Characteristic Cycles and the Microlocal Geometry of the Gauss Map I, Arxiv:1604.02389

arXiv:1807.01929v2 [math.AG] 4 Nov 2019 fteadvsr,adacieinfrtesmlct fteaiigLealg Lie arising the of simplicity Pop the and for Pareschi criterion of a conjecture and a divisors, theta of five, so weak of genus half A in second applications: problem sample the some Schottky In with methods new classes. these illustrate Chern-Mather with computations by lost nld hrceitccce ihabtaymlilcte n r and multiplicities group arbitrary Weyl with the on cycles th not of characteristic natural rings focus include representation we with to t that working allows Consequently is on orbits. feature cycles its new on Lagrangian main conic The with bundle. weights theoretic representation bla aite a mre nymr eety nti ae efu we paper this In holonomic recently. to more approach many bee only the from has emerged others this has and varieties tori Dettweiler affine abelian Sabbah, convolut on Katz, while a Loeser, description; with Tannakian Gabber, endowed a are to they leads varieties that semiabelian On geometry. yl,Gusmp ht divisor. theta map, Gauss cycle, 2010 .Acieinframs simplicity almost for criterion References A subvarieties of 6. Summands problem Schottky 5. The classes 4. Chern-Mather 3. Holonomic .Apiainto cycles Application clean of ring 2. The 1. Introduction e od n phrases. and words Key HRCEITCCCE N H MICROLOCAL THE AND CYCLES CHARACTERISTIC ahmtc ujc Classification. Subject Mathematics λ aite n rtro o h ipiiyo h rsn L arising the of subvari simplicity of the summands for of criterion existence a the and varieties for Schottk the obstruction to an solution cy weak five, Lagrangian a conic include a Applications by realized bundle. is cover universal their of holonomic Abstract. rn tutr ntegopo uhcce,wihcnotnb co be often can which cycles, such of group the on structure -ring D mdlsadpres hae lya nraigrl nalgebraic in role increasing an play sheaves perverse and -modules EMTYO H AS A,II MAP, GAUSS THE OF GEOMETRY D eso htfrterdcieTnaagop fsemisimple of groups Tannaka reductive the for that show We mdlso bla aite,eeyWy ru ri fwei of orbit group Weyl every varieties, abelian on -modules D bla ait,cnouin esrcategory, tensor convolution, variety, Abelian -modules D mdlso ope bla aite n[6,relating [26], in varieties abelian complex on -modules HMSKR THOMAS Introduction Contents rmr 41;Scnay1F0 81,20G05. 18D10, 14F10, Secondary 14K12; Primary AMER ¨ ealgebras. ie l ntecotangent the on cle rbe ngenus in problem y te nabelian on eties D mdl,characteristic -module, ie,tecs of case the sides, nsummands on a rsn groups arising e h ae we paper the te develop rther ecotangent he tde by studied n uint the to lution ghts o product ion ebras. tefbut itself vasa eveals ntrolled 37 29 24 21 17 12 7 1 2 THOMAS KRAMER¨

Let us now discuss the contents of the paper in some more detail. Let A be a complex abelian variety of dimension g. As a motivating problem for what follows, suppose that we want to know whether a given subvariety Z ⊂ A decomposes nontrivially as a sum Z = X + Y where X, Y ⊂ A are irreducible subvarieties of positive dimension such that the addition morphism X × Y ։ Z is generically ﬁnite. Famous examples are theta divisors1 on the Jacobian of a smooth projective curve or on the intermediate Jacobian of a smooth cubic threefold. A conjecture of Pareschi and Popa [34, p. 222], in a reformulation by Schreieder [39, conj. 19] who in loc. cit. proved it for curve summands, says that the existence of such summands characterizes the locus of Jacobians in the moduli space Ag of principally polarized abelian varieties (ppav’s):

Conjecture 1. Let (A, Θ) ∈ Ag(C) be an indecomposable ppav. If Θ = X + Y decomposes nontrivially as a sum, then (1) (A, Θ) is the Jacobian of a smooth projective curve, or (2) (A, Θ) is the intermediate Jacobian of a smooth cubic threefold, with g =5. In both cases the only decompositions of the theta divisor are the obvious ones, writing it as a sum of copies of the curve or a diﬀerence of two copies of the Fano surface of lines on the threefold. Somewhat surprisingly, without a priori assumptions on the dimension or fundamental classes of the summands as in the work of Casalaina-Martin, Popa and Schreieder [7, 39, 40], the only proof we know uses the representation theory of reductive groups and the Tannakian methods developed below, see [27]. The idea is to reformulate the existence of a nontrivial decomposition in a sum of subvarieties as a statement in the abelian category Hol(DA) of holonomic DA-modules as follows. Recall that for M ∈ Hol(DA) its de Rham complex

an 1 an 2 an DR(M ) = M → ΩA ⊗OA M → ΩA ⊗OA M → · · · is a perverse sheaf, andh by the Riemann-Hilbert correspondence any perversei sheaf arises as the de Rham complex of a unique regular holonomic module. For a closed subvariety Z ⊂ A we denote by δZ ∈ Hol(DA) the unique regular holonomic module such that

DR(δZ ) ≃ ICZ [dim(Z)] is isomorphic to the perverse intersection complex with support Z. If Z = X + Y and the sum map a : X × Y → Z is generically ﬁnite, then by the decomposition theorem of Beilinson, Bernstein, Deligne and Gabber [3, 9] we get an embedding as a direct summand

( δZ ֒→ δX ∗ δY = a†(δX ⊠ δY where ∗ is the convolution product from [26]. The Tannakian formalism allows to control such summands via representations of certain reductive groups. Let us brieﬂy recall the setup: We work in the quotient M(A) = Hol(DA)/S(A) by the Serre subcategory S(A) ⊂ Hol(DA) of negligible modules, where a module is said to be negligible if each of its simple subquotients is stable under translations by a

1Note that summands of indecomposable theta divisors are nondegenerate, so by [40, th. 1] the generic finiteness of the addition morphism X × Y ։ Z is automatic in this case. 3 non-zero abelian subvariety of A. On the quotient the convolution product descends to a product ∗ : M(A) × M(A) −→ M(A) with the usual associativity, commutativity and duality properties making M(A) a rigid abelian tensor category. Here by a tensor category we mean a symmetric monoidal Q-linear pseudoabelian category. The abelian tensor category from above is Tannakian: For any given M ∈ M(A), let us denote by hM i ⊂ M(A) the smallest rigid abelian tensor subcategory containing it, then we have an equivalence of abelian tensor categories ∼ ω : hM i −→ Rep(G) with the tensor category of finite-dimensional linear representations of an affine algebraic group G = G(M ), see section 2.1. In particular, to any subvariety Z ⊂ A this attaches

• a reductive group G = G(δZ ), • a representation ω(δZ ) ∈ Rep(G).

If Z = X +Y decomposes as a sum, then the above embedding δZ ֒→ δX ∗δY yields an embedding

.(( ω(δZ ) ֒→ ω(δX ) ⊗ ω(δY ) in Rep(G(δX ⊕ δY This is a very strong condition on the three representations. One way to exploit it is to look at the adjoint representation. Recall that any object of the quotient category M(A) has a unique representative M ∈ Hol(DA) which is clean in the sense that it has no negligible sub- or quotient objects. Let AdZ ∈ Hol(DA) be the unique clean module such that ω(AdZ ) is the adjoint representation of the group G(δZ ) on its Lie algebra. Then the dimensions dX = dim X and dY = dim Y of the summands are bounded as follows: Theorem 2 (=5.4.1). Let Z ⊂ A be an irreducible subvariety which is not stable under any translation on the abelian variety. Then for any nontrivial decomposition as a sum of geometrically nondegenerate subvarieties Z = X +Y with dX +dY = dZ one has

1 . min{dX , dY } ≥ δ = 2 min dim Supp(M ) | M ֒→ AdZ , dim Supp(M ) > 0

In particular, there can be no such decomposition if δ > ⌊dZ /2⌋. To apply this one only needs to compute a single convolution product: We always ∨ have AdZ ֒→ δZ ∗ δ−Z since Lie(G) ֒→ V ⊗ V for any faithful V ∈ Rep(G). For theta divisors on ppav’s the resulting bound is sharp on Jacobians of curves or intermediate Jacobians of smooth cubic threefolds, in all other known cases it rules out the existence of summands. Conjecture 1 would follow from

Conjecture 3. If (A, Θ) ∈ Ag(C) is an indecomposable ppav which is not the Jacobian of a curve or the intermediate Jacobian of a cubic threefold, then any .summand M ֒→ AdΘ is either a skyscraper sheaf or has support A Let us take a look at a few examples. After a translation we may assume Θ ⊂ A is symmetric. Then ω(δΘ) is a symplectic or orthogonal representation depending on whether g is even or odd [30, lemma 2.1]. Very often G(δΘ) is the full symplectic 4 THOMAS KRAMER¨ or orthogonal group. The reason will become clear later, here we only give a simple example: Let 2n n S− = n | n∈ / 2Z ∪ 2 | n ≡ 1, 2mod4 ∪ 56 2n n S+ = n | n ∈ 2Z ∪ 2 | n ≡ 0, 3mod4 ∪ 7 be the dimensions of symplectic minuscule resp. orthogonal weight multiplicity free representations of the simple complex Lie algebras other than the standard representations, see tables 2 and 3 in the appendix. A special case of theorem 6.3.2 then gives

Theorem 4. Let (A, Θ) ∈ Ag(C) be a ppav whose theta divisor Θ = −Θ ⊂ A is smooth except for ﬁnitely many ordinary double points e1,...,ek.

(1) If g is even with g! − 2k∈ / S− then G(δΘ)= Spg!−2k(C). (2) If g is odd with g! − k∈ / S+ and no two double points diﬀer by a torsion point, then

SOg!−k(C) for e1 + ··· + ek = 0, G(δΘ) = (Og!−k(C) for e1 + ··· + ek 6= 0. The above assumptions on the singularities are only made for simplicity. For example, lemma 6.4.2 illustrates how in some cases the condition on torsion points can be dropped, and similarly one can often treat nonisolated singularities. One should compare this with the situation for Jacobians of curves or intermediate Jacobians of cubic threefolds where G(δΘ) is much smaller [28]. The exceptions that we have taken out already appear for nonhyperelliptic Jacobians of genus g =4 where k = 2. This is also the only possibly missing case in the stratification of the moduli space Ag defined by G(δΘ) for g = 4; this stratification will be discussed in section 6.4, it refines both the Andreotti-Mayer stratification by the dimension of the singular locus of the theta divisor and the one by the degree of the Gauss map in [8]. For conjecture 3 it is good news that in so many cases the group G(δΘ) is the full symplectic or orthogonal group, since then the adjoint representation is just the symmetric or exterior square of the standard representation. Thus in section 5.4 we will verify the conjecture in many cases:

Theorem 5 (=5.4.3). Let (A, Θ) ∈ Ag(C). If G(δΘ) is a symplectic group and ω(δΘ) is its standard representation, then

dim Supp(AdΘ) ≥ g − 1, and hence Θ cannot be written as a sum of positive-dimensional subvarieties. One reason why the support estimate works so easily in this case is that here the adjoint representation is irreducible. In general, it is a very important problem to ﬁnd conditions on a clean module M ∈ Hol(DA) which ensure that the Lie algebra of the group G(M ) is simple modulo its center. In theorem 6.2.1 we will give various criteria which in particular apply to M = δΘ for many ppav’s, in fact we do not know any example of an indecomposable ppav for which the adjoint module AdΘ is not simple. Like all results in this paper, our criteria for the simplicity of the adjoint module rely on the notion of characteristic cycles. Recall that the sheaf DA has a natural order ﬁltration whose associated graded is the symmetric algebra of the tangent bundle. For M ∈ Hol(DA) the associated graded for any good 5

ﬁltration can be viewed as a coherent sheaf on the cotangent bundle T ∗A. Its support is independent of the chosen good ﬁltration and is called the characteristic cycle CC(M ) [21, def. 2.2.2]. It is a conic Lagrangian cycle on T ∗A and hence of the form CC(M ) = nZ ΛZ with nZ ∈ N0, Z⊆A ∗ X where ΛZ ⊂ T A denotes the conormal variety to a closed subvariety Z ⊂ A. As our Tannakian construction kills negligible modules, it only sees part of the above cycle: Let us say that a component ΛZ is negligible if its base Z ⊂ A is stable under translations by a non-zero abelian subvariety, and clean otherwise. A conic Lagrangian cycle is called negligible resp. clean if all components of its support are so. Thus any conic Lagrangian cycle on splits uniquely as a sum of a negligible and a clean part. We denote2 by

cc(M ) = nZ ΛZ ΛZXclean the clean part of the characteristic cycle. In section 1.1 we attach to each ΛZ a Gauss map generalizing the Gauss map for divisors on abelian varieties. We denote its degree by deg(ΛZ ) ∈ N0 and extend the degree additively to cycles. Coming back to the question if the adjoint module is simple, we show in section 6.1 that if the group G(M ) is isogenous to a product of two groups, then after applying an isogeny on the abelian variety one has a similar decomposition of cc(M ). This is excluded in many cases:

Theorem 6 (⊂ 6.2.1). Let (A, Θ) ∈ Ag(C). Then the Lie algebra of G(δΘ) is simple modulo its center in each of the following cases: 1 (1) If deg(ΛZ ) > 3 deg(cc(δZ )). (2) If the theta divisor has only isolated singularities.

(3) If the image of cc(δZ ) under any isogeny of abelian varieties is reduced. The main technical point for the above is to introduce a λ-ring L (A) of clean conic Lagrangian cycles on T ∗A in such a way that on the Grothendieck ring of representations of the Tannakian group G = G(M ), the assignment that sends a holonomic DA-module to the clean part of its characteristic cycle is a λ-ring homomorphism cc : K0(Rep(G)) −→ L (A). The resulting correspondence between Weyl group orbits of weights for G = G(M ) and characteristic cycles is based on the microlocalization from [26], but with three important novelties: First, rather than considering the monodromy of Gauss maps we only use that the representation ring of a connected reductive group is the Weyl group invariant part of the character ring; this allows to treat cycles with multiplicities. Next, we emphasize λ-rings to control tensor constructions such as exterior powers or more general Schur functors, where our methods are inspired by the work of Capell, Maxim, Sch¨urmann, Shaneson and Yokura in [6]. Finally, we discuss the problem whether an abstract representation of a reductive group is realized by a DA-module if some tensor construction of it is so. For our applications we want tensor constructions in several variables: By a tensor construction in r

2In definition 2.2.1 this notation will be used in a more general framework where clean cycles need not be conic Lagrangian, but for this introduction the present setup suffices. 6 THOMAS KRAMER¨ variables on a tensor category C we mean any functor S : C r → C that can be obtained as a composition of direct sums, tensor products and Schur functors. The same definition works in any λ-ring, see section 1.4. We then have:

Theorem 7 (⊂ 2.3.1). Let M ∈ Hol(DA). Let G be an abstract connected reductive group and p : G ։ G = Gb(M ) an isogeny. Let S be a tensor construction in r variables. If U1,...,Ur ∈ Rep(G) are representations with b ∗ b p (ω(N )) ≃ S(U1,...,Ur) for some N ∈ hM i = Rep(G), then there exist clean cycles Λ1,..., Λr ∈ L (A) with

[e]∗ cc(N ) = S(Λ1,..., Λr) for e = deg(p). We emphasize that a priori the group G need not be the Tannakian group of any holonomic DA-module and so the above is an inverse Galois problem for these groups. For instance, it would be interestingb to study the spin covers of orthogonal groups arising for theta divisors. Another case is the Schottky problem; we here only state the nonhyperelliptic case and refer to section 4 for the hyperelliptic case: Let us say that a ppav (A, Θ) ∈ Ag(C) is a nonhyperelliptic fake Jacobian if the pair (G(δΘ),ω(δΘ)) looks like the one for the theta divisor on the Jacobian of a nonhyperelliptic curve. We want to understand whether any fake Jacobian is ∗(g−1) indeed the Jacobian of such a curve C ⊂ A. For Jacobians δΘ = Alt (δC ) is an exterior convolution power. Reconstructing the curve from the theta divisor amounts to reversing the eﬀect of this tensor construction. So theorem 7 tells us where to look for the curve also in the case of fake Jacobians:

Theorem 8 (⊂ 4.1.2). If (A, Θ) ∈ Ag(C) is a nonhyperelliptic fake Jacobian, then we have g−1 [n]∗ cc(δΘ) = Alt (Λ) for some effective cycle Λ ∈ L (A) and some n ∈ N. When applied to Jacobians this gives a constructive Torelli theorem. For fake Jacobians the cycle Λ might however be supported not over a curve but over a higher-dimensional subvariety. Trying to exclude this, we find in section 4.2 the following solution to the weak Schottky problem in genus five: Corollary 9 (=4.2.1). For g = 5 the locus of fake Jacobians is contained in the Andreotti-Mayer locus

N1 = {(A, Θ) ∈ A5 | dim Sing(Θ) ≥ 1} ⊂ A5, hence it contains the locus of true Jacobians of curves as an irreducible component. The proof is a computation with Chern-Mather classes via Vogel’s intersection algorithm [44, 15, 16], inspired by [8]. Since these techniques are also useful in other contexts, we gather the relevant properties of Chern-Mather classes in section 3; in particular, we show that the total Chern-Mather class gives a ring homomorphism from certain subrings of L (A) to truncations of the Chow ring CH•(A) endowed with the Pontryagin product. It seems a very interesting problem to understand excess intersections related to positive-dimensional ﬁbers of the Gauss map. 7

1. The ring of clean cycles We now set up our basic framework for the discussion of characteristic cycles by introducing a λ-ring of clean cycles on the cotangent bundle. Let A be a complex 0 1 abelian variety of dimension g ≥ 1 and V = H (A, ΩA).

1.1. Motivation: Kashiwara’s formula. The cotangent bundle T ∗A = A × V is canonically trivialized using translations on the abelian variety. We then define the Gauss map of a subvariety Λ ⊂ T ∗A = A × V of pure dimension g = dim(A) to be the projection ∗ γΛ :Λ ֒→ T A = A × V ։ V to the second factor. We will mostly be interested in the case where Λ ⊂ T ∗A is conic, i.e. invariant under rescaling in the fibers of the cotangent bundle. We then write γ PΛ : PΛ → PV for the projectivized Gauss map. If Λ is taken to be the conormal variety to a divisor Z ⊂ A, this recovers the classical notion of Gauss maps on abelian varieties. Coming back to the general case, for dimension reasons the map γΛ is either generically finite or non-dominant. We say that Λ is clean in the first case and negligible in the second case. In what follows we denote the generic degree of the Gauss map by deg(Λ) = deg(γΛ) ∈ N0, so Λ is negligible iff deg(Λ) = 0. We extend the degree additively to cycles of pure dimension g ∗ on the cotangent bundle T A. Then for M ∈ Hol(DA) we can read off the Euler characteristic i i χ(A, DR(M )) = (−1) dimC H (A, DR(M )) i Z X∈ from the characteristic cycle: Theorem 1.1.1. We have χ(A, DR(M )) = deg(CC(M )) ≥ 0. Proof. This is a special case of Kashiwara’s index formula and was shown more generally for semiabelian varieties by Franecki and Kapranov [17]. The above suggests that the most interesting part of characteristic cycles are their clean components, so we should take a closer look at the group of clean cycles on T ∗A. In what follows we will endow it with a natural λ-ring structure tailored to fit the convolution product on holonomic DA-modules in section 2 below.

1.2. A categoriﬁcation. Motivated by the microlocalization of DA-modules, we categorify the notion of clean cycles by considering germs of vector bundles on their support as in [26, sect. 4]: Deﬁnition 1.2.1. For u ∈ V let VB(A, u) be the C-linear exact pseudoabelian category of germs

α = (Fα,Uα), an where Fα ∈ Coh(OA×Uα ) is an analytic coherent sheaf and u ∈ Uα ⊂ V a Zariski open such that

• the support Supp(Fα) ⊂ A × Uα is algebraic,

• the projection γα : Supp(Fα) → Uα is ﬁnite and ﬂat,

• the direct image γα∗(Fα) is a coherent analytic vector bundle. 8 THOMAS KRAMER¨

Morphisms of germs are deﬁned by

Oan HomVB(A,u)(α, β) = colim Hom A×U (Fα|A×U , Fβ|A×U ) where the colimit runs over all Zariski open U ⊂ Uα ∩ Uβ containing u, and short exact sequences are taken to be those that are represented by short exact sequences an in Coh(OA×U ) for U ∋ u small enough. Each germ in the above category gives rise to a clean cycle as follows: Definition 1.2.2. For ∅ 6= U ⊆ V Zariski open, a clean cycle on A × U is a finite formal sum Z = mΛ · Λ with coefficients mΛ ∈ Z, XΛ where Λ ⊂ A × U ranges over all irreducible Zariski closed subvarieties for which the projection to the second factor γΛ :Λ −→ U is a generically finite dominant map. Thus a cycle is clean iff the Zariski closure of each irreducible component of its support is a clean subvariety of the cotangent bundle. Put Z (A) = { clean cycles on T ∗A }, Z (A, U) = { clean cycles on A × U }, then the pullback of cycles via the embedding j : A × U ֒→ A × V gives a group isomorphism ∼ −1 Z (A) −→ Z (A, U), Λ 7→ Λ|U = j (Λ) whose inverse is the map that sends a clean cycle Z = Λ mΛ · Λ on A × U to its Zariski closure P Z = mΛ · Λ. XΛ on T ∗A. This Zariski closure does not change if we replace U ⊂ V by a smaller Zariski open dense subset. Sending the germ of a sheaf to the Zariski closure of its support we get a homomorphism 0 supp : K (VB(A, u)) −→ Z (A), α = (Fα,Uα) 7→ Supp(Fα) on the additive Grothendieck group modulo split exact sequences. 1.3. Lambda rings. The category VB(A, u) is naturally a tensor category for the product −1 (Fα,Uα) ∗ (Fβ,Uβ) = (̟∗ρ (Fα ⊠ Fβ),U = Uα ∩ Uβ), where ρ and ̟ are the maps in the correspondence of cotangent bundles induced by the addition on the abelian variety:

ρ ∗ 2 2 2 ̟ ∗ T A ⊇ A × Uα × Uβ o A × U / A × U ⊆ T A

((x, y),u,u)o ✤ ((x, y),u) ✤ / (x + y,u)

This deﬁnition is adapted to the convolution product on holonomic DA-modules in section 2. To see what it means for characteristic cycles, recall that for any exact tensor category its Grothendieck group modulo split exact sequences is a ring via the tensor product and a λ-ring3 via exterior powers [20, lemma 4.1]. In

3We use the term λ-ring for what is called a special λ-ring in loc. cit. 9 proposition 1.3.3 we will upgrade supp to a λ-ring homomorphism. Any λ-ring R has a natural operation by the ring of symmetric functions in countably many formal variables x1, x2,... [2, §I.3]: If ei = ei(x1, x2,... ) are the elementary symmetric functions, the map n µR : Z[e1,e2,... ] −→ Maps(R, R), en 7→ λ is a ring homomorphism when the target is endowed with pointwise addition and multiplication, and for any homomorphism f : R → S of λ-rings and any symmetric function σ the diagram

µR(σ) R / R

f f

µS (σ) S / S commutes. For instance the Adams operations on a λ-ring R can be defined as the maps n n n Ψ = µR(pn): R −→ R for the power sums pn = x1 + x2 + ··· As the elementary symmetric functions can be expanded as polynomials in power sums with rational coefficients, the Adams operations determine the λ-operations on any λ-ring without Z-torsion (see also section 1.4). Example 1.3.1. The representation ring of a complex algebraic group G is by definition the Grothendieck ring R(G) = K0(Rep(G)) of the category of its finite dimensional complex algebraic representations. By restricting representations to a multiplicative subgroup T ⊂ G we get a homomorphism R(G) → R(T ) = Z[X] to the group ring of the character group X = Hom(T, Gm). For maximal tori in connected reductive groups this restriction homomorphism is an isomorphism onto ∼ W the invariants R(G) −→ Z[X] of the Weyl group W = NG(T )/ZG(T ). Since any representation of a torus is a sum of characters and on characters the Adams operation Ψn acts by the n-th power map [n]: X → X,χ 7→ χn [5, II.7.4], we can then describe the Adams operations on the representation ring via the following commutative diagrams: R(G) / Z[X]

n Ψ [n]∗ R(G) / Z[X]

Example 1.3.2. By deﬁnition of the tensor structure on VB(A, u), the support map supp : K0(VB(A, u)) → Z (A) in deﬁnition 1.2.2 is a ring homomorphism for the product

−1 ◦ : Z (A) × Z (A) −→ Z (A), Λ1 ◦ Λ2 = ̟∗ρ (Λ1 × Λ2)|U .

Here U ⊆ V is any Zariski open dense subset over which the support of both Λi is finite and flat, and the closure and the pushforward are again taken in the sense of cycles. There is also a natural structure of a λ-ring on Z (A). To describe it we need some notations: For Λ ∈ Z (A), let hΛi ⊂ Z (A) be the smallest subring which contains Λ and is stable under taking irreducible components of its members; 10 THOMAS KRAMER¨ subrings of this form will be called finitely generated. For u ∈ V (C) consider the subgroup Γ(Λ,u) = ha ∈ A(C) | (a,u) ∈ Supp(Λ)i ⊂ A(C). Since Λ is a clean cycle, this group is finitely generated for general u ∈ V (C), and its isomorphism type jumps at most for u in a countable union of proper Zariski closed subsets of V (C). For n ∈ Z we write [n]: A × V → A × V, (a, v) 7→ (na,v) for the multiplication by n on the abelian variety and denote the induced group homomorphism on clean cycles by [n]∗ : Z (A) → Z (A). Proposition 1.3.3. With notations as above we have: (1) There exists a unique structure of a λ-ring on Z (A) such that for all u ∈ V the map supp : K0(VB(A, u)) −→ Z (A) n is a λ-ring homomorphism. This λ-ring structure is given by Ψ = [n]∗. (2) For any given Λ the subring hΛi⊂ Z (A) is stable under the λ-operations, and by taking the fiber over a very general closed point v ∈ V (C) we get an embedding

[v : hΛi ֒→ Z[Γ(−) as a λ-subring in the group ring of the abelian group Γ = Γ(Λ, v). Proof. If η ∈ V denotes the generic point, then for each closed point u ∈ V (C) we have an exact inclusion functor ι : VB(A, u) ֒→ VB(A, η) which preserves the tensor structure and is compatible with supports. This provides a factorization

supp K0(VB(A, u)) / Z (A) ▼▼ ✈: ▼▼ ✈✈ ▼▼▼ ✈✈ ι∗ ▼▼ ✈✈ supp ▼▼& ✈✈ K0(VB(A, η)) where ι∗ is a λ-ring homomorphism, so for the existence and uniqueness statement it suffices to take u = η. Note that we can verify these statements entirely in terms of the Adams operations because Z (A), being a ring of cycles with integer coefficients, has no Z-torsion. Moreover it will suffice to consider finitely generated subrings: To check whether a given set of ring endomorphisms Ψn : Z (A) → Z (A) are Adams operations of some λ-ring structure, we must check a set of universal identities that each only involves finitely many elements at a time. So let us fix Λ ∈ L (A). The subring hΛi ⊂ Z (A) is countable, so the reduced supports of all its members are finite and étale over a very general v ∈ V (C). In particular, distinct irreducible components of members of hΛi cannot meet over such a point v since then the union of their reduced supports would not be étale over that point. Hence taking the fiber of the occuring cycles over v defines an injective map (−)v : hΛi ֒→ Z[Γ] where Γ = Γ(Λ, v), and the definitions easily imply that this is a ring homomorphism. To see that its image is a λ-subring, let us denote by VectΓ(C) the category of finite-dimensional Γ-graded complex vector spaces. The full tensor subcategory

VB(A, v)Λ = { α ∈ VB(A, v) | supp(α) ∈ hΛi} ⊂ VB(A, v) 11

ﬁts in the commutative diagram

0 0 K (VB(A, v)Λ) / K (VectΓ(C))

supp (−)v hΛi / Z[Γ] where the top row is induced by the functor that sends α = (Fα,Uα) ∈ VB(A, v)Λ to the ﬁber (γα∗(Fα))(v) of the corresponding locally free sheaf, with the grading coming from the inclusion {a ∈ A(C) | (a, v) ∈ Supp(Fα)} ⊆ Γ = Γ(Λ, v). In particular, the top row of the diagram is a homomorphism of λ-rings and hence it follows that 0 supp : K (VB(A, v)Λ) −→ hΛi is a homomorphism of λ-rings when the target is endowed with the λ-ring structure 0 induced by the one of the group ring Z[Γ] = K (VectΓ(C)). For the latter the Adams operations Ψn are induced by the multiplication by n map on the character group Γ as we recalled in example 1.3.1. It follows that on the subring hΛi⊂ Z[Γ] n we have Ψ = [n]∗ as claimed.

1.4. Schur functors. In any exact tensor category C we have exterior powers or more generally Schur functors [12, sect. 1.4]. By a tensor construction in r variables we mean any functor S : C r = C ×···× C −→ C that can be obtained as a composition of tensor products, direct sums and Schur functors. Any tensor construction descends to an operation on K0(C ); in fact we can deﬁne tensor constructions more generally as natural operations on the category of λ-rings, with Schur functors replaced by the natural operation of the Schur polynomials [25, p. 43 and sect. III.3]. Thus any homomorphism of λ-rings is compatible with tensor constructions. Once we know the Adams operations, the eﬀect of Schur functors is easy to describe: For any partition β = (β1,...,βℓ) we have the power sum polynomials

ℓ βi pβ = xj . i=1 j YX These power sums are a basis for the ring of symmetric functions with rational coeﬃcients, so for any partition α = (α1, α2,... ) the Schur polynomial sα has an expansion

sα = mαβpβ with unique coeﬃcients mαβ ∈ Q β X where β = (β1,...,βℓ) runs over all partitions of degre deg(β) = deg(α). It follows that the operation of the Schur polynomial sα is given in terms of the Adams operations by

β1 β2 βℓ sα(x)= mαβ · Ψ (x) · Ψ (x) ··· Ψ (x). β X We thus obtain 12 THOMAS KRAMER¨

Lemma 1.4.1. For any Λ ∈ Z (A) and any partition α we have in Z (A) ⊗Z Q the formula

sα(Λ) = mαβ · Λ[β] where Λ[β] = [β1]∗(Λ) ◦···◦ [βℓ]∗(Λ) β X Proof. Immediate since by proposition 1.3.3 the Adams operations on the ring n of clean cycles are given by Ψ = [n]∗.

Example 1.4.2. The elementary symmetric polynomials en = s1,1,...,1 can be written in terms of the power sum polynomials pβ from above by expanding the generating series ∞ ∞ ν n (−1) +1 ν enX = exp ν pν X . n=0 ν=1 X X So the above lemma expresses exterior powers of clean cycles in terms of convolution products: 2 1 λ (Λ) = 2 Λ[1,1] − Λ[2] 3 1 λ (Λ) = 6 Λ[1,1,1] − 3Λ[2,1] + 2Λ[3] 4 1 λ (Λ) = 24 Λ[1,1,1,1] − 6Λ[2,1,1] + 3Λ [2,2] + 8Λ[3,1] − 6Λ[4] . .

2. Application to D-modules

We now apply the above to characteristic cycles of holonomic DA-modules along the lines of [26] to get a correspondence between such cycles and Weyl group orbits of weights in representation theory that allows for arbitrary multiplicities.

2.1. Fiber functors. A module M ∈ Hol(DA) satisﬁes χ(A, DR(M )) = 0 iﬀ it is negligible in the sense that each of its simple subquotients is stable under translations by a non-zero abelian subvariety [31, 38]. Let S(A) ⊂ Hol(DA) be the Serre subcategory of negligible modules. The quotient category

M(A) = Hol(DA)/S(A) is a rigid abelian tensor category for the convolution product ∗ induced by the group law A × A → A [26]. The discussion in the previous section is motivated by Theorem 2.1.1. Let η ∈ V denote the generic point. Then there exists a C-linear exact tensor functor µ : M(A) −→ VB(A, η) such that CC(M ) ≡ supp(µ(M )) modulo negligible cycles for all M ∈ M(A). Proof. Microlocalization even allows to define a tensor functor to an analogous category of local systems [26]. We will not use the monodromy operation in what follows, so we replace local systems by their sheaves of holomorphic sections. One reason for working with germs of vector bundles rather than local systems is that there are other interesting tensor functors to which the following discussion applies. So for the rest of this section we fix an arbitrary C-linear exact tensor functor F : M(A) → VB(A, η) where η ∈ V is the generic point; for the moment the specific choice will not matter, although in our applications starting from section 4 13 we will always use the functor F = µ from theorem 2.1.1. Fix M ∈ Hol(DA), and let hM i⊂ M(A) be the smallest rigid abelian tensor subcategory containing M . It can be described explicitly as the full subcategory of all subquotients of convolution powers of M ⊕M ∨ ∈ M(A), and as such has at most countably many isomorphism classes of objects. So for a very general closed point u ∈ V (C) the restriction of F to this subcategory factors over a C-linear exact tensor functor

Fu : hM i −→ VB(A, u).

Taking ﬁbers over u we then obtain a C-linear exact tensor functor ωu such that the diagram ωu hM i / Vect(C) ❏❏❏ r9 ❏❏ rrr Fu ❏❏ rrfiber ❏$ rr over u VB(A, u) commutes. Thus we recover the following Tannakian description [31, 26]:

Corollary 2.1.2. There is a linear algebraic group G = G(M ,u) such that ωu underlies an equivalence ∼ ωu : hM i −→ Rep(G) with the tensor category of ﬁnite-dimensional complex linear representations of G.

Proof. The functor ωu is faithful since any exact tensor functor between rigid abelian categories with End(1)= C is so [13, prop. 1.19]. By th. 2.11 in loc. cit. the ⊗ group G(M ,u) = Aut (ωu) of its tensor automorphisms is then represented by a linear algebraic group which does the job. Since we work over an algebraically closed field, any two fiber functors on our category are non-canonically isomorphic, and so are the corresponding Tannakian groups. We simply write G(M ) = G(M ,u) and ω = ωu when the specific choice of the fiber functor is either clear from the context or does not matter.

2.2. Characteristic cycles and weights. We want to understand the relation between the above Tannakian groups and the clean cycles that arise as supports of objects in the image of our chosen functor F : M(A) −→ VB(A, η): Definition 2.2.1. By the clean characteristic cycle of M ∈ M(A) we mean the cycle cc(M ) = supp(F (M )) ∈ Z (A). If F = µ is chosen as in theorem 2.1.1, then this definition agrees with the usual characteristic cycle modulo negligible cycles, but unlike the latter it is well-defined not only on Hol(DA) but also on the quotient category M(A); we use the lowercase notation to distinguish our clean characteristic cycle from the usual one. For other choices of F there is no obvious relation between the two but cc still behaves like a characteristic cycle, for instance Kashiwara’s index formula still holds: Corollary 2.2.2. For G = G(M ,u) as in corollary 2.1.2, we obtain a λ-ring homomorphism 0 cc : R(G) = K (Rep(G)) −→ hcc(M )i ⊂ Z (A), ωu(N ) 7→ cc(N ) 14 THOMAS KRAMER¨ with

dim(ωu(N )) = deg(cc(N )) = χ(A, DR(N )) for all N ∈ hM i.

Proof. By assumption we have a tensor functor Fu : hM i −→ VB(A, u). Any such functor induces a λ-ring homomorphism between the Grothendieck rings, so the first claim holds by proposition 1.3.3. The formula dim(ωu(N )) = deg(cc(N )) follows from our construction of the fiber functor. On the other hand, since any two fiber functors are isomorphic, we have dim(ωu(N )) = χ(A, DR(N )) by [31]. This gives a dictionary between characteristic cycles and representation-theoretic weights as follows. With notations as in proposition 1.3.3, take the finitely generated subgroup Γ = Γ(Λ,u) ⊂ A(C) for Λ = cc(M ). Its torsion part contains the subgroup

K = K(M ,u) = a ∈ Γtors | Supp(N )= {a} for some N ∈ hM i . ◦ By [26, th. 1.3] this subgroup is Cartier dual to the group of components G/G of the Tannakian group G = G(M ,u), and the direct image of D-modules under the isogenies in the diagram

h = g◦f A / A/Γ ❇ tors ❇ ✉: ❇❇ ✉✉ ❇ ✉✉ f ❇❇ ✉✉ g ❇! ✉✉ A/K

∼ ∼ ◦ induces isomorphisms G(h∗M ,u) −→ G(f∗M ,u) −→ G ⊆ G onto the connected component of our Tannakian group. This is particularly useful in the reductive case because representations of connected reductive groups are determined completely by their weights with respect to a maximal torus (example 1.3.1). In any case we obtain: (Theorem 2.2.3. In the above setting, for any maximal torus T ֒→ G = G(M ,u one has an epimorphism p : X = Hom(T, Gm) ։ Γfree such that the following diagram commutes:

(−)|G◦ R(G) / R(G◦) ∼ / R(G◦) / Z[X]

cc cc cc p

f∗ g∗ (−)u hcc M i / / hf∗ cc M i / / hh∗ cc M i / Z[Γfree ]

(−)u

h∗ Z[Γ] / / Z[Γfree ]

◦ Proof. By the above discussion f∗ ◦cc and h∗ ◦cc factor over R(G ) as in the two upper left squares of the diagram. Furthermore, by construction we have a ﬁber functor ωu : Rep(G) → VectΓ(C) with values in Γ-graded vector spaces. The latter are naturally representations of the Cartier dual Hom(Γ, Gm), which is a subgroup of multiplicative type. Its connected component is a subtorus Hom(Γfree , Gm) ֒→ G and up to conjugacy we may assume that it sits in a given maximal torus. We then get the desired epimorphism X ։ Γfree on character groups. 15

If M ∈ Hol(DA) is semisimple, then the group G = G(M ,u) is reductive. In this case, let us denote by W (G) = NG◦ (T )/ZG◦ (T ) the Weyl group of its connected component. The above dictionary then takes the following form:

Corollary 2.2.4. In the above setting, any orbit O ⊆ X of the Weyl group W (G) is realized geometrically by a unique clean eﬀective conic Lagrangian cycle in the sense that p(O)=Λu for a unique Λ ∈ L (A).

[Proof. By the theory of connected reductive groups, the image of R(G◦) ֒→ Z[X are the Weyl group invariants (example 1.3.1).

2.3. An inverse Galois problem. Applying tensor constructions like exterior powers to a faithful representation will usually destroy the faithfulness. In good cases the outcome will still be a faithful representation of an isogenous group, n 2n e.g. Alt (C ) is a faithful representation of the quotient Sl 2n(C)/µn by the n-th roots of unity. In this case we can recover the original group as the universal cover. It is a natural to ask if this abstract Lie-theoretic procedure has a geometric incarnation: Let M ∈ Hol(DA). If G = G(M ,u) is connected and semisimple, is its universal cover realized as

G = G(M ,u) ։ G for some M ∈ Hol(DA) with M ∈ hM i?

In this generalitye thise is not true, one cane ﬁnd examples on elliptic curve es where one ﬁrst needs to replace M by its direct image under an isogeny A ։ B. On the level of characteristic cycles this is the only obstruction:

Theorem 2.3.1. Let M ∈ Hol(DA), and let p : G ։ G = G(M ,u) be an isogeny from an abstract connected reductive group onto its Tannakian group. Let S be a tensor construction in r variables. If N ∈ hM i andb U1,...,Ur ∈ Rep(G) are such that

∗ b p (ωu(N )) ≃ S(U1,...,Ur), then this isomorphism is realized geometrically by cycles Λi ∈ hcc(M )i ⊂ Z (A) with

[e]∗ cc(N ) = S(Λ1,..., Λr) for e = deg(p).

Proof. For any isogeny G ։ G the image T ⊂ G of a maximal torus T ⊂ G is again a maximal torus, and the degree of the induced isogeny between these tori is again e. Then X = Hom(T,bGm) ֒→ X = Hom(T, Gm) is an index e subgroupb b and hence

b [e]∗ b Z[eX] = im Z[X] / Z[X] ⊆ Z[X]. b b b e Recalling from the proof of lemma 1.4.1 that on character rings we haveΨ = [e]∗, we get a factorization

Ψe : R(G) −→ R(G) ⊆ R(G)

b b 16 THOMAS KRAMER¨ from the diagram

R(G) ∼ / Z[X]W

e ⑤= O ①< O Ψ ⑤ [e]∗ ①① ⑤⑤ b ①① b ⑤⑤ ①① ⑤⑤ ①① R(G) ∼ / Z[X]W

b b ∃ ∃ ! ? # ? R(G) ∼ / Z[X]W where W = W (G) = W (G) denotes the Weyl group of our two groups. Hence we put e b Vi := Ψ (Ui) ∈ R(G). In general these are only virtual representations, formal Z-linear combinations of representations, but this is enough for our purpose. Since the Adams operations are homomorphisms of λ-rings, the naturality of tensor constructions with respect to such homomorphisms gives e e e e Ψ (ωu(N )|Gb)= Ψ (S(U1,...,Ur)) = S(Ψ (U1),..., Ψ (Ur)) = S(V1,...,Vr) in R(G). Applying the λ-ring homomorphism cc : R(G) → hM i from corollary 2.2.2 we obtain

[e]∗ cc(N ) = S(Λ1,..., Λr) forthecycles Λi := cc(Vi) ∈ hcc(M )i as desired, where for the left hand side we have used the description of the Adams operations on clean cycles in proposition 1.3.3. Note that we have not touched the Adams operations on the right hand side: The homomorphism cc is only deﬁned e on the subring R(G) ⊆ R(G), so we can apply it to Vi =Ψ (Ui) but not to Ui.

Remark 2.3.2. In the aboveb result the group G = G(M ) must be connected. To reduce to the connected case we pass to A/K for the subgroup K = K(M ) ⊆ Γtors :

f g A / A/K / A/Γtors : h

The cycles Λi in the above result need not be eﬀective, but the g∗Λi are eﬀective since they come from nonnegative linear combinations of weights in the following diagram: ch _ R(G) / Z[X] o ? N0[X]

h∗ cc p (−)u _ hh∗ cc(M )i / Z[Γfree ] o ? N0[Γfree ]

We refer to this situation by saying that the Λi ∈ L (A) are eﬀective up to isogeny.

We will apply the above results for the tensor construction S(V ) = Altn(V ) in theorem 4.1.2 and for S(V1, V2)= V1 ⊠ V2 in proposition 6.1.1. 17

3. Chern-Mather classes For the rest of this paper we will always use the ﬁber functors from theorem 2.1.1, so we work in the subring L (A) ⊂ Z (A) of clean conic Lagrangian cycles. This allows to use tools from conormal geometry. We will see that Chern-Mather classes deﬁne a ring homomorphism from certain subrings of L (A) to truncations of the Chow ring CH•(A) endowed with the Pontryagin product.

3.1. Generic transversality. Recall that any conic Lagrangian subvariety of the cotangent bundle arises as the conormal variety Λ = ΛZ ⊂ A × V to a closed subvariety Z ⊂ A [23, lemma 3]. We denote its image in the projective cotangent bundle by PΛ ⊂ A × PV and extend the notation additively to conic Lagrangian cycles. Let p : A×PV ։ A be the projection.

Deﬁnition 3.1.1. The Chern-Mather classes of a conic Lagrangian cycle Λ on T ∗A are deﬁned by

cM,d(Λ) = p∗ [PΛ] · [A × Hd] ∈ CHd(A) where Hd ⊆ PV is a general linear subspace of dimension d ∈{0, 1,...,g − 1}.

For conormal varieties Λ = ΛZ with Z 6= A, the triviality of the cotangent bundle implies that these classes coincide with the dual Chern-Mather classes of Z in the sense of [37, lemme 1.2.1] and [23, lemma 1], except that we consider them as classes on the ambient abelian variety. By Kleiman’s generic transversality theorem they ∗ are all represented by eﬀective cycles, and they are non-trivial if ΛZ ⊂ T A is clean in the sense of section 1.1:

Lemma 3.1.2. Let Z ⊂ A be an irreducible subvariety of dimension d

reg (1) For any open subset U ⊂ Z the intersection PΛZ ∩(U ×Hi) is transversal or empty, so

cM,i(ΛZ ) = p∗ PΛZ ∩ (U × Hi) . (2) In particular d = max i | cM,i(ΛZ ) 6=0 and cM,d(ΛZ ) = [Z].

(3) If the map γZ : PΛZ ։ PV is dominant, then cM,i(ΛZ ) 6=0 for all i ≤ d.

Proof. For the ﬁrst two parts see [41, prop. 2.8] [24]. If the map γZ is dominant, then

Wi = PΛZ ∩ (A × Hi) is nonempty for all i < g.

But we know from the second part that the projection p : Wi ։ p(Wi) ⊂ A is generically ﬁnite over its image for i = d, and then the same holds for all i ≤ d because the Hi ⊂ PV have been chosen generically. 18 THOMAS KRAMER¨

3.2. Vogel’s intersection algorithm. In order to compute Chern-Mather classes for divisors one can apply Vogel’s intersection algorithm [44, 15, 16]. Let us brieﬂy recall how this works in the special case required here: Consider an irreducible projective variety Z of dimension n. Fix L ∈ Pic(Z), and let W ⊆ H0(Z, L ) be a subspace such that the map

|W | : Z 99K PW ∗ is generically ﬁnite and dominant. Then inductively for n ≥ i ≥ 0 we get eﬀective cycles Vi and Ri of pure dimension i such that

• Vn = ∅ and Rn = [Z], 0 • Vi + Ri = Ri+1 ∩ div(si) for the divisor of a generic si ∈ W ⊂ H (Z, L ),

• Supp(Vi) is entirely contained in the base locus of |W |,

• Supp(Ri) does not have any component contained in this base locus.

The Vi are called Vogel cycles and the Ri are the residual Vogel cycles. Note that their classes

vi(L , W ) = [Vi], ri(L , W ) = [Ri] ∈ CHi(Z) modulo rational equivalence do not depend on the choice of the generic sections si in the construction. If Z ⊂ A is a subvariety of our abelian variety, we use the same notation for the images of these cycles in CHi(A). They can be expressed via Segre classes but have the advantage of being represented by eﬀective cycles. As in [8] we have

Corollary 3.2.1. Let L = OA(Z)|Z where Z ⊂ A is an irreducible reduced ample divisor, and denote by

0 W = ∂ν ϑ | ν =1, 2,...,g ⊆ H (Z, L )

0 the subspace spanned by the derivatives of a section ϑ ∈ H (A, OA(Z)) with zero locus div(ϑ)= Z. Then

ri(L , W ) for all i ∈{0, 1,...,g − 1}, cM,i(ΛZ ) = g−1 ( [Z] for all i> dim Sing(Z).

Proof. For a divisor the projection p : PΛZ → Z is a birational map, and the composite

γZ : Z 99K PΛZ ֒→ A × PV ։ PV is the Gauss map sending a smooth point of the divisor to the conormal direction 0 at that point. Up to a scalar there is a unique ϑ ∈ H (A, OA(Z)) with div(ϑ)= Z, and γZ (z) = ∂1ϑ(z): ··· : ∂gϑ(z) for any smooth point of the divisor. So the Gauss map is the rational map deﬁned by the linear series |W |, its base locus is the singular locus of the divisor, and for an ample divisor it is dominant and generically ﬁnite [4, sect. 4.4]. Now apply part (1) of the previous lemma. 19

3.3. The Pontryagin product. Let us now consider Chern-Mather classes of convolutions. For a conic Lagrangian subvariety Λ ⊂ T ∗A its total Chern-Mather class is

cM (Λ) = cM,0(Λ) + cM,1(Λ) + ··· + cM,g−1(Λ) ∈ CH•(A). By the above this class is given by an eﬀective cycle, which vanishes iﬀ Λ ⊂ T ∗A is the zero section. So if A has nontrivial proper abelian subvarieties, then there are negligible conic Lagrangian subvarieties whose total Chern-Mather class does not vanish. In what follows we restrict ourselves to clean cycles and consider the group homomorphism cM : L (A) −→ CH•(A) which sends a clean cycle to its total Chern-Mather class. Ideally we would like this to be a ring homomorphism when the target is endowed with the Pontryagin product

∗ : CHa(A) × CHb(A) −→ CHa+b(A), α ∗ β = a∗(α ⊠ β), but this cannot be true: In the definition of the convolution product ◦ on L (A) we have only worked with germs of subvarieties, whereas in general the Pontryagin product of clean cycles may involve negligible contributions. To control the latter we need estimates on positive-dimensional fibers of Gauss maps. For1 ≤ d ≤ g − 1, let us denote by L >d(A) ⊆ L (A) the subgroup generated by all clean conic Lagrangian subvarieties Λ ⊂ T ∗A whose Gauss map γPΛ : PΛ → PV restricts to a finite morphism over the complement of a proper closed subset S ⊂ PV of codimension codim(S, PV ) > d. Since the class of finite morphisms is stable under base change, one easily checks that this defines a filtration by subrings L (A) = L >1(A) ⊇ L >2(A) ⊇ ··· ⊇ L >g−1(A) where the final step on the right is generated as a group by the conic Lagrangian subvarieties whose Gauss map is finite. Similarly, for any d consider the quotient ring

CH≤d(A) = CH•(A)/CH>d(A) by the ideal CH>d(A) = CHi(A), i>d M then we obtain: Lemma 3.3.1. For any d ∈{1, 2,...,g − 1} the Chern-Mather class gives rise to a ring homomorphism >d cM : L (A) −→ CH≤d(A). More precisely, we have >d cM (Λ1 ◦ Λ2) − cM (Λ1) ∗ cM (Λ2) ∈ CH>d(A) for all Λ1 ∈ L (A), Λ2 ∈ L (A). Proof. The first claim follows from the second. For the latter we may assume ∗ by additivity that Λ1, Λ2 ⊂ T V are irreducible subvarieties. The main point will then be to give an intersection-theoretic expression for the class of Λ1 ◦ Λ2 in CHg−1(A × PV ). The diagonal δ : ∆= A2 × PV ֒→ A2 × PV × PV 20 THOMAS KRAMER¨ is a regular embedding of codimension g − 1, so by [18, lemma 7.1] any irreducible component −1 W ⊂ δ PΛ1 × PΛ2 = PΛ1 ×PV PΛ2 has dim(W ) ≥ g − 1. Strict inequality can only occur if the image of W in PV is contained in the locus where at least one of the Gauss maps PΛi → PV has positive-dimensional fibers. Extending the notions from section 1.1 to this case, we say that • W is clean if the projection W → PV is dominant and hence restricts to a finite étale cover over an open dense subset of the target. • W is negligible otherwise, in particular whenever dim W >g − 1.

Now Λ1 ◦ Λ2 is a linear combination of clean conic Lagrangian subvarieties. After projectivization we obtain by deﬁnition of the product ◦ that P(Λ1 ◦ Λ2)= ̟∗(α), where ̟ : A2 × PV −→ A × PV is the addition map and

2 α = [W ] ∈ CHg−1(A × PV ) W clean X −1 is the sum of all clean irreducible components W ⊂ δ (PΛ1 × PΛ2). Note that each of these components is reduced and enters with multiplicity one because the projection W → PV is ﬁnite ´etale over an open dense subset of the target. On the other hand the class ! 2 β = δ ([PΛ1 × PΛ2]) ∈ CHg−1(A × PV ) is represented by a cycle which might contain also negligible summands, but it follows from basic intersection theory that α and β agree on all clean components. So we may write

2 α − β = mW · βW with βW ∈ Im CHg−1(W ) → CHg−1(A × PV ) W negligible X for certain multiplicities mW ∈ Z.

Now let S ⊂ PV be a closed subset such that the Gauss map γ : PΛ1 → PV restricts to a finite morphism over U = PV \ S. Since finite morphisms are stable under base change, −1 γ × id : δ (PΛ1 × PΛ2)|U −→ PΛ2|U is then also a finite morphism. Since the target is irreducible of dimension g − 1, it follows that −1 (γ × id)(W ) ⊆ S for any negligible component W ⊂ δ (PΛ1 × PΛ2).

If codim(S, PV ) > d, it follows that via the K¨unneth decomposition for the Chow ring of a projective bundle

2 2 α − β ∈ CHi(A ) ⊗ CHg−1−i(PV ) ⊂ CHg−1(A × PV ). i>d M 21

To conclude the proof, one easily checks from the deﬁnition of the classes α and β that i ̟∗(α)= cM,i(Λ1 ◦ Λ2) ⊗ H i X i+j ̟∗(β)= (cM,i(Λ1) ∗ cM,j (Λ2)) ⊗ H i,j X where H ∈ CHg−2(PV ) denotes the class of a hyperplane.

Corollary 3.3.2. If A is a simple abelian variety, the total Chern-Mather class gives a ring homomorphism

cM : L (A) −→ CH•(A)/CHg(A).

Proof. If there exists a subvariety Z ⊂ A whose Gauss map γZ : PΛZ ։ PV is −1 dominant but admits a positive-dimensional ﬁber γZ (ξ), then this ﬁber generates a nontrivial abelian subvariety. On a simple abelian variety this is impossible, so we get L (A)= L >g−1(A) and the previous lemma applies.

4. The Schottky problem One motivation for the inverse Galois problem from section 2.3 is the Schottky problem whether the Tannakian formalism detect Jacobians among all ppav’s. In this section we discuss a possible approach to this question combining theorem 2.3.1 with the results of the previous section about Chern-Mather classes.

4.1. Tannakian Schottky. Let A be a ppav with theta divisor Θ ⊂ A. The theta divisor is determined by the polarization only up to a translation, and in what follows we ﬁx one of the 22g symmetric translates; in section 4.3 we will see that the speciﬁc choice does not matter too much. If Θ is smooth, we know by [26] that

Spg!(C) if g is even, G(δΘ,u) ≃ (SOg!(C) if g is odd, and ωu(δΘ) is the standard representation. For theta divisors on special ppav’s the group can be much smaller: Example 4.1.1. If (A, Θ) = Jac(C) is the Jacobian of a smooth projective curve C of genus g = n+1, we have the resolution C(n) ։ Θ. By the decomposition theorem we get an inclusion ∗n ( δΘ ֒→ Alt (δC and by [29, sect. 6] [45] it follows that we have an isogeny G = G(δC ,u) ։ G(δΘ,u) with n e 2n Sl 2n(C) Alt (C ) e (⋆) G ≃ acting via ωu(δΘ)|G ≃ n 2n n−2 2n (Sp2n(C) (Alt (C )/Alt (C ) in the nonhyperelliptice resp. hyperelliptic case. It is natural to ask whether this characterizes Jacobians. We call a ppav (A, Θ) of dimension g = n + 1 a nonhyperelliptic or hyperelliptic fake Jacobian if for some symmetric translate of the theta divisor the semisimple group G(δΘ,u) is connected 22 THOMAS KRAMER¨ and its universal cover acting on ωu(δΘ) has the form (⋆). In the moduli space Ag consider the loci

Jg ⊆ Jg,fake of Jacobians of smooth projective curves and of fake Jacobians; the latter is a locally closed algebraic subset by [30, prop. 7.4]. We would like to see if the two loci are the same, so for any fake Jacobian we need to ﬁnd a candidate for the curve whose Jacobian it should be. A natural guess is provided by

Theorem 4.1.2. If (A, Θ) is a fake Jacobian, then there exists a cycle Λ ∈ hcc(δΘ)i, eﬀective up to isogeny, such that

Alt∗(g−1)(Λ) with e = g − 1 [e]∗ cc(δΘ) = ∗(g−1) ∗(g−3) (Alt (Λ) − Alt (Λ) with e = gcd(2,g − 1) ∈{1, 2} in the nonhyperelliptic resp. hyperelliptic case.

Proof. Apply theorem 2.3.1 to the connected reductive group G = G(δΘ,u). We have f = id A and by direct inspection the degree e of the universal covering map is the given one. So the claim follows by taking S(U) = Altg−1(U) respectively S(U) = Altg−1(U)/Altg−3(U); remark 2.3.2 says that the resulting cycle is eﬀective up to isogeny. If (A, Θ) is the Jacobian of a smooth projective curve, then the above cycle Λ is indeed the conormal variety to the image of the curve under the Abel-Jacobi map and this gives a constructive proof of Torelli’s theorem [26]. For fake Jacobians the situation is less clear. We would be done if we could show that the image Z ⊂ A of Λ ⊂ A×V is a curve, as then Θ = Z +···+Z is a sum of g−1 copies of this curve and hence a Jacobian by [39]. Unfortunately it is hard to control convolutions with excess dimension: How to show dim(Z)=1?

4.2. A computation of Chern-Mather classes. One way to control dimensions is to look at Chern-Mather classes. Let us illustrate this technique by a simple example. For any g the locus of Jacobians Jg is an irreducible component of the Andreotti-Mayer locus

Ng−4 = {(A, Θ) ∈ Ag | dim Sing(Θ) ≥ g − 4}, see [1]. So for abelian ﬁvefolds we get

Corollary 4.2.1. For g = 5, any nonhyperelliptic fake Jacobian lies in Ng−4 and so the closure of the locus of fake Jacobians in Ag contains the closure of the Jacobian locus as an irreducible component. Proof. Let (A, Θ) be a nonhyperelliptic fake Jacobian of genus g = 5. In what follows we work up to numerical equivalence and therefore view Chern-Mather classes cM (−) ∈ H2•(A, Z) as classes in the even homology ring without further ′ ′ notice. This being said, let us write cc(δΘ)=ΛΘ +Λ where Λ is supported over the singular locus of the theta divisor. If the latter were empty or ﬁnite, we would get that 4 cM,1(cc(δΘ)) = cM,1(ΛΘ) = [Θ] in H2(A, Z), 23 where the second equality again uses our assumption on isolated singularities and corollary 3.2.1. For Λ ∈ L (A) as in theorem 4.1.2, we compute the Chern-Mather classes of ∗(g−1) [g − 1]∗ cc(δΘ) = Alt (Λ) in two diﬀerent ways, looking at both sides of the above equation. For the right hand side put ci = cM,i(Λ). By construction c0 = 2g − 2 = 8. Now consider the cycles

Λ[β] = [β1]∗(Λ) ◦···◦ [βℓ]∗(Λ) for partitions β = (β1,...,βℓ). Lemma 3.3.1 allows to easily compute their Chern-Mather classes in degree d = 1, yielding 3 cM,1(Λ[1,1,1,1])= 4 · c0 · c1 = 2048 · c1, 2 cM,1(Λ[2,1,1])= 6 · c0 · c1 = 384 · c1,

cM,1(Λ[2,2])= 8 · c0 · c1 = 64 · c1,

cM,1(Λ[3,1]) = 10 · c0 · c1 = 80 · c1,

cM,1(Λ[4]) = 16 · c1 = 16 · c1. By example 1.4.2 then

4 1 cM,1(Alt (Λ)) = 24 cM,1 Λ[1,1,1,1] − 6Λ[2,1,1] + 3Λ[2,2] + 8Λ[3,1] − 6Λ[4] = 20 · c1.

On the other hand

2 4 cM,1([g − 1]∗ cc(δΘ)) = (g − 1) · cM,1(ΛΘ) = 16 · [Θ] by corollary 3.2.1, again using our assumption that the theta divisor has at most 4 isolated singularities. Altogether then 16 · [Θ] = 20 · c1. But this leads to the contradiction 4 c = · [Θ]4 ∈/ H (A, Z) 1 5 2 4 since [Θ] is not divisible by ﬁve, being 24 times a primitive class in H2(A, Z).

4.3. Dependence on the translate. Let us brieﬂy explain to what extent the above depends on the translate of the theta divisor. Taking Θ ⊂ A to be symmetric determines G(δΘ,u) except for a small issue about connected components:

Example 4.3.1. For g = 1 the theta divisor is a 2-torsion point, so the group G(δΘ) will be trivial if we choose the point to be the origin, while otherwise it will have order two. A more interesting example is the intermediate Jacobian of a smooth cubic threefold. Here g = 5 and the theta divisor has a unique singularity x ∈ A[2], whence E6(C) for x =0, G(δΘ) ≃ (E6(C) × Z/2Z for x 6=0, as one easily sees in terms of the action of Hom(hxi, Gm) ⊂ G(δΘ) in [26, ex. 2.2]. In the above examples the groups arising from diﬀerent translates have the same connected component, which is a semisimple group since by the symmetry of the theta divisor its irreducible faithful representation ωu(δΘ) must be self-dual. More 24 THOMAS KRAMER¨ generally, let ta : A → A, x 7→ x + a denote the translation by a point a ∈ A(C) and put ∗ Ma = ta(M ) = M ∗ δa for M ∈ Hol(DA). If G = G(M ) is reductive, let G′ = [G◦, G◦] be the derived group of its connected component. The following shows that even for nonsymmetric theta divisors the situation does not seriously depend on the chosen translate:

′ ′ Lemma 4.3.2. If M is semisimple, then G(M ) ≃ G(Ma) for all a ∈ A(C). ⊠ Proof. Put N = (M δ0) ⊕ δ(0,a) ∈ Hol(DA×A). The group law a : A × A → A gives rise to the following commutative diagram between tensor categories, which by Tannakian duality translates to a diagram of reductive groups: a∗ hM ⊕ δaio o hN i G(M ⊕ δa) / G(M ) × G(δa) e❏❏ O PP ❏ ⊠ PP ❏❏ (−) δ0 PP ❏❏ PPP ❏ PP( ( hM i G(M ) The functor which sends a reductive group to the derived group of its connected component clearly preserves embeddings, epimorphisms and direct products and it sends the multiplicative group G(δa) to the trivial group, hence from the above we get a diagram ′ ′ G(M ⊕ δa) / G(M ) ▼ ▼▼▼ ▼▼▼ ▼▼& & G(M )′ which shows that the diagonal arrow must be an isomorphism. On the other hand, ′ ′ since hM ⊕ δai = hMa ⊕ δai, we know that G(M ⊕ δa) ≃ G(Ma ⊕ δa) and hence the claim follows by symmetry.

5. Summands of subvarieties Our next application is a criterion to detect nontrivial summands in a given subvariety Z ⊂ A. For this we will consider the adjoint representation, but before doing so we collect some general facts about conormal varieties. 5.1. Geometric nondegeneracy. An irreducible subvariety X ⊂ A is said to be geometrically nondegenerate if for any epimorphism A ։ B to an abelian variety the induced morphism X → B is either surjective or generically finite onto its image [36, II.12]. This is weaker than being nondegenerate in Ran’s sense [36, II.1] but still forces the stabilizer Stab(X) = { a ∈ A | X + a = X } to be finite. The converse does not hold in general: For example, replacing A by a larger ambient abelian variety will destroy the geometric nondegeneracy but does not affect the stabilizer. By [47], the stabilizer Stab(X) is finite if and only if the Gauss map ∗ γ :ΛX ⊂ T A = A × V ։ V is generically finite, i.e. iff the conormal variety is not negligible. 25

5.2. Two lemmas about conormal varieties. The following two lemmas allow to control the convolution product of conormal varieties:

Lemma 5.2.1. If Z1,Z2 ⊂ A are irreducible subvarieties with ﬁnite stabilizer, every nonnegligible irreducible component ΛY ⊆ Supp(ΛZ1 ◦ ΛZ2 ) of their convolution satisﬁes

dY ≥ |dZ1 − dZ2 |, and equality implies

Z1 = Y − Z2 or Z2 = Y − Z1.

Proof. If ΛY is an irreducible component of the convolution, then by definition of the convolution there exists a component Λ of ΛZ1 ×V ΛZ2 that dominates Y via the sum map: Λ / ΛZ1 ×V ΛZ2 ((z1, ξ), (z2, ξ)) ❴ Y / A z1 + z2 Composing the inclusion of this component in the fiber product with the projection on the second factor we get a map Λ → ΛZ2 , ((z1, ξ), (z2, ξ)) 7→ (z2, ξ) which is dominant because both the source and the target are irreducible and over some open dense subset of V both are finite étale. In particular, for general z2 ∈ Z2 we can always find a point z1 ∈ Z1 with y = z1 + z2 ∈ Y . We therefore have an inclusion Z2 ⊆ Y − Z1 and the same holds after interchanging Z1 and Z2.

Lemma 5.2.2. If Z1,Z2 ⊂ A are irreducible subvarieties with ﬁnite stabilizer and ΛY is a component of multiplicity one in the cycle ΛZ1 ◦ ΛZ2 , then one has dominant rational maps

Y ❴❴❴❴ / / Zi for i = 1, 2.

Proof. As above there exists an irreducible component Λ ⊆ ΛZ1 ×V ΛZ2 which dominates ΛY . Here ̟ :Λ ։ ΛY is generically ﬁnite of degree one because of our multiplicity one assumption. Hence ̟ is birational, and composing a rational inverse with the projection on the i-th factor we get a rational map fi as indicated below: Λ / ΛZ1 ×V ΛZ2 / / ΛZi

∃fi ΛY ❴❴❴❴❴❴❴❴❴❴ / / ΛZi

As in the previous proof fi is dominant since it commutes with the projection to V and both the source and the target are irreducible generically ﬁnite covers of the latter. Composing with the map to the abelian variety we get a dominant rational map ΛY 99K Zi. Now ΛY contains the conormal bundle to the smooth locus of Y as a Zariski open dense subset, and as such it is birational to the product Y × Cd where d = g−dY and Y → Y is a resolution of singularities. Since any rational map from a smooth variety to an abelian variety extends to a morphism [32, th. 3.1],e we e 26 THOMAS KRAMER¨

d get a morphism Y × C → A with image Zi ⊂ A. By the universal property of the Albanese the latter factors as e Y × Cd / Alb(Y × Cd) / A

pr e e Y / Alb(Y ) where pr is the projection.e So we get a morphisme Y → A with image Zi ⊂ A.

5.3. Summands of subvarieties. Now let Z ⊂eA be an irreducible subvariety with trivial stabilizer Stab(Z)= {0} and suppose that there exists a decomposition as a sum Z = X + Y of irreducible geometrically nondegenerate subvarieties of positive dimension. Then by [11, sect. 8.2] the addition morphism X × Y ։ Z is generically ﬁnite and Z is geometrically nondegenerate. By the decomposition theorem [3] we have an inclusion . δZ ֒→ δX ∗ δY ∈ hδXY i where δXY := δX ⊕ δY

So we have a fully faithful embedding as a tensor subcategory hδZ i ֒→ hδXY i stable under subquotients and compatible with our ﬁber functors. Any such embedding gives an epimorphism p : G(δXY ) ։ G(δZ ) [13, cor. 2.9, prop. 2.21], so we may consider VS = ω(δS) ∈ Rep(G(δXY )) for S = X,Y,Z. Lemma 5.3.1. The above three representations remain irreducible when restricted ◦ to the connected component G(δXY ) ⊆ G(δXY ), and replacing X,Y,Z by translates we may assume ◦ GS := G(δS ) is semisimple for S ∈ {X,Y,Z,XY }. Proof. The triviality of the stabilizer Stab(Z) implies that Stab(X) and Stab(Y ) are trivial. Hence VX , VY , VZ restrict to irreducible representations of the connected component of the respective Tannaka groups by [26, cor. 1.4]. All these groups are quotients of G(δXY ), hence the connected component of the latter also acts irreducibly on the three representations.

For semisimplicity, recall that any one-dimensional representation of G(δXY ) is given by a skyscraper sheaf δa on a point a ∈ A(C) [26, sect. 3.c] [46, lemma 13], so det(VX ) = ω(δx) and det(VY ) = ω(δy) for some x, y ∈ A(C). Replacing X, Y ⊂ A by suitable translates we may assume x = y = 0 so that the above two determinants become trivial. But by Schur’s lemma the center Z ⊂ GXY acts via scalars on both VX , VY , and the triviality of the determinant forces these scalars to be roots of unity. Since the action of GXY on VX ⊕ VY is faithful, it follows that Z is ﬁnite and so the group GXY is semisimple. Since GX , GY , GZ are quotients of the group GXY , they are then semisimple as well.

By semisimplicity, the universal cover πS : GS ։ GS of each of the above groups is again an algebraic group. By the correspondence between simply connected e 27 semisimple groups and their Lie algebras there exists a unique embedding i making the diagram ∃!i GZ / GXY ●● ●● ●● p◦πXY e πZ ●● e ●# GZ commute, and we get

Proposition 5.3.2. The subgroup GZ acts nontrivially on both ω(δX ), ω(δY ). Proof. We have a splitting into G = Im(i) and G = ker(p ◦ π ) as indicated e1 2 XY below: ∼ G1 × G2 / GXY / / GXY

e pr1 e p p ∼ G1 / GZ / / GZ So any irreducible representation U ∈ Rep(G ) can be written as U ≃ U ⊠ U e XY 1 2 where the Ui ∈ Rep(Gi) are irreducible representations of the two factors. In our case

VX ≃ VX,1 ⊠ VX,2, VZ ⊆ VX,1 ⊗ VY,1,

VY ≃ VY,1 ⊠ VY,2, 1 ⊆ VX,2 ⊗ VY,2, where 1 ∈ Rep(G2) denotes the trivial representation. Note that the last inclusion is equivalent to VX,2 ≃ Hom(VY,2, 1) because both these representations are irreducible. For suﬃciently divisible n ∈ N and S = X, Y theorem 2.3.1 and remark 2.3.2 yield eﬀective cycles ΛS,i ∈ hcc(δXY )i so that

cc(δnX ) =ΛX,1 ◦ ΛX,2, cc(δnZ ) ⊆ ΛX,1 ◦ ΛY,1,

cc(δnY ) =ΛY,1 ◦ ΛY,2, ΛX,2 = [−1]∗(ΛY,2). Note that these are identities of cycles, including any possible multiplicities. Now take any irreducible subvariety X2 ⊂ A such that ΛX2 ⊆ ΛX,2. By the above equations then

ΛY2 ⊆ ΛY,2 for Y2 := −X2, and looking at weights we ﬁnd irreducible subvarieties X1, Y1 ⊂ A with ΛX1 ⊆ ΛX,1 and ΛY1 ⊆ ΛY,1 such that

(5.3.3) ΛnZ ⊆ ΛX1 ◦ ΛY1 ⊆ ΛX1 ◦ ΛY1 ◦ ΛX2 ◦ ΛY2 where the second inclusion comes from the antidiagonal Λ{0} ⊆ ΛX2 ◦ ΛY2 . Now for S = X, Y we have

(5.3.4) ΛS1 ◦ ΛS2 ⊆ ΛS,1 ◦ ΛS,2 = cc(δnS ). In particular, any component of the left hand side maps to a subvariety of A of dimension ≤ dim(S). By (5.3.3) it follows that equality must hold for at least one such component, and then by looking at the right hand side of (5.3.4) it follows 28 THOMAS KRAMER¨ that this component must be equal to ΛnS. Since this component enters with multiplicity one, lemma 5.2.2 implies

dim(Si) ≤ dim(S) for i = 1, 2. Taking i = 1 and recalling that dim(X) + dim(Y ) = dim(Z) < g, we obtain from the ﬁrst inclusion in (5.3.3) that

nZ = X1 + Y1 and dim X1 = dim X, dim Y1 = dim Y.

In particular, for both S = X, Y we have dim S1 > 0, hence deg(ΛS1 ) > 1 and then a fortiori

dim VS,1 = deg(ΛS,1) ≥ deg(ΛS1 ) > 1.

Since the VS,1 are irreducible representations, it follows that they are nontrivial.

The above argument should give more: It seems likely that p : GXY ։ GZ is an isogeny. For this we would like to argue that dim(X2) = 0 by showing that one has an inclusion ΛX1+X2 ⊆ ΛX1 ◦ ΛX2 . The problem is that a priori ΛX1+X2 could be negligible and in this case it is not taken into account by the clean product ◦ from example 1.3.2. However, in favorable cases this issue does not occur:

Corollary 5.3.5. With notations as above, if the Gauss map γ PΛ : PΛ → PV for the projectivization of the characteristic variety Λ = Supp(cc(δZ )) is a ﬁnite morphism, then the epimorphism p : GXY ։ GZ is an isogeny. Proof. This is shown in [27] by computing Chern-Mather classes via lemma 3.3.1; >g−1 here the ﬁniteness of the Gauss map means cc(δZ ) ∈ L (A).

5.4. The adjoint representation. Any linear algebraic group has a distinguished representation, the adjoint representation on its Lie algebra. If Z ⊂ A is a closed subvariety, let AdZ ∈ Hol(DA) be the unique clean module which corresponds to the adjoint representation of G(δZ ). Combining the argument of [26, prop. 2.3] with the above we get

Theorem 5.4.1. If Z ⊂ A is irreducible with Stab(Z)= {0}, then any nontrivial decomposition into a sum of geometrically nondegenerate subvarieties Z = X + Y with dZ = dX + dY satisfies 1 . min{dX , dY } ≥ δ = 2 min dim Supp(M ) | M ֒→ AdZ , dim Supp(M ) > 0 Proof. If G is a connected semisimple group that is simple modulo its center, then for any nontrivial U ∈ Rep(G) the homomorphism G → Gl(U) has finite kernel and so Lie(G) ⊆ U ⊗ U ∗ = End(U). If G is not simple modulo its center, the same argument still shows that U ⊗ U ∗ contains an irreducible summand of the adjoint representation. We apply this as follows: Up to an isogeny on the abelian variety we may assume all occuring Tannaka groups to be connected, so that the representations of their universal cover determine the corresponding clean DA-modules. For S = X, Y the proposition says that GZ ×{1} acts nontrivially on VS . So by the above remarks on the adjoint representation there must be a nontrivial submodule M ⊆ AdZe with M ⊆ δS ∗ δ−S. Since for a connected semisimple 29 group the adjoint representation has no one-dimensional summands, we must have dim Supp(M ) > 0 and the claim follows because Supp(M ) ⊆ S − S. If we fix a maximal torus, the nontrivial weights in the adjoint representation of a connected reductive group are by definition the roots of the group. The following criterion helps to estimate the support of AdZ :

Lemma 5.4.2. Let Z ⊂ A be irreducible with Stab(Z)= {0}. If the weights of VZ with respect to a maximal torus contain some nonzero rational multiple of a root of GZ , then with notations as in theorem 2.2.3

dim Supp(AdZ ) ≥ dim W for any ΛW ⊂ h∗ cc(δZ ) in the image of the Weyl group orbit of this weight.

Proof. The assumptions imply [a]∗ΛW ⊆ [b]∗ cc(AdZ ) for some a,b ∈ N. For ample divisors with at most rational singularities we know that any summand is geometrically nondegenerate [40, th. 1], hence the above in particular applies to summands of theta divisors on indecomposable ppav’s. The bound on the dimension of such summands is sharp whenever such summands are known to exist:

• If A is the Jacobian of a smooth projective curve C, then AdΘ = δC−C .

• If A is the intermediate Jacobian of a smooth cubic threefold, AdΘ = δΘ. By the conjecture of Pareschi and Popa these should be the only theta divisors with summands of positive dimension. For curve summands this has been established in [39]. In higher dimension it remains open, but the above rules out various cases such as the following:

Theorem 5.4.3. Let (A, Θ) ∈ Ag(C). If G(δΘ) is a symplectic group and ω(δΘ) is its standard representation, then we have dim Supp(AdΘ) ≥ g − 1 and hence the theta divisor Θ ⊂ A cannot be written as a sum of positive-dimensional subvarieties. Proof. For symplectic groups the adjoint representation is the symmetric square of the standard representation. So twice any weight of the standard representation is a root, and the previous lemma shows dim Supp(AdΘ) ≥ g − 1. Thus any summand of the theta divisor must have dimension either zero or ≥ (g − 1)/2, which concludes the proof since 2 | g if ωu(δΘ) is symplectic.

6. A criterion for almost simplicity In order to control the adjoint modules from above and to compute the arising groups in more general cases, one needs a way to decide whether their Lie algebras are simple. We now give a suﬃcient criterion using characteristic cycles. 6.1. Decomposable characteristic cycles. We want to show that under the dictionary between weights and characteristic cycles in theorem 2.2.3, any nontrivial decomposition of the Lie algebra of the Tannakian group induces a decomposition on characteristic cycles. Let deg : L (A) → Z be the homomorphism sending a ∗ clean subvariety Λ ⊂ T A to the degree deg(Λ) = deg(γΛ) of its Gauss map.

Proposition 6.1.1. Let G = G(M ) for a simple holonomic module M ∈ Hol(DA) with Stab(M ) = {0}. If the Lie algebra Lie(G) is not simple modulo its center, then [n]∗ cc(M )= Λ1 ◦ Λ2 30 THOMAS KRAMER¨ for some n ∈ N and effective clean cycles Λi ∈ hcc(M )i of degree deg(Λi) > 1. Proof. Under the assumptions of the proposition, the connected component G◦ is not simple modulo its center Z = Z(G◦), and by the theory of connected reductive groups we can find nontrivial connected semisimple groups G1, G2 6= {1} and an epimorphism ◦ G = G1 × G2 × Z ։ G whose kernel is a finite central subgroup. As Stab(M )= {0}, we know that ωu(M ) restricts to an irreducible representationb of the connected component [26, cor. 1.4], so U = ωu(M )|Ge is an irreducible representation of the covering group. As such it splits as a product U ≃ U1 ⊠ U2 of irreducible U1 ∈ Rep(G1),U2 ∈ Rep(G2 × Z), and the claim then follows from theorem 2.3.1 and remark 2.3.2. 6.2. A criterion for divisors. Now let Z ⊂ A be a divisor and p : Z → Z a resolution of singularities. The universal property of the Albanese variety gives a uniquep ˆ making the following diagram commute: e

alb Z / Alb(Z)

p e ∃e!p ˆ Z / A We wantp ˆ to be an isomorphism. As the Albanese variety is a birational invariant, this condition does not depend on the chosen resolution and we abbreviate it by writing A = Alb(Z). The following criterion applies in particular to M = δZ in many cases: Theorem 6.2.1. Let Z ⊂ A be a geometrically nondegenerate irreducible reduced symmetric divisor such that Stab(Z)= {0} and Alb(Z)= A. Let M ∈ Hol(DA) be simple and

cc(M )= ΛZ + cW · ΛW with cW ∈ N0. dW 3 deg(cc(M )). (2) If nW 6=0 at most for dim(W )=0.

(3) If γZ : PΛZ → PV is ﬁnite and [2m]∗ cc(M ) 6=ΛmZ ◦ ΛmZ for all m ∈ N.

(4) If [n]∗ cc(M ) is reduced for all n ∈ N.

Proof. If the connected component is not simple modulo its center, pick n ∈ N and Λi ∈ L (A) with deg(Λi) > 1 as in proposition 6.1.1. Since Stab(Z)= {0}, the map [n]: Z −→ A, z 7→ nz is birational onto its image. So our assumption implies that the conormal variety to this image Y = nZ enters the clean conic Lagrangian cycle [n]∗ cc(M )=Λ1 ◦ Λ2 with multiplicity one, and all other components of this cycle are conormal varieties to subvarieties of codimension > 1 in A. Let ΛZi ⊂ Supp(Λi) be the unique irreducible components with

ΛY ⊆ Supp(ΛZ1 ◦ ΛZ2 ), 31

′ ′ and write Λi =ΛZi +Λi with eﬀective cycles Λi ∈ L (A), possibly zero. Then we have

(6.2.2) ΛZ1 ◦ ΛZ2 = ΛY + ··· ′ ′ (6.2.3) (ΛZ1 +Λ1) ◦ (ΛZ2 +Λ2) = ΛY + ··· ′ ′ (6.2.4) deg(ΛZ1 +Λ1) · deg(ΛZ2 +Λ2) = deg(cc(M )) where ··· stands for a nonnegative linear combination of conormal varieties to subvarieties of codimension > 1 in A. We now argue by contradiction.

First we claim that dim(Zi) > 0 for both i =1, 2. If not, then by symmetry we can assume Z2 = {p} is reduced to a single point. Then Z1 = Y − p by (6.2.2), ′ and by (6.2.3) it follows that Supp(Λ2) cannot contain the conormal variety to any point on the abelian variety. Since conormal varieties to points are the only ones ′ whose Gauss map has degree one, we get deg(Λ2) ≥ 2. But deg(ΛY −p) = deg(ΛZ ), so case (1) is excluded by (6.2.4). In case (2) any irreducible component of the ′ convolution ΛY −p ◦ Λ2 6= 0 would be supported over a point. By the last part of ′ lemma 5.2.1 and the symmetry of Y ⊂ A then ΛY +q ⊆ Λ2 for some q ∈ A(C), but then the diagonal in the ﬁbered product would give an irreducible component of ′ the form Λ2Y +q−p ⊆ Supp(ΛY −p ◦Λ2) contradicting (6.2.2)-(6.2.3). Case (3) would result in [Y ] = cM,g−1(Λ1 ◦ Λ2) ≥ deg(Λ2) · [Y ] > [Y ] by (6.2.3) and lemma 3.3.1. In the remaining case (4) the right hand side of (6.2.3) ′ is a reduced cycle, hence so is the convolution product ΛZ1 ◦ Λ2 6= 0. Since Z1 ⊂ A ′ is a divisor, lemma 5.2.2 then forces ΛZ1 ◦ Λ2 to be a sum of conormal varieties to divisors on A. But this is not the case for any of the summands ··· in (6.2.3).

So dim(Zi) ≥ 1 for both i =1, 2. By (6.2.2) and lemma 5.2.2 we have dominant rational maps [n] fi : Z ։ Y 99K Zi for i = 1, 2. By the extension property of rational maps from smooth projective varieties to abelian varieties [32, th. 3.1], these maps can be extended to surjective morphisms fi : Z ։ Zi ⊂ A where Z ։ Z is a resolution of singularities. Our assumption A = Alb(Z) then implies that the original rationale mapse fi are defined everywhere and fit into a commutativee diagram fi Z / Z / Zi _ _ alb e ∃!gi Alb(Z) A / A where gi comes from the universal property of the Albanese variety. Replacing Zi by a translate we may assume Zi is symmetric and gi is a homomorphism. Then the image Bi = gi(A) ⊆ A is an abelian subvariety. On the other hand Zi ⊂ A cannot be a translate of an abelian subvariety, indeed deg(ΛZi ) > 0. So the geometric nondegeneracy of Z and the above diagram imply that the map fi : Z ։ Zi is generically finite, and gi : A ։ Bi = A must then be an isogeny for dimension reasons. Since Stab(Z)= {0}, it follows that fi : Z → Zi is birational. 32 THOMAS KRAMER¨

Note that since we have adjusted our translates such that the gi : A → A are homomorphisms, the symmetry of Z implies that Zi is symmetric. On the other hand fi 6= −fi since Zi 6⊂ A[2]. It follows that in the ﬁber product PΛZ1 ×PV PΛZ2 we have components ± ∆ = {(f1(z), ±f2(z), ξ) ∈ A × A × PV | (z, ξ) ∈ ΛZ } and these two components are distinct since the fi are birational. Both components are dominant over PV and so their image under the addition morphism ̟ gives two components ± ± ΛY ⊂ ΛZ1 ◦ ΛZ2 where Y = (g1 ± g2)(Z). + Here Y = Y because g1 +g2 = [n]. Looking at the multiplicities of the components − − in equation (6.2.2) we then obtain dim Y < dim Z. So g1 − g2 : Z ։ Y is not generically ﬁnite, and since Z is geometrically nondegenerate, it follows that the − − image Y = (g1 −g2)(A) is an abelian subvariety. But deg(ΛY − ) > 0, so Y = {0} must be a single point. Thus g1 = g2. Since f1 + f2 = [n], it follows by looking at the induced map on the universal cover of the abelian variety that n =2m is even and g1 = g2 = [m], so

Z1 = Z2 = mZ. To show that this is impossible in each of the four cases listed in the theorem, recall that ′ ′ [n]∗ cc(M ) =Λ1 ◦ Λ2 = (ΛZ1 +Λ1) ◦ (ΛZ2 +Λ2).

So case (4) is excluded by observing that this cycle contains the component Λ{0} − ։ with multiplicity at least deg(̟ : ∆ Λ{0}) = deg(ΛZ ) ≥ 2. The remaining three cases can be excluded by a computation of Chern-Mather classes: In degree zero we 2 get deg(ΛZ ) ≤ deg(cc(M )), so case (1) could only happen if deg(cc(M )) ≤ 3 and this is irrelevant for us since any reductive Lie algebra with a faithful representation of dimension at most three is simple modulo its center. In degree one lemma 3.3.1 implies 2 2 2m deg(ΛZ )cM,1(ΛZ ) ≤ n cM,1(cc(M )), where the inequality means that the right hand is equal to the left hand side plus an effective cycle. In case (2) we would have cM,1(cc(M )) = cM,1(ΛZ ) and it would follow from the above inequality that deg(ΛZ ) = 2. Then the Gauss map for the symmetric divisor Z ⊂ A would induce a birational map Z/h±id Ai 99K PV which is impossible for the same reason as in [8, lemma 3.1]: Ran’s definition of geometric nondegeneracy in [36, p. 466] implies that on any geometrically nondegenerate divisor Z ⊂ A there are non-zero holomorphic 2-forms invariant under −id A. So it only remains to exclude the case (3). Taking Chern-Mather classes in degree g − 1 in (6.2.3) we get ′ ′ ′ ′ cM,g−1 ΛmZ ◦ Λ2 +Λ1 ◦ ΛmZ +Λ1 ◦ Λ2 = 0. In case (3) the finiteness of the Gauss map allows to use lemma 3.3.1. Effective clean cycles have effective Chern-Mather classes and the Pontryagin product preserves ′ ′ effectivity, so we get deg(Λ1) · cM,g−1(ΛmZ ) = deg(Λ2) · cM,g−1(ΛmZ ) =0. This ′ ′ implies Λ1 =Λ2 = 0 because we are dealing with clean effective cycles. Altogether it would follow that [n]∗ cc(M ) =ΛmZ ◦ ΛmZ which we had excluded in (3). 33

Corollary 6.2.5. For M as in theorem 6.2.1 with [n]∗ cc(M ) reduced for all n ∈ N, we have ◦ G ⊆ G(M ) ⊆ Gm · G ⊆ Gl(W ), where (G, W ) is one of the pairs that are listed in tables 2, 3 of the appendix. Proof. By theorem 6.2.1 the connected component G(M )◦ is simple modulo its center, and since M ∈ Hol(DA) is simple with trivial stabilizer Stab(M ) = {0}, this connected component acts irreducibly on W = ω(M ). So by Schur’s lemma its ◦ center acts via a scalar, which means that we have inclusions G ⊆ G(M ) ⊆ Gm ·G where G is the derived group of the connected component. Our assumption on the reducedness of [n]∗ cc(M ) for all n ∈ N implies by theorem 2.2.3 that W is weight multiplicity free, so the claim follows from lemma 6.5.1. 6.3. Application to theta divisors. By [14] theta divisors on indecomposable ppav’s are normal and irreducible so that the Albanese assumption from above is satisﬁed:

Lemma 6.3.1. For any indecomposable ppav (A, Θ) ∈ Ag(C) of dimension g > 1, the theta divisor satisﬁes A = Alb(Θ). Proof. Let p : Θ → Θ be a resolution of singularities, and put B = Alb(Θ). We have a diagram e alb e Θ / B

p e ∃!p ˆ Θ / A wherep ˆ is an isogeny [14, rem. 3.4]. As p is birational, we know D = alb(Θ) ⊂ B is an irreducible divisor which is not stable under any translation: If e ∈ Stab(D), thenp ˆ(e) ∈ Stab(Θ) = {0},sop ˆ(d)=ˆp(d + e) for all d ∈ D(C); since d, d +e e ∈ D, the birationality of alb pˆ p : Θ / / D / / Θ then forces e = 0. Hence Stab(D)= {0} as claimed. Writing the preimage of Θ as a union e pˆ−1(Θ) = (D + e), e∈ker(ˆ[ p) we get that on the right hand side the irreducible divisors D + e ⊂ B are pairwise distinct. All these divisors are ample since they are all numerically equivalent and their unionp ˆ−1(Θ) ⊂ B is an ample divisor. Hence if there exists e ∈ ker(ˆp) \{0}, it follows that ∅ 6= D ∩ (D + e) ⊂ D is nonempty of codimension one, in which case the preimagep ˆ−1(Θ) will be singular in codimension one. But this is impossible sincep ˆ is an isogeny so that the preimage of any normal divisor must be normal.

We can thus apply theorem 6.2.1 to the module M = δΘ ∈ Hol(DA). Notice that while the polarization determines the theta divisor only up to a translate, the connected semisimple group ◦ ◦ GΘ = [G(δΘ) , G(δΘ) ] is independent of the chosen translate due to lemma 4.3.2. Thus we get: 34 THOMAS KRAMER¨

Theorem 6.3.2. Let (A, Θ) ∈ Ag(C) be a ppav such that [n]∗ cc(δΘ) is reduced for all n ∈ N. Then the connected semisimple group GΘ and its irreducible faithful representation W = ω(δΘ) appear in table 2 or 3. Proof. Up to a translation we may assume that the theta divisor is symmetric, so that the representation ω(δΘ) is isomorphic to its dual. Since the restriction of this ◦ representation to the connected component G(δΘ) remains irreducible, it follows from Schur’s lemma that this connected component must be semisimple and hence equal to GΘ. So the claim follows from corollary 6.2.5. This contains theorem 4 from the introduction: If Θ is smooth except for k ordinary double points, then the classical Gauss map has degree deg(ΛΘ)= g! − 2k by [8]; on the other hand dim(ω(δΘ)) = deg(cc(δΘ)) is always the degree of the characteristic cycle, and the difference between the two is given in lemma 6.5.2. 6.4. Stratifications of moduli spaces. It follows from [30, prop. 7.4] that the assignment (A, Θ) 7→ (G(δΘ),ω(δΘ)) defines a constructible stratification of the moduli space of ppav’s. This stratification is related to the stratification by the ind degree of the Gauss map in [8] but includes finer information. Let Ag ⊆ Ag be the locus of indecomposable ppav’s. For g = 4 we have 10 ind sm nh h k A4 = A4 ⊔ J4 ⊔ J4 ⊔ Θnull,4 k G=1 where

sm • A4 is the locus of ppav’s with a smooth theta divisor, nh • J4 is the locus of Jacobians of nonhyperelliptic curves, h • J4 is the locus of Jacobians of hyperelliptic curves, k • Θnull,4 is the locus of non-Jacobians with precisely k vanishing thetanulls. Indeed, the only singularities of indecomposable theta divisors on non-Jacobian abelian fourfolds are ordinary double points given by vanishing thetanulls [35]; the maximum number of k = 10 such vanishing theta nulls is obtained for a unique abelian fourfold discovered by Varley [43] [10]. Table 1 lists

(1) the degree deg(ΛΘ) of the classical Gauss map,

(2) the degree deg(cc(δΘ)) = dim(ω(δΘ)) of the characteristic cycle,

(3) the representation ω(δΘ) of the reductive group G(δΘ), where for the latter we denote by ̟1,̟2,... the fundamental representations.

deg(γΘ) dim(ω(δΘ)) ω(δΘ) G(δΘ) sm A4 24 24 ̟1 Sp24(C) nh J4 20 20 ̟3 Sl 6(C)/µ3 h J4 8 14 ̟3 Sp6(C) k Θnull,4 24 − 2k 24 − 2k ̟1 Sp24−2k(C) (assuming γΘ is ﬁnite if k = 4) Table 1. Invariants of principally polarized abelian fourfolds 35

Note that for k = 2 neither the degree of the classical Gauss map nor the one of the characteristic cycle suﬃces to characterize Jacobians among all ppav’s, but the pair (G(δΘ),ω(δΘ)) does. Let us brieﬂy check the new part of the table: k Lemma 6.4.1. For k 6= 2, theorem 4 implies that every ppav (A, Θ) ∈ Θnull,4(C) has

G(δΘ) ≃ Sp24−2k(C) and ω(δΘ) ≃ ̟1.

For k =2 this remains true at least if the Gauss map γΘ : PΛΘ → PV is finite. Proof. For k 6= 2 this is clear by direct inspection. For k = 2 the only possible alternative would be that G(δΘ) ≃ Sl 6(C)/µ3, ω(δΘ) ≃ ̟3. In this case theorem 4.1.2 shows 3 [3]∗ cc(δΘ) = Alt (Λ) where Λ ∈ L (A) is effective up to isogeny. If the Gauss map is finite, a computation of Chern-Mather classes via lemma 3.3.1 gives 6 for i =0, 2 3 cM,i(Λ) =  3 [Θ] for i =1, 0 for i> 1, by example 1.4.2 and corollary 3.2.1. For any n ∈ N such that the cycle [n] (Λ) is  ∗ effective, it then follows from lemma 3.1.2 that this cycle must be supported over a curve Z ⊂ A. Then the image of the theta divisor under the isogeny [3n] is a sum of copies of this curve. Taking preimages under this isogeny we get that the theta divisor is a sum of curves, so (A, Θ) is a Jacobian by [39]. Finally, let us give an example to illustrate how the dictionary between weights and characteristic cycles can be used beyond theorem 6.3.2. We only give the simplest case but the method obviously generalizes:

Lemma 6.4.2. If (A, Θ) ∈ A5(C) is a ppav with a symmetric theta divisor that is smooth except for two distinct ordinary double points e1,e2, then one of the following two cases occurs:

(1) Either e1,e2 ∈ A[2], in which case G(δΘ) ≃ O118(C).

(2) Or e1 = −e2 ∈/ A[2], in which case G(δΘ) ≃ SO118(C). Proof. Since the theta divisor is symmetric and the dimension g = 5 is odd, the representation ω(δΘ) is orthogonal. By lemma 6.5.2 the characteristic cycle is given by

cc(δΘ) =ΛΘ +Λe1 +Λe2 . If the weights of the representation form a single Weyl group orbit, then we are done by lemma 6.5.1 because the distinction between the orthogonal and special orthogonal group can be read oﬀ from det(δΘ) = δe1+e2 . So it only remains to exclude the case that there are more than one Weyl group orbits. Since one of them must be of size deg(ΛΘ), the remaining orbits would have to be of size at most two. But for any simple complex Lie algebra, any nontrivial Weyl group orbit has size at least the rank of the Lie algebra, which in our case must be bigger than two as otherwise the ﬁrst orbit could not be so big. So all nontrivial weights are in the same Weyl group orbit, i.e. the representation is quasi-minuscule. But these are all known and there is no such of dimension deg(cc(δΘ)) = 118. 36 THOMAS KRAMER¨

6.5. Appendix: Multiplicity free representations. Although the dictionary between Weyl group orbits and characteristic cycles works in general, the simplest situation is when M is essentially multiplicity free in the sense that the clean cycle [n]∗(cc(M )) is reduced for all n ∈ N, see [26]. Then any Weyl group orbit of weights that enters the corresponding representation must do so with multiplicity one, i.e. ωu(M ) is a weight multiplicity free representation. For the simple Dynkin types there are only very few such representations, most of them are minuscule in the sense that their weights form a single orbit under the Weyl group: Lemma 6.5.1. If G is a connected semisimple group which is simple modulo its center and acts faithfully on some weight multiplicity free irreducible W ∈ Rep(G), then (G, W ) appears on table 2 or 3 below. Proof. The weight multiplicity free irreducible representations of the simple Lie algebras are classiﬁed in [22, th. 4.6.3] [42, sect. A], and the tables list the images of the corresponding simply connected groups under these representations. Whether an irreducible representation is orthogonal, symplectic or not self-dual can be read oﬀ from its highest weight [19, sect. 3.2.4 and exercise 9].

G Sl n/µ(k,n) Spin2n+1 Sp2n SO 2n Spin2n E6 E7 n n n−1 dim W k for 1 ≤ k ≤ n 2 2n 2n 2 27 56 symplectic? n = 2k∈ / 4Z n ≡ 1, 2(4) yes no n ≡ 2(4) no yes orthogonal? n = 2k ∈ 4Z n ≡ 0, 3(4) no yes n ≡ 0(4) no no

Table 2. The minuscule representations W and their images G ⊂ Gl(W ) for the simple Dynkin types. All these are fundamental representations. In the last two rows we list whether they are symplectic, orthogonal or not self-dual.

G Sl n/µ(k,n) SO 2n+1 Sp6 G2 n+k−1 dim W k for k> 1 2n + 1 14 7 symplectic? no no yes no orthogonal? no yes no yes

Table 3. The weight multiplicity free nonminuscule representations W and their images G ⊂ Gl(W ) for the simple Dynkin types. All these are fundamental k n representations except for the symmetric powers W = Sym (C ) ∈ Rep(Sl n).

Lemma 6.5.2. If Z ⊂ A is a reduced divisor whose singular locus consists of ﬁnitely many ordinary double points ei, then

ΛZ if g is even, CC(δZ ) = (ΛZ + i Λei if g is odd. So δZ is essentially multiplicity free iﬀ P • Stab(Z)= {0}, and

• no two of the ei diﬀer by a torsion point if 2 ∤ g. 37

Proof. The characteristic cycle can be computed locally in the classical topology, so we can apply the result of [33, th. 4.2]: If X ⊆ Cg is an open neighborhood of the origin and f : X → C is a holomorphic function such that Z = f −1(0) is smooth outside the origin, then CC(δZ )=ΛZ + (µ0 − N1) · Λ0 where µ0 is the Milnor number of a generic hyperplane section of Z through the origin and where N1 denotes the number of Jordan blocks for the eigenvalue one in the Milnor g−1 −1 monodromy on H (f (t), C) for small t 6= 0. For ordinary double points µ0 = 1, and the local Picard-Lefschetz formula says that the Milnor monodromy acts on the one-dimensional space Hg−1(f −1(t), C) by (−1)g.

Acknowledgements. I would like to thank Giulio Codogni for valuable comments on a previous version of the paper, and Michael Dettweiler, Gavril Farkas, Javier Fres´an, Bert van Geemen, Bruno Klingler, Mihnea Popa and Stefan Schreieder for stimulating discussions on various related topics.

References 1. Andreotti, A. and Mayer, A. L., On period relations for abelian integrals on algebraic curves, Ann. Sc. Norm. Super. Pisa Cl. Sci. 21 (1967), 189–238. 22 2. Atiyah, M. F., and Tall, D. O., Group representations, λ-rings and the j-homomorphism, Topology 8 (1969), 253–297. 9 3. Beilinson, A., Bernstein, J. and Deligne, P., Faisceaux Pervers, Astérisque 100 (1982). 2, 26 4. Birkenhake, C. and Lange, H., Complex abelian varieties (second ed.), Grundlehren Math. Wiss., vol. 302, Springer Verlag, 2004. 18 5. Bröcker, T. and tom Dieck, T., Representations of Compact Lie Groups, Graduate Texts in Math., vol. 98, Springer, 1985. 9 6. Cappell, S. E., Maxim, L., Schürmann, J., Shaneson, J. L. and Yokura, S., Characteristic classes of symmetric products of complex quasi-projective varieties, J. Reine Angew. Math. 728 (2017), 35–63. 5 7. Casalaina-Martin, S., Popa, M. and Schreieder, S., Generic vanishing and minimal cohomology classes on abelian fivefolds, J. Alg. Geom. 27 (2018), 553–581. 2 8. Codogni, G., Grushevsky, S. and Sernesi, E., The degree of the Gauss map of the theta divisor, Algebra & Number Theory 11 (2017), 983–1001. 4, 6, 18, 32, 34 9. De Cataldo, M. and Migliorini, L., The decomposition theorem, perverse sheaves and the topology of algebraic maps, Bull. Amer. Math. Soc. 46 (2009), 535–633. 2 10. Debarre, O., Annulation de thêtaconstantes sur les variétés abéliennes de dimension quatre, C. R. Acad. Sci. Paris Sér. I Math. 305 (1987), 885–888. 34 11. , Complex tori and abelian varieties, SMF/AMS Texts and Monographs, vol. 11, 2005. 26 12. Deligne, P., Catégories tensorielles, Moscow Math. J. 2 (2002), 227–248. 11 13. Deligne, P. and Milne, J. S., Tannakian categories, Hodge Cycles, Motives, and Shimura varieties, Lecture Notes in Math., vol. 900, Springer Verlag, 1982, pp. 101–228. 13, 26 14. Ein, L. and Lazarsfeld, R., Singularities of theta divisors and the birational geometry of irregular varieties, Journal of the AMS 10, 243–258. 33 15. Flenner, H., Join varieties and intersection theory, Recent progress in intersection theory (Eddingsrud, G., Fulton, W. and Vistoli, A., ed.), Trends in Math., Birkhäuser, 2000, pp. 129– 198. 6, 18 16. Flenner, H., O’Carroll, L. and Vogel, W., Joins and intersections, Springer Monographs in Math., Springer, 1999. 6, 18 17. Franecki, J. and Kapranov, M., The Gauss map and a noncompact Riemann-Roch formula for constructible sheaves on semiabelian varieties, Duke Math. J. 104 (2000), 171–180. 7 18. Fulton, W., Intersection theory (second ed.), Springer Verlag, 1998. 20 19. Goodman, R. and Wallach, N. R., Symmetry, representations and invariants, Springer Verlag, 2009. 36 38 THOMAS KRAMER¨

20. Heinloth, F., A note on functional equations for zeta functions with values in Chow motives, Ann. Inst. Fourier Grenoble 57 (2007), 1927–45. 8 21. Hotta, R., Takeuchi, K. and Tanisaki, T., D-modules, perverse sheaves and representation theory, Progress in Math., vol. 236, Birkhäuser, 2008. 5 22. Howe, R., Perspectives on invariant theory: Schur duality, multiplicity-free actions and beyond., Israel Math. Conf. Proc., vol. 8, 1995, The Schur lectures (1992), pp. 1–182. 36 23. Kennedy, G., MacPherson’s chern classes of singular algebraic varieties, Communications in Alg. 18 (1990), 2821–2839. 17 24. Kleiman, S. L., The transversality of a general translate, Compos. Math. 28 (1974), 287–297. 17 25. Knutson, D., λ-rings and the representation theory of the symmetric group, Lecture Notes in Math., vol. 308, Springer Verlag, 1973. 11 26. Krämer, T., Characteristic cycles and the microlocal geometry of the Gauss map I, arXiv:1604.02389. 1, 2, 5, 7, 12, 13, 14, 21, 22, 23, 26, 28, 30, 36 27. Krämer, T., Summands of theta divisors on Jacobians, arXiv:1811.01072. 2, 28 28. , Cubic threefolds, Fano surfaces and the monodromy of the Gauss map, Manuscripta Math. 149 (2016), 303–314. 4 29. Krämer, T. and Weissauer, R., Semisimple super Tannakian categories with a small tensor generator, Pacific J. Math. 276 (2015), 229–248. 21 30. , The Tannaka group of the theta divisor on a generic principally polarized abelian variety, Math. Z. 281 (2015), 723–745. 3, 22, 34 31. , Vanishing theorems for constructible sheaves on abelian varieties, J. Alg. Geom. 24 (2015), 531–568. 12, 13, 14 32. Milne, J. S., Abelian varieties, Arithmetic geometry (Cornell, G. and Silverman, J. H., ed.), Springer Verlag, 1986. 25, 31 33. Nang, P. and Takeuchi, K., Characteristic cycles of perverse sheaves and Milnor fibers, Math. Z. 249 (2005), 493–511. 37 34. Pareschi, G. and Popa, M., Generic vanishing and minimal cohomology classes on abelian varieties, Math. Ann. 340 (2008), 209–222. 2 35. R. Smith and R. Varley, Components of the locus of singular theta divisors of genus 5, Algebraic Geometry Sitges (Barcelona), Lecture Notes in Math., vol. 1124, Springer Verlag, 1985, pp. 338–416. 34 36. Ran, Z., On subvarieties of abelian varieties, Invent. Math. 62, 459–480. 24, 32 37. Sabbah, C., Quelques remarques sur la geometrie des espaces conormaux, Astérisque 130 (1985), 161–192. 17 38. Schnell, C., Holonomic D-modules on abelian varieties, Publ. Math. Inst. Hautes Etudes´ Sci. 121 (2015), 1–55. 12 39. Schreieder, S., Theta divisors with curve summands and the Schottky problem, Math. Annalen 365 (2016), 1017–1039. 2, 22, 29, 35 40. , Decomposable theta divisors and generic vanishing, Int. Math. Res. Notices 16 (2017), 4984–5009. 2, 29 41. Schürmann, J. and Tib˘ar, M., Index formula for Macpherson cycles on affine algebraic varieties, Tohoku Math. J. 62 (2010), 29–44. 17 42. Stembridge, J. R., Multiplicity-free products and restrictions of Weyl characters, Representa- tion Theory 7 (2003), 404–439. 36 43. Varley, R., Weddle’s surfaces, Humbert’s curves, and a certain 4-dimensional abelian variety, Amer. J. Math. 108 (1986), 931–952. 34 44. Vogel, W., Lectures on results on Bezout’s theorem, Tata Institute of Fundamental Research Lectures on Mathematics and Physics, vol. 74, Springer, 1984. 6, 18 45. Weissauer, R., Brill-Noether sheaves, arXiv:math/0610923. 21 46. , Degenerate perverse sheaves on abelian varieties, arXiv:1204.2247. 26 47. , On subvarieties of abelian varieties with degenerate Gauss mapping, arXiv:1110.0095. 24

Institut fur¨ Mathematik, Humboldt-Universitat¨ zu Berlin Unter den Linden 6, 10099 Berlin (Germany) E-mail address: [email protected]