Semicontinuity of Gauss maps and the Schottky problem

We show that the degree of Gauss maps on abelian varieties is semicontinuous in families, and we study its jump loci. As an application we obtain that in the case of theta divisors this degree answers the Schottky problem. Our proof computes the degree of Gauss maps by specialization of Lagrangian cycles on the cotangent bundle. We also get similar results for the intersection cohomology of varieties with a finite morphism to an abelian variety; it follows that many components of Andreotti-Mayer loci, including the Schottky locus, are part of the stratification of the moduli space of ppav's defined by the topological type of the theta divisor.


Introduction
The Gauss map of a hypersurface in projective space is the rational map that sends any smooth point of the hypersurface to its normal direction in the dual projective space. The analogous notion of Gauss maps for subvarieties of abelian varieties appears already in Andreotti's proof of the Torelli theorem [2]. In contrast to the case of projective hypersurfaces, the Gauss map for any ample divisor on an abelian variety is generically finite of degree > 1, and its degree is related to the singularities of the divisor. We show that this degree is lower semicontinuous in families, and we study its jump loci. As an application we get that in the moduli space of principally polarized abelian varieties, the degree of the Gauss map refines the Andreotti-Mayer stratification and answers the Schottky problem as conjectured in [11]. We work over an algebraically closed field k with char(k) = 0. In section 7 we obtain similar results for the intersection cohomology of complex varieties with a finite morphism to an abelian variety. In particular, many Andreotti-Mayer loci such as the Schottky locus are determined over the complex numbers already by the topological type of the theta divisor.

1.A. Gauss maps and their jump loci. Let
A be an abelian variety over k. By translations we may identify its tangent spaces at all points, hence the cotangent bundle T ∨ A = A × V is trivial with fiber V = H 0 (A, Ω 1 A ). The Gauss map of a reduced effective divisor D ⊂ A is the rational map that sends a smooth point of the divisor to its conormal direction at that point; it coincides with the rational map given by the linear series PV ∨ = |O D (D)|. For an irreducible divisor this is a generically finite dominant map iff the divisor is ample, which happens iff the divisor is not stable under translations by any positive dimensional abelian subvariety [32,cor. II.11,lem. II.9]. Even in the generically finite case the Gauss map can have positive dimensional fibers [4].
For algebraic families of generically finite maps the generic degree always defines a constructible stratification of the parameter space, but in general it can jump in both directions (see example 4.2). Our first semicontinuity result says that for Gauss maps on abelian varieties this does not happen: Theorem 1.1. Let A → S be an abelian scheme over a variety S, and let D ⊂ A be a relatively ample divisor which is flat over S. Let D s ⊂ A s denote their fibers over s ∈ S, and let γ Ds be the corresponding Gauss map. Then for each d ∈ N the subsets S d = s ∈ S | deg(γ Ds ) ≤ d ⊆ S are closed in the Zariski topology.
The above result does not show where the degree actually jumps. Let us say that an irreducible subvariety of an abelian variety is negligible if it is stable under translations by a positive dimensional abelian subvariety. Simple abelian varieties have no negligible subvarieties other than themselves. More generally, by [1, th. 3] an irreducible closed subvariety of an abelian variety is negligible iff it is not of general type. Our second result says that in the setting of theorem 1.1 the degree of the Gauss map jumps whenever a new component of general type appears in the singular locus Sing(D s ). To make this precise we specify a curve along which we move inside the parameter space: The above in particular applies if all components of the singular locus are of general type and dim(Sing(D 0 )) > dim(Sing(D s )) for all s = 0. This last condition is motivated by the case of theta divisors and the Schottky problem.
1.B. Application to the Schottky problem. Let A g be the moduli space of principally polarized abelian varieties of dimension g. Inside it, consider for d ∈ N the Gauss loci [11, sect. 4] The above results show that these loci are closed (cor. 6.1) and refine the Andreotti-Mayer stratification (cor. 6.3). Thus the Gauss loci provide a solution for the Schottky problem to characterize the closure of the locus of Jacobians in the moduli space of principally polarized abelian varieties: (a) The locus of Jacobians is a component of The above corollary is shown in section 6 together with an analogous statement for Prym varieties. It confirms a conjecture by the first author, Grushevsky and Sernesi [11, conjecture 1.6] who verified it for g ≤ 4 by an explicit description of the Gauss loci. As pointed out in loc. cit., this is only a weak solution to the Schottky problem: In general the Gauss loci in the above corollary have more than one irreducible component and the Jacobian locus is only one of them. The theory of D-modules allows to refine the degree of the Gauss map to representation theoretic invariants that might distinguish the Jacobian locus [26].
1.C. The degree of conormal varieties. For the proof of theorem 1.1 and 1.2 we interpret the degree of the Gauss map as an intersection number of Lagrangian cycles on the cotangent bundle of the abelian variety and apply specialization for such cycles [19,34], which we can do because char(k) = 0. To explain how this works, let us forget about abelian varieties for a moment and fix any ambient smooth variety W over k. The conormal variety to a subvariety X ⊂ W is defined as the closure of the conormal bundle to the smooth locus Sm(X), where the closure is taken in the total space of the cotangent bundle of the ambient smooth variety. This conormal variety always has pure dimension n = dim(W ), in fact it is Lagrangian with respect to the natural symplectic structure on the cotangent bundle. It is also conic, i.e. stable under the natural action of the multiplicative group on the fibers of the cotangent bundle. Conversely, any closed conic Lagrangian subvariety of the cotangent bundle arises like this [23, lemma 3]. So the map X → Λ X induces an isomorphism where by a conic Lagrangian cycle we mean a Z-linear combination of closed conic Lagrangian subvarieties. In the case of projective varieties we can talk about the degree of conormal varieties: If W is projective, the degree homomorphism on conic Lagrangian cycles is the map which is given by the intersection number with the zero section i : W ֒→ T ∨ (W • deg(Λ X ) ≥ 0 for any X ⊂ A, • deg(Λ X ) > 0 if and only if X is of general type, • deg(Λ X ) = deg(γ X ) for divisors X ⊂ A with Gauss map γ X .
This easily implies theorem 1.1 when combined with the principle of Lagrangian specialization which we recall in section 2: For any flat family of subvarieties in a smooth ambient 1-parameter family, the limit of their conormal varieties is an effective conic Lagrangian cycle whose support contains the conormal variety to the central fiber as a component, and the total degree of the limit cycle equals the degree of a general fiber. The same argument shows that our semicontinuity result holds not only for divisors but for subvarieties of any codimension: Theorem 1.7. Let A → S be an abelian scheme over a variety S, and let X ⊂ A be an arbitrary family of subvarieties which is flat over S. Then for each d ∈ N the subsets S d = s ∈ S | deg(Λ Xs ) ≤ d ⊆ S are closed in the Zariski topology.
It remains to prove theorem 1.2. Given the interpretation for the degree of Gauss maps in prop. 1.6, the proof has nothing to do with abelian varieties: In section 3 we show that for any flat family of divisors on a smooth 1-parameter variety, the specialization of their conormal varieties contains an extra component whenever the singular locus of the fiber jumps. While the final criterion is phrased only for divisors, we formulate our arguments as far as possible for subvarieties in arbitrary codimension to get beyond theorem 1.2 (see example 3.6). This is important even if one only wants to study singularities of divisors: In the theory of Chern classes for singular varieties one attaches to any subvariety X ⊂ A a characteristic cycle of the form where Z runs through certain strata in Sing(X) [23,34], and the topologically meaningful invariant that appears in generalizations of the Gauss-Bonnet index formula is the total degree deg(Λ) involving all the strata.
1.E. A topological view on jump loci. In section 7, which is not used in the rest of the paper, we deduce from our previous results a general semicontinuity theorem for the intersection cohomology of varieties over the complex numbers. Recall that for a complex variety X, the intersection cohomology IH • (X) only depends on its homeomorphism type in the Euclidean topology; it coincides with Betti cohomology in the smooth case but is better behaved in general [6,20,21,24,30]. We denote by the Euler characteristic of the intersection cohomology. This Euler characteristic is usually not semicontinuous in families, it can jump in both directions. But for families of finite branched covers of subvarieties in complex abelian varieties this does not happen (see lemma 7.6 and corollary 7.7): Theorem 1.8. Let f : X → S be a family of varieties such that each fiber X s is generically reduced and admits a finite morphism to an abelian variety. Then for each d ∈ N 0 the loci are closed in the Zariski topology.
This puts our results in a topological context, since the intersection cohomology of a complex variety only depends on its homeomorphism type. For instance, it follows from the above that a singular theta divisor cannot be homeomorphic to a smooth one (recall that there are examples of normal varieties which are singular but homemorphic to smooth varieties, such as those by Brieskorn [9,10]). In corollary 7.9 we will see that the Jacobian locus appears in the stratification of A g by the intersection cohomology of the theta divisor, so we obtain: Corollary 1.9. The locus of Jacobian varieties in A g is an irreducible component of the closure of the locus of all ppav's whose theta divisor is homeomorphic to a theta divisor on a Jacobian variety.
It seems an interesting problem to study the topology of theta divisors on abelian varieties in more detail.

Lagrangian specialization
For convenience we include in this section a self-contained review of some basic facts about the specialization of Lagrangian cycles, which was introduced in relation with Chern-MacPherson classes [34] and nearby cycles for D-modules and perverse sheaves [19]. We work in a relative setting over a smooth curve S. The family of our ambient spaces is given by a smooth dominant morphism of varieties f : W → S where dim(W ) = n + 1. Let X ⊂ W be a reduced closed subvariety. The relative smooth locus of the relative cotangent bundle, we define the relative conormal variety to X as the closure Remark 2.1. In [8] the relative conormal variety is instead defined as the closure inside the absolute cotangent bundle. This notion of relative conormal variety is obtained from ours by base change via the quotient map Indeed, both sides are irreducible closed subvarieties of T ∨ (W )| X . For the right hand side this holds by definition, for the left hand side it follows from the fact that Λ X/S ⊂ T ∨ (X/S) is an irreducible closed subvariety and T ∨ W → T ∨ (X/S) is a fibration with irreducible fibers. So it suffices to show that both sides agree over some open dense U ⊂ X. We can assume X is flat over S and take U = Sm(X/S), in which case the claim becomes obvious.
Lemma 2.2. If X is flat and irreducible over S, then so is Λ X/S . Proof. Λ X/S is defined as the schematic closure of a locally closed subscheme V of the relative cotangent bundle T ∨ (A/S). The subscheme V is the total space of a vector bundle over a smooth variety, so it is a smooth variety as well. Its schematic closure is integral, and a morphism from an integral scheme to a smooth curve is flat iff it is dominant [22,chapter III,prop. 9.7].
Relative conormal varieties can be seen as families of conormal varieties. In what follows we denote by L (W/S) = X⊂W Z · Λ X/S the free abelian group on relative conormal varieties to closed subvarieties X ⊂ W that are flat over S. By the specialization of Λ ∈ L (W/S) at s ∈ S(k) we mean the cycle sp s (Λ) = Λ · f −1 (s) which underlies the schematic fiber of the morphism Λ X/S → S at s. This is again a conic Lagrangian cycle by the following classical result, see [18, prop sending effective cycles to effective cycles. On Chow groups it induces the Gysin map in the bottom row of the following commutative diagram: where m Xs , m Z > 0 and the sum runs over finitely many subvarieties Z ⊂ Sing(X s ).
Proof. Note that T ∨ (W s ) is an effective Cartier divisor in T ∨ (W ). It intersects properly any relative conormal variety to a subvariety which is flat over S. Hence it is clear that the specialization induces on Chow groups the Gysin map defined in [17, sect. 2.6] and sends effective cycles to effective cycles. Now take an irreducible subvariety X ⊂ W which is flat over S. By Lemma 2.2 the morphism Λ X/S → S is flat and hence all its fibers are pure dimensional of the same dimension. Furthermore the action of the multiplicative group preserves the fibers of T ∨ (W/S) → S and so the fibers of Λ X/S → S are unions of conic subvarieties. As the canonical relative symplectic form on T ∨ (W/S) restricts to the canonical symplectic form on T ∨ (W s ) for every s, we conclude that the fibers of Λ X/S → S are also Lagrangian and hence a union of conormal varieties, since the conic Lagrangian subvarieties of the cotangent bundle are precisely the conormal varieties [23, lemma 3]. The coefficients are non-negative as the specialization of effective cycles is effective. Hence sp s (Λ X/S ) is a sum of conormal varieties, and since under the morphism T ∨ (A s ) → A s its support surjects onto X s , we conclude that one of the appearing components must be Λ Xs .
As Λ X/S is irreducible and we work over a field of characteristic zero, there exists a Zariski open dense subset of S over which the fibers of the morphism Λ X/S → S are reduced and irreducible. We conclude that for s in this Zariski open dense subset of S we have sp s (Λ X/S ) = Λ Xs . Moreover, the specialization cannot have any further components over the relative smooth locus Sm(X/S) ⊆ X, since on that locus also the morphism Λ X/S → S restricts to a smooth morphism.
We have the following consequence of flatness: Proof. The cycle class of the specialization sp s (Λ X/S ) is the image of [Λ X/S ] under the Gysin map in lemma 2.3, and its degree is defined as the intersection number of this image with the zero section X = X ֒→ W = T ∨ (W/S). As X and W are flat over S and X ֒→ W is a regular embedding, the degree is therefore constant by [17, th. 10.2] applied to the relative conormal variety V = Λ X/S . Note that V ֒→ W is not required to be a regular embedding; in order to apply loc. cit. we only need that its base change to X is proper over S, which is true.
Remark 2.5. For char(k) > 0 the specialization of a family of conormal varieties need not be a sum of conormal varieties, see [25, p. 215]. Similarly, for char(k) = 0 we need dim(S) = 1, otherwise we would have to restrict the class of morphisms as in [33]. For instance, for

Jump loci for the degree
Let f : W → S be a smooth dominant morphism from a smooth variety to a smooth curve as above. For any S-flat subvariety X ⊂ W and s ∈ S(k) we have seen that sp s (Λ X/S ) − Λ Xs ≥ 0, where the inequality means that the cycle on the left hand side is effective or zero. It is natural to ask for which s ∈ S(k) the above inequality is strict. In the notation of lemma 2.3 this happens iff m Z > 0 for some Z ⊆ Sing(X s ). The following provides a sufficient criterion for this to happen for families of divisors: We divide the proof in several steps. Most of the argument works in arbitrary codimension d, so for the moment we do not yet assume d = 1. It will be enough to prove the claim over some open dense subset of W . Fixing a general point p ∈ Z(k) and working locally near that point, we can assume that • Z = Sing(X/S) (equality as a scheme), By the first item each x ∈ (X \ Z)(k) is a smooth point of X t for t = f (x). Fixing a trivialization as in the second item, we can furthermore identify the conormal space to X t ⊂ W t at x with a subspace in V = T ∨ x (W t ) of codimension d. Consider the relative Gauss map which sends each point to the corresponding conormal space. This is a rational map whose locus of indeterminacy is precisely Z. Let γ X :X → Gr(d, V ) denote its resolution of indeterminacy which is obtained by blowing up the base locus Z ⊂ X as in [17, sect. 4.4]:X We want to control the image of the exceptional divisor E X = π −1 X (Z) ⊂X under the map Lemma 3.2. The morphism α X is a closed embedding.
Proof. By assumption X ⊂ W is cut out by a regular sequence f 1 , . . . , f d . The same then holds for each fiber X t ⊂ W t . Hence it follows that the relative singular locus Sing(X/S) is cut out as a closed subscheme of W by f 1 , . . . , f d and by the d × d minors where we fix an arbitrary basis ∂ 1 , . . . , ∂ n ∈ V ∨ for the fiber of the relative tangent bundle and regard the basis vectors as relative derivations for the smooth morphism W → S. Now let ι : X × Gr(d, V ) ֒→ X × P(Λ d V ) be the Plücker embedding of the Grassmannian as a closed subvariety of projective space. We want to show that the composite is a closed embedding. For this let I O X be the ideal sheaf of Z ⊂ X. Then we havê X = Proj X R I for the graded Rees algebra where t is a dummy variable to keep track of degrees. The homomorphism for 1 ≤ i 1 < · · · < i k ≤ n. But we have seen above that the O X -module I is generated by the minors on the right hand side. Hence it follows that β * X is an epimorphism in all degrees and so β X is a closed immersion. In particular, one may look at the scheme-theoretic fiber ofX over s ∈ S(k) to compute the multiplicities in sp s (Λ X/S ). For d > 1 the situation becomes more complicated, so in what follows we restrict ourselves to set-theoretic arguments. To pass back to conormal varieties we look at the projection Fl(d, 1, V ) −→ Gr(d, V ) from the partial flag variety. This projection is a smooth equidimensional morphism of relative dimension d−1. On the fiber product Taking the preimage of the previous exceptional divisor E X ⊂X we get the following lower bound on the specialization: Proof. The statement about the dimension holds because dim(Y ) = n = dim(V ) and since the subvariety E Y ⊂ Y is a divisor, being the preimage of the exceptional divisor E X  Proof. Each irreducible component of P(Supp(sp s (Λ X/S ))) is the projectivization of some conormal variety. Each of them has dimension n − 1, so lemma 3.4 and our generic finiteness assumption imply that α Y (E Y ) must appear as one of the components. But then this component is Λ Z because it maps onto Z ⊂ X.
Note that by lemma 3.2 the morphism Y → X × Fl(d, 1, V ) is a closed immersion and hence generically finite onto its image. So corollary 3.5 finishes the proof of proposition 3.1 since for codimension d = 1 the morphism Fl(d, 1, V ) → PV is an isomorphism. This is the only point where we use d = 1. For higher codimension the morphism α Y : E Y → X ×PV is not always a closed embedding, as the following example shows, but it may still be generically finite onto its image as needed for corollary 3.5: Here Z = Sing(X/S) ⊂ X is a fat point with ideal sheaf I = (xy, xz, yz) O X and looking at the minors of the Jacobian matrix we see that the relative Gauss map is given in Plücker coordinates on the Grassmannian Gr(2, V ) = Proj k[w 1 , w 2 , w 3 ] by Note that the right hand side does not involve the parameter s. Furthermore, we have (2x 2 +y 2 +z 2 )| X = (f +g)| X = 0 and hence the relative Gauss map γ X factors over Q X = {2w 2 2 w 2 3 + w 2 1 w 2 3 + w 2 2 w 2 3 = 0} ⊂ Gr(2, V ). Write PV = Proj k[v 1 , v 2 , v 3 ] for the dual coordinates v i where the flag variety is given by we get the following diagram where the squares are Cartesian and the hooked arrows are closed immersions: The composite of the arrows in the top row is the morphism α Y : E Y → X × PV that we are interested in. The diagram shows that it is not a closed immersion, over Z red = {0} ⊂ X we have the factorization where Q Y → PV is an irreducible cover of generic degree four! However, since the left diagonal arrow is a closed immersion and hence birational for dimension reasons, the morphism α Y : E Y → X × PV is generically finite over its image.

Generalities about families of rational maps
Before we apply the above to Gauss maps, let us recall some generalities about families of rational maps. Let f : X → S be a faithfully flat morphism of varieties of relative dimension n with irreducible fibers. Let L ∈ Pic(X) be a line bundle and V a rank n+1 vector subbundle of f * L . Then for each point s ∈ S(k) we get a linear series V s ⊂ H 0 (X s , L s ) and we denote by φ s : X s PV s the corresponding rational map. Note that since the source and the target of this map have the same dimension, the map is a generically finite cover iff it is dominant. So we consider the degree map deg : Proof. By [17, prop. 4.4] we can compute the degree in terms of Segre and Chern classes as where B s ⊂ X s denotes the base locus of the linear series V s ⊂ H 0 (X s , L s ). We can put together all these fiberwise base loci into a relative base locus and consider the flattening stratification of this relative base locus. This is a stratification of S such that on each stratum the above intersection number is constant, hence the function deg is constructible.
However, in general the degree is neither upper nor lower semi-continuous, as the following variation of [11, ex. 2.3] shows: Example 4.2. Let S be a smooth affine curve with two marked points s ± ∈ S(k) and fix positive integers n ± ≤ n < 27. Let p j : S −→ P 3 for j = 1, . . . , n be such that • for t = s ± the points p j (t) are in general position, • for t = s ± they consist of n ± general points on a given line ℓ and n − n ± points in general position not on that line.
For t ∈ S(k) let f t : P 3 P 3 be the generically finite rational map defined by a linear system of four generic cubics passing trough the p j (t). By loc. cit. its degree is deg(f t ) = 27 − n for t = s ± , 20 − (n − n ± ) for t = t ± and n ± ≥ 4, since in the second case the indeterminacy locus of f ± consists of the chosen line ℓ together with the remaining n−n ± points. So the degree is a constructible function as predicted by lemma 4.1. However, taking for example (n, n + , n − ) = (20, 10, 4) we obtain an example where the generic value of the degree is seven, jumps up to ten at one point and down to four at another point. Hence in the same family the degree can both decrease and increase under specialization.

Gauss maps on abelian varieties
We now apply the above to an abelian scheme f : A → S, so in this section we take W = A. Let X ⊂ A be a closed subvariety which is flat over S. For s ∈ S(k) we have the Gauss map γ Xs : PΛ Xs −→ PV where V = H 0 (A s , Ω 1 As ). If this map is dominant, then for dimension reasons it is a generically finite cover and we denote by deg(γ Xs ) its generic degree. If the map is not dominant we put deg(γ Xs ) = 0; this happens iff the subvariety X s ⊂ A s arises by pull-back from some smaller dimensional abelian quotient variety [37, th. 1], which by [1, th. 3] happens iff X s is not of general type. The degree of the above Gauss map is related to the degree of conormal varieties as follows: Proof. The degree of our Gauss map γ Xs : PΛ Xs → PV coincides with the degree considered by Franecki and Kapranov terms of tangent rather than cotangent spaces in [16, sect. 2]: Up to the duality Gr(d, V ) ≃ Gr(g − d, V ∨ ) they study the map p • q defined by the diagram In particular deg(Λ Xs ) ≥ 0. Together with the preservation of the total degree under Lagrangian specialization this leads to our first semicontinuity result: Proof. By Lemma 4.1, we know that the map is constructible. We have to show that its values decreases under specialization. For this we may assume that S is a curve, and after base change to its normalization we may assume this curve to be smooth. Let d be the value from proposition 2.4. With notations as in lemma 2.3 then where δ(s) = Z⊂Xs m Z · deg(Λ Z ) with multiplicities m Z ≥ 0. Now in the case of abelian varieties the occuring degrees coincide with the degrees of the corresponding Gauss maps by lemma 5.1. In particular, since the degrees of Gauss maps are obviously nonnegative, we have deg(γ Z ) ≥ 0 and therefore δ(s) ≥ 0, which proves that the degree of the Gauss map is constant on an open dense subset and can only drop on the finitely many points of the complement.
To see where the function in the previous corollary actually jumps, recall that a subvariety Z ⊂ A s has Gauss degree deg(γ Z ) = 0 iff it is not of general type. Thus we obtain the following sufficient jumping criterion: Corollary 5.3. Suppose dim(S) = 1. Let X ⊂ A be a divisor which is flat over S and let 0 ∈ S(k) be a point such that Sing(X 0 ) has an irreducible component Z which is of general type and not contained in the closure of t =0 Sing(X t ). Then there is an open dense subset U ⊂ S such that Proof. The first inequality follows from prop. 3.1, the second from our assumption that the subvariety Z ⊂ A 0 is of general type.

Application to the Schottky problem
The moduli space A g of principally polarized abelian varieties of dimension g over the field k admits the finite filtration · · · ⊆ G d ⊆ G d−1 ⊂ · · · ⊆ G g! = A g by the Gauss loci Our semicontinuity result implies: Proof. The moduli space A g has a finite cover by a smooth quasi-projective variety over which there exists universal theta divisor. On this cover the Gauss maps fit together in a family of rational maps as in the setting of section 4. The Gauss loci are the level sets of the degree map, and we have to show that this map is lower-semicontinuous. This follows from corollary 5.2.
Using our sufficient criterion for jumps in the degree of Gauss maps, we can now show that the stratification by the Gauss loci refines the stratification by the Andreotti-Mayer loci [3]. Some care is needed because the singular locus of the theta divisor may have components which are negligible, i.e. not of general type: Remark 6.2. There are indecomposable ppav's (A, Θ) ∈ A g with a theta divisor for which Sing(Θ) ⊂ A is negligible. This even happens for generic ppav's on certain irreducible components of Andreotti-Mayer loci: For instance, for g = 5 one can show that for a generic ppav on the component E 5,1 ⊂ N 1 from [13, thm. 4.1(ii)] the singular locus Sing(Θ) is an elliptic curve.
A more detailed discussion will be given in a forthcoming work by Constantin Podelski. In any case negligible components can only appear on decomposable abelian varieties, hence the following corollary of our jumping criterion covers all Andreotti-Mayer strata whose general point is a simple abelian variety: Corollary 6.3. Let c ∈ N, and let N ⊂ N c be an irreducible component whose general point is a ppav whose singular locus of the theta divisor has no negligible components. Then N is an irreducible component of G d for some d ∈ N.
Proof. Let s ∈ N (k) be a general point on the given component. Since the moduli space of ppav's is a quasiprojective variety, we may pick an affine curve S ⊂ A g such that S ∩ N c = {s} and S meets N transversely. After passing to a finite cover we may assume that there exists an abelian scheme f : A → S and a universal theta divisor Θ ⊂ A over this curve, and by our choice of the curve we have dim Sing(Θ t ) < dim Sing(Θ s ) for all t = s. As an application we get that the stratification by the degree of the Gauss map gives a solution to the Schottky problem as conjectured in [11], where for the Prym version we denote by D(g) the degree of the varieties of quadrics in P g−1 of rank at most three: (c) The locus of Prym varieties is a is a component of G d for d = D(g) + 2 g−3 .
Proof. The locus of Jacobians is a component of an Andreotti-Mayer locus by [3], and by [31,Proposition 3.4] a general Jacobian variety is a simple abelian variety and thus in particular has no negligible subvarieties other than itself. Furthermore, it is well-known that the degree of the Gauss map of a Jacobian is d = 2g−2 g−1 , see e.g. [2, proof of prop. 10]. Hence part (a) follows from corollary 6.3. If we replace Jacobians by hyperelliptic Jacobians, the above arguent works also in the hyperelliptic case, with the same references. For Prym varieties the argument is again the same but now one has to replace reference [3] with [14], and reference [2] with [36,Main Theorem]. In the last reference the reader can also find an explicit expression for the number D(g).

A topological view on jump loci
In this section we work over the complex numbers with the Euclidean topology. Let W be a smooth complex projective variety. For a closed subvariety X ⊂ W the singular locus Sing(X) and the conormal degree deg(Λ X ) are not topological invariants of the subvariety, as example 1.5 shows. But both are related to the intersection cohomology IH • (X) which only depends on the homeomorphism type of the subvariety in the Euclidean topology; see [6,20,21,24,30]. The Euler characteristic can be read off from a generalization of the Gauss-Bonnet theorem: The Kashiwara index formula [19, th. 9.1] writes it as a degree in the sense of definition 1.4. More precisely χ IC (X) = deg(CC(δ X )), where δ X ∈ Perv(W ) denotes the perverse intersection complex of X ⊆ W [5,12] and where the characteristic cycle CC(δ X ) ∈ L (W ) is an effective conic Lagrangian cycle which contains Λ X as a component of multiplicity one but may also have as components the conormal varieties to certain Z ⊆ Sing(X). Passing from conormal varieties to characteristic cycles restores topological invariance of the degree: Example 7.1. In W = P 2 the conormal degree for a smooth rational curve differs from the one for a cuspical cubic, see example 1.5. But this is compensated by a difference in which in both cases gives the total degree deg(CC(δ X )) = −2.
In what follows we want to understand how for a morphism X → S of complex varieties the intersection cohomology Euler characteristic of the fibers varies. Basic stratification theory implies:  3 if Q is a cone over a smooth rational curve, 6 if Q is a union of two projective planes. So for a family of quadrics whose general member is smooth, the number χ IC (Q) jumps down on nodal quadrics but jumps up on reducible quadrics.
Note that here the size of the jumps is precisely the Euler characteristic of the singular locus. This fits with the following sheaf-theoretic version of the Lagrangian specialization principle: Theorem 7.4. Let f : W → S be a smooth proper family over a curve S, and let X ⊂ W be a closed subvariety such that the morphism f : X → S is flat with generically reduced fibers. Then there exists d ∈ Z and a finite subset Σ ⊂ S such that where Λ(s) = sp s (CC(δ X )) − CC(δ Xs ) ∈ L (W s ) is an effective cycle.
Proof. We interpret Lagrangian specialization via perverse sheaves. For s ∈ S(C) one has the functor of nearby cycles Ψ s : Perv(W ) → Perv(W s ), which is an exact functor with CC(Ψ s (P )) = sp s (CC(P )) for all P ∈ Perv(W ) by [19, th. 5.5]. Here we abuse notation and view CC(P ) as an element of the group of relative conic Lagrangian cycles L (W/S) via remark 2.1, discarding any component that is not flat over S. The last part of the specialization lemma 2.3 has a sheaf-theoretic version: For any closed S-flat subvariety X ⊂ W such that the map f : X → S has generically reduced fibers, there exists a finite subset Σ ⊂ S such that the semisimplification (Ψ s (δ X )) ss of the perverse sheaf Ψ s (δ X ) has the form where P (s) ∈ Perv(X s ) is a perverse sheaf with support contained in Sing(X s ). So we get where Λ(s) = CC(P (s)) ∈ L (W s ) is effective, being the characteristic cycle of a perverse sheaf. Hence the result follows by noting that if f : W → S is proper, then by proposition 2.4 the degree d = deg(sp s (CC(δ X ))) is independent of s.
In the case of abelian varieties the positivity of conormal degrees then gives an analog of theorem 1.7. The same argument works for a much wider class of varieties, we only need the following positivity property: Definition 7.5. A variety X satisfies the signed Euler characteristic property if we have χ(X, P ) := i∈Z (−1) i dim H i (X, P ) ≥ 0 for all P ∈ Perv(X).
The terminology is borrowed from [15]. The above property holds for semiabelian varieties [16] and hence also for any finite cover of closed subvarieties of them: Lemma 7.6. If a variety A has the signed Euler characteristic property, then so does any variety with a finite morphism to A. In particular, any variety with a finite morphism to a semiabelian variety has the signed Euler characteristic property.
Proof. If f : X → A is a finite morphism, then for any perverse sheaf P ∈ Perv(X) the direct image is a perverse sheaf Rf * (P ) ∈ Perv(A). If A has the signed Euler characteristic property, which holds for instance for abelian varieties [16], then we get χ(A, P ) = χ(A, Rf * (P )) ≥ 0.
The above theorem shows that for any family of such varieties the intersection cohomology Euler characteristic is semicontinuous: Corollary 7.7. Let f : W → S be a smooth proper morphism to a variety S, and let X ⊂ W be a closed subvariety such that f : X → S is flat and all its fibers are generically reduced and have the signed Euler characteristic property. Then for each d ∈ N the subsets S d = s ∈ S | χ IC (X s ) ≤ d ⊆ S are Zariski closed.
Proof. We must show that χ IC (X s ) cannot increase under specialization. For this we can assume S is a smooth curve. Then theorem 7.4 applies, here χ(X s , P (s)) ≥ 0 since X s has the signed Euler characteristic property.
In deciding where the Euler characteristic actually jumps, we need to be more careful. Proposition 3.1 gives a way to see extra components in sp s (CC(δ X )) but does not guarantee that these enter in a new summand Λ(s), a priori they could also appear in CC(δ Xs ); however, this second case can only happen if CC(δ Xs ) is reducible, which one can often exclude by a direct computation.
Let us illustrate this again with theta divisors. Corollary 7.7 says that for d ∈ N the loci are closed, and by the homeomorphism invariance of intersection cohomology they only depend on the topology of the theta divisor. This provides a topological view on Andreotti-Mayer loci, for instance: Then N is also an irreducible component of X d for some d ∈ N.
Proof. Use the same argument as in corollary 6.3, together with the remark after the proof of corollary 7.7.
This in particular applies to the locus of Jacobians. In the following corollary we do not mention hyperelliptic Jacobians because for them CC(δ Θ ) is reducible, and we haven't checked what happens for a generic Prym variety. However, we include the Andreotti-Mayer locus N 0 ⊂ A g of ppav's with a singular theta divisor: Corollary 7.9. Inside the moduli space A g we have: (1) The locus N 0 is equal to X d for d = g! − 1 if g is odd, g! − 2 if g is even.
(2) The locus of Jacobians is a component of X d for d = 2g−2 g−1 . Proof. (1) By definition A g \ N 0 consists of all ppav's (A, Θ) with a smooth theta divisor and for those we know that χ IC (Θ) = g! because for a smooth variety intersection cohomology equals Betti cohomology. But at a generic point (A, Θ) of each of the two components of N 0 the theta divisor has one respectively two nodes, and then where k ∈ {1, 2} denotes the number of nodes [27, proof of prop. 4.2 (2)]. Hence the claim follows by corollary 7.7. Note that the degree of the classical Gauss map is deg(Λ Θ ) = g! − 2k in both cases, but for odd g the cycle CC(δ Θ ) = Λ Θ + Λ Sing(Θ) is reducible and we cannot directly apply corollary 7.8.
(2) For Jacobians of nonhyperelliptic curves we know that CC(δ Θ ) = Λ Θ is irreducible by [7, th. 3.3.1], so if we specialize to such a Jacobian, then any new component of the specialization must enter in Λ(s). So the same argument as in the proof of corollary 6.4 shows that the locus of Jacobians is a component of X d where d = χ IC (Θ) = deg(CC(δ Θ )) = deg(Λ Θ ) is the degree of the Gauss map for the theta divisor on a general Jacobian variety as in corollary 6.4.
The above is still only a weak solution to the Schottky problem, though χ IC (Θ) also appears as the dimension of an irreducible representation of a certain reductive group which gives more information [26, sect. 4]. The following example for g = 4 illustrates the difference between the various numerical invariants: So there are non-homeomorphic theta divisors whose Gauss maps have the same degree. Are there also homeomorphic theta divisors with different Gauss degrees?