 + 24 ) .
(ii) Since θ j 3 = 0 , θ 2 m = ( 2 m ) ! 2 m θ 1 2 ⋯ θ m 2 , and ∫ X θ 2 m / ( 2 m ) ! = 1 gives ∫ X θ 1 2 ⋯ θ m 2 = 2 m and ∫ X exp ( ∑ j c j θ j ) = ∏ j c j 2 . The dual of e λ θ j is e − λ θ j , so ∫ X e ε ∨ e ε ′ = ∏ j λ j ( ε j ′ ) − λ j ( ε j ) 2 , which is 0 unless ε ′ = − ε , where it is ∏ j ( 2 s j ) 2 = ( − 4 ) m N ( q ) . Hence χ ( v , v ) = ( − 4 ) m N ( q ) ∑ ε w ε w − ε = ( − 4 ) m N ( q ) ∑ ε | w ε | 2 .
(iii) The evaluation map and the semiregularity map of Setting 9 in degree k satisfy σ F ∘ ev F = ( · ) ⌟ ch ( F ) , so the kernel of ev F : H T k ( X ) → Ext k ( F , F ) lies in the annihilator of v and e k ≥ r k ( v ) . Serre duality on the sixfold gives e k = e 6 − k , so χ ( F , F ) = 2 ( e 0 − e 1 + e 2 ) − e 3 = 2 r 2 ( v ) − 22 − e 3 for the least values, and by Riemann–Roch on an abelian variety χ ( F , F ) = ∫ X v ∨ v = χ ( v , v ) ; this determines e 3 , and e 3 ≥ r 3 ( v ) is the stated inequality. With the maximal ranks it reads χ ≤ − ( 112 − 108 + 22 ) = − 26 , and with N w = 2 it reads χ ≤ − ( 40 − 60 + 22 ) = − 2 .    □
Proposition  A18
(Integral flat characters in degree six). Keep the notation of Proposition A17.
(i)
For every perfect complex F on a complex torus of dimension six, χ ( F , F ) is even. So the condition of Proposition A17(iii) holds for every F whose character has N w = 2 .
(ii)
For every nonzero real v ∈ S ( t , q ) , r 3 ( v ) − 2 r 2 ( v ) ≤ 4 , with equality only if ( r 2 ( v ) , r 3 ( v ) ) = ( 54 , 112 ) . So for every shape the condition of Proposition A17(iii) is implied by χ ( F i , F i ) ≤ − 26 .
(iii)
The rational points of S ( 0 , q ) are the classes v ( c 0 , f , g , c 3 ) , and
χ ( v , v ) = − 8 N ( q ) c 0 2 + Tr N ( q ) q − 1 f 2 + Tr ( q g 2 ) + c 3 2 .
(iv)
For q = 1 , the classes Re e − 1 θ = v ( 1 , 0 , − 1 , 0 ) and Im e − 1 θ = v ( 0 , 1 , 0 , − 1 ) are integral points of S ( 0 , 1 ) with χ = − 32 and N w = 2 , so they satisfy the condition of Proposition A17(iii).
Proof. 
(i) The Chern character v = ch ( F ) of a perfect complex on a complex torus is integral, since by the Künneth theorem the topological K-theory of a real torus is generated by products of classes pulled back from its circle factors, whose Chern characters are integral; and χ ( F , F ) = ∫ X v ∨ v by Riemann–Roch. Write v = ∑ k v 2 k , so that ∫ X v ∨ v = ∑ k ( − 1 ) k ∫ X v 2 k v 12 − 2 k . The terms k and 6 − k are equal, since ( − 1 ) k = ( − 1 ) 6 − k and classes of even degree commute, so they contribute an even number when k ≠ 3 . For k = 3 write v 6 = ∑ I a I e I in the basis of ⋀ 6 H 1 ( X , Z ) given by the products e I of six elements of a basis of H 1 ( X , Z ) ; since e I e I = 0 and e I e J = e J e I , v 6 2 = 2 ∑ I < J a I a J e I e J is even. For N w = 2 the condition of Proposition A17(iii) is χ ( F , F ) ≤ − 2 , and χ ( F , F ) is a negative even integer by Proposition A17(ii).
(ii) For j and σ = ± 1 let W j , σ be the 2 × 2 slice ( w ε ) ε j = σ , with rows and columns indexed by the signs at the two other places. The matrices M σ ( j ′ ; j ) are W j , σ and its transpose, so ∑ j ≠ j ′ rk M + ( j ; j ′ ) + rk M − ( j ; j ′ ) = 2 ∑ j , σ rk W j , σ . Since w − ε = w ¯ ε , W j , − is the complex conjugate of W j , + with its rows and its columns reversed, so the two slices have the same rank k j . Let S be the support of w and N = N w = | S | . By Proposition A17(i),
r 3 ( v ) − 2 r 2 ( v ) = 8 N − 8 ( A 12 + A 13 + A 23 ) + ∑ j 8 k j − 2 ρ j .
The set S is stable under ε ↦ − ε , which exchanges the two slices along each place; so N is even, each slice has N / 2 nonzero entries, k j ≥ 1 , and A j j ′ ∈ { 2 , 4 } , the projection of S to the signs at j and j ′ being stable under negation. Also 1 ≤ ρ j ≤ 2 and k j ≤ 2 , and a slice whose nonzero entries lie in one row or one column has rank one. If ρ j = 1 , then w ε = a ε j W ( ε ′ ) , with ε ′ the signs at the two other places, so every slice along a place j ′ ≠ j is a ⊗ ( a row of W ) and k j ′ ≤ 1 . Now take the four values of N in turn.
If N = 8 , every A j j ′ is 4 and (A6) is − 32 + ∑ j ( 8 k j − 2 ρ j ) . When every ρ j is 2 this is at most − 32 + 3 · 12 = 4 , with equality exactly when every k j is 2, and then r 2 = 48 + 6 = 54 and r 3 = 64 + 48 = 112 . When some ρ j is 1, the two other k are at most 1 and (A6) is at most − 32 + 14 + 6 + 6 = − 6 .
If N = 6 , S misses one pair { ± ε 0 } . As ε 0 and − ε 0 differ at every place, they never have the same signs at two places, so every sign pair at two places occurs in S, every A j j ′ is 4, and (A6) is at most 48 − 96 + 3 · 14 = − 6 .
If N = 4 , S = { ± ε , ± ε ′ } , and after replacing ε ′ by − ε ′ if necessary, ε and ε ′ differ at exactly one place j. For the two other places k and l one has A k l = 2 and A j k = A j l = 4 . The two nonzero entries of a slice along k are w ± ε and w ± ε ′ with the same sign, whose indices agree at l; they lie in one row or one column, so k k = 1 , and likewise k l = 1 . Hence (A6) is at most 32 − 80 + 8 ( 2 + 1 + 1 ) − 6 = − 22 .
If N = 2 , S = { ± ε 0 } , every A j j ′ is 2, every k j is 1, and every ρ j is 2, the two slices along j having their nonzero entries at different positions; so (A6) is 16 − 48 + 24 − 12 = − 20 , the value 40 − 60 of Proposition A17(i).
So r 3 − 2 r 2 ≤ 4 , with equality only at ( 54 , 112 ) , and the threshold r 3 − 2 r 2 + 22 of Proposition A17(iii) is at most 26.
(iii) Since θ j 3 = 0 and s j 2 = − τ j ( q ) , e ± s j θ j = B j ± s j θ j , so the eight exponentials and the eight products ∏ j β j with β j ∈ { B j , θ j } span the same complex space, the change of basis being the tensor cube of an invertible 2 × 2 matrix. The Galois group permutes the τ j , the θ j and the B j compatibly, so ∑ a C a ∏ j β j , indexed by the subsets a of { 1 , 2 , 3 } with β j = θ j for j ∈ a , is rational exactly when C π ( a ) = σ ( C a ) for every σ inducing π ; the stabiliser of { j } , and of its complement, is the stabiliser of τ j , which gives the form v ( c 0 , f , g , c 3 ) . The involution v ↦ v ∨ and the integral are multiplicative over the three factors, with ∫ X θ 1 2 θ 2 2 θ 3 2 = 8 ; on one factor B j ∨ B j = 1 − τ j ( q ) θ j 2 , θ j ∨ θ j = − θ j 2 and the mixed products are odd in θ j . So χ ( v , v ) = − 8 ∑ a C a 2 ∏ j ∉ a τ j ( q ) , which is the formula.
(iv) For q = 1 , s j = − 1 and e ( 1 , 1 , 1 ) = ∏ j e − 1 θ j = e − 1 θ . The polarisation is principal, so each θ k / k ! is integral, and so are the real and imaginary parts ∑ k ( − 1 ) k θ 2 k / ( 2 k ) ! and ∑ k ( − 1 ) k θ 2 k + 1 / ( 2 k + 1 ) ! . They are 1 2 ( e ( 1 , 1 , 1 ) + e ( − 1 , − 1 , − 1 ) ) and 1 2 − 1 ( e ( 1 , 1 , 1 ) − e ( − 1 , − 1 , − 1 ) ) , so N w = 2 and χ = − 64 ( 1 4 + 1 4 ) = − 32 by Proposition A17(ii); expanding ∏ j ( B j + − 1 θ j ) gives the coordinates.    □

Appendix A.8. The Rules of the Closure Graph

The statements (L), (M), (V) and (F3) are those of Section 3.3. Recall André’s definition [6]: a class ξ ∈ H 2 p ( X , Q ) on a smooth projective variety X is motivated if there are a smooth projective variety Y and algebraic classes α , β on X × Y with ξ = pr X * ( α ∪ * β ) , where * is the Lefschetz involution of X × Y for a product polarisation. An algebraic class is motivated (take Y a point and β = ± h dim X ).
Proposition  A19
(The conjecture for the varieties that are not abelian is the conjecture). Let A be a complex abelian variety of dimension g ≥ 1 and let X = A × P 1 . Then X is a smooth complex projective variety that is not an abelian variety, and if every Hodge class of degree 2 p on X is algebraic, then so is every Hodge class of degree 2 p on A. Consequently the statement
the Hodge conjecture holds for every smooth complex projective variety that is not an abelian variety
implies the Hodge conjecture for every abelian variety, and it is equivalent to HC .
Proof. 
X is smooth and projective of dimension g + 1 . By the Künneth formula h 1 , 0 ( X ) = h 1 , 0 ( A ) + h 1 , 0 ( P 1 ) = g , while an abelian variety of dimension g + 1 has h 1 , 0 = g + 1 ; so X is not an abelian variety. (Equally, X contains the rational curves { a } × P 1 , and every morphism from P 1 to a complex torus is constant, since it lifts to the universal cover C g + 1  [19, Chapter 1].)
Let pr : X → A be the projection and D = A × { 0 } , a smooth divisor that pr maps isomorphically onto A, so that pr * [ D ] = 1 in H 0 ( A , Q ) . Let α be a Hodge class of degree 2 p on A. Pull-back along a morphism is a morphism of Hodge structures, so pr * α is a Hodge class of degree 2 p on X, and by hypothesis pr * α = [ Z ] for an algebraic cycle Z on X with rational coefficients. The projection formula gives
pr * [ Z ] ∪ [ D ] = pr * pr * α ∪ [ D ] = α ∪ pr * [ D ] = α .
The cycle class map is compatible with intersection products and with proper push-forward [44, Chapter 19]], so the left side is the class of the cycle pr * ( Z · D ) and α is algebraic. A point, the abelian variety of dimension zero, satisfies the conjecture trivially.
So the displayed statement gives the conjecture for every abelian variety, and with it the conjecture for every smooth complex projective variety. The converse is immediate: the displayed statement is a special case of HC .    □
Theorem A22 
(The rules with premises (L), (M), (V)).
(i)
(L)and (M) together imply HC .
(ii)
(L) implies (V).
(iii)
(V) implies AB .
(iv)
(L) implies AB .
Proof. 
(i) Let ξ be a Hodge class on X. By (M), ξ = pr X * ( α ∪ * β ) with α , β algebraic on X × Y . By (L) the Lefschetz standard conjecture holds for X × Y , so the involution * is induced by an algebraic correspondence [63, Proposition 1.4.4 and Theorem 2.9]; hence * β , the product α ∪ * β and its push-forward ξ are algebraic. This is André’s observation that motivated classes are algebraic as soon as the Lefschetz involutions are [6, §2.1].
(ii) Let f : X → S be a smooth projective morphism over a smooth connected base, ξ a flat section of R 2 p f * Q that is a Hodge class at every point, and s ∈ S with ξ s algebraic. Then ξ s is motivated, so ξ t is motivated for every t ∈ S by André’s deformation theorem [6, Théorème 0.5], and a motivated class is algebraic under (L) by the proof of (i).
(iii) A polarised abelian scheme over a smooth connected base is a smooth projective family, so (V) gives statement (iii) of Theorem 9, which is equivalent to AB by that theorem. Its proof uses Theorem A2, André’s theorem on Hodge classes of CM type [5] and Proposition 5, and no statement of this subsection.
(iv) By (ii) and (iii).    □

Appendix A.9. Mumford Fourfolds

The notation of this subsection refines Setting 1; the polarisation class of X is θ , and Hdg k ( Y ) denotes the space of Hodge classes in H 2 k ( Y , Q ) .
Proposition  A20
(The Hodge ring is not generated by divisor and Weil classes). Let X be a complex abelian fourfold whose Hodge group is a Q -form of an almost direct product of three copies of SL 2 , acting on V = H 1 ( X , Q ) through the tensor product V 1 ⊗ V 2 ⊗ V 3 of the three standard representations, with End 0 ( X ) = Q ; such fourfolds exist and form families of dimension one, by Mumford [87], and by [91, 2.5(i)] they are exactly the fourfolds with End 0 ( X ) = Q and Hodge group smaller than Sp 8 . Then:
(i)
the Hodge ring of X is generated by the polarisation: dim Q Hdg 1 ( X ) = dim Q Hdg 2 ( X ) = 1 ;
(ii)
dim Q Hdg 1 ( X × X ) = 3 and dim Q Hdg 2 ( X × X ) = 8 , while the products of divisor classes span a subspace of dimension 6 in H 4 ( X × X , Q ) ;
(iii)
the two remaining classes lie in H 4 ( X × X , Q ) and are Weil classes of no CM field acting on X × X , on a quotient of it, or on an abelian subvariety of it.
In particular the statement that every Hodge class on an abelian variety lies in the subring generated by divisor classes and Weil classes of quotients is false in dimension eight, though true through dimension five [91].
Proof. 
Hodge classes are the invariants of the Hodge group, so Hdg k ( Y ) ⊗ C is the space of Hg ( Y ) C -invariants in ⋀ 2 k H 1 ( Y , C ) , and Hg ( X × X ) = Hg ( X ) acting diagonally on V ⊕ V . Over C the group is SL 2 3 and the weights of V under its maximal torus are the eight sign vectors ( ± 1 , ± 1 , ± 1 ) . Write V = V 1 ⊗ W with W = V 2 ⊗ V 3 , on which SL 2 2 = Spin ( 4 ) acts orthogonally. The Cauchy formula ⋀ k ( V 1 ⊗ W ) = ⨁ | λ | = k S λ V 1 ⊗ S λ ′ W  [41, §6.1], with S 2 V 1 = S ( 3 , 1 ) V 1 = V ( 2 ) , ⋀ 2 V 1 = S ( 2 , 2 ) V 1 = 1 and S 4 V 1 = V ( 4 ) , gives ⋀ 2 V = V ( 2 ) ⊗ ⋀ 2 W ⊕ S 2 W and ⋀ 4 V = V ( 4 ) ⊗ ⋀ 4 W ⊕ V ( 2 ) ⊗ S ( 2 , 1 , 1 ) W ⊕ S ( 2 , 2 ) W . Under Spin ( 4 ) , ⋀ 2 W = V ( 2 , 0 ) ⊕ V ( 0 , 2 ) , S 2 W = V ( 2 , 2 ) ⊕ 1 , ⋀ 4 W = 1 , S ( 2 , 1 , 1 ) W ≅ sl ( W ) = ⋀ 2 W ⊕ S 0 2 W , and S ( 2 , 2 ) W = V ( 4 , 0 ) ⊕ V ( 0 , 4 ) ⊕ V ( 2 , 2 ) ⊕ 1 , the decomposition of the algebraic curvature tensors in dimension four into the self-dual and anti-self-dual Weyl parts, the traceless Ricci part and the scalar part. Hence
⋀ 2 V = V ( 2 , 2 , 0 ) ⊕ V ( 2 , 0 , 2 ) ⊕ V ( 0 , 2 , 2 ) ⊕ 1 , ⋀ 4 V = V ( 2 , 2 , 2 ) ⊕ ⨁ cyc V ( 4 , 0 , 0 ) ⊕ ⨁ cyc V ( 2 , 2 , 0 ) ⊕ 1 ,
each with one trivial summand, which is (i). For (ii), ⋀ 4 ( V ⊕ V ) = ⨁ a ⋀ a V ⊗ ⋀ 4 − a V and the invariants in ⋀ a V ⊗ ⋀ 4 − a V are Hom SL 2 3 ( ⋀ a V , ⋀ 4 − a V ) , since every summand is self-dual; the counts are 1 , 1 , 4 , 1 , 1 , total 8, the 4 coming from the four distinct irreducible summands of ⋀ 2 V . Likewise the invariants in ⋀ 2 ( V ⊕ V ) number 1 + 1 + 1 = 3 . The divisor classes of X × X are the three symplectic forms ψ 11 , ψ 12 , ψ 22 on V ⊕ V built from the polarisation ψ of V. Their six pairwise products are linearly independent: they lie in the summands ⋀ a V ⊗ ⋀ 4 − a V with a = 4 , 3 , 2 , 2 , 1 , 0 , each is nonzero, and in the summand a = 2 the classes ψ 11 ψ 22 and ψ 12 2 correspond, as maps ⋀ 2 V → ⋀ 2 V , to a multiple of the projection onto C ψ and to a multiple of the identity. The Sp 8 -invariants of ⋀ 4 ( V ⊕ V ) number 1 + 1 + 2 + 1 + 1 = 6 by the same count, since ⋀ 1 V = V , ⋀ 2 V = ⋀ 0 2 V ⊕ 1 and ⋀ 3 V = ⋀ 0 3 V ⊕ V under Sp 8 ; so the six products are all the classes a Hodge group equal to Sp 8 would allow. For (iii), the Weil classes of an abelian variety Y with respect to a CM field F ⊆ End 0 ( Y ) are, by definition, the classes in ⋀ F 2 k H 1 ( Y , Q ) ⊆ H 2 k ( Y , Q ) with 2 k = dim F H 1 ( Y , Q ) (Setting 8; here F denotes a CM field). A CM field acting on X × X lies in End 0 ( X × X ) = M 2 ( Q ) , so it is imaginary quadratic, 2 k = 8 , and its Weil classes lie in H 8 , while the two classes of (ii) have degree four. Every proper abelian subvariety and every proper quotient of X × X is isogenous to X, because X is simple, and End 0 ( X ) = Q contains no CM field, so X carries no Weil classes at all. This is the statement of [91, 0.1(iii)] that the exceptional classes of X × X are not of Weil type. The last sentence is (ii) and (iii) together with the theorem of [91] for dimension at most five.    □
Setting  15
(Mumford type). Let X be as in Proposition A20, with V = H 1 ( X , Q ) , polarisation ψ and polarisation class θ ∈ H 2 ( X , Q ) , and let G = Hg ( X ) . Over Q ¯ the Hodge cocharacter is nontrivial on exactly one of the three factors of G: on V i it has weights ± a i / 2 with a i ≥ 0 , and V has only the two weights ± 1 2 , so exactly one a i is nonzero. Since End 0 ( X ) = Q , it follows that G is Q -simple: the Lie algebra of a Q -factor not containing that geometric factor would consist of Hodge classes in End ( V ) , that is, of endomorphisms of X. So G is the image in GL ( V ) of Res F / Q D 1 , where F is a totally real cubic field, D is a quaternion algebra over F split at one real place σ 3 and definite at the other two, σ 1 and σ 2 , with Cor F / Q D ≅ M 8 ( Q ) , and D 1 is the group of elements of reduced norm one [45,87]; the kernel is the central subgroup of triples of signs with product one. Over Q ¯ , V = V 1 ⊗ V 2 ⊗ V 3 , where V i is the standard representation of D ⊗ F , σ i Q ¯ = M 2 ( Q ¯ ) ; the Galois group permutes the three factors as it permutes the three embeddings, and H 1 , 0 ( X ) = V 1 ⊗ V 2 ⊗ V 3 1 , 0 . Put
T = Lie G = D 0 , q λ ( x , y ) = Tr F / Q λ trd ( x y ) ( λ ∈ F × ) ,
where D 0 is the space of quaternions of reduced trace zero, regarded as a Q -vector space of dimension nine, and trd is the reduced trace. Over Q ¯ , T = T 1 ⊕ T 2 ⊕ T 3 with T i = sl ( V i ) , and a ∈ F acts on T i by σ i ( a ) . Write m a for multiplication by a. The space T ⊂ End ( V ) is a sub-Hodge structure of weight zero.
Proposition  A21
(The two classes are a real multiplication). In Setting 15:
(i)
H 2 ( X , Q ) = Q θ ⊕ U , where U is the primitive part, of dimension 27, and over Q ¯
U = U 12 ⊕ U 13 ⊕ U 23 , U i j = S 2 V i ⊗ S 2 V j ⊗ ⋀ 2 V k , { i , j , k } = { 1 , 2 , 3 } ,
each of dimension nine. The summand U 12 is of type ( 1 , 1 ) , and U 13 and U 23 have Hodge numbers ( 3 , 3 , 3 ) .
(ii)
The algebra of Hodge endomorphisms of H 2 ( X , Q ) is Q × F , acting by scalars on Q θ , with a ∈ F acting on U i j by σ k ( a ) . Through Poincaré duality and the Lefschetz isomorphism θ 2 : H 2 → H 6 this algebra is the space of Hodge classes in the Künneth component H 2 ( X ) ⊗ H 2 ( X ) of H 4 ( X × X , Q ) ; the subalgebra Q × Q is spanned by the products of divisor classes that lie there, and the two classes of Proposition A20(ii) may be taken to be the elements of F of trace zero.
(iii)
T ( − 1 ) is a Hodge structure of K3 type with Hodge numbers ( 1 , 7 , 1 ) , and its Hodge endomorphisms are the m a , a ∈ F . The G-invariant symmetric forms on T are the q λ , and q λ polarises T ( − 1 ) if and only if λ is totally positive.
Proof. 
(i) is the decomposition of ⋀ 2 ( V 1 ⊗ V 2 ⊗ V 3 ) in the proof of Proposition A20, where U i j appears as the summand of highest weight 2 on the i-th and j-th factors. The Hodge cocharacter acts trivially on V 1 and V 2 and splits V 3 = V 3 1 , 0 ⊕ V 3 0 , 1 , so ⋀ 2 V 3 has type ( 1 , 1 ) and S 2 V 3 has one dimension of each of the types ( 2 , 0 ) , ( 1 , 1 ) , ( 0 , 2 ) .
(ii) A Hodge endomorphism commutes with the Mumford–Tate group G · G m , and G m acts on H 2 by scalars, so the Hodge endomorphisms are the G-endomorphisms. Over Q ¯ the four summands Q θ , U 12 , U 13 , U 23 are irreducible and pairwise non-isomorphic, so the G-endomorphisms are the combinations of the four projections. The Galois group fixes the projection onto Q θ and permutes the other three as it permutes the indices k, that is, as it permutes the embeddings of F. For a ∈ F an endomorphism acting on U i j by σ k ( a ) is therefore Galois invariant, hence rational, and with the scalars on Q θ these exhaust the rational points, which form Q × F . Poincaré duality identifies End H 2 ( X , Q ) with H 6 ( X ) ⊗ H 2 ( X ) , and θ 2 ⊗ 1 carries H 2 ⊗ H 2 onto it; both are given by algebraic classes, and on an abelian variety so is the inverse of the Lefschetz isomorphism [63, §2]. The products of divisor classes in H 2 ⊗ H 2 are ψ 11 ψ 22 and ψ 12 2 in the notation of the proof of Proposition A20. They are invariant under the whole of Sp ( V , ψ ) , whose endomorphisms of H 2 = Q θ ⊕ U are spanned by the identity and the projection onto Q θ , and they are linearly independent, being a multiple of the projection and a multiple of the identity (proof of Proposition A20); so they span Q × Q , and any complement of Q × Q in Q × F may serve as the two remaining classes.
(iii) The adjoint action of the Hodge cocharacter is trivial on T 1 and T 2 and has weights − 1 , 0, 1 on T 3 , so T has Hodge numbers ( 1 , 7 , 1 ) in the types ( − 1 , 1 ) , ( 0 , 0 ) , ( 1 , − 1 ) , and T ( − 1 ) is of K3 type. Its Hodge endomorphisms are again the G-endomorphisms, and the argument of (ii), applied to the three pairwise non-isomorphic summands T i , gives F acting by multiplication. On each T i the invariant symmetric forms are the multiples of tr ( x y ) , and tr ( x i y i ) = σ i ( trd ( x y ) ) for x , y ∈ D 0 , so the Galois-invariant ones are the q λ . At σ 1 and σ 2 the algebra D ⊗ R is Hamilton’s quaternions, T i is of type ( 1 , 1 ) , and trd ( x 2 ) = − 2 nrd ( x ) is negative on nonzero pure quaternions. At σ 3 the form tr ( x y ) on sl 2 ( R ) is negative on the line of type ( 1 , 1 ) , the Lie algebra of the rotations fixed by the Hodge cocharacter, and positive on its orthogonal complement, the real points of the types ( 2 , 0 ) and ( 0 , 2 ) . The Hodge–Riemann relations for a structure of K3 type ask for exactly these signs after multiplication by σ i ( λ ) , so q λ polarises T ( − 1 ) precisely when every σ i ( λ ) is positive.    □
Lemma  A27
(The product map). For x , y ∈ T let μ ( x ⊗ y ) ∈ H 2 ( X , Q ) be the class that corresponds under ψ to the alternating form ( v , w ) ↦ ψ ( x v , y w ) + ψ ( y v , x w ) on V. Then μ is a G-equivariant map T ⊗ T → H 2 ( X , Q ) , and:
(i)
for i ≠ j it restricts to an isomorphism T i ⊗ T j → U i j , and it carries T i ⊗ T i onto Q θ ;
(ii)
for e , λ ∈ F × put μ e = μ ∘ ( m e ⊗ m e ) and let μ e † be its adjoint for the form q λ ⊗ q λ on T ⊗ T and the Lefschetz pairing B ( α , β ) = ∫ X α β θ 2 on H 2 ( X , Q ) . Then μ e μ e † preserves Q θ and acts on U as the element
ν N F / Q ( g ) g − 1 ∈ F , g = e 2 / λ ,
where ν is a nonzero rational number independent of e and λ.
Proof. 
Equivariance is clear, since μ is built from ψ and the action of End ( V ) on V. An element x ∈ T i acts on V through the i-th factor and lies in the symplectic Lie algebra of ψ = ϵ 1 ⊗ ϵ 2 ⊗ ϵ 3 , so ψ ( x v , y w ) = − ψ ( v , x y w ) and μ ( x ⊗ y ) is the form − ψ ( v , ( x y + y x ) w ) . If x ∈ T i and y ∈ T j with i ≠ j , then x y = y x acts as x ⊗ y on V i ⊗ V j , and the form is − 2 times the product of ϵ i ( · , x · ) , ϵ j ( · , y · ) and ϵ k ; the first two are symmetric because x and y lie in sp ( ϵ ) = sl 2 , and x ↦ ϵ ( · , x · ) is an isomorphism sl 2 → S 2 V ∨ , so μ maps T i ⊗ T j isomorphically onto U i j . If i = j , then x y + y x = tr ( x y ) on V i , because x and y are trace-free 2 × 2 matrices, and μ ( x ⊗ y ) = − tr ( x y ) θ .
For (ii), m e is self-adjoint for q λ , and the adjoint of a map out of ( T ⊗ T , q λ ⊗ q λ ) is ( m λ ⊗ m λ ) − 1 times its adjoint for q 1 ⊗ q 1 , so μ e μ e † = μ ( m g ⊗ m g ) μ † 1 with μ † 1 the adjoint for q 1 ⊗ q 1 . This map is G-equivariant, so it preserves Q θ and each U i j , and μ † 1 carries U i j into ( T i ⊗ T j ) ⊕ ( T j ⊗ T i ) , on which m g ⊗ m g is multiplication by σ i ( g ) σ j ( g ) = N ( g ) / σ k ( g ) . Hence μ e μ e † is N ( g ) σ k ( g ) − 1 times μ μ † 1 on U i j . The summands of H 2 ( X , Q ¯ ) are pairwise non-isomorphic and self-dual, so B is nondegenerate on U i j , and by Schur’s lemma the pull-back of B along the isomorphism μ 0 = μ | T i ⊗ T j is c k q 1 ⊗ q 1 with c k ≠ 0 ; then μ 0 μ 0 † 1 = c k on U i j . As μ is symmetric in its two arguments and the exchange of the factors is an isometry of q 1 ⊗ q 1 , μ † 1 u = μ 0 † 1 u + s μ 0 † 1 u with s that exchange, and μ μ † 1 = 2 c k on U i j . On the split model V = Q ¯ 2 ⊗ Q ¯ 2 ⊗ Q ¯ 2 , ψ = ϵ ⊗ ϵ ⊗ ϵ , the permutations of the three factors preserve ψ , hence θ , B and μ , and also q 1 = ∑ i tr ( x i y i ) , while they permute the U i j ; so c 1 = c 2 = c 3 , and μ μ † 1 is one and the same nonzero number ν on the three summands U i j , rational because μ μ † 1 is defined over Q . By Proposition A21(ii) an endomorphism of U acting on U i j by ν N ( g ) / σ k ( g ) is the element ν N ( g ) g − 1 of F.    □
Theorem  A23
(One K3 surface for the two classes). In Setting 15, let λ ∈ F be totally positive.
(i)
There is a projective K3 surface S λ of Picard number thirteen and an isomorphism of Hodge structures T ( S λ ) Q ≅ T ( − 1 ) carrying the intersection form to q λ . Its Kuga–Satake variety is isogenous to X 32 .
Proof. 
(i) The quadratic space ( T , q λ ) has dimension nine and signature ( 2 , 7 ) by Proposition A21(iii), and H 2 of a K3 surface with its intersection form has dimension 22 and signature ( 3 , 19 ) . Over Q p a nondegenerate quadratic space embeds isometrically in every nondegenerate space of dimension at least three larger, and over R the signatures allow it, so by the Hasse–Minkowski theorem for representations of forms by forms [95, §§63, 66] there is an isometric embedding of ( T , q λ ) into Λ Q , where Λ is the K3 lattice. The line ℓ = T ( − 1 ) 2 , 0 ⊂ T C ⊂ Λ C satisfies ( ℓ , ℓ ) = 0 and ( σ , σ ¯ ) > 0 for 0 ≠ σ ∈ ℓ , so by the surjectivity of the period map there is a K3 surface S λ with a marking H 2 ( S λ , Z ) ≅ Λ that carries H 2 , 0 ( S λ ) to ℓ [58, Chs. 6–7]. A rational class is of type ( 1 , 1 ) exactly when it is orthogonal to ℓ, and its component in T is then a rational class of type ( 1 , 1 ) in T ( − 1 ) , that is a G-invariant vector of T = Lie G ; there is none, since G is semisimple. So NS ( S λ ) Q = T ⊥ , of dimension 13 and signature ( 1 , 12 ) , and T ( S λ ) Q = T with the form q λ . A class of positive square in NS ( S λ ) makes S λ projective [58, Ch. 8].
For the Kuga–Satake variety, note first that over Q ¯ the space T is the orthogonal sum of the three ternary spaces T i , so C ( T ) is the graded tensor product of the C ( T i ) . Scaling a ternary form does not change its even Clifford algebra, so D 1 is the spin group of ( D 0 , λ trd ) for every λ , and Res F / Q D 1 maps to Spin ( T , q λ ) by multiplication in the Clifford algebra, with the same kernel as its map to GL ( V ) , so that G is a subgroup of Spin ( T , q λ ) ; as a representation of the i-th factor, C ( T i ) is four copies of V i . Hence C + ( T ) is V ⊕ 32 as a representation of G over Q ¯ , and over Q as well, because V is absolutely irreducible. The Hodge structures on C + ( T ) = H 1 ( KS ( S λ ) , Q ) and on V ⊕ 32 = H 1 ( X 32 , Q ) are both given by lifts of weight one of the same homomorphism S → MT ( T ) R through G · G m , and two such lifts coincide.29 So H 1 ( KS ( S λ ) , Q ) ≅ H 1 ( X 32 , Q ) as Hodge structures, and KS ( S λ ) is isogenous to X 32 . For the form used by Galluzzi this is her theorem [45]; van Geemen proves the corresponding splitting for every K3 surface with real multiplication [118].    □
Remark  A6.
In Theorem A23 the field F is the totally real cubic field of the Mumford fourfold, an arbitrary one admitting a quaternion algebra as in Setting 15; it need not be Galois over Q , so the Galois group of its normal closure may be S 3 . By Proposition A21(iii) the field of Hodge endomorphisms of T ( S λ ) Q is F, so S λ has real multiplication by F with dim F T ( S λ ) Q = 3 . The Kuga–Satake variety has dimension 2 9 − 2 = 2 7 = 128 = dim X 32 .
Theorem  A24
(Every power of a Mumford fourfold). In Setting 15, suppose that the two exceptional classes of Proposition A20(ii) are algebraic on X × X . Then every Hodge class on X n is algebraic, for every n ≥ 1 .
Proof. 
Over Q ¯ keep V = V 1 ⊗ V 2 ⊗ V 3 and ψ = ϵ 1 ⊗ ϵ 2 ⊗ ϵ 3 , as in the proof of Lemma A27. For m ≥ 1 regard V ⊗ m = H 1 ( X ) ⊗ m as the Künneth component of H m ( X m , Q ) of multidegree ( 1 , ⋯ , 1 ) , whose projector is algebraic [63, §2]. Call an endomorphism of V ⊗ m algebraic if it is the action on V ⊗ m of an algebraic cycle on X m × X m that preserves V ⊗ m . Composing with the Künneth projector, the algebraic endomorphisms form a Q -algebra, and they carry algebraic classes in V ⊗ m to algebraic classes.
Step 1: reduction to multilinear invariants. Let n ≥ 1 and p ≥ 0 . The map V ⊗ 2 p ⊗ ( Q n ) ⊗ 2 p → ⋀ 2 p ( V ⊗ Q n ) = H 2 p ( X n , Q ) sending ( v 1 ⊗ ⋯ ⊗ v 2 p ) ⊗ ( a 1 ⊗ ⋯ ⊗ a 2 p ) to ( v 1 ⊗ a 1 ) ∧ ⋯ ∧ ( v 2 p ⊗ a 2 p ) is surjective and G-equivariant for the trivial action on Q n , so it is surjective on G-invariants, G being reductive. For a 1 , ⋯ , a 2 p ∈ Z n the image of t ⊗ ( a 1 ⊗ ⋯ ⊗ a 2 p ) is f * t , where f : X n → X 2 p is the homomorphism with rows a 1 , ⋯ , a 2 p , and such tensors span ( Q n ) ⊗ 2 p . So Hdg p ( X n ) is spanned by the classes f * t with f a homomorphism and t ∈ ( V ⊗ 2 p ) G , and it suffices to show that every element of ( V ⊗ 2 p ) G is algebraic.
Step 2: degree four. The space ( V ⊗ 4 ) G has dimension eight, the product of the dimensions of the three spaces of SL 2 -invariants in V k ⊗ 4 . Let P 0 ⊂ ( V ⊗ 4 ) G be the span of the multilinear components of the classes f * ξ , for ξ ∈ Hdg 2 ( X × X ) and f : X 4 → X 2 a homomorphism, and of the products of two divisor classes of X 4 . By Proposition A20(ii) Hdg 2 ( X × X ) is spanned by the six products of divisor classes and the two exceptional classes, so under the hypothesis every element of P 0 is algebraic. For t ∈ V ⊗ 4 , read as a correspondence from the first two factors of X 4 to the last two, let t ^ be the endomorphism v ↦ t * ( θ 3 × θ 3 ) ∪ v of V ⊗ V . The pairing ( v , w ) ↦ ∫ X v ∪ w ∪ θ 3 on V is nondegenerate by hard Lefschetz and G-invariant, so t ↦ t ^ is an isomorphism ( V ⊗ 4 ) G → End G ( V ⊗ V ) , and t ^ is algebraic when t is. We show that the t ^ and the composites u ^ t ^ , for t , u ∈ P 0 , span End G ( V ⊗ V ) . Over Q ¯ , End G ( V ⊗ V ) = ⨂ k End SL 2 ( V k ⊗ V k ) is the algebra of functions on the eight sign patterns ε ∈ { ± 1 } 3 , where ε k is the eigenvalue of the exchange s k of the two copies of V k ; the patterns with ε 1 ε 2 ε 3 = − 1 , the odd ones, make up ⋀ 2 V .
(a) Let f = ( x 1 + x 2 , x 3 + x 4 ) : X 4 → X 2 . The multilinear component of f * ξ depends only on the Künneth component of ξ in H 2 ( X ) ⊗ H 2 ( X ) , and it is that component read as a tensor t ∈ ⋀ 2 V ⊗ ⋀ 2 V , so t ^ vanishes on S 2 V and takes values in ⋀ 2 V . These components fill ( H 2 ⊗ H 2 ) G = End G ( H 2 ) , of dimension 4 by Proposition A21(ii); so the t ^ span the functions supported on the four odd patterns.
(b) The map g = ( x 1 + x 3 , x 2 + x 4 ) is f composed with the exchange of the factors 2 and 3 of X 4 , which permutes the factors of V ⊗ 4 up to sign. On V k ⊗ 4 the SL 2 -invariants are spanned by ϵ 13 ϵ 24 , ϵ 14 ϵ 23 and ϵ 12 ϵ 34 = ϵ 13 ϵ 24 − ϵ 14 ϵ 23 , where ϵ a b is ϵ k on the factors a and b; up to a common sign their hats are 1, s k and 1 − s k . The exchange of the factors 2 and 3 swaps the first and the third and changes the sign of the second, so on functions of ε k it carries 1 2 ( 1 − s k ) to 1 2 and 1 2 ( 1 + s k ) to 1 2 − s k . Applied to the indicator of the odd pattern ( 1 , 1 , − 1 ) it gives 1 2 ( 1 2 − s 1 ) ( 1 2 − s 2 ) . The identity is t ^ for the multilinear component of a product of two divisor classes, the pull-backs of the class ψ on X × X along the projections of X 4 to the factors ( 1 , 3 ) and ( 2 , 4 ) , so h = ( 1 − 2 s 1 ) ( 1 − 2 s 2 ) − 1 lies in the span of the t ^ , t ∈ P 0 ⊗ Q ¯ . As ( ε 1 , ε 2 ) runs through ( 1 , 1 ) , ( 1 , − 1 ) , ( − 1 , 1 ) , ( − 1 , − 1 ) , h takes the values 0 , − 4 , − 4 , 8 , so h ( h + 4 ) / 96 is the indicator of ε 1 = ε 2 = − 1 ; subtracting the indicator of ( − 1 , − 1 , − 1 ) , from (a), leaves that of ( − 1 , − 1 , 1 ) . The same argument with the factors V k permuted gives the other two even patterns with two minus signs, and the indicator of ( 1 , 1 , 1 ) is 1 minus the other seven.
So the t ^ and u ^ t ^ span End G ( V ⊗ V ) ⊗ Q ¯ . Since composition is bilinear and P 0 ⊗ Q ¯ is spanned by the same pull-backs and products of a basis of Hodge classes, the Q -algebra of algebraic G-endomorphisms of V ⊗ V has dimension 8: every G-endomorphism of V ⊗ V is algebraic.
Step 3: the partial swaps. Over Q ¯ , End G ( V ⊗ V ) = ⨂ k End SL 2 ( V k ⊗ V k ) , and End SL 2 ( V k ⊗ V k ) is spanned by 1 and the exchange s k of the two copies of V k . In a basis of V k the matrix units E a b and E b a are dual for the trace form of gl ( V k ) , and ∑ a , b E a b ⊗ E b a = s k ; removing the centre, spanned by the identity 1 with tr ( 1 · 1 ) = 2 , gives s k = Cas k + 1 2 , with Cas k = ∑ c x c ⊗ x c for dual bases ( x c ) , ( x c ) of sl ( V k ) under the trace form. So the s k come from the three summands T k of T. The Galois group permutes the s k as it permutes the embeddings σ k , so for a ∈ F the endomorphism r a = ∑ k σ k ( a ) s k is rational, hence algebraic by Step 2. For m ≥ 2 and 1 ≤ i < m let r a ( i ) and s k ( i ) act as r a and s k on the factors i and i + 1 of V ⊗ m and as the identity on the others; r a ( i ) is algebraic, being induced by the product of a cycle inducing r a with the diagonals of the other factors. By Schur–Weyl duality End SL 2 ( V k ⊗ m ) is the image of Q ¯ [ S m ] , generated by the adjacent transpositions, and the commutant of SL 2 3 in ⨂ k V k ⊗ m is the tensor product of the three commutants; so End G ( V ⊗ m ) ⊗ Q ¯ is generated by the s k ( i ) . For a basis a 1 , a 2 , a 3 of F the matrix ( σ k ( a j ) ) is invertible, the square of its determinant being the discriminant of F, so each s k ( i ) is a Q ¯ -linear combination of the r a j ( i ) . The Q -algebra generated by the r a ( i ) therefore lies in End G ( V ⊗ m ) and spans it over Q ¯ , so it is End G ( V ⊗ m ) : every G-endomorphism of V ⊗ m is algebraic. Its dimension is the cube of dim End SL 2 ( V k ⊗ m ) , the sum of the squares of the multiplicities in V k ⊗ m : 8 = 2 3 for m = 2 and 125 = 5 3 for m = 3 , where V k ⊗ 3 = S 3 V k ⊕ V k ⊕ 2 .
Step 4: all multilinear invariants. By the first fundamental theorem for SL 2 the invariants in V k ⊗ 2 m are spanned by the images of ϵ k ⊗ m under the permutations of the 2 m factors. So ( V ⊗ 2 m ) G ⊗ Q ¯ is spanned by the images of ψ ⊗ m = ⨂ k ϵ k ⊗ m under the operators that permute the factors of the three V k ⊗ 2 m separately, which lie in End G ( V ⊗ 2 m ) ⊗ Q ¯ , and the map End G ( V ⊗ 2 m ) → ( V ⊗ 2 m ) G , e ↦ e ( ψ ⊗ m ) , is surjective. The class ψ in V ⊗ V is that of the divisor m * θ − p 1 * θ − p 2 * θ on X × X (proof of Theorem A23(ii)), up to a nonzero factor, and ψ ⊗ m is, up to sign, the product of its pull-backs along the projections of X 2 m to the pairs of factors ( 2 j − 1 , 2 j ) ; so it is algebraic, and by Step 3 so is every element of ( V ⊗ 2 m ) G . Step 1 concludes.    □
Proposition  A22
(The Hodge locus of an exceptional class). In Setting 15, let H be the period domain of polarised Hodge structures of weight one on ( V , ψ ) , a Siegel upper half space of dimension ten, let h X ∈ H be the point of X, and let e = m a , a ∈ F ∖ Q , be an exceptional class, regarded by Proposition A21(ii) as an endomorphism of ⋀ 2 V . The connected component through h X of the locus in H where e is a Hodge class is the orbit G ( R ) + · h X , of dimension one, and its image in the moduli space of polarised abelian fourfolds is a compact curve. In particular no fibre of a degenerating family of polarised abelian fourfolds at which e is still of Hodge type is singular.
Proof. 
Let M ⊆ Sp ( V , ψ ) be the stabiliser of e and m its Lie algebra. Since e is G-invariant, m is a G-stable subspace of sp ( V ) = S 2 V containing g = Lie G . Over Q ¯ ,
S 2 ( V 1 ⊗ V 2 ⊗ V 3 ) = ⨁ i S 2 V i ⊗ ⋀ 2 V j ⊗ ⋀ 2 V k ⊕ S 2 V 1 ⊗ S 2 V 2 ⊗ S 2 V 3 ,
the sum of g = T 1 ⊕ T 2 ⊕ T 3 and of an irreducible representation of dimension 27. So m is g or sp ( V ) . In the second case e would commute with Sp ( V , ψ ) on ⋀ 2 V = Q θ ⊕ U , where U is irreducible for Sp ( V , ψ ) , and would act on U by a rational scalar, against a ∉ Q . Hence m = g and the identity component of M is G. The class e is a Hodge class at h ∈ H exactly when the homomorphism h : U ( 1 ) → Sp ( V , ψ ) ( R ) takes values in M ( R ) , hence in G ( R ) . Homomorphisms from the compact group U ( 1 ) to G ( R ) are locally rigid up to conjugation, so the connected component of this locus through h X is the orbit G ( R ) + · h X . Since G ( R ) is SL 2 ( R ) × SU ( 2 ) × SU ( 2 ) modulo a finite central subgroup and the stabiliser of h X contains the compact factors and a circle of the first factor, the orbit is a copy of the upper half plane. Its image in the moduli space is the quotient by an arithmetic subgroup of G ( Q ) . The algebra D is a division algebra, being definite at two real places, so G is anisotropic over Q and the quotient is compact [101, Theorem 4.12]; this is the compactness of Mumford’s curves [87]. A singular fibre of a degenerating family gives a point of the boundary of the moduli space in the closure of the locus, and a compact curve has no such point.    □

Appendix A.10. The Very General K3 Surface with Real Multiplication

Setting  16
(A family with real multiplication). Let E be a totally real field of degree r ≥ 2 and T a finite-dimensional Q -vector space with an action of E and a nondegenerate symmetric form q for which E acts by self-adjoint maps, so that q = Tr E / Q q E for a unique E-bilinear form q E . Put m = dim E T ≥ 3 , and suppose that q E has signature ( 2 , m − 2 ) at one real place σ 1 of E and is negative definite at the others. Let T σ = T ⊗ E , σ R , T 1 = T σ 1 and
D E = { [ ω ] ∈ P ( T 1 ⊗ R C ) : q ( ω , ω ) = 0 , q ( ω , ω ¯ ) > 0 } + ,
one of its two components, a bounded symmetric domain of dimension m − 2 . For ω ∈ D E let T ω be the Hodge structure of K3 type on T with T ω 2 , 0 = C ω , T ω 0 , 2 = C ω ¯ and T ω 1 , 1 their orthogonal complement; it is polarised by − q . Each e ∈ E acts on T 1 ⊗ C by the scalar σ 1 ( e ) and is self-adjoint, so it is an endomorphism of the Hodge structure T ω . Let G = Res E / Q SO ( T , q E ) , so that G ( R ) = SO ( T 1 ) × ∏ σ ≠ σ 1 SO ( T σ ) , and write G ( R ) + for the identity component, which acts transitively on D E through its first factor. When ( T , q ) is embedded isometrically in Λ Q for the K3 lattice Λ, the surjectivity of the period map gives for every ω a K3 surface S ω with H 2 , 0 ( S ω ) = C ω , and then T ( S ω ) Q ⊆ T is the orthogonal complement in T of the rational classes of type ( 1 , 1 ) of T ω .
Proposition  A23
(The very general member). In Setting 16 there is a countable union Z of proper closed analytic subsets of D E such that for every ω ∉ Z the Hodge structure T ω has no nonzero rational class of type ( 1 , 1 ) and End Hdg ( T ω ) = E . In the geometric situation, T ( S ω ) Q = T and the field of Hodge endomorphisms of T ( S ω ) Q is E for ω ∉ Z . The proof uses the hypothesis on m only through m ≥ 3 .
Proof. 
The Hodge structure T ω is given by u ω : U ( 1 ) → SO ( T R ) , acting by z 2 on C ω , by z − 2 on C ω ¯ and trivially on their orthogonal complement; it commutes with E, lies in G ( R ) + , and u g ω = g u ω g − 1 for g ∈ G ( R ) + .
(a) No nonzero t ∈ T is orthogonal to every ω ∈ D E . The lines C ω , ω ∈ D E , span T 1 ⊗ C , since D E is a nonempty open subset of the nondegenerate quadric q ( ω , ω ) = 0 in P ( T 1 ⊗ C ) , so such a t would have zero image in T 1 ; but T → T 1 is injective, since T is a free E-module and σ 1 is injective.
(b) An endomorphism f of T that is an endomorphism of Hodge structures at every ω ∈ D E is multiplication by an element of E. Such an f commutes with every u ω , hence with the elements X ω of the Lie algebra of G ( R ) that span the Lie algebras of the groups u ω ( U ( 1 ) ) . They lie in so ( T 1 ) ⊕ 0 , which acts on T ′ = ⨁ σ ≠ σ 1 T σ by zero, and their span is stable under G ( R ) + , since u g ω = g u ω g − 1 , so it is an ideal of so ( T 1 ) . That ideal has nonzero projection to every simple ideal, as the eigenvalues 2 i , − 2 i , 0 , ⋯ , 0 of X ω on T 1 ⊗ C show also for m = 4 , where so ( T 1 ) ⊗ C ≅ sl 2 ⊕ sl 2 and T 1 ⊗ C is the tensor product of the two standard representations; so it is all of so ( T 1 ) , and f commutes with so ( T 1 ) . For m ≥ 3 the representation of so ( T 1 ) on T 1 ⊗ C is irreducible and nontrivial, so f preserves T 1 = so ( T 1 ) · T R and T ′ and acts on T 1 ⊗ C = T ⊗ E , σ 1 C by a scalar λ . An automorphism τ of C , acting on T ⊗ Q C through the second factor, commutes with f and maps T ⊗ E , σ 1 C onto T ⊗ E , τ σ 1 C , so f acts on the latter by τ ( λ ) ; and every embedding of E is of the form τ σ 1 . So f ⊗ 1 is multiplication by an element e of E ⊗ Q C on T ⊗ Q C = T ⊗ E ( E ⊗ Q C ) , and e is fixed by Aut ( C ) because f is defined over Q ; hence e ∈ E .
For t ∈ T the set of ω with t ⊥ ω is a closed analytic subset of D E , and for f ∈ End Q ( T ) so is the set where f is an endomorphism of Hodge structures, namely where f ( ω ) ∈ C ω and f * ( ω ) ∈ C ω , f * the adjoint. By (a) and (b) these sets are proper for t ≠ 0 and f ∉ E . There are countably many of them; let Z be their union. A rational class of T is of type ( 1 , 1 ) for T ω exactly when it is orthogonal to ω , so outside Z there is none, and End Hdg ( T ω ) = E . In the geometric situation this gives T ( S ω ) Q = T .    □

Appendix B. The Point of Complex Multiplication in Coordinates

This appendix works out, in coordinates, the objects of Lemma 15 and Proposition 25 on A 0 = E 0 6 : the Hodge ring, the Weil plane, an explicit family of abelian threefolds whose classes combine to the Weil class, their intersection numbers, and the dimensions of every Ext group between the pieces built from them.

Appendix B.1. The Forms and the Hodge Ring

Write z 1 , ⋯ , z 6 for the coordinates of A 0 = E 0 6 , E 0 = C / Z [ i ] , group the factors into the three surfaces B j = E 0 2 with coordinates ( z j , z 3 + j ) , j = 1 , 2 , 3 , and let K = Q ( i ) act by i on z 1 , z 2 , z 3 and by − i on z 4 , z 5 , z 6 . Put
a j = i 2 d z j ∧ d z ¯ j , b j = i 2 d z 3 + j ∧ d z ¯ 3 + j , c j = i 2 d z j ∧ d z ¯ 3 + j , c ¯ j = i 2 d z 3 + j ∧ d z ¯ j .
In the real coordinates z j = x j + i y j one has a j = d x j ∧ d y j , the volume form of the j-th factor, of integral one over C / Z [ i ] ,30 and c j = 1 2 ( d x j ∧ d y 3 + j − d y j ∧ d x 3 + j ) + i 2 ( d x j ∧ d x 3 + j + d y j ∧ d y 3 + j ) , so that the real and imaginary parts of c j are rational classes and c ¯ j is the complex conjugate of c j . All four forms have type ( 1 , 1 ) . The polarisation is
θ = ∑ j = 1 3 ( a j + b j ) , θ 6 = 6 ! a 1 b 1 a 2 b 2 a 3 b 3 , ∫ A 0 θ 6 6 ! = 1 ,
the principal polarisation of the product; the hermitian forms ∑ j ≤ 3 a j ′ z j w ¯ j − ∑ j > 3 a j ′ z j w ¯ j of Lemma 15 with all a j ′ = 1 give θ itself.
The Hodge group of A 0 is the circle acting by z on H 1 , 0 and by z ¯ on H 0 , 1 , so the Hodge classes of degree 2 k are exactly the rational classes of type ( k , k ) , of dimension 6 k 2 :
dim Hdg 2 k ( A 0 ) = 1 , 36 , 225 , 400 , 225 , 36 , 1 ( k = 0 , ⋯ , 6 ) .
In degree two these are the 36 hermitian forms on Z [ i ] 6 , the Picard number of E 0 6 . The K-eigenspaces of H 1 ( A 0 , C ) are spanned by d z 1 , d z 2 , d z 3 , d z ¯ 4 , d z ¯ 5 , d z ¯ 6 (eigenvalue i) and by their conjugates, so the Weil plane W = ⋀ K 6 H 1 ( A 0 , K ) ⊆ H 6 ( A 0 , Q ) is spanned over C by
α = c 1 c 2 c 3 ∝ d z 1 d z 2 d z 3 d z ¯ 4 d z ¯ 5 d z ¯ 6 and α ¯ = c ¯ 1 c ¯ 2 c ¯ 3 ,
and over Q by α + α ¯ and i ( α − α ¯ ) . The relation W θ = 0 is visible term by term: c j a j contains d z j twice and c j b j contains d z ¯ 3 + j twice, so α ( a k + b k ) = 0 for each k, since α already contains d z k and d z ¯ 3 + k .

Appendix B.2. The Elliptic Curves C ζ and the Threefolds Z ζ

For ζ ∈ μ 4 = { 1 , i , − 1 , − i } let C ζ ⊆ B j be the kernel of ( u , v ) ↦ u + ζ v , an elliptic curve isomorphic to E 0 , and for ζ = ( ζ 1 , ζ 2 , ζ 3 ) ∈ μ 4 3 let Z ζ = C ζ 1 × C ζ 2 × C ζ 3 ⊆ A 0 , an abelian threefold.
Lemma A28.
(i)
[ C ζ ] = a j + b j + ζ ¯ c j + ζ c ¯ j in H 2 ( B j ) , and [ Z ζ ] = ∏ j [ C ζ j ] .
(ii)
( C ζ · C ζ ′ ) = | ζ − ζ ′ | 2 ∈ { 0 , 2 , 4 } on B j ; Z ζ ∩ Z ζ ′ is finite exactly when ζ j ≠ ζ j ′ for all j, and then | Z ζ ∩ Z ζ ′ | = ∏ j | ζ j − ζ j ′ | 2 ∈ { 8 , 16 , 32 , 64 } .
(iii)
θ | Z ζ has degree two on each factor, ( θ | Z ζ ) 3 = 48 and χ Z ζ = 8 ; in particular [ Z ζ ] θ ≠ 0 .
(iv)
Writing S ( ω ) = ∑ ζ 1 ζ 2 ζ 3 = ω [ Z ζ ] for ω ∈ μ 4 ,
∑ ω ∈ μ 4 ω m S ( ω ) = ∏ j = 1 3 T m , j , T m , j = ∑ ζ ∈ μ 4 ζ m [ C ζ ] , T m , j = 4 ( a j + b j ) , 4 c j , 0 , 4 c ¯ j for m ≡ 0 , 1 , 2 , 3 ( mod 4 ) ,
and therefore
S ( 1 ) − S ( − 1 ) = ∑ ζ 1 ζ 2 ζ 3 = 1 [ Z ζ ] − ∑ ζ 1 ζ 2 ζ 3 = − 1 [ Z ζ ] = 32 ( α + α ¯ ) .
Proof. 
(i) C ζ is the fibre over 0 of the surjective homomorphism f ζ : B j → E 0 , w = u + ζ v , so [ C ζ ] = f ζ * [ 0 ] with [ 0 ] = i 2 d w ∧ d w ¯ , and
d w ∧ d w ¯ = ( d u + ζ d v ) ∧ ( d u ¯ + ζ ¯ d v ¯ ) = d u ∧ d u ¯ + d v ∧ d v ¯ + ζ ¯ d u ∧ d v ¯ + ζ d v ∧ d u ¯ ,
which is the formula. A product of subvarieties of the factors has the product class.
(ii) On B j the only nonzero products of two of the four forms are a j b j and c j c ¯ j = c ¯ j c j = − a j b j (reorder d u d v ¯ d v d u ¯ to − d u d u ¯ d v d v ¯ ), and ∫ B j a j b j = 1 . Hence [ C ζ ] [ C ζ ′ ] = 2 − ζ ¯ ζ ′ − ζ ζ ¯ ′ = | ζ − ζ ′ | 2 . Directly: a point of C ζ ∩ C ζ ′ has ( ζ − ζ ′ ) v = 0 , so v runs over the kernel of ζ − ζ ′ on E 0 , of order | ζ − ζ ′ | 2 , and u = − ζ v is then determined. The statement for Z ζ follows factor by factor.
(iii) ∫ C ζ ( a j + b j ) = ∫ B j ( a j + b j ) [ C ζ ] = 2 , since a j c j , a j c ¯ j , b j c j , b j c ¯ j all repeat a differential. So θ | Z ζ = 2 ( p 1 + p 2 + p 3 ) with p j the point class of the j-th factor, ( θ | Z ζ ) 3 = 3 ! · 8 = 48 , and χ Z ζ = 48 / 6 = 8 .
(iv) The left side of (A11) is ∑ ζ ( ζ 1 ζ 2 ζ 3 ) m ∏ j [ C ζ j ] , which factors as the product of the three sums T m , j ; these are read off from (i), using ∑ ζ ∈ μ 4 ζ m = 4 or 0 according as 4 ∣ m or not. Then S ( 1 ) − S ( − 1 ) = 1 2 ∑ ω ( ω + ω 3 ) S ( ω ) = 1 2 ∏ j T 1 , j + ∏ j T 3 , j = 1 2 ( 64 α + 64 α ¯ ) .    □
Equation (A12) is the analogue, with abelian threefolds in place of line bundles, of the signed sum of the 32 line bundles on E 0 6 in Proposition A4: the 32 threefolds Z ζ with ζ 1 ζ 2 ζ 3 = ± 1 , taken with the sign ζ 1 ζ 2 ζ 3 , have classes adding up to 32 times the real Weil class. Together with three powers of L this is a Chern character of polarised shape and Hankel rank three at A 0 , and by Lemma 15(i) it is one of the shapes a convolution must realise. Two of these threefolds with ζ j = ζ j ′ for one or two indices j meet along a curve or a surface, so Lemma 15(ii) does not apply to them and Lemma 16 gives their groups; Proposition 26 shows that no convolution of these thirty-five pieces meets the criterion. Figure A1 draws the curves, the dimensions of the Ext groups and the chains.
Figure A1. The point E 0 6 of complex multiplication. Left: the three Weil surfaces B j = E 0 2 and the four elliptic curves C ζ in each, kernels of u + ζ v ; a threefold Z ζ picks one curve in each surface, and the 32 threefolds with ζ 1 ζ 2 ζ 3 = ± 1 , signed by that product, add up to 32 ( α + α ¯ ) . Middle: the dimensions of Ext k ( O Z ⊗ L a , O Z ⊗ L b ) by degree for b − a = 0 (the 6 k ), b − a = 1 (degrees 0 to 3) and b − a = − 1 (degrees 3 to 6). Right: the closed chains of Proposition 25; every two-step chain has total degree at least three, and the shortest chains reaching degree two run through three twists of one threefold with degrees 0 , 0 , 3 .
Figure A1. The point E 0 6 of complex multiplication. Left: the three Weil surfaces B j = E 0 2 and the four elliptic curves C ζ in each, kernels of u + ζ v ; a threefold Z ζ picks one curve in each surface, and the 32 threefolds with ζ 1 ζ 2 ζ 3 = ± 1 , signed by that product, add up to 32 ( α + α ¯ ) . Middle: the dimensions of Ext k ( O Z ⊗ L a , O Z ⊗ L b ) by degree for b − a = 0 (the 6 k ), b − a = 1 (degrees 0 to 3) and b − a = − 1 (degrees 3 to 6). Right: the closed chains of Proposition 25; every two-step chain has total degree at least three, and the shortest chains reaching degree two run through three twists of one threefold with degrees 0 , 0 , 3 .
Preprints 244139 g0a1

Appendix B.3. The Dimensions in Lemma 15

For the threefolds Z = Z ζ the numbers of Lemma 15 are explicit. Table A1 lists, for every pair of piece types, the degrees k in which Ext k is nonzero and its dimension; here m = b − a is the difference of the twists, χ Z = 8 , Z ′ = Z ζ ′ with ζ j ≠ ζ j ′ for all j, and ( Z · Z ′ ) = ∏ j | ζ j − ζ j ′ | 2 .
Table A1. The Ext groups between two pieces at A 0 = E 0 6 , from Lemma 15; m = b − a , χ Z = 8 for the threefolds Z ζ .
Table A1. The Ext groups between two pieces at A 0 = E 0 6 , from Lemma 15; m = b − a , χ Z = 8 for the threefolds Z ζ .
from to degrees k dim Ext k
O Z L a O Z L a 0 , ⋯ , 6 6 k : 1 , 6 , 15 , 20 , 15 , 6 , 1
O Z L a O Z L b , m > 0 0 , ⋯ , 3 3 k 8 m 3
O Z L a O Z L b , m < 0 3 , ⋯ , 6 3 k − 3 8 | m | 3
O Z L a O Z ′ L b 3 ( Z · Z ′ ) ∈ { 8 , 16 , 32 , 64 }
O Z L a L b , m > 0 3 8 m 3
O Z L a L b , m < 0 6 8 | m | 3
O Z L a L a 3 , ⋯ , 6 3 k − 3
L a O Z L b , m > 0 0 8 m 3
L a O Z L b , m < 0 3 8 | m | 3
L a O Z L a 0 , ⋯ , 3 3 k
L a L b , m > 0 0 m 6
L a L b , m < 0 6 | m | 6
L a L a 0 , ⋯ , 6 6 k
Each row of the table restates the corresponding item of Lemma 15, with χ ( L m | Z ) = m 3 χ Z = 8 m 3 from (iii) above and χ ( L m ) = m 6 θ 6 / 720 = m 6 from (A8). Two features of the table drive Proposition 25. Every group between two pieces of different twist is concentrated in degree 0, in degree 3 or 6, or in the range 0 to 3 for two twists of one threefold; and a group in degree at most two between distinct pieces occurs only from a lower twist to a higher one, or from L a to O Z L a . The bookkeeping of the proposition is then the following count. A closed chain of m pieces whose edges have Ext degrees k 1 , ⋯ , k m carries an m-fold product into Ext ∑ k i − m + 2 of the first piece. For m = 2 the table gives, for each unordered pair of distinct piece types, the two degrees and their sum:
( O Z L a , O Z ′ L b ) 3 + 3 = 6 , ( O Z L a , O Z L b ) , b > a ( 0 to 3 ) + ( 3 to 6 ) ≥ 3 , ( O Z L a , L b ) , b > a 3 + 3 = 6 , ( O Z L a , L b ) , b < a 6 + 0 = 6 , ( O Z L a , L a ) ( 3 to 6 ) + ( 0 to 3 ) ≥ 3 , ( L a , L b ) 0 + 6 = 6 ,
so no two-step product reaches degree two. For m = 3 the sum must be 3 with two edges of degree zero, and the table leaves exactly the ascending chains O Z L k → O Z L k ′ → O Z L k ′ ′ , k < k ′ < k ′ ′ , closed by the degree-three edge H 3 ( Z , L k − k ′ ′ | Z ) , of dimension 8 ( k ′ ′ − k ) 3 . The program julia/cmpointext.jl enumerates the chains of two and three pieces over four twists and two threefolds and confirms both statements, together with the identities of Lemma A28. Figure A2 shows the table as the matrix of the twelve pieces, and the distribution of the total degree over every closed chain of two and of three distinct pieces: the two-piece chains begin at total degree three, and the twenty-four ordered three-piece chains of total degree three are the eight ascending triples of twists of one threefold, each read from its three starting pieces.
Figure A2. The degrees of the Ext groups between the twelve pieces O Z L k , O Z ′ L k , L k , k = 0 , ⋯ , 3 , of Table A1. Left: the degrees in which Ext k ( F P , F Q ) is nonzero, shaded by the least one; degree zero occurs only from a lower twist to a higher one and from L a to O Z L a . Right: the number of closed chains of two and of three distinct pieces by total degree, on a logarithmic scale; no two-piece chain has total degree below three, and the three-piece chains of total degree three are those of Proposition 25.
Figure A2. The degrees of the Ext groups between the twelve pieces O Z L k , O Z ′ L k , L k , k = 0 , ⋯ , 3 , of Table A1. Left: the degrees in which Ext k ( F P , F Q ) is nonzero, shaded by the least one; degree zero occurs only from a lower twist to a higher one and from L a to O Z L a . Right: the number of closed chains of two and of three distinct pieces by total degree, on a logarithmic scale; no two-piece chain has total degree below three, and the three-piece chains of total degree three are those of Proposition 25.
Preprints 244139 g0a2

Appendix C. Finite Computations

This appendix carries out the finite computations used in the proofs of the text and of Appendix A. The programs named in each paragraph repeat them, in exact arithmetic except in Appendix C.6 (Appendix D); the hand computations given here are complete, except for the enumerations that are stated as such.

Appendix C.1. Lemma 11

Rank at most five. A rational quadratic space of rank at least five that is indefinite is isotropic by Meyer’s theorem [108, Chap. IV, §3.2], so a space T of signature ( 2 , 3 ) is U Q ⊕ 〈 x , y , z 〉 with x > 0 > y , z after reordering, and
T ⊕ 〈 − x , − y 〉 ≅ U Q ⊕ 〈 x , − x 〉 ⊕ 〈 y , − y 〉 ⊕ 〈 z 〉 ≅ U Q 3 ⊕ 〈 z 〉 ,
since 〈 a , − a 〉 ≅ U Q ; this is V m with m = − z up to squares. Spaces of rank 2, 3, 4 embed in one of rank five of signature ( 2 , 3 ) , by adding negative squares. The program k3wittindex.jl confirms this on all 34606 diagonal forms 〈 a , b , − c , − e , − f 〉 with entries in { 1 , 2 , 3 , 5 , 6 , 7 , 10 , 14 , 23 , 46 , 69 } and on 2000 random isotropic forms. The example after Corollary 3: T = 〈 1 , 1 , − 1 , − 23 , − 23 〉 = U Q ⊕ 〈 1 , − 23 , − 23 〉 with U Q = 〈 1 , − 1 〉 , so x = 1 , y = z = − 23 give T ⊕ 〈 − 1 , 23 〉 ≅ U Q 3 ⊕ 〈 − 23 〉 = V 23 .
Rank six. For λ ∈ Q > 0 and d > 0 , λ U Q ≅ U Q and 〈 − λ , λ 〉 ≅ U Q , so
λ U Q 2 ⊕ 〈 − 1 , − d 〉 ⊕ 〈 λ 〉 ≅ U Q 2 ⊕ 〈 − λ , λ 〉 ⊕ 〈 − λ d 〉 ≅ U Q 3 ⊕ 〈 − λ d 〉 = V λ d ,
which is (c)⇒(a) with m ≡ λ d . The witnesses printed by k3wittindex.jl are 〈 1 , 1 , − 1 , − 1 , − 7 , − 7 〉 = U Q 2 ⊕ 〈 − 7 , − 7 〉 , of discriminant 49 ≡ 1 , with λ = 7 and m ≡ 7 ; 〈 1 , 1 , − 1 , − 1 , − 5 , − 23 〉 , of discriminant 115, with λ = 5 , since 〈 − 5 , − 5 · 115 〉 = 〈 − 5 , − 23 · 5 2 〉 ≅ 〈 − 5 , − 23 〉 , and m ≡ 5 · 115 ≡ 23 ; and 〈 1 , 1 , − 1 , − 1 , − 7 , − 23 〉 , of discriminant 161, with λ = 7 and m ≡ 7 · 161 ≡ 23 . Over the 1980 diagonal forms 〈 a , b , − x 1 , − x 2 , − x 3 , − x 4 〉 with 1 ≤ a ≤ b ≤ 3 and x 1 ≤ ⋯ ≤ x 4 in { 1 , 2 , 3 , 5 , 6 , 7 , 10 , 23 } , the local criterion (d) and the existence of λ < 3000 in (c) agree; 1484 forms have Witt index two and 496 Witt index one, among them 〈 1 , 1 , − 1 , − 2 , − 3 , − 3 〉 and 〈 1 , 1 , − 1 , − 2 , − 5 , − 10 〉 , which the program confirms to embed in no V m with m ≤ 2000 squarefree, as (a)⇔(b) predicts for every m. No space of signature ( 2 , 5 ) embeds in a space of signature ( 3 , 4 ) , which is (iii).

Appendix C.2. Table 4

For α = a + b e ∈ E the binary form Tr E / Q 〈 α 〉 ( x + y e ) = Tr α ( x + y e ) 2 has, in the basis 1 , e , the Gram matrix
2 a 2 b e 2 b e 2 a e , det = 4 e ( a 2 − e b 2 ) = 4 e N ( α ) ,
which is 〈 2 a , 2 e N ( α ) / a 〉 for a ≠ 0 and the hyperbolic plane for a = 0 . It is positive definite if α is totally positive, negative definite if α is totally negative, and of signature ( 1 , 1 ) if N ( α ) < 0 . So T = Tr W for W = 〈 α 1 , α 2 , α 3 〉 with α 1 , α 2 of negative norm and positive at s 1 and α 3 totally negative has signature ( 2 , 4 ) , and W has signature ( 2 , 1 ) at s 1 and ( 0 , 3 ) at s 2 . The data are:
e N ( α 1 ) , N ( α 2 ) , N ( α 3 ) diagonal form of T d = e N ( δ ) p ε p ( T ) , ε p ( U 3 )
2 − 9 , − 28 , 1 〈 6 , − 12 , 4 , − 56 , − 6 , − 4 3 〉 504 = 14 · 6 2 3 − 1 , 1
3 − 44 , − 27 , 1 〈 4 , − 132 〉 ⊕ U ⊕ 〈 − 2 , − 6 〉 3564 = 11 · 18 2 3 − 1 , 1
5 − 4 , − 1 , 11 〈 − 2 , 40 , 4 , − 5 , − 8 , − 55 2 〉 220 = 55 · 2 2 2 1 , − 1
6 − 15 , − 54 , 16 〈 − 6 , 60 〉 ⊕ U ⊕ 〈 − 8 , − 48 〉 77760 = 15 · 72 2 2 1 , − 1
7 − 28 , − 54 , 1 U ⊕ 〈 6 , − 252 , − 2 , − 14 〉 10584 = 6 · 42 2 7 − 1 , 1
In each row − d is a square in Q p : − 14 ≡ 1 ( mod 3 ) , − 11 ≡ 1 ( mod 3 ) , − 55 ≡ 1 ( mod 8 ) , − 15 ≡ 1 ( mod 8 ) , − 6 ≡ 1 ( mod 7 ) . At such p the spaces T p and U p 3 have the same rank and discriminant, so they are isometric if and only if their Hasse invariants agree [108, Chap. IV, Thm. 7]; they do not, so T p ≅ U p 3 , condition (d) of Lemma 11(ii) fails, and T has Witt index one. For instance for e = 2 , p = 3 : modulo squares in Q 3 the diagonal form is 〈 6 , − 3 , 1 , − 14 , − 6 , − 3 〉 ; the Hilbert symbols ( u , v ) 3 of units are 1, ( 3 , u ) 3 = u 3 and ( 3 , 3 ) 3 = ( 3 , − 1 ) 3 = − 1 , and the product over the fifteen pairs is − 1 , while ε 3 ( 〈 1 , − 1 , 1 , − 1 , 1 , − 1 〉 ) = 1 since all entries are units. For e = 2 , 3 , 7 the program k3wittindex.jl also evaluates ∑ P ∣ p inv P ( Q W ) by tame symbols and finds it nonzero at p, as Theorem 17 requires; on 7000 random spaces W for e ∈ { 2 , 3 , 5 , 6 , 7 , 10 , 13 } the criterion of Theorem 17 agrees with the Hasse invariants at all 3160 pairs ( W , p ) with p odd.

Appendix C.3. Split Polarisations

Lemma 28 is proved for every totally real sextic L; splitpolarisation.jl exhibits the polarisations in four fields: L = Q ( ζ 21 ) + , Q ( ζ 28 ) + , Q ( ζ 36 ) + , defined by x 6 − x 5 − 6 x 4 + 6 x 3 + 8 x 2 − 8 x + 1 , x 6 − 7 x 4 + 14 x 2 − 7 , x 6 − 6 x 4 + 9 x 2 − 3 , of polynomial discriminants 453789 = 3 3 7 5 , 1075648 = 2 6 7 5 and 1259712 = 2 6 3 9 , and the field defined by x 6 + x 5 − 9 x 4 − 4 x 3 + 8 x 2 + 3 x − 1 , of discriminant 17609533 and Galois group S 6 , which contains Frobenius elements of cycle types ( 6 ) , ( 1 , 5 ) and ( 1 , 1 , 1 , 1 , 2 ) at p = 5 , 3 , 227 . Each polynomial has six real roots by Sturm sequences and is irreducible modulo one prime. For K = Q ( − d ) , d ∈ { 1 , 2 , 3 , 5 , 7 } , and each of the 20 sign vectors ε with three entries − 1 , an element a = ∑ i c i x i with | c i | ≤ 2 is found with sign s i ( a ) = ε i and − det H a = − N L / Q ( a ) d L ∈ N ( K × ) up to squares, checked by Hilbert symbols at the primes dividing 2 d N ( a ) d L ; the Gram matrix Tr L / Q ( a ω i ω j ) is recomputed exactly and has signature ( 3 , 3 ) . Elements with the same signs but non-split H a also occur, which is why the choice in Lemma 28 is needed.

Appendix C.4. Sixfolds with Complex Multiplication by KL

cmsixfolds.jl lists the transitive subgroups of S 6 by closing generating pairs under composition: there are 279, in 16 conjugacy classes, distinguished by order and cycle statistics. For each Γ and each k = | I | ∈ { 0 , ⋯ , 6 } , 1953 pairs in all, it forms the CM type Φ I , I = { 1 , ⋯ , k } , and for p ≤ 5 enumerates the subsets S of the twelve embeddings with | S | = 2 p and | g S ∩ Φ I | = p for every g ∈ Z / 2 × Γ . It then compares their span B p with the span D p of products of elements of B 1 , the divisor classes: B 2 = D 2 , B 4 = D 4 , B 5 = D 5 in all cases, and B 3 is spanned by D 3 and the sections of Theorem 29; the Weil classes are needed in 332 of the 1953 cases. By Poincaré duality B 4 , B 5 give H 8 , H 10 . For Γ = S 6 and A 6 the dimensions are those stated after Theorem 29. For M = Q ( ζ 21 ) , Q ( ζ 28 ) , Q ( ζ 36 ) , with Galois groups ( Z / 21 ) × , ( Z / 28 ) × , ( Z / 36 ) × acting regularly, each field has two imaginary quadratic subfields, and of the 64 CM types 40 have divisor-generated Hodge rings and 24, in two orbits of twelve, need exactly the two Weil classes of one of them.

Appendix C.5. Hodge Classes on X a ×Z K b

At a very general point the Hodge group of X a × Z K b is the image of G × K 1 , which over Q ¯ acts on H 1 ( X ) = V 1 ⊗ V 2 ⊗ V 3 and on H 1 ( Z K ) = ⨁ i ( P i ⊕ P i ∨ ) , P i = V i χ i . Powers of Z K . By the Cauchy formula ⋀ ( P ⊗ Q ¯ b ) = ⨁ λ S λ P ⊗ S λ ′ Q ¯ b , over partitions λ with at most two rows, and the GL ( P ) -invariants of S λ P ⊗ S μ P ∨ are one-dimensional for λ = μ and zero otherwise. So the invariants of one factor in ⋀ 2 k ( P i b ⊕ P i ∨ b ) have dimension ∑ | λ | = k ( dim S λ ′ Q ¯ b ) 2 . For b = 1 this is 1 , 1 , 1 in k = 0 , 1 , 2 , and ( 1 + t 2 + t 4 ) 3 = 1 + 3 t 2 + 6 t 4 + 7 t 6 + 6 t 8 + 3 t 10 + t 12 gives dim Hdg p ( Z K ) = 1 , 3 , 6 , 7 , 6 , 3 , 1 . For b = 2 the partitions in a 2 × 2 box give 1 , 4 , 1 + 9 , 4 , 1 , so one factor contributes q ( t ) = 1 + 4 t 2 + 10 t 4 + 4 t 6 + t 8 , and the coefficients of q ( t ) 3 in degrees 0 , 2 , 4 , 6 , 8 are 1, 12, 3 · 10 + 3 · 4 2 = 78 , 3 · 4 + 6 · 4 · 10 + 4 3 = 316 and 3 · 1 + 6 · 4 · 4 + 3 · 10 2 + 3 · 10 · 4 2 = 879 . Products. linkinvariants.jl computes the invariants of the torus ∏ i ( G m × G m ) and the kernels of the raising and lowering operators of sl 2 3 on the weight-zero part of ⋀ * ( H 1 ( X ) a ⊕ H 1 ( Z K ) b ) , Künneth component by Künneth component, and compares with the dimension of B * ( X a ) ⊗ B * ( Z K b ) computed from the factors. With dim Hdg p ( X ) = 1 for p = 0 , ⋯ , 4 and dim Hdg p ( X 2 ) = 1 , 3 , 8 , 16 , 28 for p = 0 , ⋯ , 4 , the dimension of B * ( X a ) ⊗ B * ( Z K b ) in degree 2 p is ∑ i + j = p dim Hdg i ( X a ) dim Hdg j ( Z K b ) : on X × Z K in degrees 4 , 6 , 8 , 10 it is 6 + 3 + 1 = 10 , 7 + 6 + 3 + 1 = 17 , 6 + 7 + 6 + 3 + 1 = 23 and 3 + 6 + 7 + 6 + 3 = 25 ; on X 2 × Z K in degrees 4 , 6 , 8 it is 6 + 3 · 3 + 8 = 23 , 7 + 3 · 6 + 8 · 3 + 16 = 65 and 6 + 3 · 7 + 8 · 6 + 16 · 3 + 28 = 151 ; on X 2 × Z K 2 in degrees 4 , 6 it is 78 + 3 · 12 + 8 = 122 and 316 + 3 · 78 + 8 · 12 + 16 = 662 . The program finds 10, 20, 29, 35; 26, 98, 293; 134, 995 Hodge classes. In degree six on X × Z K the Künneth components ( 2 , 4 ) , ( 0 , 6 ) , ( 4 , 2 ) , ( 6 , 0 ) have dimensions 9 = 6 + 3 , 7, 3, 1, the 3 of ( 2 , 4 ) being F ι ′ of Theorem 39(i), and in degree four there is no mixed class; on X × X × Z K in degree four the total is 26, with 6 = 3 + 3 in the component ( 1 , 1 , 2 ) . The multiplicities used in Lemma 36 and Theorem 39(i) are: S 2 V 1 occurs once in ( ⋀ 2 D ) K 1 and in V ⊗ V , and not in ⋀ 2 V or ⋀ 4 V ; S 2 V 1 ⊗ S 2 V 2 occurs once in ( ⋀ 4 D ) K 1 and in ⋀ 2 V . Finally, for F defined by t 3 + t 2 − 2 t − 1 , t 3 − 3 t + 1 and t 3 − t 2 − 3 t + 1 , of discriminants 49, 81, 148, the elements N ( b ) 2 b − 2 for b = 1 + t , 2 + t , t + t 2 are irrational, so with 1 they span F, as in Theorem 39(ii).
Products with E c . Let t be a point that is not a CM point, and E the elliptic curve with complex multiplication by k of Theorem 44, with H 1 ( E , Q ¯ ) = Q ¯ y ι ⊕ Q ¯ y ι ¯ . The Hodge group of X × Z K × E projects onto those of X × Z K and of E, and its derived group is G, so it is the image of G × T ′ with T ′ ⊆ K 1 × k 1 a torus; write its characters as ( b ; e ) = ( b 1 , b 2 , b 3 ; e ) , so that H 1 ( X ) has character 0, the summands P i and P i ∨ of H 1 ( Z K ) have characters ± f i , with f 1 , f 2 , f 3 the standard basis, and y ι , y ι ¯ have e = − 1 and e = 1 . The projection of T ′ to K 1 is onto, because the Hodge group of Z K is the image of G × K 1 , so T ′ has rank at least 3; and the classes of Theorem 44(a), which are Hodge classes by the proof of that theorem, have over Q ¯ a nonzero component in V ∨ ⊗ D Φ 0 ⊗ ( y ι ⊗ y ι ′ ) ( y ′ on the second factor E), of character ( 1 , 1 , 1 ; − 2 ) when the signs of the f i are chosen so that D Φ 0 has character f 1 + f 2 + f 3 ; this character is therefore trivial on T ′ . So the characters trivial on T ′ are the multiples of ( 1 , 1 , 1 ; − 2 ) , and a vector in the cohomology of a product of copies of X, Z K and E is a Hodge class exactly when it is G-invariant of character ( d , d , d ; − 2 d ) for some d ∈ Z . H 1 ( X ) ⊗ H 3 ( Z K ) . The characters of H 3 ( Z K ) are sums of three vectors ± f i , whose coordinate sum is odd, and e = 0 forces d = 0 ; so there is no Hodge class. By Theorem 42(i) the V-isotypic part of H 3 ( Z K ) is ⨁ Φ D Φ , with one factor above each place, so dim Hom G ( V , H 3 ( Z K ) ) = 2 3 = 8 . Adding two factors E. Now e ∈ { − 2 , 0 , 2 } , and e = 0 is excluded as before. For e = − 2 , d = 1 , the factor y ι ⊗ y ι ′ is a line and the part of H 3 ( Z K ) of character ( 1 , 1 , 1 ) , one vector of P i above each place, is the summand D Φ 0 ≅ V , whose G-invariants in V ⊗ V form a line; the case e = 2 is conjugate. So the Hodge classes in H 1 ( X ) ⊗ H 3 ( Z K ) ⊗ H 1 ( E ) ⊗ H 1 ( E ) form a Q -plane, the k-line of γ ; replacing H 1 ( E ) ⊗ H 1 ( E ) by H 2 ( E × E ) adds two summands H 2 ( E ) of character e = 0 and nothing else. X × E c . The cohomology of X has character 0, so a Hodge class has e = 0 and d = 0 ; the part of character e = 0 of H * ( E c ) consists of Hodge classes of E c , so every Hodge class on X × E c is a sum of products of Hodge classes on the two factors, and there are no mixed classes. Z K × E c . For d = 0 the same argument applies. Otherwise | e | = 2 | d | ≤ c . For d = ± 1 the part above each place i must have character ± f i , hence lie in an odd exterior power of P i ⊕ P i ∨ = V i ⊗ W i , W i = χ i ⊕ χ i − 1 ; both ⋀ 1 and ⋀ 3 ≅ ( V i ⊗ W i ) ∨ ⊗ det are of type V i and have no invariants, and H * ( E c ) is trivial for G, so there is no Hodge class. For d = ± 2 , possible only for c = 4 , the factor of E 4 is the line spanned by the product of the four classes y ι in degree four, resp. its conjugate, and the part above each place must have character ± 2 f i , which in ⋀ * ( V i ⊗ W i ) occurs only in ⋀ 2 P i = ⋀ 2 V i ⊗ χ i 2 , resp. ⋀ 2 P i ∨ , a trivial G-line. This gives one class for d = 2 and one for d = − 2 , both in degree 6 + 4 = 10 : no mixed class for c ≤ 3 and a plane in degree ten for c = 4 . The program r4gamma.jl counts the G-invariants of each character by the multiplicities of the weights 0 and 2 of each factor SL 2 and confirms these numbers, together with the totals 26 and 20 computed above.

Appendix C.6. The Hochschild Rank on S 1 ×S 2

The proof of Proposition 16(a) reduces R ( v ) to the kernel of ( a 1 , a 2 ) ↦ ( ξ a 1 ⊗ 1 + 1 ⊗ ξ a 2 ) v , which is { ( a 1 , φ ˜ a 1 ) : φ ˜ * φ ˜ a 1 = μ a 1 } with a 1 ∈ s 1 ⊥ ∩ s ¯ 1 ⊥ . On the L-part every a 1 qualifies, which gives 2 m − 2 dimensions after removing s 1 and s ¯ 1 from the 2 m -dimensional L C ; on ( U ⊕ N ) C it gives V v . Since dim H T 2 ( Y ) = 2 · 22 = 44 , R ( v ) = 44 − ( 2 m − 2 ) − dim V v . For O Δ the map φ ˜ is the identity, the kernel is the diagonal, of dimension 22, and R = 44 − 22 = 22 ; without the bivector the two copies of C 1 are removed from the domain and one dimension from the kernel, so r = 42 − 21 = 21 . The program fullht2rank.jl realises H ˜ ( S ) = C 24 with the Mukai pairing and a period s in L C for e ∈ { 2 , 3 , 5 , 7 } and m ∈ { 3 , 4 , 5 , 7 , 10 } , checks that ξ 1 and ξ q pt agree with the contraction with the bivector and with the cup product with s ¯ , written out directly, and computes the three ranks R, r and the rank on H 1 ( T ) alone from singular values for seven shapes of v in each of the 20 cases: four with given U-blocks, one with random N-components and a random Ξ , a similitude, and the two special eigenvalue cases of (a). In all 140 cases the ranks agree with the formulas of (a), v is a Hodge class, and 〈 v , v 〉 = 2 m e c 2 + tr M * M ; for e = 2 , m = 10 the seven values of R are 25, 25, 26, 26, 22, 24, 25, so the generic value 46 − 2 m = 26 , the similitude value 22 and the special values 45 − 2 m and 44 − 2 m all occur. The controls v ( O Δ ) and its negative dual give R = 22 , r = 21 . Contracting into ch ( F ) in place of v changes the rank for some shapes, which is why v and not ch enters Proposition 16. The ranks are read from the singular values with relative threshold 10 − 8 .

Appendix C.7. Hermitian Forms of Weil Type

Theorem 27. For H = 〈 a 1 , ⋯ , a 2 n 〉 the trace form is q H = 〈 1 , d 〉 ⊗ 〈 a 1 , ⋯ , a 2 n 〉 , of discriminant d 2 n ( ∏ a i ) 2 ≡ 1 , so q H is hyperbolic exactly when its Hasse invariants are those of 2 n hyperbolic planes, by Hasse–Minkowski. The program cliffordcheck.jl draws 2400 diagonal forms of signature ( n , n ) , n ≤ 4 , with entries of absolute value in { 1 , 2 , 3 , 5 , 6 , 7 , 10 , 11 , 13 , 14 , 15 , 21 } , for d ∈ { 1 , 2 , 3 , 5 , 6 , 7 , 10 , 11 , 13 , 15 } , tests hyperbolicity of q H through the local invariants at every prime dividing 2 d ∏ a i and at infinity, and finds it equivalent to ( − d , ( − 1 ) n det H ) v = 1 at every place, in 497 split and 1903 non-split cases; for the 251 non-split forms with n ≥ 2 it confirms q H ≅ ( 2 n − 2 ) H ⊥ 〈 1 , d 〉 ⊗ 〈 1 , b 〉 with b = ( − 1 ) n − 1 det H anisotropic, the Witt index 2 n − 2 of q H ; and on 180 binary forms 〈 b 1 , b 2 〉 of signature ( 1 , 1 ) that the trace form is hyperbolic exactly when − b 1 b 2 ∈ N ( K × ) .
Theorem 28(ii) for m = 1 . Let B = ( − d , β ) with basis 1 , i , j , i j , i 2 = − d , j 2 = β , and E ( x , y ) = n trd ( x y ¯ i ) , n ∈ Q × , on V = B , with K = Q ( i ) acting by left multiplication. In the K-basis 1 , j of V, with H ( x , y ) = E ( i x , y ) + i E ( x , y ) : E ( 1 , 1 ) = n trd ( i ) = 0 and E ( i , 1 ) = n trd ( i 2 ) = − 2 d n , so H ( 1 , 1 ) = − 2 d n ; E ( j , j ) = n trd ( − β i ) = 0 and E ( i j , j ) = n trd ( − β i 2 ) = 2 β d n , so H ( j , j ) = 2 β d n ; and E ( 1 , j ) = n trd ( − j i ) = 0 , E ( i , j ) = n trd ( − i j i ) = n trd ( − d j ) = 0 , so H ( 1 , j ) = 0 . Hence H = 〈 − 2 d n , 2 β d n 〉 , of signature ( 1 , 1 ) exactly when β > 0 , and det H = − 4 d 2 n 2 β ≡ − β modulo squares, which is the case m = 1 of det H = ( − β ) m in Q × / N ( K × ) . The program qmdisc.jl takes B = ( − 1 , 3 ) , ( − 2 , 5 ) , ( − 3 , 2 ) , ( − 5 , 2 ) , ( − 7 , 3 ) , ( − 1 , 7 ) , ( − 3 , 5 ) and m = 1 , 2 , 3 , and for two random hermitian matrices ν each checks in exact arithmetic that E ν is alternating with E ν ( b x , y ) = E ν ( x , τ ( b ) y ) , that H is K-hermitian of signature ( m , m ) , and that det K H / ( − β ) m is a norm from K.

Appendix C.8. Convolutions on E 0 2 n

Lemma A9 and Theorem A9. H T 2 ( A ) = H 0 ( ⋀ 2 T A ) ⊕ H 1 ( T A ) ⊕ H 2 ( O A ) has dimension 2 n 2 + ( 2 n ) 2 + 2 n 2 , that is 15 + 36 + 15 = 66 for n = 3 and 28 + 64 + 28 = 120 for n = 4 ; the rank of the contraction with γ is r ( γ ) = 7 n 2 − 2 n , that is 57 and 104, so the bound of Theorem A9 is 58 and 105. The program contractioncheck.jl works in the exterior algebra over Q ( i ) on the 4 n forms d z j , d z ¯ j , and checks the Weil identity ∑ ζ ∏ j ζ j e c 1 ( L ζ ) = 2 · 4 n − 1 ( α + α ¯ ) ; that T D n , n , computed as the K-linear v with v ⌟ θ = 0 , has dimension n 2 and contains v 0 ⌟ ; that v 0 ⌟ c 1 ( L ζ ) ≠ 0 for all 32, resp. 128, pieces L ζ and v 0 ⌟ c 1 ( M i ) = 0 ; that the map T D n , n → H 0 , 2 , v ↦ v ⌟ c 1 ( L ζ ) , has rank 3, resp. 6, for every ζ ; and that r ( γ ) = 57 , resp. 104, with T D n , n in the kernel.
Theorem A8. With d i = t i − p the Ext degree of a step between a bundle L ζ and M i is bounded below by Lemma A7 applied to the eigenvalues d i ± 1 on each surface. The program excesscheck.jl computes, for | t i − p | ≥ 2 , the least excess of a step between bundles over all 32 2 , resp. 128 2 , ordered pairs, which is 1, and the least total excess of a closed path of distinct pieces among four bundles and the three multiples, over every order of p among the t i , which is 2; excessgeneral.jl drops the hypothesis and runs over all pairwise distinct d i ∈ { − 5 , ⋯ , 5 } , using the degree ranges of Lemma A7, and finds the least total excess 1 for n = 3 , attained at ( d 1 , d 2 , d 3 ) = ( − 5 , − 4 , − 1 ) , and 2 for n = 4 ; this is the bound n − 2 of the proof, and no placement gives total excess 0.
Proposition A16(ii). There are 128 bundles L ζ , ζ ∈ μ 4 4 with ∏ ζ = ± 1 , and three multiples, so 131 pieces. On E 0 8 the space H 2 ( O ) of one piece decomposes by the types τ ∈ { 0 , 1 , 2 } 4 , | τ | = 2 : four types 2 e j of dimension 1 and six types e j + e k of dimension 2 · 2 = 4 , in all 28; so dim D 2 = 131 · 28 = 3668 . For τ = 2 e j the components of the graph of the proof are the four sets { ζ j = z } , and for τ = e j + e k the sixteen sets { ( ζ j , ζ k ) = ( z , z ′ ) } , so the kernel has dimension at least 3 · 28 + 4 · 4 · 1 + 6 · 16 · 4 = 84 + 16 + 384 = 484 ; cupkernel.jl recomputes the components by union–find on the potential edges.
Proposition A16(iii). The program mchains.jl enumerates the chains s = a 0 → ⋯ → a K = e of distinct pieces among one bundle class L and the multiples M 1 , M 2 , M 3 , with steps up (excess − 1 ), down (excess 7) and between bundles (excess at least 1), for each of the four positions of p among the values, keeps those that pass the degree test of the proof and removes those with two consecutive steps up through a multiple whose value is the only one strictly between. For each position five chains pass the degree test; one, two, two and one survive, namely M 2 → M 3 → L → M 1 ; M 2 → M 3 → M 1 → L and M 3 → M 1 → L → M 2 ; L → M 3 → M 1 → M 2 and M 2 → L → M 3 → M 1 ; M 3 → L → M 1 → M 2 , with the multiples numbered by increasing value. These are the six types: each has four leaves, the three multiples and one bundle, and ϵ = 5 .
Proposition A14. On one surface B j the forms are L ζ = p ζ j ζ ¯ j p and M i = t i I , and g a fixes a piece L exactly when Im ( a * F l ) ∈ Z for all l ∈ Z [ i ] 2 , F = L − L 1 . For F = ( t i − p ) I − 0 z z ¯ 0 and F = 0 z − z ′ z − z ′ ¯ 0 this is a congruence on a modulo Z [ i ] 2 ; the program equivgroup.jl solves it over the 8-torsion points for eight placements at n = 3 , 4 and finds G j = Z / 2 , generated by ( h , h ) , exactly when all t i − p are odd, and G j = 0 otherwise, G isotropic for the commutator pairings, and | G S | = 2 | S | dividing D ( ζ , ζ ′ ) = ∏ k ∈ S | ζ k − ζ k ′ | 2 , since | 1 − u | 2 ∈ { 2 , 4 } for u ∈ μ 4 ∖ { 1 } .
Theorem A21: the slices. In the slice ( j , z ) a piece ζ is joined to ζ ′ when they differ in the three other coordinates by units u k = ζ k ′ / ζ k ≠ 1 with ∏ u k = − 1 , so that the parities differ. Among the 27 triples in { i , − 1 , − i } 3 the product is − 1 exactly for ( − 1 , − 1 , − 1 ) and the six orderings of ( − 1 , i , − i ) : an odd number of factors ± i gives a product ± i , and two factors ± i with one − 1 give − 1 only when they are i and − i . So the graph is 7-regular with 32 · 7 / 2 = 112 edges, and the weights D / 8 are 4 3 / 8 = 8 and 4 · 2 · 2 / 8 = 2 , of sum 8 + 6 · 2 = 20 at each vertex. A single vertex is a cut of weight 20; slices.jl computes the global minimum cut of each of the 16 slice graphs by the algorithm of Stoer and Wagner [113] in exact integers and finds 20, and checks connectedness.
Theorem A21: the totals. There are ∑ N ≤ 5 16 N = 1 + 16 + 120 + 560 + 1820 + 4368 = 6885 sets Σ of at most five slices. For Σ = ∅ no edge survives, each piece is its own component and κ = 84 + 28 · 128 = 3668 ; with all edges the bound is the 484 of Proposition A16(ii). The program twolevel.jl computes min κ ( Σ ) = 3668 , 3668, 3632, 3545, 3428, 3311 for N = 0 , ⋯ , 5 by union–find on the edges of the theorem, so that 20 N + κ − ( 84 + 28 · 64 ) = 20 N + κ − 1876 is 1792, 1812, 1796, 1729, 1632, 1535; for N ≥ 6 the slices give at least 20 · 6 = 120 .

Appendix D. Machine Checks

Every finite quantity used in the proofs is recomputed by the Julia programs of the repository [9], and the combinatorial and arithmetic statements among them are also proved in Lean 4. No statement proved in the text rests on these checks, except the enumerations that are stated as such: those of Proposition 27, Theorem 29, Proposition A16(iii) and Theorem A21, the strong connectivity in Proposition A12(ii) and the list of Proposition A13, Remarks 9 and 14, the Picard number of Remark 1 and the counts of Appendix C that are marked as computed; and the rest guard against slips in the hand computations.

Lean.

The file lean/HodgeClosure.lean (Lean 4.24.0, no Mathlib) contains ninety-nine theorems. Section 1 encodes the eight statements and the rule set of Table 2 as data, computes closures by iteration, and proves by evaluation that the closures of { F 2 , F 3 ′ } , { F 3 } , { L , M } , { L , F 3 ′ } and { V , F 3 ′ } contain HC , that the closures of { F 2 } , { F 3 ′ } , { L } , { M } , { V } , { AB } , { L , V } and { M , F 3 ′ } do not, and that the list of minimal sufficient sets, computed from the rule set alone, has exactly six members, those of Proposition 4. Section 2 proves the dimension counts of Section 6: 3 n < n 2 for n ≥ 4 , m 2 + m ′ 2 < ( m + m ′ ) 2 for m , m ′ ≥ 1 with 2 m m ′ mixed directions, the values r = 57 and 104 of the lower bound, and the Hodge class count 26 of Theorem 37(i) as a sum of Künneth contributions. Section 3 proves the polynomial identities behind the Fricke families of Section 5.4: c 4 3 − c 6 2 = 1728 t ( t 2 + 13 t + 49 ) 2 at level 7 and 1728 t ( t 2 + 22 t + 125 ) 3 at level 5, by exact arithmetic on coefficient lists, together with the root products 7 2 and 5 3 , the Euler numbers 12 and 24, the Shioda–Tate counts giving Picard number at least 16, the transcendental rank 6, the dimensions ( n + 1 ) / 2 of the dihedral Hecke algebras, and the polynomial identity behind the sign of Lemma 9: with A ′ , B ′ the Vélu coefficients of the quotient of (2) by C t , computed from the kernel polynomial by Newton’s identities on coefficient lists, and A ˜ ( t ) = t 4 A ( N k / t ) , B ˜ ( t ) = t 6 B ( N k / t ) , the identity N k − 1 B ′ A ˜ + A ′ B ˜ = 0 holds at both levels, which is λ ( t ) 2 = − t 2 / N k − 1 . Section 4 proves the Dickson identities p 7 ( t + t − 1 ) = t 7 + t − 7 and p 9 ( t + t − 1 ) = t 9 + t − 9 behind the families of Theorem 14, that 2 has order three in ( Z / 7 ) × / ± 1 and in ( Z / 9 ) × / ± 1 , so that 2 is inert in both cubic fields, and the sign pattern of the conjugates of θ at the three real places. Section 5 proves the finite statements of Appendix B: the binomial identity ∑ q 3 q 3 k − q = 6 k , the intersection numbers ( C ζ · C ζ ′ ) = | ζ − ζ ′ | 2 and θ · C ζ = 2 in the basis of the ( 1 , 1 ) -forms of B j with the pairing a j b j = 1 , c j c ¯ j = − 1 , the twisted sums T m , j , the identity S ( 1 ) − S ( − 1 ) = 32 ( α + α ¯ ) as an equality of the 64 coefficients of the triple products over Z [ i ] , the counts 48, 8, 75 and 18 with the Euler characteristics, and the two chain statements of Proposition 25 over the twelve pieces of Appendix B.3, each by decide. Section 6 encodes the pieces of Table 6 and the Ext dimensions of Lemmas 15 and 16, and proves that the isolated blocks of Lemma 17 sum to 7213, 2605, 2605, 0, 0 and 5325 at the six configurations named in the table; the reachability relation along blocks of total degree one is supplied as a certificate, which the kernel checks to be reflexive, to contain every such block and to be closed under one more step, before the sum is evaluated by decide +kernel. From the same certified relation it proves that the configurations (a) and (c) of the table have 32 and 22 threefold pieces outside every directed cycle of the graph Γ of Section 6.6; and, by evaluating the Ext dimensions on every pair, that no edge of Γ enters a piece at shift 1 from the thirty-two pieces or from the pieces O Z ζ ⊗ L k [ H ] , ζ ∈ μ 4 3 , 1 ≤ k ≤ 3 , and L l [ H ] , 1 ≤ l ≤ 4 , with − 3 ≤ t ≤ 2 , that L 1 [ 3 ] , L 4 [ 3 ] and O Z ( 1 , 1 , 1 ) ⊗ L 1 [ 3 ] do enter, and that L 1 [ − 2 ] , ⋯ , L 6 [ 3 ] put each of the thirty-two pieces on a cycle of length six. Section 7 proves further arithmetic of Section 6 and Section 5: the four values r = 57 , 48, 39, 30 of ( 4 + ρ ) n 2 − 2 n at n = 3 (Remark 7), the two Gram blocks 0 2 d 2 d 0 and − 1 0 0 − d of T d and its determinant 16 d 5 for the squarefree d ≤ 30 (Lemma 12), the numerical conditions of Nikulin’s criterion and the order 16 d of the discriminant group of L d , the discriminants 49, 81 and 148 of the cubic fields of Theorems 14 and 41, and the factorials in the coefficients of θ h k used in Remark 7. Section 8 proves the arithmetic of Section 5.8, Section 6.9 and Section 6.14 and of Appendix C, with Hilbert symbols and Hasse invariants computed from valuations and Euler’s criterion: the Gram matrices, signatures, discriminants and the failing Hasse invariant of the five spaces of Table 4; the witnesses of Witt index two and one after Corollary 4 and the rank-five example 〈 1 , 1 , − 1 , − 23 , − 23 〉 ; the count of 1484 forms of Witt index two among the 1980 diagonal forms, evaluated by decide +kernel from condition (d) of Lemma 11(ii); the Hodge numbers of Z K and Z K 2 of Proposition 41 and the dimensions of the products of the algebras B * after Theorem 38; the count 9 = 6 + 3 of Theorem 39; the identity 2 b = ( b + 1 ) 2 − b 2 − 1 and the norm identity behind Theorem 37(iii), and the linear system of the type conditions in the proof of Theorem 40(ii), on finite grids; the counts 279 · 7 = 1953 and 64 = 40 + 24 of Theorem 29; the discriminants of the sextic fields of Appendix C; and the degrees of Proposition 43. Section 9 proves the finite statements of Section 5.9, Section 6.9 and Section 6.16, of Appendix A.6 and of Appendices Appendix C.5 and Appendix C.8: the identity Ω m = m ! a 1 b 1 ⋯ a m b m for m ≤ 4 behind Theorem 28(iii), as an equality of all coefficients in the exterior algebra; that 2 | S | divides D ( ζ , ζ ′ ) for all pairs of the 128 pieces (Proposition A14); that every slice graph is 7-regular of weighted degree 20; the type count 28, dim D 2 = 3668 and the kernel bound 484 of Proposition A16(ii); the threshold 1876 = 84 + 28 · 64 , the totals 1792, 1812, 1796, 1729, 1632, 1535, the number 6885 of slice sets and the bound 120 of Theorem A21; the least excess 1 of a step between two bundles L ζ over all ordered pairs at n = 3 , 4 (Theorem A8); the bounds 58 and 105 of Theorem A9; the targets 46 − 2 m of Proposition 16 and the bound 44 of Corollary 9; the table of Proposition 18, m ( e + 1 ) ≤ 23 for squarefree e ≤ 60 ; the self-intersection 8 e + 16 of Proposition 19; and the counts 8, 0 and 2 of Appendix C.5, by evaluating the characters and SL 2 -weights. Section 10 proves the finite statements of Section 6.11 and of Section A.5: the contraction ranks 1 , 6 , 6 , 1 and 1 , 8 , 12 , 8 , 1 of Lemma 30, with dim Ann 2 = n 2 and Ann 1 = 0 at n = 3 , 4 ; the Euler characteristic 1 − 8 + 12 − 8 + 1 = − 2 of the forced dimensions and the rank 48 of the Künneth contraction in Theorem 30; the equivalence n ( n − 1 ) = 2 n ⇔ n = 3 for n ≥ 1 ; for the lattice of Theorem 32, that − ( 128 a 2 + 32 b 2 ) lies in 8 Z and is never − 2 , and that the least nonzero value of | χ | on the box | a | , | b | ≤ 20 is 32; the positivity of the cost bounds of Theorem A18 for 3 ≤ n ≤ 30 up to the thresholds 2 n − 3 and 2 n − 2 ; the 46 edges inside a cube, the edges into M { 1 , 2 , 3 } and out of M ⌀ , and the 32 cubes, sixteen with c = 0 , of Lemma A21; and the acyclicity of the collapsed graph of Theorem A19. Every theorem is checked by #print axioms: eighty-five use no axiom; the two inequalities of Section 2 and the two statements of Section 10 proved with the arithmetic tactic use propext and Quot.sound; and the six isolated-block counts and the two chain-free counts of Section 6, the identity Ω m of Section 9 and the cube-window count of Section 10 use propext; native_decide is not used. The file compiles in about five minutes on one core.

Julia.

The directory julia/ holds fifty-two programs for Julia 1.11.7 that use only the standard modules LinearAlgebra, Random and Printf. Each prints one line PASS or FAIL per check and a final count, and exits with a nonzero code if a check fails; julia/runall.jl runs all of them. Together they make 904 checks, all of which pass, in about nine minutes; the outputs are deterministic, with fixed seeds where generic data are drawn. Arithmetic is exact (integers, Rational{BigInt}, Gaussian rationals, Q ( − 3 ) , residues modulo primes with rational reconstruction) except in three places: the ranks of productobject.jl, convolutionchecks.jl and fullht2rank.jl, computed from singular values with relative tolerance 10 − 7 , resp. 10 − 8 in the last, which agree with the closed formulas, and the roots of j ( t ) = j ( τ ) in frickefamilies.jl, computed with 256-bit floating point. Table A2 lists the programs.
Table A2. The Julia programs of the repository, the results they check and their number of checks.
Table A2. The Julia programs of the repository, the results they check and their number of checks.
program what it checks checks
quadraticforms.jl Hilbert symbols, Hasse invariants and the Hasse–Minkowski classification, shared by the lattice programs 10
closuregraph.jl closures of the 256 subsets of the eight statements, the six minimal sets (Prop. 4), the rules with the literature adjoined and the blow-up bookkeeping of Prop. 46 18
sqrt6chain.jl Thm. 11: the 165 forms over Q ( 6 ) (98 embed, 67 do not), the complements, both orders of the chain, Varesco’s 2 , 3 as controls 25
frickefamilies.jl fibre types, Euler numbers and ρ ≥ 16 of the Fricke families; the Fricke involution on j; the twist identity; the Hecke algebras for n = 5 , 7 , 9 , 11 41
frickesign.jl Lemma 9: kernel polynomials, Vélu’s quotient, λ ( t ) 2 , the twist constants, Λ 2 = ∓ N 32
frickepicard.jl Rem. 1: point counts over F ℓ i , the characteristic polynomial of Frobenius on T and its factorisation over Q ( − N ) 12
frickelattice.jl Prop. 15(i): T ≅ U Q 2 ⊕ 〈 − 1 〉 ⊕ 〈 − N 〉 and the rational isometries of root lattices 21
zetainvariants.jl Thm. 14: the trace forms of ( − 1 , ± θ ) F against the transcendental lattices for n = 7 , 9 , three contrasting algebras, the Dickson identity 19
kummertype.jl Prop. 14 against a direct search; the counts 1084 / 1296 and 1304 / 4320 15
ranksix.jl Lemma 12 and Thm. 16 for the 305 squarefree d ≤ 500 ; the embeddings in V F 35
productobject.jl Thm. 21 in the exterior-algebra model at ( m , m ′ ) = ( 2 , 1 ) , ( 2 , 2 ) : Hodge loci, cross terms, Hankel rank, r = 57 , 104 32
convolutionchecks.jl Prop. 23: contraction ranks, the annihilators under the orthogonality hypotheses, the block trace 26
annihilatorexact.jl Thm. 22 over Q ( − 3 ) : formula (11) and the annihilator of dimension m 2 + m ′ 2 34
riemannrochpieces.jl the polynomial (12) and the thirty-one supports of Prop. 24 8
cmpointext.jl Lemmas 15, A28 and Prop. 25: Koszul cohomology, the twisted groups, the closed chains of two and three pieces 14
cmpointcounts.jl Table 6 and Lemma 17: the Chern characters, C 1 , C 2 , C 3 and the isolated blocks 13
cmpointchains.jl Prop. 28, Cor. 12, Rem. 6: the graph Γ and its directed cycles 12
Table A3. The Julia programs of the repository (continued).
Table A3. The Julia programs of the repository (continued).
program what it checks checks
linebundlechains.jl Thm. 23, Rem. 9: the arrangements of n ≤ 8 line bundles and their counts 6
linebundlefine.jl rule E3* (Lemma 23), conditions C1–C4 in two implementations, the 165 and 269 states, the nine-bundle arrangements 17
chernh.jl Rem. 7: the weight-zero Chern character against θ h for h ∈ { 1 , 2 , 3 } 6 , the ranks 57, 48, 39, 30 9
threefoldsupports.jl Section 6.8: Thm. 25, Lemmas 25, 27, 26 and Thm. 26 in exact arithmetic 16
coverboundary.jl Prop. 39: smoothness of B λ , the divisor of y 3 + 1 / λ , η ≠ 0 , the rank nine 10
sexticlattice.jl Rem. 14: the integral points of S ( 0 , q ) over the twelve cubic fields and their least values of − χ 24
sixfoldlink.jl Thm. 37: ψ c , the K 1 -invariants, the count 26, the weights of ξ ; the shared isogeny factor at L = K 55
mumfordcm.jl Thms. 40, 41: invariants by Künneth type, Pohlmann’s criterion on E a × A b and on X t 51
mumforddisc.jl Thm. 40(iii): det H = b N F / Q ( a ) d F and discriminant one, for four cubic fields 24
mumfordresidual.jl Rem. 15: the residual case L = K , the Hodge torus, the lattices and dimensions, the fields of degree twelve, discriminant-one polarisations 86
k3wittindex.jl Lemma 11, Cors. 3, 4, Thm. 17, Cors. 5, 6: embeddings in V m , the Witt index in rank six, the counts 1484 / 1980 , the local criterion for trace forms with the local symbols over E computed independently 20
splitpolarisation.jl Lemma 28 on four totally real sextic fields: for each of the twenty sign patterns a split H a , its Gram determinant, signature and symbols 37
cmsixfolds.jl Thm. 29: the 279 transitive subgroups of S 6 , the Hodge rings of the 1953 pairs ( Γ , I ) , the cyclotomic fields of degree twelve 7
linkinvariants.jl Prop. 41, Thms. 38, 39: Hodge numbers of X a × Z K b and of the subalgebras B * 14
r4gamma.jl Thms. 42, 44: Hodge classes on X a × Z K b × E c by characters and SL 2 -weights (App. Appendix C.5) 12
fullht2rank.jl Prop. 16: the operators ξ a , the ranks R, r for seven shapes in 20 cases, the control O Δ 25
cliffordcheck.jl Thm. 27: hyperbolicity of q H against the quaternion algebra, the binary criterion, the Witt index 3
qmdisc.jl Thm. 28(i),(ii): the polarisations of type II, signature ( m , m ) , det H = ( − β ) m modulo norms 21
contractioncheck.jl Lemma A9, Thm. A9: the Weil identity, T D n , n , v 0 c 1 ( L ζ ) ≠ 0 , r = 57 , 104 10
excesscheck.jl Thm. A8 for | t i − p | ≥ 2 : excess of steps and of closed paths 4
excessgeneral.jl Thm. A8: closed paths for all placements of the t i 2
equivgroup.jl Prop. A14: the groups G j , isotropy, 2 | S | ∣ D 8
cupkernel.jl Prop. A16(ii): the components and the kernel bound 484 4
mchains.jl Prop. A16(iii): the chains through the multiples and the six surviving types 5
slices.jl Thm. A21: the 16 slice graphs and their minimum cuts 2
twolevel.jl Thm. A21: κ ( Σ ) over the 6885 slice sets, the totals 1792 , ⋯ , 1535 and 120 3
secantranks.jl Lemmas 30, 31, Thm. 30: contraction ranks into secant classes, the forced dimensions, χ = 8 q ( q r 2 + s 2 ) Pf ( h ) , the ideal sheaf of a point 16
kunnethrank.jl Lemma 30 for q < 0 ; the rank 48 of the Künneth contraction on X × X (Thm. 30) 7
kappa.jl Thm. 35 on E 4 : Orlov’s equivalence on cohomology, κ 1 = 0 and κ 2 = ( 2 P 0 Q 0 / r ) Ω 2 for six pairs, the rank-zero case 10
lattice.jl Thm. 32: D k = d 1 ⋯ d k for forty forms, χ ∈ 8 Z on the integral generators, the lattice 16 Z λ 1 ⊕ 4 Z λ 2 4
invcheck.jl Lemma A15: the joint kernel of T D and its conjugate on H * ( E 0 2 n , C ) , n = 3 , 4 4
windowsearch.jl Thm. A18: least cost of a cycle through a proper piece, the thresholds 2 n − 3 and 2 n − 2 4
cubeexample.jl Prop. A12: ch = − 32 ( α + α ¯ ) + p ( θ ) , Hankel rank 3, the Koszul windows, strong connectivity 6
collapse.jl Lemma A21, Thm. A19: the window graph in two models, the collapsed graph, entries and exits of cubes 7
subsets.jl Prop. A13: the 44 unions of whole cubes with Chern character in Q θ 3 ⊕ W 4
total 904

Acknowledgements

The paper builds on the theorems of Eyal Markman on Weil classes and on Hodge isometries, of Salvatore Floccari on sixfolds of generalised Kummer type, of Kenji Koike on Weil sixfolds, on the work of Bert van Geemen, Matthias Schütt and Mauro Varesco on real multiplication of K3 surfaces, on Nikolay Buskin’s theorem on Hodge isometries, on the semiregularity theory of Spencer Bloch, Ragnar-Olaf Buchweitz and Hubert Flenner, and on the descriptions of Hodge classes on abelian varieties by Henry Pohlmann, Ben Moonen, Yuri Zarhin, Yves André and James Milne. We thank Priyabrata Mandal for the earlier collaboration on the closure graph and the polarised criterion, whose results are reproved in Appendix A.

Author Contributions

Deep Bhattacharjee, who leads the project, posed the problem in the form treated here, carried out the research and wrote the paper. The proofs are by Deep Bhattacharjee: the reduction and the dichotomy, the theorems on real multiplication and on the Fricke families, the transcendental rank six construction, the Witt index criteria and the dense curves, the results on Weil sixfolds with a split polarisation, the sixfold link and its behaviour at the points of complex multiplication, the constructions and exclusions at the points E 0 6 and E 0 8 , and the appendices. Deep Bhattacharjee made all the computations, the Lean 4 file and the Julia checks of Appendix D and the figures, and brought the work to its present state. Ushashi Bhattacharya read the manuscript and verified its arguments and references. Both authors approved the final version.

Open Gluon and Anaxia

Two tools built by D. B. stand behind the way this project was organised. Open Gluon is a workspace D. B. assembled and maintains, which keeps pinned copies of the mathematical libraries, the source record of the project and the checks in one reproducible place. Anaxia, the engine of the eBLLM project, is the construction engine of D. B.: it stores each construction as a route with its input, its output, the obligations that must be discharged and a test that makes the route fail when an obligation is skipped, and it certifies finite statements in exact rational arithmetic. D. B. wrote both and developed them from the constructions of earlier papers. Three of their constructions were used here. The first is the encoding of the stopping points of the paper as typed targets, from which Anaxia built the candidate configurations at E 0 6 of Table 6, the quotient chains of Proposition 20 and the first enumerations of the chains of line bundles behind Remark 9. The second is the reading of supports from a localised obstruction, which led to the statements of Section 6.8. The third is a Frobenius test on elliptic K3 surfaces, which first showed that a Fricke self-map is a complex and not a real multiplication, the observation behind Lemma 9. The engine proposes routes and tests them; every statement of this paper is proved in the text.

References

  1. D. Arinkin and A. Căldăraru, When is the self-intersection of a subvariety a fibration?, Adv. Math. 231 (2012), no. 2, 815–842. [CrossRef]
  2. A. Andreotti and T. Frankel, The Lefschetz theorem on hyperplane sections, Ann. of Math. (2) 69 (1959), 713–717. [CrossRef]
  3. D. Abramovich, K. Karu, K. Matsuki and J. Włodarczyk, Torification and factorization of birational maps, J. Amer. Math. Soc. 15 (2002), no. 3, 531–572. [CrossRef]
  4. Y. André, Mumford–Tate groups of mixed Hodge structures and the theorem of the fixed part, Compositio Math. 82 (1992), no. 1, 1–24. numdam.org/item/CM_1992__82_1_1_0.
  5. Y. André, Une remarque à propos des cycles de Hodge de type CM, in Séminaire de Théorie des Nombres, Paris, 1989–90, Progr. Math. 102, Birkhäuser, Boston, 1992, 1–7.
  6. Y. André, Pour une théorie inconditionnelle des motifs, Inst. Hautes Études Sci. Publ. Math. 83 (1996), 5–49. [CrossRef]
  7. M. Artebani, A. Sarti and S. Taki, K3 surfaces with non-symplectic automorphisms of prime order, Math. Z. 268 (2011), no. 1–2, 507–533. [CrossRef]
  8. M. F. Atiyah, Vector bundles and the Künneth formula, Topology 1 (1962), no. 3, 245–248. [CrossRef]
  9. D. Bhattacharjee and U. Bhattacharya, Lean 4 file and Julia checks for this paper, GitHub repository creelie/fermat-hodge-, release v5.0.0, 2026. https://github.com/creelie/fermat-hodge-.
  10. W. L. Baily, Jr. and A. Borel, Compactification of arithmetic quotients of bounded symmetric domains, Ann. of Math. (2) 84 (1966), no. 3, 442–528. [CrossRef]
  11. R.-O. Buchweitz and H. Flenner, A semiregularity map for modules and applications to deformations, Compositio Math. 137 (2003), no. 2, 135–210. [CrossRef]
  12. R.-O. Buchweitz and H. Flenner, The global decomposition theorem for Hochschild (co-)homology of singular spaces via the Atiyah–Chern character, Adv. Math. 217 (2008), no. 1, 243–281. [CrossRef]
  13. D. Bhattacharjee, Generalization of quartic and quintic Calabi–Yau manifolds fibered by polarized K3 surfaces, Research Square, 2022. [CrossRef]
  14. D. Bhattacharjee, Calabi–Yau saturation universality: toric mirror laws, Hodge statistics, and the higher-dimensional landscape, Int. J. Professional Studies 21 (2026), 211–256. [CrossRef]
  15. D. Bhattacharjee, N=1/F-theory resolutions, Int. J. Professional Studies 21 (2026), 945–982. [CrossRef]
  16. T. Bouchet, J. Hanselman, A. Pieper and S. Schiavone, Mumford-type Shimura curves contained in the Torelli locus, preprint, 2025. arXiv:2510.00093.
  17. W. Barth, K. Hulek, C. Peters and A. Van de Ven, Compact Complex Surfaces, 2nd ed., Ergebnisse der Mathematik und ihrer Grenzgebiete (3) 4, Springer, Berlin, 2004. [CrossRef]
  18. A. I. Bondal and M. M. Kapranov, Enhanced triangulated categories, Math. USSR-Sb. 70 (1991), no. 1, 93–107. [CrossRef]
  19. C. Birkenhake and H. Lange, Complex Abelian Varieties, 2nd ed., Grundlehren der mathematischen Wissenschaften 302, Springer, Berlin, 2004. [CrossRef]
  20. S. Bloch, Semi-regularity and de Rham cohomology, Invent. Math. 17 (1972), 51–66. [CrossRef]
  21. A. Borel, Some metric properties of arithmetic quotients of symmetric spaces and an extension theorem, J. Differential Geom. 6 (1972), 543–560. [CrossRef]
  22. A. Borel, Linear algebraic groups, second edition, Graduate Texts in Mathematics 126, Springer, New York, 1991. [CrossRef]
  23. P. Brosnan, Discriminants of Hermitian forms, maximal degenerations and Kontsevich’s tropical approach to the Hodge conjecture, preprint, 2026. arXiv:2609.14169.
  24. S. Bloch and V. Srinivas, Remarks on correspondences and algebraic cycles, Amer. J. Math. 105 (1983), no. 5, 1235–1253. [CrossRef]
  25. S. Bando and Y.-T. Siu, Stable sheaves and Einstein–Hermitian metrics, in: Geometry and Analysis on Complex Manifolds, World Sci. Publ., River Edge, NJ, 1994, pp. 39–50.
  26. N. Buskin, Every rational Hodge isometry between two K3 surfaces is algebraic, J. Reine Angew. Math. 755 (2019), 127–150. [CrossRef]
  27. E. Bayer-Fluckiger, B. van Geemen and M. Schütt, K3 surfaces with real or complex multiplication, Algebr. Geom., to appear. arXiv:2401.04072.
  28. A. Căldăraru, The Mukai pairing, II: the Hochschild–Kostant–Rosenberg isomorphism, Adv. Math. 194 (2005), no. 1, 34–66. [CrossRef]
  29. A. Căldăraru and S. Willerton, The Mukai pairing, I: a categorical approach, New York J. Math. 16 (2010), 61–98.
  30. A. Conte and J. P. Murre, The Hodge conjecture for fourfolds admitting a covering by rational curves, Math. Ann. 238 (1978), 79–88. [CrossRef]
  31. D. Calaque, C. A. Rossi and M. Van den Bergh, Căldăraru’s conjecture and Tsygan’s formality, Ann. of Math. (2) 176 (2012), no. 2, 865–923. [CrossRef]
  32. M. A. de Cataldo and L. Migliorini, The Chow groups and the motive of the Hilbert scheme of points on a surface, J. Algebra 251 (2002), no. 2, 824–848. [CrossRef]
  33. P. Deligne, Théorie de Hodge. II, Inst. Hautes Études Sci. Publ. Math. 40 (1971), 5–57. [CrossRef]
  34. P. Deligne, La conjecture de Weil pour les surfaces K3, Invent. Math. 15 (1972), 206–226. [CrossRef]
  35. P. Deligne, Hodge cycles on abelian varieties (notes by J. S. Milne), in: Hodge Cycles, Motives, and Shimura Varieties, Lecture Notes in Math. 900, Springer, Berlin, 1982, pp. 9–100.
  36. D. Eisenbud, Commutative Algebra with a View Toward Algebraic Geometry, Graduate Texts in Mathematics 150, Springer, New York, 1995. [CrossRef]
  37. A. Elagin, On equivariant triangulated categories, preprint, 2014. arXiv:1403.7027. arXiv:1403.7027.
  38. C. Faber, Prym varieties of triple cyclic covers, Math. Z. 199 (1988), 61–79. [CrossRef]
  39. S. Floccari and L. Fu, The Hodge conjecture for Weil fourfolds with discriminant 1 via singular OG6-varieties, J. Math. Pures Appl. (9) 210 (2026), 103876. [CrossRef]
  40. H. Federer and W. H. Fleming, Normal and integral currents, Ann. of Math. (2) 72 (1960), 458–520. [CrossRef]
  41. W. Fulton and J. Harris, Representation Theory: A First Course, Graduate Texts in Mathematics 129, Springer, New York, 1991. [CrossRef]
  42. S. Floccari, Sixfolds of generalized Kummer type and K3 surfaces, Compos. Math. 160 (2024), no. 2, 388–410. [CrossRef]
  43. S. Floccari, K3 surfaces associated with varieties of generalized Kummer type, Geom. Topol. 30 (2026), no. 3, 1129–1154. [CrossRef]
  44. W. Fulton, Intersection Theory, 2nd ed., Ergebnisse der Mathematik und ihrer Grenzgebiete (3) 2, Springer, New York, 1998. [CrossRef]
  45. F. Galluzzi, Abelian fourfold of Mumford-type and Kuga–Satake varieties, Indag. Math. (N.S.) 11 (2000), no. 4, 547–560. [CrossRef]
  46. W. M. Goldman and J. J. Millson, The deformation theory of representations of fundamental groups of compact Kähler manifolds, Inst. Hautes Études Sci. Publ. Math. 67 (1988), 43–96. [CrossRef]
  47. J. Grivaux, Chern classes in Deligne cohomology for coherent analytic sheaves, Math. Ann. 347 (2010), no. 2, 249–284. [CrossRef]
  48. P. A. Griffiths, On the periods of certain rational integrals. I, Ann. of Math. (2) 90 (1969), 460–495. [CrossRef]
  49. A. Grothendieck, Techniques de construction et théorèmes d’existence en géométrie algébrique IV: les schémas de Hilbert, Séminaire Bourbaki, Vol. 6 (1960/61), Exp. No. 221, 249–276; reprinted by Soc. Math. France, Paris, 1995. numdam.org/item/SB_1960-1961__6__249_0.
  50. A. Garbagnati and A. Sarti, Symplectic automorphisms of prime order on K3 surfaces, J. Algebra 318 (2007), no. 1, 323–350. [CrossRef]
  51. R. Goodman and N. R. Wallach, Symmetry, Representations, and Invariants, Graduate Texts in Mathematics 255, Springer, New York, 2009. [CrossRef]
  52. R. Hartshorne, Local Cohomology, a seminar given by A. Grothendieck, Lecture Notes in Mathematics 41, Springer, Berlin, 1967. [CrossRef]
  53. R. Harvey and H. B. Lawson, Calibrated geometries, Acta Math. 148 (1982), 47–157. [CrossRef]
  54. W. V. D. Hodge, The topological invariants of algebraic varieties, in: Proceedings of the International Congress of Mathematicians (Cambridge, Mass., 1950), Vol. 1, Amer. Math. Soc., Providence, 1952, pp. 182–192. Zbl 0048.41701.
  55. S. Huang, A note on a question of Markman, J. Pure Appl. Algebra 225 (2021), no. 9, 106673. [CrossRef]
  56. D. Huybrechts, E. Macrì and P. Stellari, Derived equivalences of K3 surfaces and orientation, Duke Math. J. 149 (2009), no. 3, 461–507. [CrossRef]
  57. D. Huybrechts and R. P. Thomas, Deformation-obstruction theory for complexes via Atiyah and Kodaira–Spencer classes, Math. Ann. 346 (2010), no. 3, 545–569. [CrossRef]
  58. D. Huybrechts, Lectures on K3 Surfaces, Cambridge Studies in Advanced Mathematics 158, Cambridge University Press, Cambridge, 2016. [CrossRef]
  59. D. Huybrechts, Motives of isogenous K3 surfaces, Comment. Math. Helv. 94 (2019), no. 3, 445–458. [CrossRef]
  60. L. Illusie, Complexe cotangent et déformations I, Lecture Notes in Math. 239, Springer, Berlin, 1971. [CrossRef]
  61. U. Jannsen, Motives, numerical equivalence, and semi-simplicity, Invent. Math. 107 (1992), no. 3, 447–452. [CrossRef]
  62. J. R. King, The currents defined by analytic varieties, Acta Math. 127 (1971), 185–220. [CrossRef]
  63. S. L. Kleiman, Algebraic cycles and the Weil conjectures, in: Dix exposés sur la cohomologie des schémas, Advanced Studies in Pure Mathematics 3, North-Holland, Amsterdam, 1968, pp. 359–386. Zbl 0198.25902.
  64. K. Koike, Algebraicity of some Weil Hodge classes, Canad. Math. Bull. 47 (2004), no. 4, 566–572. [CrossRef]
  65. J. Kollár, Trento examples, in: Classification of Irregular Varieties (Trento, 1990), Lecture Notes in Mathematics 1515, Springer, Berlin, 1992, 134–139. [CrossRef]
  66. M. Kuranishi, New proof for the existence of locally complete families of complex structures, in: Proc. Conf. Complex Analysis (Minneapolis, 1964), Springer, Berlin, 1965, pp. 142–154.
  67. N. Kurnosov, A. Soldatenkov and M. Verbitsky, Kuga–Satake construction and cohomology of hyperkähler manifolds, Adv. Math. 351 (2019), 275–295. [CrossRef]
  68. T. Y. Lam, Introduction to Quadratic Forms over Fields, Graduate Studies in Mathematics 67, Amer. Math. Soc., Providence, RI, 2005. [CrossRef]
  69. W. Landherr, Äquivalenz Hermitescher Formen über einem beliebigen algebraischen Zahlkörper, Abh. Math. Sem. Univ. Hamburg 11 (1935), no. 1, 245–248. [CrossRef]
  70. S. Lefschetz, L’analysis situs et la géométrie algébrique, Collection de monographies sur la théorie des fonctions, Gauthier-Villars, Paris, 1924. JFM 50.0663.01.
  71. M. Lieblich, Moduli of complexes on a proper morphism, J. Algebraic Geom. 15 (2006), no. 1, 175–206. [CrossRef]
  72. D. I. Lieberman, Numerical and homological equivalence of algebraic cycles on Hodge manifolds, Amer. J. Math. 90 (1968), no. 2, 366–374. [CrossRef]
  73. G. Lombardo, Abelian varieties of Weil type and Kuga–Satake varieties, Tohoku Math. J. (2) 53 (2001), no. 3, 453–466. [CrossRef]
  74. S. Ma, On K3 surfaces which dominate Kummer surfaces, Proc. Amer. Math. Soc. 141 (2013), no. 1, 131–137. [CrossRef]
  75. Ju. I. Manin, Correspondences, motifs and monoidal transformations, Math. USSR-Sb. 6 (1968), no. 4, 439–470. [CrossRef]
  76. E. Markman, The monodromy of generalized Kummer varieties and algebraic cycles on their intermediate Jacobians, J. Eur. Math. Soc. 25 (2023), no. 1, 231–321. [CrossRef]
  77. E. Markman, Rational Hodge isometries of hyper-Kähler varieties of K3[n] type are algebraic, Compos. Math. 160 (2024), no. 6, 1261–1303. [CrossRef]
  78. E. Markman, Cycles on abelian 2n-folds of Weil type from secant sheaves on abelian n-folds, preprint, arXiv:2502.03415 (2025). [CrossRef]
  79. E. Markman, Secant sheaves and Weil classes on abelian varieties, in: Proceedings of the International Congress of Mathematicians 2026, Vol. 3, 2026, pp. 586–602. [CrossRef]
  80. E. Markman, Secant sheaves on abelian n-folds with real multiplication and Weil classes on abelian 2n-folds with complex multiplication, preprint, 2025. arXiv:2509.23079.
  81. D. R. Morrison, On K3 surfaces with large Picard number, Invent. Math. 75 (1984), no. 1, 105–121. [CrossRef]
  82. A. Mostaed, McMullen’s curve, the Weil locus, and the Hodge conjecture for abelian sixfolds, preprint, 2026. arXiv:2603.20268.
  83. S. Mukai, Semi-homogeneous vector bundles on an abelian variety, J. Math. Kyoto Univ. 18 (1978), no. 2, 239–272. [CrossRef]
  84. S. Mukai, Duality between D(X) and D( X ^ ) with its application to Picard sheaves, Nagoya Math. J. 81 (1981), 153–175. [CrossRef]
  85. S. Mukai, On the moduli space of bundles on K3 surfaces, I, in: Vector bundles on algebraic varieties (Bombay, 1984), Tata Inst. Fund. Res. Stud. Math. 11, Tata Institute of Fundamental Research, Bombay, 1987.
  86. D. Mumford, On the equations defining abelian varieties. I, Invent. Math. 1 (1966), 287–354. [CrossRef]
  87. D. Mumford, A note of Shimura’s paper “Discontinuous groups and abelian varieties”, Math. Ann. 181 (1969), 345–351. [CrossRef]
  88. D. Mumford, Abelian varieties, Tata Institute of Fundamental Research Studies in Mathematics 5, Oxford University Press, London, 1970.
  89. V. K. Murty, Exceptional Hodge classes on certain abelian varieties, Math. Ann. 268 (1984), no. 2, 197–206. [CrossRef]
  90. B. Moonen and Yu. G. Zarhin, Hodge classes and Tate classes on simple abelian fourfolds, Duke Math. J. 77 (1995), no. 3, 553–581. [CrossRef]
  91. B. Moonen and Yu. G. Zarhin, Hodge classes on abelian varieties of low dimension, Math. Ann. 315 (1999), no. 4, 711–733. [CrossRef]
  92. J. Neukirch, Algebraic Number Theory, Grundlehren der mathematischen Wissenschaften 322, Springer, Berlin, 1999. [CrossRef]
  93. V. V. Nikulin, Integral symmetric bilinear forms and some of their applications, Math. USSR Izv. 14 (1980), no. 1, 103–167. [CrossRef]
  94. V. V. Nikulin, Finite automorphism groups of Kähler K3 surfaces, Trans. Moscow Math. Soc. 1980, no. 2, 71–135.
  95. O. T. O’Meara, Introduction to quadratic forms, Classics in Mathematics, Springer, Berlin, 2000, reprint of the 1973 edition (Grundlehren der mathematischen Wissenschaften 117, 1963). ISBN 3-540-66564-1. Zbl 1034.11003.
  96. D. O. Orlov, Equivalences of derived categories and K3 surfaces, J. Math. Sci. (New York) 84 (1997), no. 5, 1361–1381. [CrossRef]
  97. D. O. Orlov, Derived categories of coherent sheaves on abelian varieties and equivalences between them, Izv. Math. 66 (2002), no. 3, 569–594. [CrossRef]
  98. A. Perry, The semiregularity theorem for equivariant noncommutative varieties, preprint, 2026. arXiv:2604.00511.
  99. A. Polishchuk, Lefschetz type formulas for dg-categories, Selecta Math. (N.S.) 20 (2014), no. 3, 885–928. [CrossRef]
  100. H. Pohlmann, Algebraic cycles on abelian varieties of complex multiplication type, Ann. of Math. (2) 88 (1968), 161–180. [CrossRef]
  101. V. P. Platonov and A. S. Rapinchuk, Algebraic groups and number theory, Pure and Applied Mathematics 139, Academic Press, Boston, 1994, translated from the Russian by R. Rowen. ISBN 0-12-558180-7. Zbl 0841.20046.
  102. J. P. Pridham, Semiregularity as a consequence of Goodwillie’s theorem, Forum Math. Sigma 12 (2024), e126. arXiv:1208.3111. [CrossRef]
  103. I. Reiner, Maximal orders, London Mathematical Society Monographs. New Series 28, Oxford University Press, Oxford, 2003, corrected reprint of the 1975 edition. ISBN 0-19-852673-3. Zbl 1024.16008.
  104. J. J. Ramón Marí, On the Hodge conjecture for products of certain surfaces, Collect. Math. 59 (2008), no. 1, 1–26. [CrossRef]
  105. U. Schlickewei, The Hodge conjecture for self-products of certain K3 surfaces, J. Algebra 324 (2010), no. 3, 507–529. [CrossRef]
  106. W. Scharlau, Quadratic and Hermitian Forms, Grundlehren der mathematischen Wissenschaften 270, Springer, Berlin, 1985. [CrossRef]
  107. J.-P. Serre, Géométrie algébrique et géométrie analytique, Ann. Inst. Fourier 6 (1956), 1–42. [CrossRef]
  108. J.-P. Serre, A Course in Arithmetic, Graduate Texts in Mathematics 7, Springer, New York, 1973. [CrossRef]
  109. M. Schütt and T. Shioda, Elliptic surfaces, in: Algebraic Geometry in East Asia (Seoul, 2008), Adv. Stud. Pure Math. 60, Math. Soc. Japan, Tokyo, 2010, pp. 51–160. [CrossRef]
  110. G. Shimura, On analytic families of polarized abelian varieties and automorphic functions, Ann. of Math. (2) 78 (1963), 149–192. [CrossRef]
  111. T. Shioda and N. Mitani, Singular abelian surfaces and binary quadratic forms, in: Classification of algebraic varieties and compact complex manifolds, Lecture Notes in Math. 412, Springer, Berlin, 1974, 259–287. [CrossRef]
  112. H. Sterk, Finiteness results for algebraic K3 surfaces, Math. Z. 189 (1985), no. 4, 507–513. [CrossRef]
  113. M. Stoer and F. Wagner, A simple min-cut algorithm, J. ACM 44 (1997), no. 4, 585–591. [CrossRef]
  114. C. Soulé and C. Voisin, Torsion cohomology classes and algebraic cycles on complex projective manifolds, Adv. Math. 198 (2005), 107–127. [CrossRef]
  115. M. Varesco, Hodge similarities, algebraic classes, and Kuga–Satake varieties, Math. Z. 305 (2023), paper no. 69. [CrossRef]
  116. M. Varesco, The Hodge conjecture for powers of K3 surfaces of Picard number 16, Michigan Math. J. 75 (2025), no. 5, 1085–1116. [CrossRef]
  117. B. van Geemen, Kuga–Satake varieties and the Hodge conjecture, in: The Arithmetic and Geometry of Algebraic Cycles (Banff, 1998), NATO Sci. Ser. C Math. Phys. Sci. 548, Kluwer, Dordrecht, 2000, pp. 51–82.
  118. B. van Geemen, Real multiplication on K3 surfaces and Kuga–Satake varieties, Michigan Math. J. 56 (2008), no. 2, 375–399. [CrossRef]
  119. B. van Geemen and A. Rapagnetta, Hyperkähler sixfolds, abelian fourfolds of Weil type and a Hodge class, preprint, 2026. arXiv:2607.18341. arXiv:2607.18341.
  120. B. van Geemen and A. Sarti, Nikulin involutions on K3 surfaces, Math. Z. 255 (2007), no. 4, 731–753. [CrossRef]
  121. B. van Geemen and M. Schütt, On families of K3 surfaces with real multiplication, Forum Math. Sigma 13 (2025), Paper No. e2. [CrossRef]
  122. R. Villaflor Loyola, Periods of complete intersection algebraic cycles, Manuscripta Math. 167 (2022), 765–792. [CrossRef]
  123. C. Voisin, Hodge Theory and Complex Algebraic Geometry I, Cambridge Studies in Advanced Mathematics 76, Cambridge University Press, Cambridge, 2002. [CrossRef]
  124. C. Voisin, A counterexample to the Hodge conjecture extended to Kähler varieties, Int. Math. Res. Not. 2002, no. 20, 1057–1075. [CrossRef]
  125. A. Weil, Abelian varieties and the Hodge ring, in: Œuvres Scientifiques, Collected Papers, Vol. III (1964–1978), Springer, New York, 1979, pp. 421–429.
  126. Yu. G. Zarhin, Hodge groups of K3 surfaces, J. Reine Angew. Math. 341 (1983), 193–220. [CrossRef]
1
Mumford constructed [87] abelian fourfolds X whose Hodge group is a Q -form of SL 2 3 acting on H 1 ( X ) , of dimension eight, through the tensor product of the three two-dimensional representations; the form comes from a quaternion algebra over a totally real cubic field F that is split at one real place and ramified at the other two, and the corestriction to Q of that algebra is the matrix algebra, which is what makes the representation rational. Such an X is simple, has End 0 ( X ) = Q , is not of CM type, and its Hodge group is smaller than the full symplectic group although no endomorphism accounts for the loss; its Shimura curve is compact. The exceptional classes of [90] live in degree four on X × X (equivalently in End of H 1 as a Hodge structure of weight zero) and span the invariants of the Hodge group that are not generated by the polarisation, and no one has proved them algebraic. The fourfold enters this paper because it is the smallest variety on which (F2) has an unsettled instance that no CM point reaches directly.
2
For a smooth projective variety X of dimension d with hyperplane class h, (L) asks that the inverse of the hard Lefschetz isomorphism h d − 2 p : H 2 p ( X , Q ) → H 2 d − 2 p ( X , Q ) be induced by an algebraic cycle on X × X for every p; it is known for abelian varieties, surfaces and the varieties whose cohomology is spanned by algebraic classes, and open for a general threefold. A Hodge class is motivated in André’s sense when it is obtained from classes of algebraic cycles by cup products, pull-backs, push-forwards and the inverse Lefschetz operators, on X and on the products X × Y ; (M) asks that every Hodge class be of this kind, which is weaker than the conjecture by exactly the use of the inverse Lefschetz operator. (V) asks that a flat Hodge class in a smooth projective family be algebraic on every fibre once it is algebraic on one; Grothendieck observed that (V) follows from the conjecture, and the converse is open. (F3) concerns the varieties that are not abelian and makes no statement about abelian varieties at all; it is the complement, inside the conjecture, of (F2), and the pair (F2), (F3) is the coarsest division of the conjecture that this paper refines.
3
The theorem says that for a polarisable variation of Hodge structure V over a smooth connected quasi-projective base S, the space H 0 ( S , V ) of global flat sections, which is the subspace of a fibre V s fixed by the monodromy group, is a sub-Hodge structure of V s , and the Hodge structure it carries does not depend on s. Deligne proved it for the variations R k π * Q of a smooth projective morphism, where it follows from the degeneration of the Leray spectral sequence and the weight filtration of a smooth compactification; André proved it for every polarisable variation by showing that the connected monodromy group is a normal subgroup of the derived group of the generic Mumford–Tate group. In the situation of the text the section ξ is flat, so its value at one point is fixed by the monodromy; if that value has type ( p , p ) at one point, the sub-Hodge structure of flat sections contains it as a Hodge class, and since that Hodge structure is the same at every point, ξ s has type ( p , p ) everywhere.
4
A subset of a topological space is meagre when it is a countable union of nowhere dense sets, that is, of sets whose closure has empty interior. The Baire category theorem states that a complete metric space, or a locally compact Hausdorff space, is not meagre in itself; a connected complex manifold such as S an is locally compact, so the theorem applies to it. A proper Zariski closed subset of an irreducible variety has empty interior in the analytic topology, because a holomorphic function vanishing on an open set vanishes identically on a connected manifold; so a countable union of proper Zariski closed subsets is meagre, and the alternative in (ii) is exactly whether one of the countably many closed strata of (i) is the whole of S.
5
The curve X 0 ( N ) parametrises elliptic curves with a cyclic subgroup of order N; it has genus zero exactly for N ≤ 10 and N ∈ { 12 , 13 , 16 , 18 , 25 } , and for N = 5 and N = 7 a coordinate is the Hauptmodul t = η ( τ ) / η ( N τ ) k with k = 12 / ( N − 1 ) , so k = 3 for N = 5 and k = 2 for N = 7 ; this is why N k is 125 in one case and 49 in the other. The Fricke involution sends τ to − 1 / N τ , which exchanges the two cusps 0 and ∞ of X 0 ( N ) and, on the moduli interpretation, replaces ( E , C ) by the quotient E / C with the image of E [ N ] as its distinguished subgroup. Applying it twice gives E / E [ N ] ≅ E , so it is an involution, and in the coordinate t it reads t ↦ N k / t because η ( − 1 / τ ) = − i τ η ( τ ) turns the quotient of eta functions into its reciprocal up to the constant N k / 2 , squared to N k by the exponent.
6
Families of K3 surfaces polarised by a fixed lattice also occur as the fibres of Calabi–Yau threefolds, the quartic and quintic ones being studied in [13]; the chains of this section act on the fibres only, and the Hodge classes of such a total space are not considered here.
7
The Dickson polynomial p n , a is the unique polynomial of degree n with p n , a ( v + a / v ) = v n + ( a / v ) n ; for a ≠ 0 the map t ↦ p n , a ( t ) from P 1 to itself is a cover of degree n with monodromy group the dihedral group D n , totally ramified over ∞ [121, Lem. 4.3], since it is the quotient of v ↦ v n + ( a / v ) n , on which D n acts by v ↦ ζ n v and v ↦ a / v , by the involution. For a = 0 it is the cyclic cover t ↦ t n . The identity p n , a ( μ t ) = μ n p n , a / μ 2 ( t ) follows from the defining equation with v replaced by μ v .
8
Explicitly p 7 , a ( t ) = t 7 − 7 a t 5 + 14 a 2 t 3 − 7 a 3 t , the case n = 7 of p n , a ( t ) = ∑ k ≤ n / 2 n n − k n − k k ( − a ) k t n − 2 k ; for a = 0 it is t 7 , so the family X a contains the surfaces X b .
9
Reaching F means producing, by algebraic correspondences from S λ to itself, Hodge endomorphisms of T ( S λ ) Q that generate the cubic field E = F , which is what (F3′) asks of this surface, since the class of a generator of F in H 4 ( S λ × S λ ) is a Hodge class reached from no abelian variety as long as it is not algebraic. The correspondences built in this section, and every correspondence whose components have geometric genus at most one, act on T ( S λ ) as similarities, and a composite of similarities is a similarity; by Proposition 9(ii) a similarity lies in Q when [ E : Q ] is odd, so it generates nothing. The correspondences coming from finite group actions act through abelian fields by Proposition 10(iii), and a cubic field whose Galois closure has group S 3 is not abelian. The two propositions together say what a correspondence realising F must look like: an irreducible component with p g ≥ 2 and transcendental rank at least eighteen, not arising from a finite group; no construction in the literature produces one.
10
The Hankel matrix of a sequence ( a k ) k ≥ 0 has entries a i + j . If a k is a sum ∑ j λ j b j k of m terms with distinct b j and nonzero λ j , the matrix factors as V T diag ( λ j ) V with V the Vandermonde matrix of the b j , so its rank is m; conversely a sequence whose Hankel matrix has finite rank m satisfies a linear recurrence of order m and is a sum of m exponentials with polynomial coefficients. For the polynomial part ∑ k c k θ k of a Chern character, writing it as ∑ j λ j e b j θ means k ! c k = ∑ j λ j b j k , so the number of exponentials needed is the Hankel rank of ( k ! c k ) , provided the degree 2 n allows the recurrence to be read off, which it does for m ≤ 2 ; the cap at three records that the criterion of Theorem 20 does not distinguish further.
11
For a perfect complex E on a smooth projective variety the semiregularity map of Buchweitz and Flenner [11] is the linear map σ E : Ext 2 ( E , E ) → ∏ q H q + 1 ( X , Ω X q − 1 ) sending ξ to the trace of ξ ∘ exp − At ( E ) , where At ( E ) ∈ Ext 1 ( E , E ⊗ Ω X 1 ) is the Atiyah class; it generalises Bloch’s map for a subvariety [20], which is the case of the structure sheaf of the subvariety. Its value is that the obstruction to deforming E along a first-order deformation of X lies in Ext 2 ( E , E ) and is sent by σ E to the obstruction to keeping ch ( E ) of Hodge type; so when σ E is injective, E deforms wherever its Chern character stays a Hodge class, and the Chern character of the deformed complex is algebraic there. The dimension count of the theorem is the arithmetic of this statement: the rank of σ E is at least the rank of the contraction into ch ( E ) , so dim Ext 2 is at least r ( γ ) , and equality forces injectivity.
12
The direction matters: if H ⊆ H ′ then every H ′ -invariant is H-invariant, so containment of the Hodge group in a known group bounds the Hodge classes from below, never from above.
13
Every object obtained from P 1 , ⋯ , P m by iterated cones in this order is of this form, because the Dolbeault complexes form a differential graded enhancement of the full subcategory of line bundles. All the objects of this subsection are given in this way.
14
In the language of A ∞ -categories they are the higher products m k of the minimal model of the line bundles, evaluated on the Maurer–Cartan element δ H of the twisted complex; for k = 2 they are cup products.
15
Such solutions are far from unique: the equations of (i) are linear in each of x 1 , x 2 ζ , x 3 ζ separately, and the components on the 144 pairs do not enter them at all.
16
The ratios ζ j ′ / ζ j of two pieces of opposite parities have product − 1 ; when all three differ from 1 they are ( − 1 , − 1 , − 1 ) , with D = 4 3 , or a permutation of ( i , − i , − 1 ) , with D = 2 · 2 · 4 .
17
This is Mumford’s index theorem: a line bundle on an abelian variety whose hermitian form is nondegenerate with i negative eigenvalues has cohomology only in degree i. On B j = E 0 × E 0 the form of M i ⊗ L ζ − 1 is t i − p − ζ j − ζ ¯ j t i − p , with eigenvalues t i − p ± 1 , and | t i − p | ≥ 2 makes both of the sign of t i − p ; over the four surfaces the index is 0 or 8. The hypothesis | t i − p | ≥ 2 is what keeps the steps between a piece and a multiple out of the middle degrees.
18
The bound is attained: with p below the three values t i , the run L ζ → M 1 → M 2 → M 3 → L ζ ′ , up three times and down once, changes the shift by 3 − 7 = − 4 . With p in the middle, t 1 < p < t 2 < t 3 , the run up through M 2 and M 3 , down to M 1 and up to L ζ ′ changes it by 3 − 7 = − 4 as well.
19
The ratios ζ j ′ / ζ j in the three coordinates have product − 1 , so they are ( − 1 , − 1 , − 1 ) , with D = 64 , or a permutation of ( i , − i , − 1 ) , with D = 16 ; with four places for the coordinate where they agree this gives 4 · 7 = 28 partners and 4 · ( 64 + 6 · 16 ) = 640 classes.
20
The ratios e j / c j ≠ 1 have product 1. In two coordinates they are ( − 1 , − 1 ) , with D = 16 , or ( i , − i ) , ( − i , i ) , with D = 4 ; then H 5 = H 1 ⊗ H 1 ⊗ H 3 ( B j × B k , O ) has dimension 4 D , and the six pairs of coordinates give 6 · 64 + 12 · 16 = 576 . In three coordinates they are a permutation of ( i , i , − 1 ) or of ( − i , − i , − 1 ) , with D = 16 , and H 5 = ( H 1 ) ⊗ 3 ⊗ H 2 ( B l , O ) has dimension 16; the four triples give 4 · 6 · 16 = 384 . In four coordinates H ( c , e ) sits in degree 4.
21
The eight values are 960, 384, 192, 96, − 32 , − 64 , − 128 and − 192 . The bound 18432 is attained by the cut ζ 3 ζ 4 ∈ { 1 , i } against ζ 3 ζ 4 ∈ { − 1 , − i } , so 22528 is the best bound that this count gives.
22
A ratio that moves every coordinate has entries in { i , − 1 , − i } and product 1. With a − 1 among them it is ( − 1 , − 1 , − 1 , − 1 ) or has exactly two entries − 1 and the entries i and − i , twelve ratios; together thirteen, and 256 + 12 · 64 = 1024 . The other eight ratios have all four entries in { i , − i } and D = 16 ; (ii) makes no statement about them.
23
The orders of the subgroups of G 0 ≅ ( Z / 4 ) 3 are powers of two. A cut at one vertex weighs exactly 1024, so the bound is sharp.
24
Let the ratio move k coordinates. For k = 1 it is − 1 in one coordinate, D = 4 , and H 5 is H 1 tensored with the part of degree four of H * ( O ) on the other three surfaces, of dimension 6 4 = 15 : four partners with 60 each. For k = 2 it is ( i , i ) or ( − i , − i ) , D = 4 , and the part of degree three on the other two surfaces has dimension 4 3 = 4 : twelve partners with 16 each. For k = 3 the part of degree two on the last surface has dimension one, and the 28 partners give 640 as in Theorem A13. For k = 4 the group sits in degree 4. So 240 + 192 + 640 = 1072 .
25
Between distinct pieces H 8 = 0 . A run from a piece through m ≤ 3 distinct multiples to a piece has m + 1 steps, at least one up and one down, and lowers the shift by 7 D ′ − U ′ with U ′ + D ′ = m + 1 : by 6, 5, 4 for D ′ = 1 , by 13, 12 for D ′ = 2 and by 20 for D ′ = 3 , never by seven.
26
With the values t 1 < t 2 < t 3 and p, the path with U = 3 visits the four nodes in the order of their values; a path with U = 2 and D = 1 ending in degree eight visits the largest j values in increasing order and then the others, j = 1 , 2 , 3 . So each placement has exactly four such paths, and for each the shifts of the multiples are determined by d a .
27
The proof does not look at the classes between the two parities. For g = 5 these are 64 · 16 = 1024 classes of degree two in groups H 7 , on which d E vanishes by Lemma A12(i); but the coboundaries of the loops in H 1 ( a , a ) , and, when two multiples have suitable shifts, the products along runs W → M i → M j → V , reach them, and counting alone gives nothing there.
28
The subgroups of Γ ≅ ( Z / 4 ) 3 × Z / 2 have orders that are powers of two. A cut at a single vertex weighs exactly 640, so the bound is sharp.
29
They differ by a homomorphism from the anisotropic torus S / G m to the centre of G · G m , whose identity component is G m , and such a homomorphism is trivial.
30
The fundamental domain of Z [ i ] is the unit square, so ∫ E 0 d x ∧ d y = 1 ; this normalisation makes every form in (A7) integral on the lattice Z [ i ] 6 , and all intersection numbers below are integers computed as integrals of top forms.
Figure 1. The plan of the paper. The conjecture is equivalent to (F2) and (F3′) (Theorem 1, Proposition 4); the locus of either statement is dense, contains every CM point and is all or meagre (Section 4). For (F3′) the first class is the real multiplication of K3 surfaces: the chain of quotients for 6 (Theorem 11), the Fricke isogenies (Theorem 13), the cyclic cubic fields on the surfaces of Mumford type (Theorem 14), every real quadratic field in rank six (Theorem 16), Floccari’s theorem at Picard number at least sixteen and on dense curves (Corollary 3, Theorems 17 and 18), and the steps at which each stops (Propositions 9 and 11). For (F2) the first class is the Weil classes of a product of two fourfolds: the product object (Theorem 21), the point of complex multiplication (Lemma 15 to Remark 9), Weil sixfolds with a split polarisation (Corollary 15, Theorem 29) and the sixfold link (Theorems 37, 38 and 41); the defect survives the operations of Proposition 47.
Figure 1. The plan of the paper. The conjecture is equivalent to (F2) and (F3′) (Theorem 1, Proposition 4); the locus of either statement is dense, contains every CM point and is all or meagre (Section 4). For (F3′) the first class is the real multiplication of K3 surfaces: the chain of quotients for 6 (Theorem 11), the Fricke isogenies (Theorem 13), the cyclic cubic fields on the surfaces of Mumford type (Theorem 14), every real quadratic field in rank six (Theorem 16), Floccari’s theorem at Picard number at least sixteen and on dense curves (Corollary 3, Theorems 17 and 18), and the steps at which each stops (Propositions 9 and 11). For (F2) the first class is the Weil classes of a product of two fourfolds: the product object (Theorem 21), the point of complex multiplication (Lemma 15 to Remark 9), Weil sixfolds with a split polarisation (Corollary 15, Theorem 29) and the sixfold link (Theorems 37, 38 and 41); the defect survives the operations of Proposition 47.
Preprints 244139 g001
Figure 4. The chain of Theorem 11: the quotient of S ′ by a symplectic automorphism of order three, a Hodge isometry (algebraic by Buskin), and the quotient of Y ′ by a Nikulin involution. The multipliers compose to 6.
Figure 4. The chain of Theorem 11: the quotient of S ′ by a symplectic automorphism of order three, a Hodge isometry (algebraic by Buskin), and the quotient of Y ′ by a Nikulin involution. The multipliers compose to 6.
Preprints 244139 g004
Figure 7. Where the dihedral mechanism stops, Proposition 11. (i) Over F = Q ( ζ 7 ) + or Q ( ζ 9 ) + the Mumford algebras are D ∞ and the algebras D p ramified at two real places and at two places above a split prime p; a rank-twelve family unimodular away from p n has square discriminant and carries only D ∞ . (ii) The period domain of the rank-eighteen family is four-dimensional, the image of the family is a threefold, and the Shimura curves of the surfaces of Mumford type for the other algebras lie in the domain; whether the threefold meets them is the open transport condition. (iii) The finite subgroups of PGL 2 acting on the base produce only Q ( ζ m ) , Q ( ζ m ) + , Q , Q ( − 3 ) and Q ( 5 ) ; a cubic field arises only from the dihedral groups D 7 , D 9 , D 14 and D 18 .
Figure 7. Where the dihedral mechanism stops, Proposition 11. (i) Over F = Q ( ζ 7 ) + or Q ( ζ 9 ) + the Mumford algebras are D ∞ and the algebras D p ramified at two real places and at two places above a split prime p; a rank-twelve family unimodular away from p n has square discriminant and carries only D ∞ . (ii) The period domain of the rank-eighteen family is four-dimensional, the image of the family is a threefold, and the Shimura curves of the surfaces of Mumford type for the other algebras lie in the domain; whether the threefold meets them is the open transport condition. (iii) The finite subgroups of PGL 2 acting on the base produce only Q ( ζ m ) , Q ( ζ m ) + , Q , Q ( − 3 ) and Q ( 5 ) ; a cubic field arises only from the dihedral groups D 7 , D 9 , D 14 and D 18 .
Preprints 244139 g007
Figure 8. Projective K3 surfaces by Picard number ρ , the height of a bar being the dimension 20 − ρ of the family. Dark: the Kuga–Satake correspondence is algebraic by Floccari’s theorem and the Hodge conjecture holds on every power and Hilbert scheme of points; this is every member for ρ ≥ 17 (Corollary 3) and, for ρ = 16 , exactly the surfaces whose transcendental space has Witt index two (Corollary 4). Light: Floccari’s theorem does not apply, since T ( S ) Q embeds in no V m (Lemma 11).
Figure 8. Projective K3 surfaces by Picard number ρ , the height of a bar being the dimension 20 − ρ of the family. Dark: the Kuga–Satake correspondence is algebraic by Floccari’s theorem and the Hodge conjecture holds on every power and Hilbert scheme of points; this is every member for ρ ≥ 17 (Corollary 3) and, for ρ = 16 , exactly the surfaces whose transcendental space has Witt index two (Corollary 4). Light: Floccari’s theorem does not apply, since T ( S ) Q embeds in no V m (Lemma 11).
Preprints 244139 g008
Figure 9. The period domain D W of a family of K3 surfaces with real multiplication by a real quadratic field E, of dimension m − 2 , and the curves D W ′ of Theorem 18, one for each three-dimensional E-subspace W ′ ⊂ W of the shape of Corollary 5. On every curve the real multiplication is algebraic and the Hodge conjecture holds on all powers and Hilbert schemes of points; the union of the curves is dense and meagre, and the very general member is open.
Figure 9. The period domain D W of a family of K3 surfaces with real multiplication by a real quadratic field E, of dimension m − 2 , and the curves D W ′ of Theorem 18, one for each three-dimensional E-subspace W ′ ⊂ W of the shape of Corollary 5. On every curve the real multiplication is algebraic and the Hodge conjecture holds on all powers and Hilbert schemes of points; the union of the curves is dense and meagre, and the very general member is open.
Preprints 244139 g009
Figure 10. The lower bound r ( γ ) = ( 4 + ρ ) n 2 − 2 n of Theorem 20 for 3 ≤ n ≤ 7 and Hankel rank ρ = 1 , 2 , 3 . The two labelled bars are the product cases of Theorem 21.
Figure 10. The lower bound r ( γ ) = ( 4 + ρ ) n 2 − 2 n of Theorem 20 for 3 ≤ n ≤ 7 and Hankel rank ρ = 1 , 2 , 3 . The two labelled bars are the product cases of Theorem 21.
Preprints 244139 g010
Figure 11. The domain D n , n near a product point A × A ′ . The curve is the product locus, whose tangent space has dimension m 2 + m ′ 2 ; the 2 m m ′ mixed directions obstruct every exterior product F ⊠ G ′ , and only the combined object of Theorem 21 has the full tangent space as first-order Hodge locus.
Figure 11. The domain D n , n near a product point A × A ′ . The curve is the product locus, whose tangent space has dimension m 2 + m ′ 2 ; the 2 m m ′ mixed directions obstruct every exterior product F ⊠ G ′ , and only the combined object of Theorem 21 has the full tangent space as first-order Hodge locus.
Preprints 244139 g011
Figure 12. Left: the period domain D K , δ of abelian sixfolds of Weil type for K and discriminant δ ; the Weil classes are algebraic at every point of complex multiplication (stars, Corollary 15) and on the six-dimensional subdomains D v with a K-line as sub-Hodge structure (curves, Proposition 34(c)). Right, schematically: at a point with complex multiplication by K L the K-compatible polarisations are the forms H a , a ∈ L × with the signs of the CM type; Lemma 28 finds one with det H a = − 1 (solid), to which Markman’s theorem applies.
Figure 12. Left: the period domain D K , δ of abelian sixfolds of Weil type for K and discriminant δ ; the Weil classes are algebraic at every point of complex multiplication (stars, Corollary 15) and on the six-dimensional subdomains D v with a K-line as sub-Hodge structure (curves, Proposition 34(c)). Right, schematically: at a point with complex multiplication by K L the K-compatible polarisations are the forms H a , a ∈ L × with the signs of the CM type; Lemma 28 finds one with det H a = − 1 (solid), to which Markman’s theorem applies.
Preprints 244139 g012
Figure 14. Where the defect goes under surgery. A blow-up along Z carries the Hodge classes of X to X ˜ and adds those of Z, so the quotient Hdg / Alg grows by the defect of the centre; a dominating map Y → X carries the defect of X into Y; a Fourier–Mukai equivalence permutes degrees and preserves the total defect; a degeneration to X 0 loses the class only in the special fibre, from which no cycle lifts.
Figure 14. Where the defect goes under surgery. A blow-up along Z carries the Hodge classes of X to X ˜ and adds those of Z, so the quotient Hdg / Alg grows by the defect of the centre; a dominating map Y → X carries the defect of X into Y; a Fourier–Mukai equivalence permutes degrees and preserves the total defect; a degeneration to X 0 loses the class only in the special fibre, from which no cycle lifts.
Preprints 244139 g014
Table 1. The main results. “Corollary” marks a statement whose decisive input is a published theorem, named in the text.
Table 1. The main results. “Corollary” marks a statement whose decisive input is a published theorem, named in the text.
Statement Where Status
HC ⇔ ( F 2 ) + ( F 3 ’), ( F 2 ) ⇔ AB Thm. 1 proved
algebraic locus is all or meagre; dense for Weil classes Thm. 2 proved
real multiplication by 6 through chains of quotients Thm. 3 proved
Hodge conjecture on all powers of the Fricke families, N = 5 , 7 Thm. 13 proved
real multiplication by Q ( ζ 7 ) + , Q ( ζ 9 ) + Thm. 14 proved
every Q ( d ) in rank six, all powers Thm. 16 corollary [42]
K3 surfaces with ρ ≥ 17 : HC on all powers and Hilbert schemes Cor. 3 corollary [43]
which T ( S ) Q Floccari’s theorem reaches; Witt index in rank six Lem. 11, Thm. 17 proved
dense curves in every family with real multiplication of degree two Thm. 18 proved, via [43]
Weil classes of sixfolds with complex multiplication Cor. 15 corollary [78]
HC for sixfolds with complex multiplication by K L Thm. 29 proved; enumeration
real multiplication of Z K algebraic; the link is all or nothing Thms. 37, 38 proved
link classes on the dense set Σ K Thm. 41 corollary [76]
E 0 6 : at most seven line bundles; eight outside two arrangements Thm. 23, Rem. 9 proved; enumeration
E 0 6 , E 0 8 : no line-bundle convolution meets the criterion, with any group Thm. A9 proved
eightfolds from fourfolds: Markman’s Lemma 11.3 fails; indefinite secants miss the polarised locus Thms. 30, 35 proved
Weil classes on dense quaternionic loci of non-split sixfolds; Markman’s forms split Thms. 28, 27 proved
non-split sixfolds: cyclic covers reach them only on loci of dimension ≤ 7 ; no transport from hyper-Kähler or Kuga–Satake varieties Prop. 40, Thm. 36 proved
Hodge locus of degree-two real multiplication on S 1 × S 2 is the Hecke graph Thm. 19 proved, via [26]
Mumford link: one γ t suffices; divisors, Weil classes, homomorphisms do not Thm. 44, Prop. 45 proved
HC , (F2) and (F3′) Sect. 7 open
Table 2. The rule set. A rule with two premises needs both. The closure of a set of statements under these rules, and the list of minimal sets whose closure contains HC , are computed by hand in the text and checked in Lean (Appendix D).
Table 2. The rule set. A rule with two premises needs both. The closure of a set of statements under these rules, and the list of minimal sets whose closure contains HC , are computed by hand in the text and checked in Lean (Appendix D).
premises conclusion proof
HC (F3), (F2), (F3′), AB , (M) definitions; [6]
(F3) HC Prop. A19
(F2) AB Theorem 8
AB (F2) definition
AB , (F3′) HC Proposition 1
(L), (M) HC  [6]; Thm. A22
(L) AB Thm. A22
(V) AB Thm. A22
Table 3. Rational invariants used in Theorem 11(iii). Hasse invariants are + 1 at every prime not shown.
Table 3. Rational invariants used in Theorem 11(iii). Hasse invariants are + 1 at every prime not shown.
space rank signature d ε 2 ε 3 ε 5
T = Tr E / Q q E 8 ( 2 , 6 ) 10 + 1 − 1 + 1
〈 1 , − 10 〉 2 ( 1 , 1 ) − 10 + 1 + 1 + 1
T ⊕ 〈 1 , − 10 〉 10 ( 3 , 7 ) − 1 + 1 − 1 + 1
V 3 10 ( 3 , 7 ) − 1 + 1 − 1 + 1
V 2 14 ( 3 , 11 ) − 1 − 1 + 1 + 1
Table 6. The isolated-block bound of Lemma 17 over the configurations at A 0 = E 0 6 . “Polarised” counts the configurations of the family in the window whose Chern character has the polarised shape; “least” is the smallest bound among them, with the configuration attaining it. The criterion needs dim Ext 2 = 57 .
Table 6. The isolated-block bound of Lemma 17 over the configurations at A 0 = E 0 6 . “Polarised” counts the configurations of the family in the window whose Chern character has the polarised shape; “least” is the smallest bound among them, with the configuration attaining it. The criterion needs dim Ext 2 = 57 .
family (least bound attained at) polarised least
thirty-five pieces, L 1 , L 2 , L 3 at shifts 0 , 0 , 0 1 7213
the same at shifts 2 , 3 , 4 or − 3 , − 2 , − 1 2 2605
(a) four line bundles ( L 1 [ − 2 ] , L 2 [ − 1 ] , L 3 [ 0 ] , L 4 [ 1 ] ) 2401 0
(b) one threefold at a second shift 0
(c) one threefold at three twists ( Z ( 1 , 1 , 1 ) ⊗ L 1 at [ − 1 ] , [ 0 ] and ⊗ L 2 at [ 1 ] , [ 2 ] ) 432 0
(d) a second threefold meeting Z ζ along a surface, at the shift of Z ζ and a neighbouring one ( Z ( 1 , 1 , − 1 ) at [ 0 ] , [ 1 ] ) 6 5325
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.